Content uploaded by Saroj Kumar
Author content
All content in this area was uploaded by Saroj Kumar on Aug 27, 2021
Content may be subject to copyright.
Advances in Intelligent Systems and Computing 1172
SubhransuSekharDash
SwagatamDas
BijayaKetanPanigrahiEditors
Intelligent
Computing and
Applications
Proceedings ofICICA 2019
Advances in Intelligent Systems and Computing
Volume 1172
Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
Advisory Editors
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing,
Universidad Central de Las Villas, Santa Clara, Cuba
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
Hani Hagras, School of Computer Science and Electronic Engineering,
University of Essex, Colchester, UK
LászlóT. Kóczy, Department of Automation, Széchenyi István University,
Gyor, Hungary
Vladik Kreinovich, Department of Computer Science, University of Texas
at El Paso, El Paso, TX, USA
Chin-Teng Lin, Department of Electrical Engineering, National Chiao
Tung University, Hsinchu, Taiwan
Jie Lu, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, NSW, Australia
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute
of Technology, Tijuana, Mexico
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro,
Rio de Janeiro, Brazil
Ngoc Thanh Nguyen , Faculty of Computer Science and Management,
Wrocław University of Technology, Wrocław, Poland
Jun Wang, Department of Mechanical and Automation Engineering,
The Chinese University of Hong Kong, Shatin, Hong Kong
The series “Advances in Intelligent Systems and Computing”contains publications
on theory, applications, and design methods of Intelligent Systems and Intelligent
Computing. Virtually all disciplines such as engineering, natural sciences, computer
and information science, ICT, economics, business, e-commerce, environment,
healthcare, life science are covered. The list of topics spans all the areas of modern
intelligent systems and computing such as: computational intelligence, soft comput-
ing including neural networks, fuzzy systems, evolutionary computing and the fusion
of these paradigms, social intelligence, ambient intelligence, computational neuro-
science, artificial life, virtual worlds and society, cognitive science and systems,
Perception and Vision, DNA and immune based systems, self-organizing and
adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics
including human-machine teaming, knowledge-based paradigms, learning para-
digms, machine ethics, intelligent data analysis, knowledge management, intelligent
agents, intelligent decision making and support, intelligent network security, trust
management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing”are
primarily proceedings of important conferences, symposia and congresses. They
cover significant recent developments in the field, both of a foundational and
applicable character. An important characteristic feature of the series is the short
publication time and world-wide distribution. This permits a rapid and broad
dissemination of research results.
** Indexing: The books of this series are submitted to ISI Proceedings,
EI-Compendex, DBLP, SCOPUS, Google Scholar and Springerlink **
More information about this series at http://www.springer.com/series/11156
Subhransu Sekhar Dash •Swagatam Das •
Bijaya Ketan Panigrahi
Editors
Intelligent Computing
and Applications
Proceedings of ICICA 2019
123
Editors
Subhransu Sekhar Dash
Department of Electrical Engineering
Government College of Engineering
Keonjhar, Odisha, India
Bijaya Ketan Panigrahi
Indian Institute of Technology Delhi
New Delhi, Delhi, India
Swagatam Das
Electronics and Communication
Sciences Unit
Indian Statistical Institute
Kolkata, West Bengal, India
ISSN 2194-5357 ISSN 2194-5365 (electronic)
Advances in Intelligent Systems and Computing
ISBN 978-981-15-5565-7 ISBN 978-981-15-5566-4 (eBook)
https://doi.org/10.1007/978-981-15-5566-4
©Springer Nature Singapore Pte Ltd. 2021
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface
This AISC volume contains the papers presented at the Fifth International
Conference on Intelligent Computing and Applications (ICICA 2019) held during
December 6–8, 2019, at, SRM Institute of Science and Technology, Delhi-NCR
Campus, Modinagar, Ghaziabad, India.
ICICA 2019 is the Fifth International Conference aiming at bringing together the
researchers from academia and industry to report and review the latest progresses in
the cutting-edge research on various research areas of electronic circuits, power
systems, renewable energy applications, image processing, computer vision and
pattern recognition, machine learning, data mining and computational life sciences,
management of data including big data and analytics, distributed and mobile sys-
tems including grid and cloud infrastructure, information security and privacy,
VLSI, antenna, intelligent manufacturing, signal processing, intelligent computing,
soft computing, web security, privacy and e-commerce, e-governance, optimiza-
tion, communications, smart wireless and sensor networks, networking and infor-
mation security, mobile computing and applications, industrial automation and
MES, cloud computing, green IT and finally to create awareness about these
domains to a wider audience of practitioners.
ICICA 2019 received 290 paper submissions including two foreign countries
across the globe. All the papers were peer-reviewed by the experts in the area in
India and abroad, and comments have been sent to the authors of accepted papers.
Finally 75 were accepted for oral presentation in the conference. This corresponds
to an acceptance rate of 32% and is intended to maintain the high standards of the
conference proceedings. The papers included in this AISC volume cover a wide
range of topics in intelligent computing and algorithms and their real-time appli-
cations in problems from diverse domains of science and engineering.
The conference was inaugurated by Prof. Nattachote Rugthaicharoencheep,
Senior IEEE Member, Rajamangala University of Technology, Thailand, on
December 6th, 2019. The conference featured distinguished keynote speakers as
follows: Dr. Ikechi Ukaegbu, Nazarbayev University, Republic of Kazakhstan;
Prof. D. P. Kothari, FIEEE, Nagpur, India; Shri. Aninda Bose, Senior Editor,
Springer New Delhi, India; Dr. Bijaya Ketan Panigrahi, IIT Delhi, India;
v
Dr. Swagatam Das, ISI, Kolkata, India; Dr. Subhransu Sekhar Dash, Professor and
Head, GCE Keonjhar, Odisha, India.
We take this opportunity to thank the authors of the submitted papers for their
hard work, adherence to the deadlines, and patience with the review process. The
quality of a referred volume depends mainly on the expertise and dedication of the
reviewers. We are indebted to the Technical Committee members, who produced
excellent reviews in short time frames. First, we are indebted to the Hon’ble
Dr. T. R. Paarivendhar, Member of Parliament (Lok Sabha), Founder-Chancellor,
SRM Institute of Science and Technology; Shri. Ravi Pachamoothoo, Chairman,
SRM Institute of Science and Technology; Dr. P. Sathyanarayanan, President, SRM
Institute of Science and Technology; Dr. R. Shivakumar, Vice President, SRM
Institute of Science and Technology; Dr. Sandeep Sancheti, Vice Chancellor,
SRM Institute of Science and Technology for supporting our cause and encouraging
us to organize the conference there. In particular, we would like to express our
heartfelt thanks for providing us with the necessary financial support and infras-
tructural assistance to hold the conference. Our sincere thanks to Dr. D. K. Sharma,
Professor and Dean; Dr. S. Viswanathan, Deputy Registrar; Dr. Navin Ahalawat,
Professor and Dean (Campus Life), SRM Institute of Science and Technology,
Delhi-NCR Campus, Modinagar, Ghaziabad, for their continuous support and
guidance. We specially thank Dr. R. P. Mahapatra, Professor and Head (CSE),
Convener of ICICA 2019 and Dr. Dambarudhar Seth, Co-converner, SRM Institute
of Science and Technology, Delhi-NCR Campus, of this conference for their
excellent support and arrangements, without them it is beyond imagination to
conduct this conference. We thank the International Advisory Committee members
for providing valuable guidelines and inspiration to overcome various difficulties in
the process of organizing this conference. We would also like to thank the partici-
pants of this conference. The members of faculty and students of SRM Institute of
Science and Technology, Delhi-NCR Campus, Modinagar, Ghaziabad, deserve
special thanks because without their involvement, we would not have been able to
face the challenges of our responsibilities. Finally, we thank all the volunteers who
made great efforts in meeting the deadlines and arranging every detail to make sure
that the conference could run smoothly. We hope the readers of these proceedings
find the papers inspiring and enjoyable.
Keonjhar, India Subhransu Sekhar Dash
Kolkata, India Swagatam Das
New Delhi, India
December 2019
Bijaya Ketan Panigrahi
vi Preface
Contents
Performance Analysis of Smart Meters for Enabling a New Era for
Power and Utilities with Securing Data Transmission and Distribution
Using End-to-End Encryption (E2EE) in Smart Grid .............. 1
M. Manimegalai and K. Sebasthirani
Energy Efficient Data Centre Selection Using Service
Broker Policy ............................................. 13
Sameena Naaz, Iffat Rehman Ansari, Insha Naz, and Ranjit Biswas
Co-ordinate Measurement of Roll-Cage Using Digital Image
Processing ............................................... 23
Ritwik Dhar, Parth Kansara, Sanket Shegade, Atharv Bagde,
and Sunil Karamchandani
Split-Ring Resonator Multi-band Antenna for WLAN/WIMAX/X
Standard ................................................ 35
Vinita Sharma, Santosh Meena, and Ritesh Kumar Saraswat
Analyses on Architectural and Download Behavior of Xunlei ........ 43
Bagdat Kamalbayev, Nazerke Seidullayeva, Adilbek Sain,
Pritee Parwekar, and Ikechi A. Ukaegbu
Advanced Driver Assistance System Technologies and Its Challenges
Toward the Development of Autonomous Vehicle ................. 55
Keerthi Jayan and B. Muruganantham
Application of Artificial Intelligence-Based Solution Methodology
in Generation Scheduling Problem ............................. 73
Shubham Tiwari, Vikas Bhadoria, and Bharti Dwivedi
Discomfort Analysis at Lower Back and Classification of Subjects
Using Accelerometer ........................................ 83
Ramandeep Singh Chowdhary and Mainak Basu
vii
Design of Lyapunov-Based Discrete-Time Adaptive Sliding Mode
Control for Slip Control of Hybrid Electric Vehicle ................ 97
Khushal Chaudhari and Ramesh Ch. Khamari
An Improved Scheme for Organizing E-Commerce-Based Websites
Using Semantic Web Mining ................................. 115
S. Vinoth Kumar, H. Shaheen, and T. Sreenivasulu
Performance Estimation and Analysis Over the Supervised Learning
Approaches for Motor Imagery EEG Signals Classification .......... 125
Gopal Chandra Jana, Shivam Shukla, Divyansh Srivastava,
and Anupam Agrawal
Fully Automated Digita l Mammogram Segmentation ............... 143
Karuna Sharma and Saurabh Mukherjee
Empirical Study of Computational Intelligence Approaches
for the Early Detection of Autism Spectrum Disorder .............. 161
Mst. Arifa Khatun, Md. Asraf Ali, Md. Razu Ahmed,
Sheak Rashed Haider Noori, and Arun Sahayadhas
Intelligent Monitoring of Bearings Using Node MCU Module ........ 171
Saroj Kumar, Shankar Sehgal, Harmesh Kumar, and Sarbjeet Singh
Image Denoising Using Various Image Enhancement Techniques ..... 179
S. P. Premnath and J. Arokia Renjith
Energy Consumption Analysis and Proposed Power-Aware
Scheduling Algorithm in Cloud Computing ...................... 193
Juhi Singh
Relationship Between Community Structure and Clustering
Coefficient ............................................... 203
Himansu Sekhar Pattanayak, Harsh K. Verma, and A. L. Sangal
Digital Image Forensics-Image Verification Techniques ............. 221
Anuj Rani and Ajit Jain
Using Automated Predictive Analytics in an Online Shopping
Ecosystem ................................................ 235
Ruchi Mittal
Design and Development of Home Automation System ............. 245
Anjali Pandey, Yagyiyaraj Singh, and Medhavi Malik
Performance Analysis Study of Stochastic Computing
Based Neuron ............................................. 255
A. Dinesh Babu and C. Gomathy
viii Contents
Scheduling of Parallel Tasks in Cloud Environment
Using DAG MODEL ....................................... 267
Sakshi Kapoor and Surya Narayan Panda
GREENIE—Smart Home with Smart Power Transmission .......... 277
N. Noor Alleema, S. Babeetha, V. L. Hariharan, S. Sangeetha,
and H. Shraddha
ANAVI: Advanced Navigation Assistance for Visually Impaired ...... 285
Arjun Sharma, Vivek Ram Vasan, and S. Prasanna Bharathi
THIRD EYE—Shopfloor Data Processing and Visualization
Using Image Recognition .................................... 297
S. Prasanna Bharathi, Vivek Ram Vasan, and Arjun Sharma
A Hybrid Search Group Algorithm and Pattern Search Optimized
PIDA Controller for Automatic Generation Control of Interconnected
Power System ............................................. 309
Smrutiranjan Nayak, Sanjeeb Kar, and Subhransu Sekhar Dash
Fractional Order PID Controlled PV Fed Quadratic Boost
Converter TZ Source Inverter Fed Permanent Magnet Brushless
Motor Drive .............................................. 323
N. K. Rayaguru and S. Sekar
Peformance Analysis of Joysticks Used in Infotainment Control
System in Automobiles ...................................... 337
Vivek Ram Vasan, S. Prasanna Bharathi, Arjun Sharma,
and G. Chamundeeswari
Probabilistic Principal Component Analysis (PPCA) Based
Dimensionality Reduction and Deep Learning for Cancer
Classification ............................................. 353
D. Menaga and S. Revathi
Blockchain Technology for Data Sharing in Decentralized
Storage System ............................................ 369
D. Praveena Anjelin and S. Ganesh Kumar
Understanding Concepts of Blockchain Technology for Building
the DApps ............................................... 383
P. Shamili, B. Muruganantham, and B. Sriman
Blockchain Technology: Consensus Protocol Proof of Work
and Proof of Stake ......................................... 395
B. Sriman, S. Ganesh Kumar, and P. Shamili
Non-Invasive Techniques of Nutrient Detection in Plants ............ 407
Amit Singh and Suneeta V. Budihal
Contents ix
Implementation of Cryptographic Approaches in Proposed Secure
Framework in Cloud Environment ............................ 419
Manoj Tyagi, Manish Manoria, and Bharat Mishra
Home Automation With NoSQL and Node-RED Through Message
Queuing Telemetry Transport ................................ 427
Naman Chauhan and Medhavi Malik
Naïve Bayes Algorithm Based Match Winner Prediction Model
for T20 Cricket ........................................... 435
Praffulla Kumar Dubey, Harshit Suri, and Saurabh Gupta
Design of an Efficient Deep Neural Network for Multi-level
Classification of Breast Cancer Histology Images .................. 447
H. S. Laxmisagar and M. C. Hanumantharaju
Autonomous and Adaptive Learning Architecture Framework
for Smart Cities ........................................... 461
Saravanan Muthaiyah and Thein Oak Kyaw Zaw
A Predictive Analysis for Heart Disease Using Machine Learning ..... 473
V. Rajalakshmi, D. Sasikala, and A. Kala
Application of Data Mining Algorithms for Tourism Industry ........ 481
Promila Sharma, Uma Meena, and Girish Kumar Sharma
Real-Time Safety and Surveillance System Using Facial Recognition
Mechanism ............................................... 497
Sachi Pandey, Vikas Chouhan, Rajendra Prasad Mahapatra,
Devansh Chhettri, and Himanshu Sharma
Liver Disease Prediction Using an Ensemble Based Approach ........ 507
B. Muruganantham, R. P. Mahapatra, Kriti Taparia, and Mukul Kumar
Significance of Route Discovery Protocols in Wireless Sensor
Networks ................................................ 519
Guntupalli Gangaprasad, Kottnana Janakiram,
and B. Seetha Ramanjaneyulu
In-Memory Computation for Real-Time Face Recognition ........... 531
Nikhil Kumar Gupta and Girijesh Singh
A Novel Approach for Detection of Basketball Using CFD Method .... 541
G. Simi Margarat, S. Siva Subramanian, and K. Ravikumar
Ensemble Similarity Clustering Frame work for Categorical Dataset
Clustering Using Swarm Intelligence ........................... 549
S. Karthick, N. Yuvaraj, P. Anitha Rajakumari, and R. Arshath Raja
x Contents
Multi-Focus Image Fusion Using Conditional Generative Adversarial
Networks ................................................ 559
A. Murugan, G. Arumugam, and D. Gobinath
Privacy Preservation Between Privacy and Utility Using ECC-based
PSO Algorithm ............................................ 567
N. Yuvaraj, R. Arshath Raja, and N. V. Kousik
Predicting Energy Demands Constructed on Ensemble
of Classifiers .............................................. 575
A. Daniel, B. Bharathi Kannan, N. Yuvaraj, and N. V. Kousik
Use of RNN in Devangari Script .............................. 585
Madhuri Sharma and Medhavi Malik
Role of Data Science for Combating the Problem of Loan Defaults
Using Tranquil-ART1NN Hybrid Deep Learning Approach ......... 593
Chandra Shaardha and Anna Alphy
Optimized Multi-Walk Algorithm for Test Case Reduction .......... 607
U. Geetha, Sharmila Sankar, and M. Sandhya
Chennai Water Crisis—Data Analysis .......................... 617
Deepak Shankar, N. Aaftab Rehman, Sharmila Sankar, Aisha Banu,
and M. Sandhya
Generalized Canonical Correlation Based Bagging Ensembled
Relevance Vector Machine Classifier for Software Quality
Analysis ................................................. 629
Noor Ayesha and N. G. Yethiraj
A Deep Learning Approach Against Botnet Attacks to Reduce
the Interference Problem of IoT ............................... 645
Pramathesh Majumdar, Archana Singh, Ayushi Pandey,
and Pratibha Chaudhary
Predicting Movie Success Using Regression Techniques ............. 657
Faaez Razeen, Sharmila Sankar, W. Aisha Banu, and Sandhya Magesh
Vehicle Recognition Using CNN ............................... 671
V. K. Divyavarshini, Nithyasri Govind, Amrita Vasudevan,
G. Chamundeeswari, and S. Prasanna Bharathi
GLCM and GLRLM Based Texture Analysis: Application to Brain
Cancer Diagnosis Using Histopathology Images ................... 691
Vaishali Durgamahanthi, J. Anita Christaline, and A. Shirly Edward
Resource Management in Wireless IoT Using Gray Wolf
Optimisation Framework .................................... 707
S. Karthick and N. Gomathi
Contents xi
Less Polluted Flue Gases Obtained with Green Technology During
Precious Metals Recovery from Unwanted and Discarded Electrical
and Electronics Components ................................. 715
Rajendra Prasad Mahapatra, Satya Sai Srikant, Raghupatruni Bhima Rao,
and Bijayananda Mohanty
Secure Data Transmission in Mobile Networks Using Modified S-ACK
Mechanism ............................................... 721
P. Muthukrishnan and P. Muthu Kannan
An Anonymization Approach for Dynamic Dataset with Multiple
Sensitive Attributes ........................................ 731
V. Shyamala Susan
Newspaper Identification in Hindi ............................. 741
Subhabrata Banerjee
Firewall Scheduling and Routing Using pfSense ................... 749
M. Muthukumar, P. Senthilkumar, and M. Jawahar
Undefeatable System Using Machine Learning .................... 759
Anand Sharma and Uma Meena
Synchronization for Nonlinear Time-Delay Chaotic Diabetes Mellitus
System via State Feedback Control Strategy ..................... 769
Nalini Prasad Mohanty, Rajeeb Dey, Binoy Krishna Roy,
and Nimai Charan Patel
Geo/G/1 System: Queues with Late and Early Arrivals ............. 781
Reena Grover, Himani Chaudhary, and Geetanjali Sharma
Intelligent Data Analysis with Classical Machine Learning .......... 793
Sanjeev Kumar Punia, Manoj Kumar, and Amit Sharma
Author Index ................................................ 801
xii Contents
About the Editors
Prof. Subhransu Sekhar Dash is currently a Professor at the Department of
Electrical Engineering, Government College of Engineering, Keonjhar, Odisha,
India. Holding a Ph.D. from the College of Engineering, Guindy, Anna University,
Chennai, India, he has more than 22 years of research and teaching experience. His
research interests include power electronics and drives, modeling of FACTS con-
trollers, power quality, power system stability, and smart grids. He is a Visiting
Professor at Francois Rabelais University, POLYTECH, France; the Chief Editor
of the International Journal of Advanced Electrical and Computer Engineering; and
an Associate Editor of IJRER. He has published more than 200 research articles in
peer-reviewed international journals and conference proceedings.
Prof. Swagatam Das received his B.E. Tel. E., M.E. Tel. E (Control Engineering
specialization) and Ph.D. degrees from Jadavpur University, India, in 2003, 2005,
and 2009, respectively. He is currently serving as an Associate Professor at the
Electronics and Communication Sciences Unit of the Indian Statistical Institute,
Kolkata, India. His research interests include evolutionary computing, pattern
recognition, multi-agent systems, and wireless communication. Dr. Das has pub-
lished one research monograph, one edited volume, and more than 200 research
articles in peer-reviewed journals and international conference proceedings. He is
the Founding Co-Editor-in-Chief of Swarm and Evolutionary Computation, an
international journal from Elsevier. He has also served as or is serving as an
Associate Editor of IEEE Trans. on Systems, Man, and Cybernetics: Systems, IEEE
Computational Intelligence Magazine, IEEE Access, Neurocomputing (Elsevier),
Engineering Applications of Artificial Intelligence (Elsevier), and Information
Sciences (Elsevier). He is also an editorial board member for many journals. He has
been associated with the international program committees and organizing com-
mittees of several regular international conferences including IEEE CEC,
IEEE SSCI, SEAL, GECCO, and SEMCCO. He is the recipient of the 2012 Young
Engineer Award from the Indian National Academy of Engineering (INAE).
xiii
Prof. Bijaya Ketan Panigrahi is a Professor at the Electrical Engineering
Department, IIT Delhi, India. Prior to joining IIT Delhi in 2005, he has served as a
faculty in Electrical Engineering Department, UCE Burla, Odisha, India, from 1992
to 2005. Dr. Panigrahi is a senior member of IEEE and Fellow of INAE, India. His
research interest includes the application of soft computing and evolutionary
computing techniques to power system planning, operation, and control. He has
also worked in the fields of bio-medical signal processing and image processing. He
has served as the editorial board member, associate editor, and special issue guest
editor of different international journals. He is also associated with various inter-
national conferences in various capacities. Dr. Panigrahi has published more than
150 research papers in various international and national journals.
xiv About the Editors
Performance Analysis of Smart Meters
for Enabling a New Era for Power
and Utilities with Securing Data
Transmission and Distribution Using
End-to-End Encryption (E2EE) in Smart
Grid
M. Manimegalai and K. Sebasthirani
Abstract The most outstanding engineering achievement of the twenty-first century
in power systems is smart grid. The SG is an emerging technology which is revo-
lutionizing the typical electrical grid by incorporating Information Communication
Technology (ICT); this can be the main enabler of smart grid. This ICT can bring
increased connectivity with increased severe security vulnerabilities and challenges.
Cyberterrorism can be a prime target in smart grid because of its complex nature. To
eliminate such cybersecurity issues, the most reliable end-to-end security algorithm
was proposed. In this paper, information transmitted from customer through smart
meter to the distributor and from distributor to the customer via smart meter was
secured by End-to-End Encryption (E2EE). Data to and from the user and distributor
cannot be modified in between by third parties. Experimental setup has customized
design of Smart Meter (SM) for Home Area Network (HAN) and the proposed
design will monitor the smart meter data transmission online. Performance analysis
will show the reliability, integrity, and availability maintained in the communication
network.
Keywords Smart grid (SG) ·Smart meter (SM) ·Information and communication
technology (ICT) ·End-to-end encryption (E2EE) ·Home area network (HAN)
M. Manimegalai (B)
Department of Computer Science and Engineering, P.S.R. Rengasamy College of Engineering for
Women, Sivakasi, India
e-mail: mm11.1990@gmail.com
K. Sebasthirani
Department of Electrical and Electronics Engineering, Sri Ramakrishna Engineering College,
Coimbatore, India
e-mail: sebasthirani.kathalingam@srec.ac.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_1
1
2 M. Manimegalai and K. Sebasthirani
1 Introduction
By integrating electrical distribution system with communication system forms,
modern power system which is called smart grid, where power and information flow
will be bi-directional. Transformation from conventional power grid to the smart
grid will increase reliability, performance, and manageability because of full-duplex
communication. SG has advanced communication systems in order to improve the
performance. With larger benefits of advanced communication system, some secu-
rity vulnerabilities are also there in the ICT. Figure 1shows the base architecture
diagram of smart grid.
In Fig. 1, electricity generation and distribution flow are shown in the straight line.
Directed lines will show the communication network flow. In the communication
network [1], smart meter plays an important role, which collects information from
each house and sends it to the distributor. Through Internet, the data from all the
houses will be sent to the distributor side. The following diagram will show the flow
of information from customer to the distributor (Fig. 2).
Security threats [2] are more in the communication system. Security attack mainly
depends on three factors as shown in Fig. 3.
Risk can be defined as
Risk =Assets ×Vulnerabilities ×Threats (1)
where assets are smart grid devices such as smart meters, substations, data, and
network devices. Vulnerabilities will allow the attackers to reduce the reliability and
integrity of the systems information. Threats are the causes of inside or outside of the
smart grid systems. If the vulnerabilities in the system are smaller, risk in the system
will be minimized. Particularly, the assets and the threats cannot be zero. Security
policies in smart grid have three important objectives, which are confidentiality,
integrity, and availability [3]. CIA triad has been shown in Fig. 4.
Power
Consumption
Distribution
Substation
Power
Distribution
Power
Transmission
Power
Transformer
Power
Generation
Electrical
Accessories
Electrical
Vehicles
Smart Meter
RTU/DTU/
TTU/FTU
Charge Pile
Control
Terminal
Centralized
Meter
Reading
Terminal
Center Control
Power
Generation
Acquisition
Terminal
Renewable
Energy
Power
Generator
Set
Power Transmission
Control command and Alert Message
Data Measurement Values
Fig. 1 Architecture model of smart grid
Performance Analysis of Smart Meters for Enabling a New Era … 3
Collector
Households with smart meter
Collector
Individual
Consumption
Reports
Internet
Cellular
Meter Data
Management
Operation
Center
Customer
Energy
Companies
(Utilities)
Customer
Fig. 2 Flow of information in the smart meter
Fig. 3 Risk assessment in smart grid
There are different levels of attacks in the smart grid to cause security threat.
These levels can be classified based on the networks,
•Home Area Network Attacks
•Neighborhood Area Network Attacks
•Wide area Network Attacks
Based on these network attacks, different levels of security protection will be given.
This is shown in Fig. 5. Proposed method concentrates on the HAN and WAN level
attacks. In the HAN level, smart meter was designed and data transmitted from
customer household will be encrypted by E2E encryption via WAN [4]. So the
security will be given in both HAN and WAN. Online database will be used to store
4 M. Manimegalai and K. Sebasthirani
Confidentiality
Integrity
Availability
Smart grid systems ,
Assets and Operations
Fig. 4 CIA triad for smart grid systems
Data Security
Application Security
HAN level Security
NAN level Security
WAN level Security
Fig. 5 Levels of security protection in smart grid
the customer information, and distributor also can make the security protection over
this information. Distributors also do the E2E decryption to view the data in the
WAN.
2 Cybersecurity Issues in Smart Meter
From household, the information like electricity usage of the single house, amount for
the used electricity, etc. can be transmitted to the main server. The distributor also can
Performance Analysis of Smart Meters for Enabling a New Era … 5
Internet
Consumer/ Producer Metering Service Provider
Distribution Network Operator (DNO)
Generation of
Controllable
Load
LMN
(OMS)
HAN
Gateway
Low Voltage
Network
Medium
Voltage
Network
Substation Control
Center
LAN
Web Portal Central IT
Head End
Systems
Anonymized Metering Data
Fig. 6 Smart metering system
view these details of the customer and access control mechanisms applied. Security
threats which affect the confidentiality, integrity, availability, and non-repudiation
can occur in the smart meter from the customer and also from the distributor. The
information flow from the customer to the central server is shown in Fig. 6.
Smart meter information from different households is sent through Internet using
Wi-Fi, ZigBee, or HomePlug technologies. Smart meter communicates in the home
area network and the neighborhood area network. Information flow from household
to the center control is shown in Fig. 7.
When information flow from the HAN to NAN or WAN, attacker can masquerade
as a legitimate meter data management, and he can change the data or view the
confidential information. For example, the amount calculated for the used electricity
can be changed (maximized or minimized) by the attacker [5]. Attacker can also
Third Party
Home Area
Network 1
Home Area
Network 2
Home Area
Network N
Smart Meter 1
Smart Meter 2
Smart Meter N
Substation/
Data
Concentrator
Utility
Company
Data Center
SSL
SSL
SSL
WiFi
ZigBee
Homeplug
WiFi
ZigBee
Homeplug
WiFi
ZigBee
Homeplug
WiFi
ZigBee
PLC
WiFi
ZigBee
PLC
WiFi
ZigBee
PLC
WiFi
4G LTE
PLC
Fig. 7 Usage of information and communication technology in smart meter
6 M. Manimegalai and K. Sebasthirani
change the commands to control the system, deny access to the distributor system,
and meter unable to access the critical information from the legitimate system.
Security threats caused can be occurred because of the information transmitted on
the Internet. So attacker can easily hack the center control of the system. Attack can be
occurred in both physical and software forms. Physical threats can easily be detected.
But software attack can be detected after it occurred. In order to avoid such threats
[6], various security mechanisms have been proposed: encryption, digital signature,
firewall, access control, and trusted party. Proposed method has concentrated on the
E2E from customer side and also from distributor side.
3 Proposed Method
From each household, smart meter information sent via Wi-Fi technology to the
distributor. Customer and the distributor can monitor the system through online.
Customized smart meter was designed, and the details are sent to the distributor.
Proposed system consists of the following functionalities:
•User can view the usage of electricity and amount need to pay at particular interval
through online.
•Meter will show the electricity usage and amount at any time.
•Data transmitted from customer to distributor and distributor to the customer will
be encrypted by E2E algorithm.
•Other side needs to do the E2E decryption algorithm to view the details.
•Two-way communication and ICT will be implemented via Raspberry Pi and
Arduino controller.
•By monitoring systems online, user can control the usage of electricity based on
their needs.
Simplified architecture diagram of the proposed system is shown in Fig. 8.
Smart meter reads the current value from the customer household by the current
transformer using the formula,
Power =Current ×Voltage (2)
Voltage value remains constant in all the connections 230 V. For the calculated
power, amount to be paid calculated displayed in the smart meter. Current transformer
calculates current in analog form; this can be converter to digital or serial by Arduino
nanocontroller. Current transformer output is connected to the Arduino in A0 pin.
Then, Raspberry Pi B+ is connected to the broadband. Output from Arduino is
connected to the USB to TTL (Transistor-to-Transistor Logic) converter in Raspberry
Pi B+. Using Python language, digital output of the Raspberry Pi is displayed. PHP
language is used to store the output from Raspberry Pi [7] to the online database.
From the online database, both the customer and the distributor can view the details.
Performance Analysis of Smart Meters for Enabling a New Era … 7
Current Transformer Arduino Nano USB to TTL Logic
Converter
Rasperry Pi B+
Internet WAN
Power Distributor Database
Input From Household
Fig. 8 Block diagram of the proposed system
4 Cybersecurity in Smart Meter
In cryptography [8] and network security, more encryption algorithms were proposed.
Particularly, end-to-end encryption algorithm will provide better security in terms
of confidentiality, integrity, and availability. In both, the end encryption is done
when sending the information. Do the corresponding decryption to view the original
information. Figure 9shows the flow of information from customer to the distributor.
When the user login into the system, relay will be on. Power and amount values
are calculated and after the E2E encryption [9]; values are stored in online database.
Start
Login to the
Smart meter
Success
Switch on
Relay
Calculate Power
and Price
E2E Encryption
Store to database
Website user or
admin Login
Success
E2E Decryption
View the Original
Information
Try to do Correct
Login details
Yes
Yes
No
No
Fig. 9 Flow of E2E encryption and decryption
8 M. Manimegalai and K. Sebasthirani
In online, user or the distributor can view the information by doing E2E decryption
after proper login. Here, MD5 [10] algorithm is used for encryption and decryption.
The following sections will discuss the hardware setup used in the smart meter.
5 Results and Discussion
The proposed system was implemented in Python and PHP. Hardware setup of the
smart meter is shown in Fig. 10.
Connection of current transformer, relay, Raspberry Pi, Arduino controller, and
OLEDisshowninFig.10. In the proposed system, Raspberry Pi B+ model was used
as shown in Fig. 11.
Relay will be enabled only after proper login of the user in the Raspberry Pi. User
login page is shown in Fig. 12. User needs to give username, password, and service
number as input to the system. This information will be given by distributor only.
Figure 13 shows the details of entered page of user.
After relay on, power can be calculated. For each and every minute, power and
price values are shown in the smart meter. This can be shown in Figs. 13 and 14.
Data are stored in online database which is unlimited. User and the distributor
have to login into the system to view the details. Figure 16 shows the login page in
website (Fig. 15).
After login, user can view the details stored in the database as shown in Fig. 17.
Fig. 10 Hardware setup of smart meter
Performance Analysis of Smart Meters for Enabling a New Era … 9
Fig. 11 Rasperry Pi B+ model
Fig. 12 User login page in Raspberry Pi
User can filter the information based on the date as shown in Fig. 18.Dataare
in the encrypted form only. Distributor also can view the details in encrypted form;
after decryption, he can view the details.
10 M. Manimegalai and K. Sebasthirani
Fig. 13 Details of user entered in login page
Fig. 14 Calculation of power and price in smart meter
Fig. 15 Calculation of power and price
Performance Analysis of Smart Meters for Enabling a New Era … 11
Fig. 16 User and admin login page
Fig. 17 Online database viewed by user
Fig. 18 Online database viewed by user after filtering applied
12 M. Manimegalai and K. Sebasthirani
6 Conclusion and Future Work
Customized user-friendly smart meter was designed and Internet of Things (IoT)
concepts have been implemented. Power and amount to be paid have been calcu-
lated based on the user needs. This can be seen digitally in the smart meter itself.
Raspberry Pi B+ is used to transmit the data from the meter to the online database.
Python language used to convert the data from meter to the database. MD5 algo-
rithm was used to encrypt data from end to end. Service number was given as the
additional authentication information to the user. Distributor no needs to enter the
service number. Thus, information from customer to the distributor was sent through
WAN with end-to-end encryption. Encryption was given in HAN and WAN level
networks.
In future, smart meter can also include power quality improvement techniques
like the elimination of harmonics, flicker, and voltage sag or swell. Through Wi-Fi
and ZigBee technologies, power quality improvement monitoring can be done in
dynamic mode.
References
1. NIST Special publication 1108, “NIST Framework and Roadmap for Smart Grid Interoper-
ability Standards”, Release 1.0, Jan 2010
2. C. Kaufman, R. Perlman, M. Speciner, Network Security: Private Communication in a Public
World (Prentice Hall Press, 2002)
3. S. Clements, H. Krishnan, Cyber security considerations for the smart grid, in 2010 IEEE
Power and Energy Society General Meeting (2010), pp. 1–5
4. A. Giani, E. Bitar, M. Garcia, M. McQueen, P. Khargonekar, K. Poolla, Smart grid data integrity
attacks: characterizations and countermeasures, in 2011 IEEE International Conference on
Smart Grid Communications (SmartGridComm) (2011), pp. 232–237
5. S. Ranjan, R. Swaminathan, M. Uysal, A. Nucci, E. Knightly, DDoS-shield: DDoS-resilient
scheduling to counter application layer attacks. IEEE/ACM Trans. Netw. 17(1), 26–39 (2009)
6. C. Bekara, Security issues and challenges for the IoT-based Smart Grid. Procedia Comput. Sci.
34, 532–537 (2014), The 9th International Conference on Future Networks and Communica-
tions (FNC’14)
7. A. Metke, R. Ekl, Security technology for Smart Grid networks. IEEE Trans. Smart Grid 1(1),
99–107 (2010)
8. Y. Wang, D. Ruan, D. Gu, J. Gao, D. Liu, J. Xu, F. Chen, F. Dai, J. Yang, Analysis of Smart
Grid security standards, in 2011 IEEE International Conference on Computer Science and
Automation Engineering (CSAE), vol. 4 (2011), pp. 697–701
9. P. McDaniel, S. McLaughlin, Security and privacy challenges in the Smart Grid. IEEE Secur.
Priv. 7(3), 75–77 (2009)
10. V. Delgado-Gomes, P. Borza, A biological approach for energy management in smart grids and
hybrid energy storage systems, in 2014 International Conference on Optimization of Electrical
and Electronic Equipment (OPTIM) (2014), pp. 1082–1086
Energy Efficient Data Centre Selection
Using Service Broker Policy
Sameena Naaz, Iffat Rehman Ansari, Insha Naz, and Ranjit Biswas
Abstract Over the past few years human beings have become completely depen-
dent on computers and IT technologies, which in turn have led to the issue of energy
and power consumption in IT industries. Since the energy cost as well as the elec-
trical requirements of the industry has increased drastically throughout the world,
therefore it has become necessary to shift our focus to Green computing, which
refers to environmentally sustainable computing and aims to limit energy and power
consumption and shrink costs thus maximizing the efficiency of the system. Green
Computing provides the proficient use of computing power.
Keywords Green computing ·Power consumption ·Sustainable computing
1 Introduction
Owing to the escalating growth of internet users and falling price of computer hard-
ware various fields like finance, medical, education and many more are becoming
heavily dependent on the services offered by IT technologies [1]. All this has resulted
in the production of huge amount of power and energy consumption which in turn
not only produces great amount of CO2which is harming the environment badly
also the increasing energy costs is taking toll on the IT industries. Green computing
S. Naaz (B)·I. Naz ·R. Biswas
Department of Computer Science and Engineering, School of Engineering Sciences and
Technology, Jamia Hamdard, New Delhi, India
e-mail: snaaz@jamiahamdard.ac.in
I. Naz
e-mail: inshanaz_sch@jamiahamdard.ac.in
R. Biswas
e-mail: rbiswas@jamiahamdard.ac.in
I. R. Ansari
University Women’s Polytechnic, Aligarh Muslim University, Aligarh, Uttar Pradesh, India
e-mail: iffat_rehman2002@yahoo.co.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_2
13
14 S. Naaz et al.
tries to provide solution to these alarming issues. One of the vital segments of green
computing is to focus on how to reduce the carbon footprint and save the energy. It
covers the complete lifespan of the computing starting from designing to the disposal
of e-waste in such a way that there is very less or no effect on the environment [2].
Green computing provides various approaches that help the IT technology firms to
tackle crucial computing requirements in a more sustainable way so that the damage
on environment and resources is either reduced or eliminated hence reducing the
carbon emission and energy consumptions.
Rest of the paper is organized as follows: Sect. 2discusses the reasons for the need
of going green, Sect. 3gives a review of the latest literature in this area. Section 4
is dedicated to current trends in green computing, Sect. 5highlights the challenges
or hurdles in the way of going green and finally Sect. 6gives the conclusions drawn
from this study.
2 Need for Greening
We need to go green to increase the energy efficiency and reduce the resource
consumption so that we can achieve the goal of limiting the carbon footprint. The
main aim of green computing is to provide a healthier computing environment. No
doubt the use of computing services and IT has made our lives simpler and our tasks
easier. But the fact that due to increase in usage, power consumption also increases
can’t be ignored, which in turn increases the production of greenhouse gasses (CO2).
Thus a data centre needs an efficient cooling system, but if the data centre doesn’t have
an appropriate and capable cooling system then there will be a loss of energy which
will lead to the environmental degradation [3]. Some of the other issues that amplify
the green IT movement are managing harmful e-waste, rising gasoline expenditures
and growing real estate costs [4]. As a result, there is a dire need to achieve or go
green to save our environment.
3 Literature Review
A basic transformation happening in the area of IT nowadays is cloud computing. The
key element of cloud computing is virtualization which brings ease and efficiency.
Virtualization increases the security of cloud computing, shielding the reliability of
guest VM’s and the cloud infrastructure components [5,6].
Since data centres that host the Cloud applications consume a large amount of
energy, which results in increased costs and carbon emission rates, therefore, we
require green cloud computing that reduces functional costs and minimizes the
environmental impact [7,8].
Energy Efficient Data Centre Selection Using Service … 15
Due to the increasing demand of cloud computing, a large number of companies
are investing in constructing huge data centres to host Cloud services. These data
centres consume huge amount of energy and are very complex in the infrastructure.
In order to make data centre energy efficient various technologies like virtualization
and consolidation are employed [9].
4 Current Trends in Green Computing
In order to address the impact of IT on the environment, we need to implement
an integrated procedure that can solve the problems using different paths from the
manufacture to disposal of the devices. Everything needs to be green like green use,
green designing, green manufacturing and green disposal thereby making the whole
lifecycle of IT greener [10].
Different trends in green computing given in Fig. 1are discussed.
4.1 Virtualization
It refers to the creation of virtual version of computer resources, for example, creation
of multiple logical versions of a single physical resource or hardware [11]. With
virtualization, solo equipment can operate as multiple equipment’s running individ-
ually. It is the best possible way to greening and conserving the space, resources and
environment. Management software and virtualization software both are provided for
Fig. 1 Current trends in cloud computing
16 S. Naaz et al.
efficient working of a virtualized environment.Virtualization directs to Server consol-
idation. The security of the system is also enhanced. It permits complete consump-
tion of resources, reduces the overall quantity of hardware used, switches off the
idle machines to conserve energy and also the cost of space and rent is decreased
[2]. Cooling expenses can also be curtailed by making the proficient use of available
resources [12].
4.2 Green Data Centres
The term “Green Data centres” means designing an algorithm and infrastructure
for data centres that provide efficient use of the resources that are available, help
in reduction of cooling costs and tackle energy consumption issues [13]. Owing
to the good environmental and financial impact, green data centres have gained
substantial focus. As the usage of the internet is increasing day by day so is the
power consumption in our data centres which results in higher energy costs and
environmental degradation. So in order to overcome this issue, IT companies need
to necessarily go green and build a sustainable data centres to save the energy, costs
and environment. According to the US department of energy, an efficient data centre
should focus on five basic areas while designing a data centre, cooling system,
conditions of the environment, management of air, IT systems and electrical system,
other things that need to be kept in mind is e-waste recycling [11,14].
4.3 Power Consumption
Reducing the power consumption is one of the major issues these days so in order
to tackle this IT industries have started making use of the devices that are efficient
in saving the energy. According to the environment protection agency, “around 30–
40% of personal computers are kept ‘ON’ after office hours and during the weekend
and even around 90% of those computers are idle” [15]. There are different ways in
which the power usage can be limited like switching off the system when it’s not in
use or sending the monitor in low power state. The use of LCD and LED can also
be helpful. Hibernate and standby mode saves the 80–90% of the power so devices
should be put in these modes when idle [16]. Combining the efficient coding and
efficient algorithms can also provide great results in saving energy [17].
4.4 E-Waste Recycling
Since electronic devices and computer systems contain toxic metals and poisonous
substances like lead and mercury that can be harmful to the environment so we need
Energy Efficient Data Centre Selection Using Service … 17
to find the proper ways for their disposal [16]. Reusing the discarded devices or old
computers can save a lot of energy and also help in combating the harmful effect of
e-waste on the environment [15]. Recycling can be done in many ways like taking
the components of old computer systems and using them for repairing or upgradation
purpose. Also the practice of changing the computers every 2–3 years needs to be
changed for the benefit of cost and environment [17].
Growing use of technology nowadays has led to the creation of an enormous
quantity of electronic wastes resulting in environmental degradation; therefore, the
safety of environment and keeping the check on environmental pollution has become
the chief concern of scientists all over the world. The most important concern related
to e-wastes is that these are non-biodegradable and their dumping has led to the
accretion of toxic material like lead, cadmium, etc. in the environment resulting in
global warming and contamination of the soil and groundwater. Thereby, disturbing
the plant and animal life which in turn has an effect on the entire living organisms
yielding harsh health risks and disorders. Increasing global warming and mounting
energy expenses has led the government as well as the private organizations to think
and examine different ways to safeguard the environment worldwide [18].
4.5 IT-Practises and Eco-Labelling
For companies to create products to be given eco-label, different policies need to be
introduced all over the world. Many organizations in the world support eco-labelling
and provide certificates to IT products on the basis of several features like energy
consumption, recycling, power consumption, etc. [12].
Eco-labelling basically originated because of the increasing environmental
concerns worldwide. Labels like eco-friendly, recyclable, energy efficient, reusable,
etc. attracted consumers who were finding ways to reduce the impact of hazardous
material on the environment.
Eco-labels give the information regarding the existence or nonexistence of partic-
ular feature in any product. It enables the customers to get an insight about the
environmental quality of items at the time of purchasing, allowing them to pick
or buy products that are suitable from an environmental point of view. Therefore,
minimizing the utilization of harmful substances that may be detrimental to the
environment [19].
Companies have to make sure that they manufacture and design products in such
a way that they can obtain the eco-label. Many organizations grant certificates to IT
products after reviewing the features like energy and resource consumption, recycling
and refurbishing, etc. thereby enabling the customers to make the environmentally
suitable decision at the time of purchasing a product [20].
18 S. Naaz et al.
4.6 Green Cloud Computing
The research carried by Pike Research, “the wide-spread adoption of cloud computing
could lead to a potential 38% reduction in worldwide data centre energy expenditures
by 2020” [21]. Cloud computing has proved to be a vital and sound means for
virtualization of data centres and servers so that they can be a resource as well as
energy efficient. Due to the high consumption of power and energy in IT firms, which
produces the harmful gases in the environment resulting in global climate changes.
As a result, there is a dire need of cloud computing to go green [22,23].
Cloud computing is one of the most important paradigm in modern world because
of the fact it has dynamic, high-powered computing abilities, with access to intri-
cate applications and data archiving, with no requirement of any extra computing
resources. Since cloud computing offers reliability, scalability and high performance
at fewer expenses thus cloud computing technologies have a diversity of applica-
tion domains. By offering promising environmental protection, economic and tech-
nological advantages, cloud computing has revolutionized the modern computing.
These technologies have the potential to improve. The various technologies of cloud
computing like energy efficiency, reduced carbon footprints and e-waste can convert
cloud computing into green cloud computing [24,25] (Table 1).
Tabl e 1 Summary
Approach Observation
Virtualisation Creates virtual version of computer resources like creating
multiple logical versions of a single physical resource or
hardware
Green data centre To design an infrastructure for data centres which makes
efficient use of the available resources thus helping in
reduction of cooling costs
E-Waste Recycling Recycling can be done in many ways like taking the
components of old computer systems and using them for
repairing or upgradation purpose thereby saving the cost and
tackling the hazards related to e-wastes
IT-Practises and Eco-Labelling Labels like eco-friendly, recyclable, energy efficient, reusable,
etc. attracted consumers who were finding ways to reduce the
impact of hazardous material on the environment
Green cloud computing The various technologies of cloud computing like energy
efficiency, reduced carbon footprints and e-waste can convert
cloud computing into green cloud computing
Power consumption There are different ways in which the power usage can be
limited like switching off the system when it’s not in use or
sending the monitor in low power state
Energy Efficient Data Centre Selection Using Service … 19
5 Proposed Technique for Energy Efficiency in Virtual
Machines
This work is an extension of the work carried out in [26]. Here various load balancing
algorithms have been implemented on Cloud Analyst to study the energy efficiency.
The proposed algorithm has been simulated on Cloud Analyst, and its result has been
compared with that of throttled and round-robin algorithm.
There are three parts in this load balancing algorithm.
(i) The energy consumption of each VM is calculated.
(ii) The most efficient VM is found.
(iii) ID of the most efficient VM is returned.
Simulation package CloudSim [27] has been used for the experiments. A single
data centre having 100 virtual machines (VM memory—1024 Mb, PM speed—100
MIPS, Data Centre OS—Linux) has been used.
6 Results and Discussions
CloudSim toolkit has been used for simulating the three algorithms. The parameters
studied here are the overall response time, data centre processing time and the energy
consumption.
Average Response Time and Energy Consumption for various algorithms are
shown in Table 3.
Table 2shows that the average processing time at data centre is much higher for
round robin and throttled algorithm as compared to that of Service Broker-based VM
load balancing policy. Response time for any user query in round-robin scheduling
and throttled policy is much higher as compared to VM load balancing policy as can
be seen from Table 3. The same results are also depicted in Figs. 2and 3, respectively.
Tabl e 2 Request processing
time for different algorithms Algorithm Data centre Request processing time (ms)
Round-robin DC1 20.499
DC2 56.296
Throttled DC1 20.631
DC2 55.859
Service broker DC1 10.575
DC2 28.659
20 S. Naaz et al.
Tabl e 3 Average response
time and energy consumption
for various algorithms
Algorithm Average response time
(ms)
Energy consumption
(mW)
Round robin 123.98 10.49
Throttled 123.85 8.33
Service broker 109.54 4.01
DC1 DC2 DC1 DC2 DC1 DC2
Round Robin Throttled Service Broker
20.499
56.296
20.631
55.859
10.575
28.659
Request Processing Time (ms)
Fig. 2 Request processing time for different algorithm
0
50
100
150
Round Robin Trottled Service Broker
Chart Title
Average Response Time (ms) Energy Consumption (mW)
Fig. 3 Average response time and energy consumption for various algorithms
7 Conclusion
In this paper, an energy-efficient service broker-based policy for data centre has been
implemented in CloudSim cloud computing environment. The proposed technique
first calculates the expected response time of each virtual machine and then sends
the identity of that machine to the data centre for allocation of any new request. The
Energy Efficient Data Centre Selection Using Service … 21
average request processing time and energy consumption are also calculated for the
three algorithms, viz, round-robin, throttled and service broker based, and it has been
observed that service broker-based algorithm performs the best in terms of all the
three parameters.
References
1. Q. Li, The Survey and Future Evolution of Green Computing, pp. 0–3 (2011)
2. T.R. Soomro, M. Sarwar, Green computing: from current to future trends, vol. 6, no. 3, pp. 326–
329 (2012)
3. P. Malviya, S. Singh, A study about green computing. Int. J. Adv. Res. 3(6), 790–794 (2013)
4. R. Sharma, Approaches for green computing. Int. J. Innov. Comput. Sci. Eng. 2(3), 52–55
(2015)
5. Y. Xing, Y. Zhan, Virtualization and cloud computing, pp. 305–306 (2012)
6. F. Lombardi, R. Di, C. Nazionale, D. Informativi, P.A. Moro, Secure virtualization for cloud
computing. J. Netw. Comput. Appl. 34(4), 1113–1122 (2011)
7. A. Beloglazov, J. Abawajy, R. Buyya, Energy-aware resource allocation heuristics for efficient
management of data centers for Cloud computing. Futur. Gener. Comput. Syst. 28(5), 755–768
(2012)
8. S. Naaz, A. Alam, R. Biswas, Effect of different defuzzification methods in a fuzzy based load
balancing application. Ijcsi 8(5), 261–267 (2011)
9. S.K. Garg, C.S. Yeo, R. Buyya, Green cloud framework for improving carbon, pp. 491–502
10. S. Murugesan, Harnessing green IT: principles and practices—adopting a holistic approach
to greening IT is our responsibility toward creating a more sustaining environment. Green
Comput. 24–33 (2008)
11. R. Thomas, Approaches in green computing
12. D. Ranjith, G.S. Tamizharasi, B. Balamurugan, A survey on current trends to future trends in
green computing, in Proceedings of International on Conference Electronics, Communication
and Aerospace Technologies ICECA 2017, vol. 2017 (2017), pp. 632–637
13. X. Jin, F. Zhang, A.V. Vasilakos, Z. Liu, Green data centers: a survey, perspectives, and future
directions (2016)
14. S. Naaz, A. Alam, R. Biswas, Load balancing algorithms for peer to peer and client server
distributed environments. Int. J. Comput. Appl. 47(8), 17–21 (2012)
15. I. AlMusbahi, O. Anderkairi, R.H. Nahhas, B. AlMuhammadi, M. Hemalatha, Survey on green
computing: vision and challenges. Int. J. Comput. Appl. 167(10), 4–6 (2017)
16. N. Mumbai, N. Mumbai, Green computing: an essential trend for secure, Apr 2013, pp. 19–20
17. S. Singh, Green computing strategies & challenges, in Proceedings of 2015 International
Conference on Green Computing Internet Things, ICGCIoT 2015, no. 1, pp. 758–760 (2016)
18. R. Panda, E-waste management: a step towards, vol. 4, no. 5, pp. 417–424 (2013)
19. Green IT
20. K. Saikumar, Design of data center, vol. 2, no. 6, pp. 147–149 (2014)
21. K. Sourabh, S. M. Aqib, A. Elahi, Sustainable green computing: objectives and approaches,
pp. 672–681
22. Y.S. Patel, N. Mehrotra, S. Soner, Green cloud computing: a review on green IT areas for cloud
computing environment, pp. 327–332 (2015)
23. I. Naz, S. Naaz, R. Biswas, A parametric study of load balancing techniques in cloud
environment
24. L. Radu, SS symmetry Green Cloud computing: a literature survey (2017)
25. S. Naaz, A. Alam, R. Biswas, Implementation of a new fuzzy based load balancing algorithm
for hypercubes. Int. J. Comput. Sci. Inf. Secur. 270–274 (2010)
22 S. Naaz et al.
26. S. Naaz, I.R. Ansari, A. Khan, Load balancing of virtual machines using service broker
algorithm. Int. J. Adv. Technol. Eng. Sci. 4(9), 232–238 (2016)
27. R. Buyya, R. Ranjan, R.N. Calheiros, Modeling and simulation of scalable Cloud computing
environments and the CloudSim toolkit: challenges and opportunities, in 2009 International
Conference on High Performance Computing & Simulation (2009), pp. 1–11
Co-ordinate Measurement of Roll-Cage
Using Digital Image Processing
Ritwik Dhar, Parth Kansara, Sanket Shegade, Atharv Bagde,
and Sunil Karamchandani
Abstract This paper proposes a unique and novel approach for co-ordinate measure-
ment of an all-terrain vehicle. The manufactured roll-cage is replete with minute
errors which are introduced during the process. The suspension points on the roll-
cage have been considered as dataset for this algorithm, as they are an integral part in
the performance of the vehicle and for maintaining suspension geometry. A feasible
method using image processing techniques such as shade correction, adaptive bina-
rization, and segmentation has been used to improve fault tolerance by analyzing and
reducing the manufacturing errors with respect to the computer-aided design model.
A MATLAB script has been developed for the proposed method with the help of
image processing toolbox.
Keywords Co-ordinate measurement ·Computer vision ·Adaptive thresholding ·
Bounding box ·Object detection ·Edge detection ·Binarization ·Integral
images ·Shape detection
R. Dhar ·A. Bagde ·S. Karamchandani
Electronics & Telecommunication, Dwarkadas J. Sanghvi College of Engineering, Mumbai
400056, India
e-mail: ritwik1798@gmail.com
A. Bagde
e-mail: atharva17bagde98@gmail.com
S. Karamchandani
e-mail: skaramchandani@rediffmail.com
P. Ka n s a r a (B)
Information Technology, Dwarkadas J. Sanghvi College of Engineering, Mumbai 400056, India
e-mail: parthskansara@gmail.com
S. Shegade
Mechanical Engineering, Dwarkadas J. Sanghvi College of Engineering, Mumbai 400056, India
e-mail: sanketshegade@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_3
23
24 R. Dhar et al.
1 Introduction
Dimensional accuracy of fabricated components is an important factor, particularly
in instances where they are to be in contact with each other. The need to fabricate
within tolerances imposes accuracy in measurement. In addition, to ensure accuracy
in any outcome, traceability must be maintained. Such precision in measurements is
required in the manufacturing of all-terrain vehicles, to fulfill the desired suspension
geometry. Failure in incorporating this precision results in ample deviation in the
position of the in-board and out-board points of the control links. These control links
decide the point of application of the reaction force from the ground to the vehicle
body. Thus, co-ordinate measurement of these points in the spatial domain helps
in quantifying the manufacturing anomalies by calculating the deviations from the
computer-aided design model, which is synonymous to degradation in the overall
stability of the vehicle. Due to the limitations of the commercial co-ordinate measure-
ment machines, an unconventional yet reliable approach has been implemented with
the help of images to process the distance of the chassis points of vehicle with respect
to a reference point [3].
Wojnar [5] and Zhao et al. [6] applied Fourier Transform as a solution to reduce the
noises in periodic signals. To remove noise, they can filter high frequencies and keep
just the low frequencies, so the filters that eliminate noise are most often pointed as
low-pass filters. Thresholding is one of the most common techniques for segregating
objects from the background of the image, Slezak et al. [7]. According to Cheriet
et al. [8], Thresholding can be used to differentiate text from the background, e.g., by
filtering the areas above or below the threshold to a certain gray value in a greyscale
image.
Corners in images have distinct characteristics that clearly distinguish them from
the pixels around them. There are many algorithms for detecting corners that are
accurate even after the image is geometrically distorted, Jeong and Moon [9]. Object
detection algorithms often use the position of corners in the image, Harris corner
detector is prevalent in this class. The most successful edge detector is the Canny
edge detector which aims at the three basic criteria, i.e., good localization of edge
points, low error rate, and avoidance of similar detector results from the same edge,
Tang and Shen [10]. The algorithm’s final phase consists of selecting edge points
using a double threshold. According to Rong et al. [11] and Li et al. [12] using double
threshold is an effective way to reduce the noise in the last stages of edge detection.
Our approach has been chosen on the basis of its ease of setup and also the
optimization with the help of different algorithms involved has resulted in better
outputs. Aldasoro et al. [1] described an algorithm for removing the image shading
element by estimating the signal envelope. This technique was implemented for shade
correction to normalize the light intensity throughout the image. Bradley et al. [2]
presented an extension of Wellner’s work but reduced the computational complexity
by using integral images rather than the moving average throughout the pixels.
Followed by binarization we perform threshold-based segmentation to extract the
required interest region from image with the help of shape detection. The distance
Co-ordinate Measurement of Roll-Cage Using Digital … 25
Fig. 1 Algorithm flow chart
from the region of interest, i.e., the suspension points in the front box used for analysis
are calibrated in pixels with respect to reference object of pre-determined dimensions
placed in the images to get the final results (Fig. 1).
2 Pre-processing
2.1 Shade Correction
It is not easy to obtain a digital image with uniform illumination over the entire pixels.
The light due to reflectance and illuminance components in the image background
is not generally uniform. The technique through which images have been captured
as well as the relationship among various objects present throughout the field of
vision and camera illumination often contributes to cases in which the image shows
irregular patterns of shading throughout the image. The basic approach toward shade
correction is using simple low-pass filters or morphological operators due to the ease
of use and universal applicability. The above techniques are effective though fail in
cases when the distortion is bigger than the background, basically it assumes only
two cases with background being below or above the median pixel value than the
region of interest. Aldasoro et al. [1] described an algorithm for removing the image
shading element by estimating the signal envelope. The primary feature of using
such a technique was that it did not make any assumptions about whether the objects
were of lesser or greater intensity than the background and worked well with objects
of all sizes.
It was presumed that the shaded image used, I(x,y) was created by an additive
shading element S(x,y) that damaged an initially unbiased image U(x,y).
I(x,y)=U(x,y)+S(x,y)(1)
26 R. Dhar et al.
We can also say that the corrected image Û(x,y), which was an estimation of U(x,
y), was obtained by
U(x,y)∼
=ˆ
U(x,y)=I(x,y)−S(x,y)(2)
The effects of noise are first minimized by low-pass filtering the input signal with
a3×3 Gaussian kernel. The envelope is then obtained by comparing the intensity
of every pixel against the average of opposite eight-connectivity neighbors with
increasing the distance of pairs in 45° directions from 0° to 135°. Two new envelope
series Smax/Smin were obtained by changing the existing intensity with the intensity
of the pixel by the maximum/minimum value of the comparison of each neighboring
pixel. The upper envelope Smax can be produced by replacing the maximum value of
the averages with the pixel as given by
Si
max(x,y)=max⎧
⎪
⎪
⎨
⎪
⎪
⎩
I(x−di,y−di)+I(x+di,y+di)
2
I(x+di,y−di)+I(x−di,y+di)
2
I(x−di,y)+I(x+di,y)
2
I(x,y−di)+I(x,y+di)
2,I(x,y)
⎫
⎪
⎪
⎬
⎪
⎪
⎭
(3)
The replacement with respect to the minimum value results in the lower enve-
lope Smin. The max/min values formed two stacks from which the maximum inten-
sity projection corresponded to the current envelope estimation: Both Simax =Simin
surfaces have been filtered with a low-pass Gaussian filter of size equal to the range di
to limit the measurement of the envelope to the intermediate orientation pixels. The
method was replicated by expanding diand filter size, enabling the envelope to be
adapted to all objects. To determine a stop criterion, magnitude and local derivatives
were calculated.
∂si
max
min
∂x,
∂si
max
min
∂y(4)
In each iteration of di,MGi
tot was compared with the i−1 gradient MGi−1
tot , and
when MGi
tot−MGi−1
tot
MGi−1
tot
<0.01 the iterations were stopped. Finally, the more smoother
surface, either Si
max or Si
min which had the lower MGi
tot, was used as the shading S.
MGi=∂Simax
min
∂x2
+∂Simax
min
∂y2
MGi
tot =
x,y
MGi
1
2
(5)
Figure 2a, b here illustrates the image before and after the shading correcting
algorithm has been applied.
Co-ordinate Measurement of Roll-Cage Using Digital … 27
Fig. 2 a Input image.
bOutput image from shade
correction algorithm
3 Co-ordinate Measurement
The results from the shade correction algorithm are further used for two applications:
A. Calibration factor from reference object using edge detection
B. Binarization for co-ordinate measurement.
3.1 Calibration Factor from Reference Object Using Edge
Detection
The output as shown in Fig. 2b has been used for computing the calibration factor.
A plate with pre-determined measurements had been placed along the axis of the
roll-cage and camera to produce no optical errors. According to Kainz et al. [4], the
dimensions of an object in a 2D image can be estimated by taking into account its
position from capturing element and the positional angles involved. Taking two cases
of input image with the reference object and without the reference object gives the
alignment of the object. In our case, the reference object is carefully placed in the
same plane as the capturing element is helping in reducing the errors.
As shown in Fig. 3, the width is estimated using the following relations:
w=α×y+β(6)
28 R. Dhar et al.
Fig. 3 Positional angles and measurements
α=m1×cam
real −b1(7)
β=m2×cam
real −b2(8)
Image width =cam ×w+b1×y+b2
m1×y+m2
(9)
Here, m1,b1,m2,b2are constants analyzed from previous tests performed, wis
the object width in pixels, αand βgive the xand yintercept values, cam gives the
distance of camera from ground, and real is the actual object width.
The length of the edge is then obtained in pixels and the calibration factor is
obtained as (Eq. 10)
Calibration Factor =Image width
real (10)
3.2 Binarization For Obtaining Region of Interest &
Distance Measurement
After the shade correction to produce uniform illumination, the image from Fig. 2b
is binarized, i.e., all the pixels are given values of 0 or 1 with respect to a threshold
that is to be determined for optimum results. The task of determining the threshold
value is of utmost importance in this stage. While it is generally known that a single
fixed threshold value will not yield optimum results in non-uniformly illuminated or
damaged images, an adaptive threshold technique is used. Bradley et al. [2] presented
Co-ordinate Measurement of Roll-Cage Using Digital … 29
Fig. 4 a Reference object. bDimensions
an extension of Wellner’s work but reduced the computational complexity by using
integral images rather than the moving average throughout the pixels. Integral images
do not take into consideration the neighborhood size and help in fast summation of
pixels. These are calculated using
I(x,y)=f(x,y)+I(x−1,y)+I(x,y−1)−I(x−1,y−1)(11)
x2
z=x1
y2
y=y1
f(x,y)=I(x2,y2)−I(x2,y1−1)−I(x1−1,y2)
+I(x1−1,y1−1)(12)
The integral image using Eqs. 11 and 12 are compared with a s x s average integral
image pixel to pixel. The value is set to black, if it is t percent less than average and
white otherwise.
Figure 5a shows the results from binarization using adaptive thresholding from
which we have obtained the region of interest (ROI), i.e., the suspension points that
are to be taken into consideration for co-ordinate measurement. The white pixel
arrays are the marked points for computation of distance as shown in Fig. 5b with
30 R. Dhar et al.
Fig. 5 a Region of Interest extraction. bCentroid detection using regionprops
centroid marked with blue points. Regionprops function has been used from image
processing toolbox to measure the centroid of the white pixel arrays whose co-
ordinates are saved in matrix Centroid. Any co-ordinates are chosen as the reference
co-ordinate from Centroid matrix and the distances are computed between each point
in pixel length.
The actual distance makes use of the calibration factor computed in Sect. 3.1:
Actual distance (in cm)=Image width ×calibration factor (13)
4 Observation
The all-terrain vehicle was also appointed to industrial co-ordinate measurement for
roll-cage. The readings in Table 1given below gives the readings obtained from the
CAD design, the industrial co-ordinate measurement, and the readings from output
of MATLAB code for some of the points used to determine the accuracy of the
algorithm and technique with industrial standards and ideal requirements. For ease
of understanding, the points have been named as shown in Figure 6.
The tab points, i.e., A-B, E-F, C-D, G-H, I-J give appropriate results with compar-
atively larger deviations as seen in case of E-F, G-H, and I-J tending to >0.01%
Co-ordinate Measurement of Roll-Cage Using Digital … 31
Tabl e 1 Comparison of algorithm results with CAD and industrial CMM
Points CAD measurements
(cm)
Industrial CMM
(cm)
Algorithm results
(cm)
Deviation w.r.t
I.CMM (%)
A-B 2.4 2.88 2.90 0.0069
E-F 3.0 3.37 3.43 0.017
C-D 3.0 3.54 3.57 0.0084
G-H 3.6 3.78 4.15 0.089
I-J 3.6 3.89 4.14 0.060
A-C 28.9 31.40 31.41 0.0003
A-E 28.94 30.36 30.41 0.0016
A-G 39.62 42.25 42.27 0.0004
A-I 39.91 42.47 42.67 0.0046
Fig. 6 Labelling of
suspension points
32 R. Dhar et al.
deviations from industrial CMM reports. The inter-tabular point distances taken into
consideration, i.e., A-C, A-E, A-G, and A-I give better results from tab points and
also overall deviations of <0.005% from the industrial CMM reports.
5 Conclusion
The paper introduces a fast and accurate low computation image processing technique
that is used to detect manufacturing defects in all-terrain vehicles. The pre-processing
takes into account external factors such as non-uniform illumination which is rectified
using signal enveloping to removing shading element without any user intervention.
The use of reference object for calibration makes the technique useful over wider
range of setups. The reference object calculations consist of simple trigonometric
rules which are computed using the image area and co-ordinates. The use of bounding
box for finding centroid improves the accuracy of the technique reducing placement
redundancies. A comparative study was performed on the roll-cage front box of an all-
terrain vehicle with the technique and industrial co-ordinate measurement machine
which showed results with lesser than 0.1%.
References
1. C.C. Reyes-Aldasoro, Retrospective shading correction algorithm based on signal envelope
estimation. Electron. Lett. 45(9), 454–456 (2009)
2. D. Bradley, G. Roth, Adaptive thresholding using the integral image. J. Graph. Tools 12, 13–21
(2007). https://doi.org/10.1080/2151237X.2007.10129236
3. R. Bello, T. Suleiman, U.S. Tunga, Determination of object position and orientation using
image processing technique. Res. Rev. J. Phys. 7(2), 18–24 (2018)
4. O. Kainz, F. Jakab, M.W. Horeˇcný, D. Cymbalák, Estimating the object size from static 2D
image, in 2015 International Conference and Workshop on Computing and Communication
(IEMCON), Vancouver, BC (2015), pp. 1–5
5. L. Wojnar, Image Analysis: Applications in Materials Engineering (CRC Press, Boca Raton,
FL, 1999)
6. L. Zhao, J. Chen, J. Benesty, A multichannel widely linearwiener filter for binaural noise
reduction in the short-time-fourier-transform domain, in IEEE China Summit & International
Conference on Signal and Information Processing (ChinaSIP), July 2014, pp. 227–231
7. D. Slezak, S.K. Pal, B.-H. Kang, J. Gu, H. Kuroda, T.-H. Kim, Signal Processing, Image
Processing and Pattern Recognition (Springer Science & Business Media, Germany, 2009)
8. M. Cheriet, N. Kharma, C.-L. Liu, Ch. Suen, Character Recognition Systems: A Guide for
Students and Practitioners (Wiley, Hoboken, NJ, 2007)
9. K. Jeong, H. Moon, Object detection using FAST corner detector based on smartphone plat-
forms, in First ACIS/JNU International Conference on Computers, Networks, Systems and
Industrial Engineering (CNSI), May 2011, pp. 111–115
10. Z. Tang, D. Shen, Canny edge detection codec using VLib on Davinci series DSP, in Interna-
tional Conference on Computer Science & Service System (CSSS), Aug 2012, pp. 221–224
Co-ordinate Measurement of Roll-Cage Using Digital … 33
11. W. Rong, Z. Li, W. Zhang, L. Sun, An improved Canny edge detection algorithm, in IEEE
International Conference on Mechatronics and Automation (ICMA), Aug 2014, pp. 557–582
12. D. Li, P. Qiao, G. Zhao, A new moving object detection approach with adaptive double thresh-
olds, in Fifth International Conference on Computational and Information Sciences (ICCIS),
June 2013, pp. 102–105
Split-Ring Resonator Multi-band
Antenna for WLAN/WIMAX/X
Standard
Vinita Sharma, Santosh Meena, and Ritesh Kumar Saraswat
Abstract In this paper, we discussed split-ring resonator inspired monopole antenna
design. The proposed design is fabricated on FR-4 substrate (30 ×26.5 ×0.8 mm3)
and also covered the WLAN (5.2 GHz Band), Upper WiMAX (3.5GHz Band), and X
Band (7.5 GHz) application. For these ranges, the respectively achieved bandwidths
are 3.56 GHz, 5.19 GHz, 7.5 GHz that covered the WLAN, WIMAX, AND X Band
wireless standards, respectively. The proposed antenna shows stable radiation pattern.
Keywords Monopole antenna ·Split-ring resonator ·WiMAX ·WLAN ·X band
1 Introduction
The antenna is an appropriate device to cover the various wireless standards which
are present in microwave range. Many designs are proposed such as CPW-fed square
ring slot antenna [1] to resonate at wireless mode. Another antenna design that has
various shaped strips (S and U type) is implemented [2], achieves many wireless
communication applications. The Y-shaped monopole antenna design consisting of
split-ring slot shows various wireless resonating bands at WLAN (5.2 GHz), WiMAX
(3.5 GHz), and X band (7.5 GHz) [3]. Many more antennas which are feed in different
forms are also designed for wireless applications [4] and microstrip fed slot antenna
[5]. Recently, metamaterials loaded design by using SRR [6] and CSRR [7]having
application for wireless standards [8]. These metamaterials inspired antenna are
having compact size and with simplified geometry.
V. Sharma (B)·S. Meena
Government Women’s Engineering College, Ajmer, Rajasthan, India
e-mail: vinita95sharma@gmail.com
S. Meena
e-mail: smeena@gweca.ac.in
R. K. Saraswat
M.L.V. Textile and Engineering College, Bhilwara, Rajasthan, India
e-mail: ritesh.saraswat9@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_4
35
36 V. Sharma et al.
The proposed design having compact size patch indicates triple-band character-
istics at wireless standards WLAN (5.2 GHz), WiMAX (3.5 GHz), and X Band
(7.5 GHz). The proposed design is showing stable radiation property at respective
wireless band.
2 Design Methodology
The proposed design has hexagonal shape which is connected with feed line dimen-
sion of 7.2 ×1.8 mm2microstrip line as represented in Fig. 1. This design is fabri-
cated on a 30 ×26.5 ×0.8 mm3low-cost FR-4 substrate (εr=4.4, tan δ=0.002).
This antenna has split-ring radiating element to create multi-band nature for wireless
applications. The radiating patch area consists of pair of metallic rings that are in
parallel form. The top and side views of the proposed design are represented in Fig. 2.
(a) (b)
Fig. 1 Antenna design atop view bbottom view
Fig. 2 Configuration of the
designed antenna
Split-Ring Resonator Multi-band Antenna for WLAN/WIMAX/X Standard 37
Tabl e 1 Dimension of various proposed antenna parameters
Design
parameter
Length (L)Widt h (W)Side
length (S)
Gap (G) Feed line
width
(Wf)
Feed line
length
(Lf)
Ground
length
(Lg)
In mm 30 26.5 0.8 0.75 1.88 7.2 7
As per the antenna design, the dimension width of SRR rings (g) and slot gap of
conducting ring are the same. Due to this the proposed design indicates the sufficient
bandwidth for wireless standards. Table 1shows the optimized dimension of the
proposed antenna.
3 Result Analysis
With the help of simulating software, we obtained the simulated S parameter as shown
in Fig. 3a. The measured S parameter results are obtained in anechoic chamber with
the help of vector network analyzer. The proposed design with optimized dimension
as mentioned in Table 1covers three resonant bands at 3.5, 5.2, and 7.5 GHz for
wireless standards. By implementation of two concentric spilt rings in radiating patch,
it creates the resonant band for WLAN, WiMAX, and X band wireless applications,
respectively. The simulated and measured S parameter of the proposed design is
shown in Fig. 3.
Figure 4shows the simulated and measured gain of the proposed antenna at
different frequencies. It is noticed that at higher and WLAN/WiMAX frequencies
gain is improved with respect to reference level of gain. Figure 5represents the
comparison of radiation efficiency at various resonant frequencies. It is observed
that at wireless resonant frequency of WLAN and WiMAX efficiency is increased.
The parametric study regarding width of feed line (Wf) at 1.5 and 2.4 mm is
indicated in Fig. 6a, b, respectively. It is identified that wider bandwidth around
3 GHz is obtained for Wf=1.88 mm, i.e., optimized value of width of feed line.
(a) (b)
1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 10.00
Freq [GHz]
-30.00
-25.00
-20.00
-15.00
-10.00
-5.00
0.00
dB(S(1,1))
Fig. 3 Simulated and measured S parameter of the proposed antenna
38 V. Sharma et al.
Fig. 4 Simulated and measured gain
Fig. 5 Simulated and measured efficiency
The simulated surface current distribution at resonant mode 3.5 GHz, 5.2 GHz,
and 7.5 GHz are shown in Fig. 7, respectively. It is observed that current density is
dense across the periphery of spilt rings at 3.5 GHz as shown in Fig. 7a. Figure 7b, c
shows the current distribution characteristics at 5.2 GHz and 7.5 GHz, respectively,
indicate maximum distribution around the inner and outer part of split ring that is
responsible to create the higher resonant band.
Figure 8indicates the E and H plane radiation patterns along with 3D polar plot
at resonant frequencies 3.5, 5.2, and 7.5 GHz. It is observed that the omnidirectional
Split-Ring Resonator Multi-band Antenna for WLAN/WIMAX/X Standard 39
Fig. 6 Parametric study regarding feed line width (Wf) at 1.5, 2.4 mm
Fig. 7 Simulated current distribution a3.5 GHz, b5.2 GHz, c7.5 GHz
radiation patterns are obtained in case of H plane for all the resonant bands, whereas
bidirectional patterns are observed in case of E plane. The radiation pattern is stable
for all the resonant frequencies. The simulated peak gain of 2.6, 2.36, and 3.03 dBi
is obtained across the 3.5, 5.2, and 7.5 GHz resonant frequency,respectively.
4 Conclusion
This antenna shows triple-band characteristics at WILAN (5.2 GHz), upper WMAX
(3.5 GHz), and X Band (7.5 GHz) wireless standards. The proposed antenna has
compact size with pair of hexagonal shape concentric split ring as a radiating part.
The proposed design has appropriate surface current distribution and stable radiation
pattern for operating band of wireless applications.
40 V. Sharma et al.
Fig. 8 E-planed H-planed radiation patterns/3-D polar plot at resonant frequency at 3.5, 5.2,
7.5 GHz
Split-Ring Resonator Multi-band Antenna for WLAN/WIMAX/X Standard 41
References
1. C.Y.D. Sim, H.D. Chen, K.C. Chiu, C.H. Chao, Coplanar waveguide-fed slot antenna for
wireless local area network/world-wide interoperability for microwave access applications.
IET Microwave Antennas Propag. 6, 1529–1535 (2012)
2. Y. Xu, Y.C. Jiao, Y.C. Luan, Compact CPW-fed printed monopole antenna with triple-band
characteristics for WLAN/WiMAX applications. Electron. Lett. 48, 1519–1520 (2012)
3. P. Liu, Y. Zou, B. Xie, X. Liu, B. Sun, Compact CPW-fed tri-band printed antenna with
meandering split-ring slot for WLAN/ WiMAX applications. IEEE Antennas Wirel. Propag.
Lett. 11, 1242–1244 (2012)
4. X. Li, X.-W. Shi, W. Hu, P. Fei, J.-F. Yu, Compact triband ACS-fed monopole antenna
employing open-ended slots for wireless communication. IEEE Antennas Wirel. Propag. Lett.
12, 388–391 (2013)
5. X.Q. Zhang, Y.C. Jiao, W.H. Wang, Compact wide tri-band slot antenna for WLAN/WiMAX
applications. Electron. Lett. 48, 64–65 (2012)
6. K. Yang, H. Wang, Z. Lei, Y. Xie, H. Lai, CPW-fed slot antenna with triangular SRR terminated
feed line for WLAN/WiMAX applications. Electron. Lett. 47, 685–686 (2011)
7. S.C. Basaran, U. Olgun, K. Sertel, Multiband monopole antenna with complementary split-ring
resonators for WLAN and WiMAX applications. Electron. Lett. 49, 636–638 (2013)
8. Y. Kim, D. Kim, CPW-fed planar ultra wideband antenna having a frequency band notch
function. Electron. Lett. 40, 403–405 (2004)
9. S. Qu, J. Li, Q. Xue, A band-notched ultra wideband printed monopole antenna. IEEE Antennas
Wirel. Propag. Lett. 5, 495–498 (2006)
10. Y. Ding, G. Wang, J. Liang, Compact band-notched ultra-wide-band printed antenna.
Microwave Opt. Technol. Lett. 49, 2686–2689 (2007)
11. Y. Zhang, W. Hong, C. Yu, Z.-Q. Kuai, Y.-D. Don, J.-Y. Zhou, Planar ultra wideband antennas
with multiple notched bands based on etched slots on the patch and/or split ring resonators on
the feed line. IEEE Trans. Antennas Propag. 56, 3063–3068 (2008)
Analyses on Architectural and Download
Behavior of Xunlei
Bagdat Kamalbayev, Nazerke Seidullayeva, Adilbek Sain, Pritee Parwekar,
and Ikechi A. Ukaegbu
Abstract The following paper conducts an extended analysis on different measure-
ment studies performed on popular P2P Chinese application called Xunlei. In this
particular research paper, several previous researches have been studied and their
arguments were compared, in order to find out finalized reason behind this enor-
mous companies’ success. Its packet-level data was evaluated in order to understand
applications network architecture and download behavior. Xunlei’s core-server layer
was analyzed by division of tasks. It was discovered that application uses direct
linking and applies multi-source technology in an effective way in order to increase
the download speed. Also, it was found that Xunlei uses UDP as the main transport
protocol for both IPTV streaming and bulk file transfer. This paper provides a broad
study on Xunlei which leads to a better understanding of the Internet.
Keywords Internet measurement ·Xunlei ·P2P ·Bandwidth theft
1 Introduction
The main aim of the following paper is to investigate the P2P application called
Xunlei, a file-sharing application, utilizing different existing researches already
B. Kamalbayev ·N. Seidullayeva ·A. Sain ·I. A. Ukaegbu (B)
School of Engineering and Digital Sciences, Nazarbayev University, Nur-Sultan, Kazakhstan
e-mail: ikechi.ukaegbu@nu.edu.kz
B. Kamalbayev
e-mail: bagdat.kamalbayev@nu.edu.kz
N. Seidullayeva
e-mail: nazerke.seidullayeva@nu.edu.kz
A. Sain
e-mail: adilbek.sain@nu.edu.kz
P. Parwekar
Computer Science & Engineering Department, SRM University, Delhi-NCR, Sonepat, India
e-mail: priteep@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_5
43
44 B. Kamalbayev et al.
undertaken toward this company. Such applications have good contribution to the
modern internet traffic [1]. Studying this download manager could provide better
understanding of the modern Internet [1,2]. Xunlei is one of the most popular
streaming media and file-sharing applications in China. It can be seen from the
following statistics which are provided by Wikipedia:
•Over 80 million users installed Xunlei.
•Xunlei’s website draws over 50 million hits per day.
•Number of users on the official website exceeded 1880 million.
Xunlei helps people to easily find files and download them at high speed. It uses
multi-source technology to deliver all kinds of files and integrates resources on the
internet. Also, there is a good contribution of file-sharing and streaming protocols
such as eMule, Bit Torrent, FTP, and RTSP [3]. However, Xunlei is not only file-
sharing application, but it also affords IPTV functionality which allows people to
watch TV programs and movies online. Popularity of Xunlei among users attracts
attention of big companies such as Google, IDGVC, and Geyuan Ventures despite
that it has not English version and runs only on Windows machines. It is spread
outside China due to migrants and soon will be popular worldwide too. Until now,
not many researches have been conducted to investigate this popular company.
1.1 Resource Searching and Sharing Mechanisms
In the succeeding paragraphs, Xunlei’s resource searching policy will be explained
and compared with Client/Server systems and P2P systems [4–8].
(1) Client/Server systems
•Traditional web downloads: A central server affords content to clients
directly.
(2) Peer-to-Peer Systems Peers can exchange content among each other by creating
their own overlay networks. Examples:
•Gnutella: sending many requests to find the needed file
•eDonkey: uses own servers as querying organs.
•Bit Torrent: uses torrent files and own servers as querying organs.
(3) Xunlei resource system is like a combination of the previous systems. Xunlei
has its own overlay network, but it also uses resources which are located on
other servers. There are examples of these resources [6]:
•Internal server: when a user seeks a file on a website, and the website gets
file from a server which belongs to this website.
•External server: when a user seeks a file on a website, and the website gets
file from another server which does not belong to this website.
•Peer: Xunlei indexes peers who got the file.
Analyses on Architectural and Download Behavior of Xunlei 45
1.2 Direct Linking or Bandwidth Leaching
Using external servers is considered as bandwidth theft in the form of direct linking
[9]. It is also called as inline linking or leeching. An example of leeching can be usage
of linked images, from one website into a webpage which belongs to the second
website. If there is no agreement between them, the second website is considered to
be leeching to the first one. In Xunlei, a client is called as leecher if it uses Xunlei
to download a file from one website and it downloads file from external servers
instead of using internal or P2P servers. Direct linking harms external sites earnings
from advertisements but allows users to download files at high speed. However,
according to the other recent researchers, Xunlei begins to compensate for this so-
called bandwidth leaching by providing membership to the Xunlei Union to these
external servers. One of the researches reveals that Xunlei also supplies with the
bandwidth support to those who are not even Xunlei’s servers and sources. For
instance, client who uses Xunlei as downloading application, starts to download some
files from an external server, and if other Xunlei’s clients have already downloaded
that file, that client can support the external server as well.
1.3 Contributions of This Paper
First of all, architecture of Xunlei’s core-system is studied by revealing its structure
and client workflow. Further, its download behavior is evaluated through observation
of several download processes for IPTV streaming and bulk file transfer. The main
results are [10–12]
•It was revealed that Xunlei’s core servers clearly divide tasks between its servers.
•It was found that Xunlei exploits external servers in order to speed up download
speeds.
•It was shown that Xunlei’s multi-source download manager works effectively.
•It was discovered that Xunlei uses UDP as the main transport protocol for both
IPTV streaming and bulk file transfer [13].
•Xunlei uses tracker and cross-protocol system to find the files requested in peers
and in other servers easily and quickly. Also, plug-in system details are provided
below.
The related work is discussed in the second section. In Sect. 2, experiment
and measurements are portrayed and based on these studies Xunlei architecture is
explained. The information about download behavior is written and the last section
concludes the paper.
46 B. Kamalbayev et al.
2 Results
2.1 Xunlei’s Architecture
Xunlei has contributed to the improvement of download session and gave great
advantages to the people. The traffic collection of Xunleis sessions has helped to
understand the basic working mechanisms of that product, and also to identify the
servers which then have been filtered based on IP addresses with the reverse DNS
protocol. The defined servers have been analyzed to understand the Xunlei’s server
infrastructure and to demonstrate the clear advantage over the servers of other popular
download software such as eDonkey [6,7].
Core servers: A Xunlei and eDonkey client/server architectures are different:
(1) The amount and IP addresses of Xunlei’s servers are stable and do not alter
frequently while those eDonkey servers’ features are changed regularly. Thus,
Xunlei’s client/server architecture aimed to have permanent IP addresses and can
be considered as a reliable source of connection. An experiment was conducted
to identify the number of Xunlei’s server. It was found that 46 IP addresses were
used and most of them are located in three provinces of China.
(2) Xunlei servers are organized in a more confounded and systematic way as
compared to those of eDonkey [6,7]. Every server of eDonkey software is
primarily responsible for indexing documents for customers. Xunlei servers
participate adequately with obvious divisions of work. These categories of work
may be listed in such a way:
a. Client initialization
b. Session tracking
c. Nodes server directory
d. Nodes server
e. Resource management (providing resource reports with external servers and
peers to clients)
f. Indexing advertisements
g. Storing advertisements
h. Picture server
i. Virus Scan
j. Statistics
k. Keep-alive server
Client-Server communication: The Client-server communication in Xunlei has
three phases:
(1) Login The client uses login session to establish a connection to the Xunlei
network. This login session can be initiated with the client request through TCP
and session tracking server on UDP/4000, and also the amount of available
nodes servers is required at the nodes server directory through UDP. In case of
Analyses on Architectural and Download Behavior of Xunlei 47
rejecting request, it continues to link to other nodes in the server directory. This
action continues till client gets positive replies and it is then informed to the
session tracker, thus login session will be stopped.
(2) Downloading: Client starts downloading from one website by requesting an
URL, then this URL is translated into a distinctive number and will be provided
with resulting resource report where it includes locations of requested file. The
translation of URL is conducted by resource management server to identify
the needed data. From the provided location, source can download the data
in parallel by applying multi-source technology that is provided by Xunlei’s
technology.
(3) Idling: In order to hold idling state, Xunlei client holds communication with
server and sends ICMP messages every 10 s and UDP in between 45 and 300 s
to the session tracker and node servers. These actions contribute to update and
report the client status.
2.2 Download Behavior
The previous analysis only gives general understandings of this huge infrastructure,
the more important things are how the resources in this enormous company are
allocated between core servers. Also, it is predicted that one of the main reasons for
popularity of Xunlei is due to this resource allocation method which increases the
rate of downloading. Therefore, in the beginning of the following subsections, the
attention was on data redundancy and finally on the protocols that Xunlei mostly
uses.
Resource allocation:
As explained in the earlier section, three main resources are used, such as Internal
servers, it is when the all the requested files are taken from the specific website which
belongs to that specific company; peer-to-peer, is the other type of the resources,
where the clients share the requested file between each other; and the last type was
external servers, where the requested files are taken from website that does not
belong to the specific website. Before analyzing the results, let us understand how
the measurements were gained through experimentation. To conduct the experiment,
10 famous files were chosen as well as 4 portal websites. Furthermore, each file
was downloaded 5 times from the same website. These totaled as 165 sessions.
The number of sessions is higher, due the nonexistence of the requested file in the
specific website. In Figs. 1and 2, the detailed allocation of the resources per session
can be seen. The horizontal line on the Figure illustrates the download processes
grouped according to websites, meanwhile the vertical line illustrates the types of
the resources. According to this figure, it is clearly shown that the most used type
of the resources is external servers. These results were collected in 2012 which was
the last research that was conducted on this subject. To see if any differences have
accrued, this work has conducted another experiment as illustrated in Fig. 2.Itis
48 B. Kamalbayev et al.
40
35
30
25
20
15
10
5
0
20 23 35 45 55 65 75 85 100 120 140 160
Peer
Internal Server
External Server
Download Sessions By Websites
Number of Resources
Fig. 1 Experimental result of resources gained from different servers and peers (conducted in 2012)
40
35
30
25
20
15
10
5
0
20 23 35 45 55 65 75 85 100 120 140 160
Peer
Internal Server
External Server
Download Sessions By Websites
Number of Resources
Fig. 2 Experimental result of resources gained from different servers and peers (conducted in 2018)
Analyses on Architectural and Download Behavior of Xunlei 49
clearly seen that even after many years behind the Xunlei mainly uses external servers
rather than internal servers and peers.
2.3 Download Redundancy
Measurements obtained were observed to ascertain how data redundancy behaves.
The attention was given especially to this feature, since the large amount of the
resources gained does not guarantee that the downloaded files are in good quality.
To examine this quality an experiment was conducted in which 210 downloads were
examined. In Fig. 3, it is clearly seen that most of the portions of the downloads have
less than 1.2 redundancy, and nearly the other rest have only redundancy between
1.2 and 1.5, which is suggested as these downloads have good quality. The results
are shown in Fig. 3.
UDP or TCP:
In this last section, bulk-files and IPTV files in Xunlei were examined. The main
reason behind it is to find out, what kind of transport protocols Xunlei uses that their
downloading rate is very high compared to the other big companies and even has good
quality of the downloaded files. It was investigated that Xunlei provides a website
where one can not only download the video files but also to watch it online. Therefore,
four popular video files were both downloaded and streamed in this experiment. The
measurements obtained are shown below on the Table. It is clearly seen that major part
of the files or it can be said almost all of them were transferred utilizing UDP rather
than TCP. In order to find out the real justification, paper has tried to become Xunlei
client and using software Wireshark downloaded several files from this application,
meanwhile were capturing files through Wireshark. The results were as was predicted
by previous researches, almost all the packets captured were using UDP rather than
Fig. 3 Redundancy of the
different downloaded files
50 B. Kamalbayev et al.
TCP as can be seen from Fig. 4.InFig.5, paper has included finalized results gained
from packet capturing, it is seen that almost 98 % is UDP. Paper has suggested that
the reason behind this good downloading speed might be in the utilization of the
right transport protocol. In case of Xunlei, it appears that UDP helps to maintain
good velocity.
Fig. 4 Packet capturing using Wireshark in order to find out which protocol is used
Fig. 5 UDP is used mainly
over TCP protocol in both
bulk file-sharing and in IPTV
Analyses on Architectural and Download Behavior of Xunlei 51
2.4 Plug-in
It was discovered while conducting the experiment that Xunlei utilizes plug-in
system. Whenever the client in first case installs the Xunlei client application, imme-
diately plug-in is also installed into the Internet Explorer browser. It allows the Xunlei
client to take over the browsers speed from other users who are also downloading
the same file. This can be applied to different types of files as well, such as video
files, pdf files, also supports different protocols HTTP, FTP, RTSP, and MMS. Also,
the suggestions made earlier in this work that Xunlei clients serve as the Bit Torrent
and eDonkey client were justified. It might be the other reason, however not the
main according to the “Xunlei: Peer-Assisted Download Acceleration on a Massive
Scale” article, that Xunlei’s downloading acceleration is the highest among Chinese
downloading applications [14].
2.5 Tracker
The recent research claims that some interesting analyses have been found during
observing the local Xunlei clients traffic. Particularly, the details between the client’s
message exchange with the Xunlei tracker when it requests the available sources for
any file. For instance, when a Xunlei client sends reference to the central tracker,
every time central tracker returns 20-byte hash values and 8-byte code for a requested
file. Actually, both hash and code values are later used to request resources for the file
from other peers and servers. Also, this hash and code values can be received from
the tracker in other ways as well, it can be obtained by sending distinctive 20-byte
identifier for the file. For Bit Torrent files, the Xunlei client utilizes the info hash of
the file by 20-byte identifier which is included to the message which is then send
to the tracker. For eDonkey files, this identifier is extracted from its link which is
similar to the file size. In case when Xunlei client sends the message to the tracker,
particularly the identifier that is not currently tracked by the Xunlei, the tracker does
not give any output with two hash and code values. Xunlei’s cross-protocol operation,
to explain it more accurately let’s take a look at an example. For instance, there is a
file that exists in HTTP server and in Bit Torrent as well, and also consider that the
tracker does not know about the file. Whenever the client downloads this file from
Bit Torrent, it begins to map two hash values internal hash within the Bit Torrent
info hash. Also, if another Xunlei client begins to download the same file from the
server in direct way, it begins to measure the same internal hash values, and then
when it informs the tracker, it will know that there is mapping between HTTP URL
and internal hash. From this, it can be concluded that the tracker knows that the same
file matches both Bit Torrent info hash and HTTP. For that reason, in case Xunlei
client begins downloading with Bit Torrent info hash, the tracker will immediately
provide the HTTP server link and the Xunlei peers that also have the same file.
52 B. Kamalbayev et al.
3 Conclusion
In this paper, attempt has been made to understand the working principal of China’s
most famous downloading application, which is known as Xunlei. This software
application was chosen since it is used by a large number of people. To find out
how the server-side of the application works, several experiments were conducted,
however, all of them are according to the packet-level feature, because data trans-
mission in Xunlei is encoded. First, it has suggested to focus on the architectural
behavior of the Xunlei, in order to do that paper have studied previous analysis done
toward in previously from different researchers, particularly this paper wanted to
find out Xunlei’s servers and client-server relations. Then, it was decided to investi-
gate Xunlei’s downloading behavior, which leads us to figure out that, this company
supply with great amount of resources, which definitely makes in its turn the speed
rise. However, a major part of the data which is requested is taken from external
servers. Furthermore, the great amount of resources does not always guarantee that
data downloaded is in good quality, for that reason, downloading redundancy as well
was examined. Also, the experiment was conducted to find out what kind of trans-
port layer Xunlei mainly uses. It was revealed that Xunlei despite the file type, bulk
file-sharing or IPTV streaming, utilizes 95 % UDP transport protocol rather than
TCP. All of these results lead us to conclude our analysis that the key factors about
this provider is that it provides a great amount of resources and mainly uses external
servers and UDP protocol, and it has great architectural structure where all the task are
accordingly divided within the server. However, the external servers are considered
as improper, since other external websites gain no advantage from it, but only Xunlei
gets the advantage. For that reason, it was suggested that the good way to correct
this is to make it an open and business-driven network architecture. However, it was
discovered that Xunlei already has solved this problem by introducing the external
servers to their Xunlei Union membership.
References
1. H. Schulze, K. Mochalski, Internet study 2008/2009. Ipoque Rep. 37, 351–362 (2009)
2. W. John, S. Tafvelin, T. Olovsson, Trends and differences in connection-behavior within
classes of internet backbone traffic, in International Conference on Passive and Active Network
Measurement (Springer, Berlin, 2008), pp. 192–201
3. M. Zhang, W. John, C. Chen, A measurement-based study of Xunlei, in Student Workshop,
PAM (2009)
4. S. Saroiu, P.K. Gummadi, S.D. Gribble, Measurement study of peer-to-peer file sharing systems,
in Proceedings of Multimedia Computing and Networking (MMCN) (2002)
5. J. Miller, Characterization of data on the gnutella peer-to-peer network. In First IEEE Consumer
Communications and Networking Conference, 2004. CCNC 2004 (IEEE, 2004), pp. 489–494
6. J. Yang, H. Ma, W. Song, J. Cui, C. Zhou, Crawling the edonkey network, in 2006 Fifth
International Conference on Grid and Cooperative Computing Workshops (IEEE, 2006),
pp. 133–136
Analyses on Architectural and Download Behavior of Xunlei 53
7. O. Heckmann, A. Bock, The edonkey 2000 protocol. Rapport technique, Multimedia
Communications Lab, Darmstadt University of Technology (2002), p. 13
8. M. Zhou, Y. Dai, X. Li, A measurement study of the structured overlay network in P2P file-
sharing applications, in Eighth IEEE International Symposium on Multimedia (ISM’06) (IEEE,
2006), pp. 621–628
9. M. Zhang, C. Chen, N. Brownlee, Behaviors of bandwidth theft in Xunlei, in 2009 2nd IEEE
International Conference on Computer Science and Information Technology (IEEE, 2009),
pp. 520–524
10. Y.J. Hall, P. Piemonte, M. Weyant, Joost: A measurement study. School of Computer Science,
Carnegie Mellon University, Technical Report (2007); J. Liang, R. Kumar, K.W. Ross, The
kazaa overlay: a measurement study. Compu. Netw. J. 49(6) (2005). (Elsevier)
11. S.A. Baset, H.G. Schulzrinne, An analysis of the skype peer-to-peer internet telephony protocol,
in INFOCOM 2006. 25th IEEE International Conference on Computer Communications
Proceedings (2006)
12. M. Zhang, M. Dusi, W. John, C. Chen, Analysis of UDP traffic usage on internet backbone
links, in 2009 Ninth Annual International Symposium on Applications and the Internet (IEEE,
2009), pp. 280–281
13. P. Dhungel, K.W. Ross, M. Steiner, Y. Tian, X. Hei, Xunlei: peer-assisted download acceleration
on a massive scale, in International Conference on Passive and Active Network Measurement
(Springer, Berlin, 2012), pp. 231–241
14. P. Dhungel, W. Ross, M. Steiner, Y. Tian, X. Hei, Xunlei: Peer-assisted download acceleration
on a Massive Scale” in SAINT 09, 9th Annual International Symposium on Applications and
the Internet, 2015
Advanced Driver Assistance System
Technologies and Its Challenges Toward
the Development of Autonomous Vehicle
Keerthi Jayan and B. Muruganantham
Abstract The automotive industry is more concerned about vehicle and road safety.
Nowadays the Govt. has introduced new safety regulations to improve the vehicle and
road safety. Most of the accidents are happening because of human errors, violation
of traffic rules, etc. To avoid or reduce these accidents and to meet the Govt. safety
regulations the automotive industries started implementing new technologies like
Advanced Driver Assistance System (ADAS), Passive and Active safety system in
vehicles. ADAS is the most important safety feature in modern automobiles. The
key functionality of ADAS is to provide more comfort to the driver and to improve
road and vehicle safety by minimizing human errors. To achieve these functionalities
the system use sensors like Radar (Radio Detection and Ranging), Camera, Light
Detection and Ranging (LIDAR), and Ultrasonic sensors. Based on sensor input
Electronic Control Unit (ECU) delivers the information to driver in terms of warnings,
alarms, etc. or system takes the longitudinal and lateral control to bring the vehicle to
safer zone. Focus of this paper is to discuss the hardware and software technologies
used for the development of ADAS systems and its challenges.
Keywords ADAS ·ECU ·LIDAR ·RADAR
1 Introduction
Since there is an intensification in popularity of automobiles over the last century,
rise in popularity of automobiles has a significant role in road accidents. As per
the Global Status Report on Road Safety 2018 statistics [1] which reveals that road
accidents on an average cause 20–50 million injuries and 1.3 million fatalities around
K. Jayan (B)·B. Muruganantham
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
Kattankulathur, Chennai, Tamil Nadu 603203, India
e-mail: keerthij@srmist.edu.in
B. Muruganantham
e-mail: muruganb@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_6
55
56 K. Jayan and B. Muruganantham
the world. Most of the accidents are due to human errors. The demand for an error-
free driving environment is increased. This facilitates the development for Advanced
Driver Assistance System (ADAS) technologies. ADAS is the combination of sensors
that aid and improve the comfort of driver and safety of passengers on the road. It
enhances the machine to human interface by detecting the set of on-road and driver
behavior by providing visual and audio information to the driver.
Nowadays most of the modern vehicles are equipped with ADAS functionalities
to improve the safety level and provide more comfort to the driver. ADAS systems are
capable of sensing environment and analyze, predict, and react to the situation. ADAS
functionalities started with providing support to the driver action, for example, Anti-
lock Braking System (ABS) [2]. This system helps to avoid wheel locking during
braking. Anti-lock breaking systems was introduced in the vehicle about 70 years
ago [2], and it was the first driver assistance system [2]. ABS first introduced as
an anti-skid system for aircraft was used in the 1950s. In 1970, the first-generation
anti-lock braking system was developed by Mercedes-Benz and TELDIX. In the
late 1960s, the first Electronic cruise control system was developed. The succeeding
developments led to adaptive cruise control systems [3].
Today the ADAS system uses the sensors like Radar, LIDAR, Ultrasonic, and
Camera to introduce new functionalities to improve driver comfort and vehicle safety.
Research and Development for the ADAS system are focusing to develop a driver-
less vehicles. The driverless vehicle needs to train with lots of data of different
road scenarios and objects. However, the system has its own limitation in terms of
technologies to handle the data. The system also needs self-learning capability to
improve safety. The Artificial Intelligence (AI) technologies aid the system for self-
learning purpose. Society of Automobile Engineers (SAE) levels were established
as a classification system for autonomous vehicle in January 2014.
2 Sae Level of Advanced Driver Assistance Systems
In May 2013, National Highway Traffic Safety Administration (NHTSA) established
a policy on automated vehicles where it defines the automation levels from zero to
four. This level shows the growth of vehicle automation from zero to fourth level,
which is the fully automated vehicle. As per the SAE classification, the ADAS
systems were classified from level 0 to 5 different stages based on the level of
automation. In both set of automation definitions (NHTSA and SAE) the first 3
levels of automations are common, but in SAE automation the fully automated 4th
level of NHTSA was subdivided into two, SAE level 4 and level 5. SAE J3016 issued
on January 2014 provides the information of all modes of automation in cars. Table 1
shows the summary of levels of driving automation.
Advanced Driver Assistance System Technologies … 57
Tabl e 1 Summary of levels of driving automation
SAE
level
SAE name SAE narrative definition Execution of
steering and
acceleration/
deceleration
Monitoring of
driving
environment
Fallback
performance of
dynamic driving
task
System
capability
(driving
modes)
BASt level NHTSA
level
Human Driver Monitor the driving Environment
0No
Automation
The full-time performance
by the human driver of all
aspects of the dynamic
driving task even when
enhanced by warning or
intervention systems
Human Driver Human Driver Human Driver NA Driver Only 0
1Driver
Assistance
The driving mode-specific
execution by a driver
assistance system of either
steering or
accelerating/deceleration
using information about the
driving environment arid
with the expectation that
the human driver performs
all remaining aspects of the
dynamic task
Human Driver
and System
Human Driver Human Driver Some
Driving
Modes
Assisted 1
(continued)
58 K. Jayan and B. Muruganantham
Tabl e 1 (continued)
SAE
level
SAE name SAE narrative definition Execution of
steering and
acceleration/
deceleration
Monitoring of
driving
environment
Fallback
performance of
dynamic driving
task
System
capability
(driving
modes)
BASt level NHTSA
level
2Partial
Automation
The driving mode-specific
execution by one or more
driver assistance system of
both steering or
accelerating/deceleration
using information about the
driving environment and
with the expectation that
the human driver performs
all remaining aspects of the
dynamic task
System Human Driver Human Driver Some
Driving
Modes
Partially
Authomated
2
Automated Driving System (“System”) Monitor the
driving Environment
System System Human Driver Some
Driving
Modes
Highly
Automated
3
3Conditional
Automation
The driving mode-specific
performance by an
automated driving system
of all aspects of the
dynamic driving task with
the expectation that the
human driver will respond
appropriately to a request to
intervene
(continued)
Advanced Driver Assistance System Technologies … 59
Tabl e 1 (continued)
SAE
level
SAE name SAE narrative definition Execution of
steering and
acceleration/
deceleration
Monitoring of
driving
environment
Fallback
performance of
dynamic driving
task
System
capability
(driving
modes)
BASt level NHTSA
level
4High
Automation
The driving mode-specific
performance by an
automated driving system
of all aspects of the
dynamic driving task, even
if a human driver does not
respond appropriately to a
request to intervene
System System System Some
Driving
Modes
Highly
Automated
3/4
5 Full
Automation
The full-time performance
by an automated driving
system of all aspects of
dynamic task under all
roadway and environmental
conditions that can be
managed by human driver
System System System All Driving
Modes
60 K. Jayan and B. Muruganantham
Fig. 1 Overview of parking sensor
2.1 SAE Level Zero: No Automation
Level zero is the basic SAE level of automation where human driver performs
all driving tasks like longitudinal control (acceleration or deceleration) and lateral
control (steering). However, this system provides information to the driver as a
warning that helps driver to monitor the surrounding environment.
2.1.1 Parking Sensor
Parking sensor system provides acoustic warning to the human driver based on the
distance between vehicle and detected obstacle that helps the human driver to avoid
collision by taking corrective measures. Generally, ultrasonic sensors are used as
parking sensors in vehicles. The ultrasonic sensors can detect the presence of neigh-
boring objects without having any physical contact. There are around 4–8 ultrasonic
sensors located on front and rear of the vehicle to detect objects.
Ultrasonic sensors determine distance from the sensor to object. The basic prin-
ciple of distance measurement in ultrasonic sensor is based on Eco [4]. The sensor
transmits the sound waves to the environment, after striking on the obstacle those
waves return to the origin. The distance between sensor and object can be calculated
based on the time of flight of an ultrasonic wave. Figure 1showstheoverviewofa
parking sensor.
2.1.2 Traffic Sign Recognition (TSR)
In Traffic Sign Recognition (TSR) [5] system, the vehicle is able to recognize traffic
signs on the road and provide information to the driver. TSR recognize traffic signs
like speed limits, turn, overtaking, slippery road, and other warnings, etc. The system
Advanced Driver Assistance System Technologies … 61
uses a forward-looking camera positioned behind the rearview mirror on the wind-
shield that detects road signs, and TSR system also uses supplementary information
from the HD maps. Basic TSR system consists of mainly three parts [5].
1. Knowledge Database: System stores information of traffic sign pattern and
representation related to it. This allows fast identification of received pattern
2. Leaning Module: Collect the image data from the received video data and improve
the knowledge database of the system.
3. Recognizing Modules: This module process the input image to identify area
of traffic sign. This processed traffic sign image is compared with available
traffic sign from the knowledge database to get the best matching result and
communicate with the driver.
2.1.3 Blind Spot Detection (BSD)
Blind spot detection system processes the sensor information and warns driver in
the form of visual audible indicator [2]. Visual indicators often integrated with the
outside side-view mirrors. Generally, this system uses two types of sensors like radar
and camera [6].
In radar-based blind spot detection system, the sensor usually mounts on rear left
and right position on the vehicle bumper fascia. This system uses radio wave to detect
the obstacle. The system sends and receives radar signal to the blind zone area and
provides visual and audible information to driver. Figure 2shows the radar-based
blind spot detection in vehicle.
Camera-based blind spot detection uses image processing technology to identify
object in the blind zone. In this system, camera is located on the rearview mirror or
near to the rearview mirror. The system identifies vehicle shape in daytime and light
in night-time. Figure 3shows camera-based blind spot detection.
Fig. 2 Radar-based blind
spot detection system
62 K. Jayan and B. Muruganantham
Fig. 3 Camera-based blind
spot detection
2.1.4 Surround View
Automotive surround view, also called “around view” or “surround vision moni-
toring system,” [2] is an emerging automotive ADAS technology that provides the
driver a 360-degree view of the area surrounded by the vehicle. The vehicle is
equipped with four or six fish-eye cameras. These camera information stiches and
get complete surround view of the vehicle. This information displayed to the driver
in the form of birds-eye view. Surround-view system helps driver to understand the
vehicle surroundings in a better way. This system is combined with other driver
assistance functions like parking assistance, blind spot detection, etc. and provides
more support to the driver. Figure 4shows illustration of surround-view camera.
2.1.5 Lane Departure Warning (LDW)
The purpose of Lane Departure Warning [7] is to alert the driver with acoustic or
haptic steering warnings during the situations where vehicle starts moving out of the
lane [7]. LDW systems are equipped with a forward-looking camera to identify the
road lane markings and detect when a vehicle starts moving out of its driving lane.
Whenever the system detects the vehicle starts moving out of its lane, then the driver
is informed about this unintentional lane change in the form of acoustic or haptic
steering warnings that helps the driver to bring vehicle back to its lane. Figure 5
shows block diagram of Lane Departure Warning system.
The prediction of a lane departure is based on certain thresholds like time-to-
line crossing, actuation of the turn signals, steering wheel angle, and in some cases
driver drowsiness and driver inattentiveness monitoring. Figure 6shows time-to-line
crossing threshold for lane departure.
Advanced Driver Assistance System Technologies … 63
Fig. 4 Illustration of
surround view camera
Fig. 5 Block diagram of lane departure warning system
2.1.6 Forward Collision Warning (FCW)
FCW systems are an active safety feature that warns the driver about an upcoming
collision with an obstacle ahead that helps the driver to start decelerating the vehicle
[8]. FCW systems use a midrange radar sensors and a forward-looking camera to
detect objects. When the distance between objects becomes too short that a crash is
forthcoming, FCW system alerts the driver so that the driver can vigilant to apply
64 K. Jayan and B. Muruganantham
Fig. 6 Time-to-line crossing threshold for lane departure
brakes or do a steering to avoid obstacle and prevent a potential crash. FCW system
uses acoustic, visual display, or other warning signals to inform driver about the
crash.
2.2 SAE Level One Automation: Driver Assistance
In this level of automation, the driver controls vehicle, but some driving assist features
are included in the vehicle design. Small elements of driving such as steering, braking
can be done automatically by the vehicle.
2.2.1 Anti-Lock Braking System (ABS)
ABS helps to prevent the wheels from locking up during hard braking situations on
low friction surface [9]. ABS brakes increase vehicle safety during braking because it
eliminates wheel lockup and avoids vehicle skidding. ABS system allows the vehicle
to stop in a straight-line manner. Anti-lock braking system provides flexibility in
lateral control of the vehicle during hard braking so that driver can steer the vehicle
and avoid the obstacle. The system consists of wheel speed sensors, ABS control
module, Hydraulic-modulating unit, Valves, and Pump. Figure 7shows schematic
illustration of Anti-lock Braking System (ABS).
Anti-lock braking system controls tire slip by observing the relative deceleration
rates of wheels during braking [10]. One or more wheel speed sensors are used to
monitor the wheel speed. If one of the wheels starts to decelerate rapidly than the
other wheels, then this shows that the wheel is losing traction and starting to slip.
Advanced Driver Assistance System Technologies … 65
Fig. 7 Schematic illustration of Anti-lock Braking System (ABS)
This leads to wheel locking situation. ABS system quickly responds to this situation
by reducing hydraulic brake pressure on the affected wheel. This helps the wheel to
accelerate to gain wheel speed and traction. Once traction is regained, the system
reapplies brake pressure to decelerate the wheel. This sequence is repeated until the
vehicle stops or driver releases the brake pedal.
2.2.2 Traction Control System (TCS)
Traction control system [11] helps to avoid wheel spin during vehicle acceleration.
It assists driver in acceleration phase, preventing wheels from spinning by cutting
engine torque, thus improving acceleration performance and keeping the vehicle
stable. Traction control system continuously monitors wheel speed using wheel speed
sensor. If one of the wheels starts to accelerate rapidly than the other wheels, then
this shows that the wheel is losing traction and starting to spin. TCS system quickly
responds to this situation by increasing hydraulic brake pressure on the affected
wheel or by reducing drive torque to the affected wheel. This helps the wheel to
decelerate to gain the desired wheel speed and traction.
2.2.3 Electronic Stability Control (ESC)
Electronic Stability Control (ESC) system helps drivers to preserve control of the
vehicles in conditions where the vehicle is starting to lose control [12]. It uses
computer-controlled technology to apply individual brakes and help bring the car
safely back on track. System uses a steering wheel angle sensor to get the desired
direction of driver input and use a yaw rate sensor to get the vehicle heading direction.
66 K. Jayan and B. Muruganantham
This information can be used to calculate the desired path. If vehicle is not moving
on the desired path, then ESC system controls the individual brake to take vehicle to
intended path and improve the vehicle stability [13].
In oversteering situation, the rear end of the vehicle moves outwards, this loses
vehicle stability. If this condition happens during a left turn, then the ESC system
applies brake on the front right side, which creates a counter torque that brings the
vehicle to the desired path.
In understeering situation, the vehicle keeps straight-line path during a turning. If
this situation detects during a right turn, the ESC system applies brakes on the rear
right side, which creates a counter torque that brings the vehicle to the desired path.
2.2.4 Adaptive Cruise Control (ACC)
Adaptive Cruise Control (ACC) [14] controls vehicle longitudinal velocity. ACC
regulates the speed of vehicle to keep a safe distance from the vehicle ahead. This
system uses radar and camera sensors for precise control of the vehicle speed. Gener-
ally, the data from sensors (radar and camera) will be fused in radar. From this fused
data, ACC System identifies the object, its relative velocity, and relative distance.
Based on this information, ECU triggers the system to accelerate or decelerate the
vehicle.
The main functional enhancement of ACC, compared to standard cruise control, is
the ability of sensing forward traffic objects [15]. Dependent on the situation, ACC
changes automatically between two basic function modes: set speed control and
follow control [15]. Figure 8shows schematic diagram of Adaptive Cruise Control
(ACC) system.
In set speed control, if no preceding vehicle is detected, speed is controlled
according to the set speed. In the following control mode, if a slower preceding
vehicle is in the same lane, then speed is reduced to that of the preceding vehicle
and an appropriate distance is controlled automatically [15]. If necessary, moderate
braking by automatic brake actuation is used for deceleration.
Fig. 8 Adaptive Cruise Control (ACC)
Advanced Driver Assistance System Technologies … 67
Fig. 9 Lane keeping system
2.2.5 Automatic Emergency Brake Assist (AEB)
AEB is an active safety feature that warns the driver about an imminent collision with
an obstacle ahead, which may be either moving or stopped, and suggests decelerating
the vehicle. If the driver does not respond to the warning or the driver’s response is
not sufficient to avoid the crash, then the system takes control and applies brakes to
decelerate the vehicle, when an unavoidable collision is identified [16]. The function
of the system is to assist the driver to evade or reduce harshness of crash by reducing
the vehicle speed. This system uses camera and radar sensor for object identification.
2.2.6 Lane Keeping Assist (LKA)
Lane Keeping Assist system supports the driver to keep e vehicle in the center of
the lane with full speed range [17]. System takes lateral control to keep the vehicle
in the current lane [18]. This system allows the vehicle to change the lane if driver
intentionally changes the lane. The system continuously monitors the driver’s action
like steering input or turn indicator to identify driver intension. The system uses
camera sensor to detect lane markings. It also requires power steering system to
provide assist torque. Figure 9shows block diagram of lane keep assist system [18].
2.2.7 Lane Cantering (LC)
This system is similar to LKA, it not only intervenes on the steering wheel when
the vehicle is exceeding lane boundaries but also continuously controls the steering
wheel to keep vehicle in center of the lane. Lane Cantering system uses mono camera.
68 K. Jayan and B. Muruganantham
2.3 SAE Level Two Automation: Partial Automation
As per SAE, in level two automation [19], the vehicle is able to take lateral and
longitudinal control simultaneously but driver must remain engaged with the driving
task and monitor the environment at all times. In this level of automation, two or
more automated functions work together to relieve the driver.
2.3.1 Highway Assist (HA)
The Highway Assist provides support to driver by controlling longitudinal and lateral
dynamics of the vehicle on highways driving situations. The Highway Assist function
can automatically control vehicle acceleration, deceleration, as well as steering to a
certain extent. The driver has to monitor the environment and is ready to take over the
vehicle control at any time. It is partially automated driving function up to a speed
of 180 kmph on highways. The system combines the radar-based Adaptive Cruise
Control (ACC) with the video-based lane keeping support.
2.3.2 Autonomous Parking
Autonomous Parking system helps driver to perform parallel, perpendicular, or
angular parking. In this function, system takes longitudinal and lateral control of
the vehicle [20]. It improves the safety and comfort of driving in a constrained envi-
ronment. This system controls steering and speed in order to achieve the desired
vehicle path with in the available space. The speed and steering angle are controlled
with respect to the input from sensor data about the environment. In general, the
system uses ultrasonic sensor and camera to get the environment information.
2.3.3 Traffic Jam Assist
Traffic Jam Assist helps the driver in driving in highway traffic jams up to 60 km/h.
The Ego vehicle automatically follows the traffic vehicle ahead. The system uses
forward-looking radar and camera to sense traffic and road ahead [21]. The fused
data from both sensors is used for adapting Ego vehicle speed to the speed of traffic
vehicle ahead, with Stop-and-Go functionality. Camera helps to detect the road lane
marking that helps to keep vehicle in the canter of occupied lane. The driver needs
to monitor both surrounding, traffic, and system status hence driver can override the
system at any point of time. Figure 10 shows the Traffic Jam Assist system and it
helps dramatically to reduce efforts and stress of highway driving in congested traffic
conditions with enhanced vehicle and occupant safety.
Advanced Driver Assistance System Technologies … 69
Fig. 10 Traffic jam assist
2.4 SAE Level Three Automation: Conditional Automation
In this level, system does all jobs like monitoring the environment and controlling
lateral and longitudinal dynamics like steering acceleration or deceleration of the
vehicle but if any dynamic situation comes, then system warns the driver to take
the control action [19]. In this level, driver needs to be vigilant to the environment
because system transfers driving control to the driver if any dynamic situation arrives.
Transfer of vehicle control to driver during dynamic situation leads to safety-related
issues. So most of the Original Equipment Manufacturer (OEM) skip this level of
automation and directly jump to level four.
2.5 SAE Level Four: High Automation
In level four, vehicle is capable of performing all driving functions under all condi-
tions. The main difference between levels three and four is that the vehicle is capable
of handling all dynamic situations without having driver interventions [19]. This
level having high automation improves vehicle safety and comfort. Level four also
provides manual driving mode so that the automation control can switch between
human and machine.
70 K. Jayan and B. Muruganantham
2.5.1 Automatic Valet Parking (AVP)
AVP driverless system helps to identify a free space in parking garage and parks
vehicle itself by means of connected car technology [22]. This system allows vehicle
to drop driver and passengers at the drop-off area of parking garage and all the driver
needs are to be activated via smartphone app. This provides digital connection with
the parking garage, and the route to the free recognized parking spot. The vehicle is
equipped with a camera that helps system to visualize the driving route and detect
obstacles and pedestrian in the route so that vehicle can respond immediately. Camera
helps system to identify the free parking slot based on the vehicle size. The system
provides a complete automatic parking feature, even if the parking space is so tight.
2.6 SAE Level Five: Full Automation
In level five, vehicle does not have any driver interventions and vehicle can perform
all driving functions under all conditions [19]. In this level, vehicle control, envi-
ronment monitoring, and dynamic responses are handled by vehicle itself. This
system helps to improve vehicle and road safety by providing matured driving and
avoid human errors. There is no need for pedals, brakes, or a steering wheel, as the
autonomous vehicle system controls all critical tasks, monitoring of the environment,
and identification of unique driving conditions like traffic jams.
3 Challenges in ADAS System
Vehicles with high levels of ADAS capability considered Autonomous Driving.
The objective of active safety or Autonomous Driving is to reduce collisions and
ensures safety of vehicles on road. The technology uses sensors and control devices
at different levels to assist or fully take over the vehicle controls depending on the
levels of automation. In order to build the trust, adequate testing is required and
perhaps one fatal incident will actually reduce the feel of safety in autonomous cars.
After two fatal incidents involving a Tesla Model X and an autonomous vehicle Volvo
XC90, consumers have become less trusting of autonomous vehicles. Since the tech-
nology behind autonomous driving is majorly dependent on machine learning, no
explicitly specified behavior can be defined.
For tier1 suppliers, the challenges would be offering the best technology with
improved performance at lower cost to OEM. Since the base price for cars is
remaining relatively stable, semiconductor and other solution providers may face
pressure from OEM and customers to keep the ADAS cost low and make the
technology as standard.
For OEM, the challenge would be establishing a record of accomplishment of
safe cars and increasing customer feeling of safety about driverless cars. It is not
Advanced Driver Assistance System Technologies … 71
easy to reach that level without adequate testing for various environmental and road
conditions. Achieving the confidence levels requires more testing time and usually
it is limited to overall vehicle development time. Front loading the testing to virtual
simulation platforms is one of the practical solution followed across the industry.
4 Conclusion
Improving automotive safety and driver comfort is a major task of Advanced Driver
Assistance System. Based on the ADAS technology implemented in the vehicle,
the level of automation maturity can be addressed using the SAE level of automa-
tion standard. This paper describes SAE levels of automation and ADAS functions
implemented at each level of automation. The hardware and software technologies
are used for the ADAS function development, and the challenges toward develop-
ment of autonomous vehicle were addressed. At present, most of the ADAS functions
are implemented in high-end vehicles. The suppliers and OEM are putting effort to
reduce ADAS system cost and use the best technologies available in the market
without compensating the performance and safety. This helps to improve the devel-
opment of autonomous vehicle in the market. In near future vehicle, the basic SAE
Level 1 and Level2 ADAS features like AEB, ACC, Parking Assist, Surround-view
systems, etc. will be implemented. As a result, the number of road accidents will be
reduced.
References
1. W. H. Organization, Global Status Report On Road Safety 2018 (2018)
2. D. Assistance, M. Galvani, History and future of driver assistance. IEEE Instrum. Meas. Mag.
22(February), 11–16 (2019)
3. A. Shaout, D. Colella, S. Awad, Advanced driver assistance systems—past, present and
future, in ICENCO’2011—2011 7th International Computer Engineering Conference Today
Information Society What’s Next? (2012), pp. 72–82 (2012)
4. A. Carullo, M. Parvis, S. Member, An Ultrasonic Sensor for Distance Measurement in
automotive applications. IEEE Sens. J. 1(2), 143–147 (2001)
5. K. Dyczkowski, P. Gadecki, A. Kułakowski, Traffic signs recognition system
6. T. National, H. Traffic, S. Administration, Blind Spot Monitoring in Light Vehicles—System
Performance, no. July, 2014
7. R.N. Mahajan, A.M. Patil, Mazda lane departure warning system. Int. J. Eng. Tech. Res.
120–123 (2014)
8. E. Dagan, O. Mano, G.P. Stein, Forward collision warning with a single camera modeling
acceleration momentary time to contact
9. A.A. Aly, E.-S. Zeidan, A. Hamed, F. Salem, An antilock-braking systems (ABS) control: a
technical review. Intell. Control Autom. 02(03), 186–195 (2011)
10. J.D. Halderman, Automotive Technology, 4 edn. (2012)
11. Y.A. Ghoneim, W.C. Lin, Y.-K. Chin, D.M. Sidlosky, Enhanced traction stability control
system. SAE Tech. Pap. Ser. 1, 724 (2010)
72 K. Jayan and B. Muruganantham
12. A.T. Van Zanten, Bosch ESP systems: 5 years of experience. SAE Tech. Pap. Ser. 1, 724 (2010)
13. D. Kim, K. Kim, W.Lee, I. Hwang, Development of Mando ESP (Electronic Stability Program).
SAE Tech. Pap. Ser. 1, 724 (2010)
14. O. Gietelink, J. Ploeg, B. De Schutter, M. Verhaegen, Development of advanced driver assis-
tance systems with vehicle hardware-in-the-loop simulations. Veh. Syst. Dyn. 44(7), 569–590
(2006)
15. T. Sauer, J. Steinle, O. Tschernoster, The BMW active cruise control ACC. SAE Int. by, no.
724 (2000)
16. T. Henriksson, Driver assistance systems with focus on automatic emergency brake (2011)
17. J. Liu, J. Wu, Y. Su, C. County, Development of an Interactive Lane Keeping Control System
for Vehicle, no. 6
18. D. Aspects, O. F. The, U. Interface, S. Vehicle, R. Departures, H. M. Interface, Jochen Pohl
Jonas Ekmark Volvo Car Corporation Sweden Paper No. 529 Lane Keeping-Module Lane
Tracker EPAS Road Lateral Position, pp. 1–8
19. B. Wendling, First revision of J3016—update on task force activities. SAE Int.
20. A. Paul, R. Chauhan, R. Srivastava, M. Baruah, Advanced Driver Assistance Systems (2018)
21. A. Driving, Automated and Autonomous Driving Regulation under uncertainty Automated and
Autonomous Driving Regulation under uncertainty
22. T.J. Titus, S.A.V Balakrishnan, J. Sam, Fully Automated Valet Car Parking System, no. Mar
2017, pp. 69–72
Application of Artificial
Intelligence-Based Solution Methodology
in Generation Scheduling Problem
Shubham Tiwari, Vikas Bhadoria, and Bharti Dwivedi
Abstract Generation Scheduling commonly known as Unit Commitment Problem
is one of the most typical optimization problems in electrical engineering. In this
paper, Particle Swarm Optimization Technique with few modifications is imple-
mented as an Artificial Intelligence-based technique for the solution of this convo-
luted nonlinear optimization problem. The proposed technique is implemented on a
standard IEEE ten generator system. The entire work is performed in MATLAB envi-
ronment. The results have been compared and found better than the other available
techniques available in literature.
Keywords Generation scheduling problem (GSP) ·Modified particle swarm
optimization technique (MPSO) ·Priority list method (PLM)
1 Introduction
Generation Scheduling Problem is a highly complex optimization problem which is
of great concern for power system engineers and researchers. Proper scheduling of
generators can lead to significant profit for generation companies [1]. The nonlinear
nature of the problem with involvement of large number of constraints which keep
on changing with time makes this problem more severe. The prime objective of UCP
is to provide proper scheduling of generators to meet load forecast with minimum
cost in a scheduled time frame under the light of various constraints [2–4]. Numerous
S. Tiwari (B)·V. Bhadoria
Electrical & Electronics Engineering Department, ABES Engineering College, Ghaziabad, India
e-mail: shubham.tiwari@abes.ac.in
V. Bhadoria
e-mail: vikas.bhadoria@abes.ac.in
B. Dwivedi
Electrical Engineering Department, Institute of Engineering & Technology, Lucknow, India
e-mail: bharti.dwivedi@ietlucknow.ac.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_7
73
74 S. Tiwari et al.
efforts have been made to solve UCP with various classical, artificial intelligence-
based and hybridized techniques [5–15] but the scope of improvement has always
been there [16,17]. The reason behind choosing PSO was that it was found to be
better in comparison to other techniques [18–21].
The main asset of PSO is that it is robust with ease in parameter tuning, moreover,
it can be combined with various classical and artificial intelligence-based techniques
[21–24]. There are two cases considered in this paper. In the first case the forecasted
load is solely satisfied by the thermal generation and the results are compared with
other solution methodologies available in literature.
2 Problem Formulation
As the main objective is minimization of operational cost. The objective function
[25] is given as Eq. (1).
CostN=
N
i=1FCi(Pih)+STCi1−Ui(h−1)Uih (1)
where, FCiis the fuel cost of ith thermal generator and Pih is the output power of
ith thermal generator at hth hour & Uih is the ON/OFF status of the ith unit at hth
hour.
The fuel cost function is given as Eq. (2).
FCi(Pih)=ai+biPih +ciP2
ih (2)
where, ai,bi&ciare the fuel cost constants.
STCiis the cost incurred during the starting of ith thermal generator. This is given
as Eq. (3).
STCi=HSci:Xi(off)≤(MDi+Csi)hrs
CSci:Xi(off)>(MDi+Csi)hrs (3)
Xoff
iis the duration in which ith unit is continuously OFF.
A. Generation Scheduling Constraints
a. Equality Constraint (Power Balance Constraint)
The power balance constraint (generated power is equal to forecasted demand) is
given as Eq. (4).
24
h=1
N
i=1
Pih=
24
h=1
LDh(4)
Application of Artificial Intelligence-Based Solution … 75
where, Pih generated power (MW) of ith unit at hth hour & LDhis the forecasted
load at hth hour.
b. Spinning Reserve Constraint
Spinning Reserve is a part of generation (offline) which serves as backup during
contingency situations. It is given as Eq. (5).
N
i=1
Pi(max)Uih ≥LDh+SRh(5)
The maximum generation limit and the spinning reserve of individual generator
at hth hour are represented by Pi(max)&SR
h, respectively. SRhis taken as five
percent.
c. Inequality Constraint (Generation Limit Constraint)
The generation limit constraint is given as Eq. (6).
Pi(min)≤Pih ≤Pi(max)(6)
where, Pi(min)&Pi(max)are the minimum and maximum generating limits of
individual thermal generator.
d. Time Constraints (Minimum Up)
Xon
i(t)≥MUi(7)
e. Time Constraint (Minimum Down)
Xoff
i(t)≥MDi(8)
where, Xon
iis the continuous off time ith unit.
f. Initial Status
Down time status is taken at the start of the problem. The data regarding thermal
generation and load profile is given in Appendix.
76 S. Tiwari et al.
3 Solution Methodology
Generation Scheduling Problem (GSP) is a two-stage problem. In the first stage the
ON/OFF status of generating units is obtained while in the latter stage economic
dispatch is done. In the proposed AI-based hybrid solution methodology the first
stage solution is obtained by the initial priority vector given in Eq. (9), this initial
vector is updated as [20]. The dispatch is obtained by Modified Particle Swarm
Optimization Technique (MPSO) [25]. The second stage solution is obtained from
MPSO. The Classical PSO (CPSO) [25] is modified as Eqs. (15, 16).
priorityvector =P(max),vec
max .P(max),vec+MDvec
max.[MDvec](9)
v(k+1)
id =⎡
⎣
ω∗vk
id +c1f−c1i∗iter
itermax ∗c1i∗Rand1() ∗Pbestid−xk
id
+c2f−c2i∗iter
itermax ∗c2i∗Rand2() ∗Gbestgd −xk
id⎤
⎦
(10)
where, ω,c1,c2are the inertia weight and acceleration coefficients respectively.
ωis obtained as Eq. (11).
ωi=[1.1−(gbesti/pbesti)](11)
where, pbestiand gbesti are the local and global best positions of ith particle. The
velocity limits are taken as [25].
4 Results and Discussion
The results obtained from the proposed techniques are given in Table 1. The ON/OFF
status obtained from stage one results are shown as “green” and “blue” colors,
respectively. The stage two results are given as respective MW values. The conver-
gence characteristics are shown as Fig. 1and the comparison with other solution
methodologies is given in Table 2.
The convergence is obtained in the tenth iteration. The time taken by proposed
method is 6 s and overall operational cost obtained is 557,090$. The comparison of
the proposed method to other methods available in literature is given in Table 2.
Application of Artificial Intelligence-Based Solution … 77
Tabl e 1 Generation schedule for case one
Hrs.
Unit No. Tot.
Thermal Generators Gen.
Tg1 Tg2 Tg3 Tg4 Tg5 Tg6 Tg7 Tg8 Tg9 Tg10 (MW)
H-1
455
245
0 0
0
0 0 0 0
0
700
H-2 295 750
H-3 395 850
H-4
455
40 950
H-5 90 1000
H-6
130
60 1100
H-7 410
130
25 1150
H-8
455
30 1200
H-9 110 20 1300
H-10
162
43 25 1400
H-11 80 25 13 1450
H-12 80 25 53 10 1500
H-13 43 25
0 0
1400
H-14 110 20
0
1300
H-15 30
0
1200
H-16 310 25 1050
H-17 260 25 1000
H-18 360 25 1100
H-19
455
30 1200
H-20 162 43
25
1400
H-21
0
162 73 1300
H-22
0
145 20 1100
H-23 425 0 20 0 900
H-24 345 0 0 800
5 Conclusion
It is evident from Table 2that the proposed AI technique gives minimum operational
cost. Although the execution time is given in the table but the comparison would
not be justified as the computer efficacies are different. The proposed technique is
implemented with the help of a computer having 2 GB RAM and Intel core processor.
78 S. Tiwari et al.
0246810 12 14 16 18 20
5.58
5.585
5.59
5.595
5.6
5.605
5.61 x 105
No of Iterations
Operational Cost (Dollars)
Fig. 1 Convergence characteristic for case one
Tabl e 2 Comparison of
operational costs Method used Cost ($) Execution time (s)
BP [15]565,804 NA
GA[13]570,781 NA
APSO [24]561,586 NA
BP [24]565,450 NA
Advance Three Stage (ATS)
PLM +PSO +SMP [13]
557,677 NA
ATS- PSO [21]557,128 8.82
ATS-WIPSO [21]557,128 8.36
ATS-CPSO [21]557,128 7.73
ATS-WICPSO [21]557,128 6.58
Proposed 557,090 06.00
Appendix
Application of Artificial Intelligence-Based Solution … 79
TGs TG1 TG2 TG3 TG4 TG5 TG6 TG7 TG8 TG9 TG10
Pmax 455 455 130 130 162 80 85 55 55 55
Pmin 150 150 20 20 25 20 25 10 10 10
a($/h) 1000 970 700 680 450 370 480 660 665 670
b($/MWh) 16.19 17.26 16.60 16.50 19.70 22.26 27.74 25.92 27.27 27.79
c($/MW2h) 0.00048 0.00031 0.002 0.00211 0.00398 0.0072 0.00079 0.00413 0.0022 0.00173
MD (h) 8 8 5 5 6 3 3 1 1 1
MU (h) 8 8 5 5 6 3 3 1 1 1
HSc ($/h) 4500 5000 550 560 900 170 260 30 30 30
CSc ($/h) 9000 10000 1100 1120 1800 340 520 60 60 60
Cs (h) 5 5 4 4 4 2 2 0 0 0
Initial status 8 8 −5−5−6−3−3−1−1−1
80 S. Tiwari et al.
References
1. A.J. Wood, B.F. Wollenberg, Power Generation, Operation and Control, 3rd edn. (Wiley, New
York, 2013)
2. R.M. Bums, C.A. Gibson, Optimization of priority lists for a unit commitment program, in
Proceedings of the IEEE Power Engineering Society Summer Meeting (1975)
3. A. Shukla, S.N. Singh, PSO for solving unit commitment problem including renewable energy
resources. Electr. India 54(12), 100–105 (2014)
4. C. Shantanu, T. Senjyu, A.Y. Saber, A. Yona, T. Funabashi, Optimal thermal unit commitment
integrated with renewable energy sources using advanced particle swarm optimization. IEEJ
Trans. 4, 617 (2009)
5. W.L. Snyder, H.D. Powell, J.C. Rayburn, Dynamic Programming approach to Unit Commit-
ment. IEEE Trans. n Power Syst. 2, 339–351 (1987)
6. P.G. Lowery, Generating unit commitment by dynamic programming. IEEE Trans. Power
Apparatus Syst. 85(5), 422–426 (1966)
7. T. Aoki, T. Satoh, M. Itoh, T. Ichimori, K. Masegi, Unit commitment in a large scale power
system including fuel constrained thermal and pumped storage hydro. IEEE Trans. Power Syst.
2(4), 1077–1084 (1987)
8. W.J.Hobbs, G. Hermon, S. Warner, G.B. Sheble, An enhanced dynamic programming approach
for unit commitment. IEEE Trans. Power Syst. 3(3), 1201–1205 (1988)
9. T. Senjyu, K. Shimabukuro, K. Uezato, T. Funabashi, A fast technique for unit commitment
problem by extended priority list. IEEE Trans. Power Syst. 18(2), 882–888 (2003)
10. D. Mukarta, S. Yamashiro, Unit Commitment scheduling by lagrange relaxation method taking
into account transmission losses. Electr. Eng. Jpn. 152, 27–33 (2005)
11. W. Ongsakul, N. Petcharaks, Unit commitment by enhanced adaptive Lagrangian Relaxation.
IEEE Trans. Power Syst. 19(1), 620–628 (2004)
12. B. Xiaomin, S.M. Shahidehpour, Y. Erkeng, Constrained unit commitment by using tabu search
algorithm, in Proceedings of International Conferenceon Electrical Engineering, vol. 2 (1996),
pp. 1088–1092
13. F. Zhuang, F.D. Galiana, Unit commitment by simulated annealing. IEEE Trans. Power Syst.
5(1), 311–317 (1990)
14. T.Matsui, T. Takata,M. Kato, M. Aoyagi, M. Kunugi, K. Shimada, J. Nagato, Practical approach
to unit commitment problem using genetic algorithm and Lagrangian Relaxation method, in
Intelligent Systems Applications to Power Systems, Proceedings ISAP (1996)
15. A.Y. Saber, Scalable Unit Commitment by memory bounded ant colony optimization with a
local search. Electr. Power Energy Syst. 30(6–7), 403–414 (2008)
16. S. Tiwari, B. Dwivedi, M.P. Dave, Advancements in unit commitment strategy, in Proceedings
of ICETEESES, 11–12 Mar 2016, pp. 259–262
17. S. Salem, Unit commitment solution methods, in World Academy of Science, Engineering and
Technology (2007), pp. 320–325
18. A. Shukla, S.N. Singh, Multi-objective unit commitment using search space based crazy PSO
and Normal boundary intersection technique. IET Gener. Transm. Distrib. 10(5), 1222–1231
(2016)
19. S. Tiwari, A. Kumar, G.S. Chaurasia, G.S. Sirohi, Economic load dispatch using particle swarm
optimization. IJAIEM 2(4), 476–485 (2013)
20. S. Khanmohammadi, M. Amiri, M. Tarafdar Haque, A new three stage method for solving unit
commitment method. Energy 35, 3072–3080 (2010)
21. V.S. Pappala, I. Erlich, A new approach for solving Unit Commitment Problem by adaptive
particle swarm optimization, in Power and Energy Society General Meeting-Conversion and
Delivery of Electrical Energy in the 21st Century (IEEE, USA, 2008), pp. 1–6
22. P. Sriyanyong, Y.H. Song, Unit commitment using particle swarm optimization combined
with Lagrangian relaxation, in Power Engineering Society General Meeting, vol. 3 (2005),
pp. 2752–2759
Application of Artificial Intelligence-Based Solution … 81
23. Z.L. Gaing, Discrete particle swarm optimization algorithm for unit commitment, in IEEE
Power Engineering Society General Meeting, vol. 1 (2003), pp. 13–17
24. Y.W. Jeong, J.B. Park, S.H. Jang, Y.L. Kwang, A new quantum-inspired binary PSO: application
to unit commitment problems for power systems. IEEE Trans. Power Syst. 25(3), 1486–1495
(2010)
25. S. Tiwari, A. Maurya, Particle swarm optimization technique with time varying acceleration
coefficients for load dispatch problem. IJRITCC 3(6), 3878–3885 (2015)
Discomfort Analysis at Lower Back
and Classification of Subjects Using
Accelerometer
Ramandeep Singh Chowdhary and Mainak Basu
Abstract Lower back pain problems are increasing these days because of a seden-
tary lifestyle and working habits such as using laptops or computers for long hours
and sitting on chairs for continuous duration. Back injury occurrence is also very
common in athletes and workers where lifting of heavyweights are required. This
research work aims to classify subjects and analyze discomfort at lower back by
performing wireless data acquisition using accelerometer sensor. The designed hard-
ware consisting of accelerometer sensor, NRF wireless module, and micro-controller
was used to implement node-hub architecture. This research work shows the clas-
sification of subjects based on lower back vibration data. Multiple classification
algorithms such as decision trees, random forest, naive Bayes, and support vector
machine were applied to perform subject classification after performing experiments.
The analysis of data shows that the subjects could be classified based on the discom-
fort level of the lower back using accelerometer data. Such a kind of study could
be used for the prediction of the core strength of lower back and treatment of lower
back problems by analyzing vibrational unrest.
Keywords Accelerometer ·Back analysis ·Classification ·Discomfort analysis ·
Micro-controllers ·Wearable sensors ·Wireless sensor node
1 Introduction
Human lower back discomfort and pain analysis is an important domain for gait
researchers as it gives various parameters for classification of subjects and prediction
of injuries related to the lower back area. Oliverio et al. [1] stated in their research
work that pain in the pelvic region is a very common problem found in almost all
R. S. Chowdhary (B)·M. Basu
GD Goenka University, Gurgaon, Haryana, India
e-mail: raman85friends@gmail.com
M. Basu
e-mail: mainak.basu87@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_8
83
84 R. S. Chowdhary and M. Basu
parts of the world specially in developed countries where people had a sedentary
lifestyle. This problem raises a huge expense of approximately 2 billion dollars
in a developed country like the United States in a year and hence could have a
significant impact on the social and economic state of any nation. Taghvaei et al.
[2] discussed another important lower back issue which is locomotion for daily
activities. Elderly persons face difficulties in performing daily activities due to lack
of muscular strength, musculoskeletal injuries, lower back pain, etc. In some of the
recent research work, sensor measurement units based on MEMS were used for lower
body posture analyses. Urukalo et al. [3] designed a wearable device to find a solution
to chronic problems such as back pain and musculoskeletal disorders which had no
technological or physiological hurdles. Wenjing et al. [4] stated in their research
work that medicines or surgeries are required for persons with acute lower back pain
in order to improve the health condition. So to avoid the use of such medicines or
undergo surgeries it is always better to analyze the discomfort at the lower back of
the patients and opt for corrective measures before it converts to severe back pain.
Some recent systems used to analyze body movements and capture data using
sensors were used extensively by researchers for classification and predictions.
Molnar et al. [5] used 6D inertial measurement units to access lower back motions.
Xu et al. [6] proposed an algorithm which used the angle between two IMU sensors
modules and did not use ground reference for angle calculation. Chutatape et al. [7]
had implemented a system using one smartphone and measured joint angles at the
hip location which led to inappropriate body postures and could result in disloca-
tion of joint. Kam et al. [8] developed a plastic optical fiber-based sensor and used
intensity interrogation technique to analyze backbone bending in the sagittal plane.
Dobrea et al. [9] designed a wearable warning system which identified an improper
sitting posture which could generate cervical pains. This was done by defining several
triggering zones and identifying the presence of the system in those zones. Sandag
et al. [10] presented machine learning algorithm to classified subjects with back
pain using K-Nearest Neighbor (KNN) classification algorithm on non-real-time
data and found that Degree Spondylolisthesis had the most important effect on lower
back pain syndrome. Further, Galbusera et al. [11] described various techniques of
artificial intelligence including machine learning that were being developed with
special focus on those used in spine research.
To overcome these problems a single sensor-based instrument was designed in
this research work and was used for discomfort analysis of lower back. The collected
data was also used to train different classifiers which will eventually perform clas-
sification of the test sample. Multiple classifiers were used to compare performance
metrics for classification of subjects and hence propose the best-suited classifier for
accelerometer related data.
Discomfort Analysis at Lower Back and Classification … 85
2 Methodology
Complete flow and steps involved in the system architecture are shown and explained
in Fig. 1. In this the first step was to collect data from the sensor. The data collection
was done when the subject had started performing experiments while wearing wire-
less sensor node at the lower back. Sensor calibration was carried out by keeping
the sensor node stationary before it starts capturing data from the sensor. As soon
as the node starts getting data it performs wireless transfer of data to the hub. With
the help of a Java software the received data at the hub side was stored in the form
of CSV files for creating logs and preprocessing. In the process of preprocessing the
concept of data segmentation and data windowing was used. After completion of the
preprocessing task, feature extraction was performed to extracts features from the
collected data from the subjects. The entire set of data was split into two sets viz.
training set and testing set. In the present research work various classifiers were used
to classify subjects based on their discomfort level at the lower back. A comparison
of performance metrics of different classifiers was carried out to find the best-suited
classifier for the given set of data. Both training set and testing set were fed to the
classifiers to get the final predicted result as shown in Fig. 1. Shabrina et al. [12]in
their research work used visual analogue scales and pain questionnaire methods to
analyze lower back pain due to prolonged standing on inclined surface. Some of the
Fig. 1 Flow and steps
required for the complete
system
86 R. S. Chowdhary and M. Basu
important features in such kind of classification process are gender, age, vibrational
data in all three axis, nature of job of the subjects, fitness level of the subjects, and
hold time of the upper body while performing experiments in the sagittal plane.
3 Hardware Design
The implemented system was designed, fabricated, and was used to perform experi-
ments on 20 different subjects. The complete system consists of 1 NRF based sensor
node, which was mounted at the lower back of the subjects while performing exper-
iments. Secondly, it consists of 1 NRF based hub, which was connected with the
laptop to creates data logs of the received data. Following sections explain hardware
infrastructure and software algorithms used in this research work.
Real-time data acquisition was implemented in this system at the time of
performing experiments. Arduino Nano micro-controller board was used in node and
hub design. In the case of node architecture the micro-controller board was interfaced
with the accelerometer sensor. The data which was transmitted from node to hub was
transferred through wireless communication. To perform wireless communication an
NRF chip was used as it is a low power device and can operate on battery. The block
diagram of wireless sensor node is shown in Fig. 2where a battery-powered Arduino
Nano is interfaced with I2C based inertial sensor and NRF wireless communication
chip.
The hub is connected at the remote site with a laptop which received sensor
data and created log files in CSV format. The hub architecture consists of Arduino
Nano micro-controller board and NRF wireless communication module. It also has
a synchronization switch which was used to send synchronization signal to the node
to start sending sensor data to the hub for data acquisition.
Fig. 2 Block diagram of
wireless sensor node
Discomfort Analysis at Lower Back and Classification … 87
4 Software Algorithms
Python programming language was used to perform discomfort analysis at lower
back and classification of subjects based on extracted features. The detailed steps
and algorithm are explained in the following sections.
A. Data Preprocessing
Sensor data acquisition frequency is directly proportional to the resolution of the
acquired signal value. Typical consumer accelerometer sensor like MPU6050 can
support communication using I2C at 400 kHz. It is known that fsampling is approxi-
mately equal to 50 Hz is the most suitable value for sampling frequency. The present
research work is focused on the detection of accelerometer data at the lower back in
sitting upright and forward lean position. The classification algorithms were written
in python which detects the subjects with pain in lower back while performing the
experiment in forward lean position. It uses various features extracted from the accel-
eration data after preprocessing. This is possible after segmentation of data and it is
stored in with window size. The captured data is processed by various classification
algorithms after the data window is filled completely. It is always good to have large
window size as it helps in providing stable status. Large window sizes sometimes
come with problem of data overlapping for different activities as it was true for sitting
upright position and lean forward position while performing experiments. To avoid
overlapping problem window size was selected to fit in two experimental positions
separately.
B. Classification Method
Classification of subjects based on discomfort analysis of lower back was done
in this research work. Various classification algorithms were used in this research
to get the predicted output. Four classification algorithms viz. Navie Bayes (NB),
Decision Trees (DT), Random Forest (RF), and Support Vector Machine (SVM)
were used and their performance metrics were compared. This gives the best-suited
classification algorithm for related classification problems. Nezam et al. [13]showed
in one of the related research work that support vector machine (SVM) had higher
classification accuracy of 83.5% for three pain levels than K-nearest neighbor (KNN)
classifier having 80.5% classification accuracy. Abdullah et al. [14] also performed
classification of spinal abnormalities and showed that the classification accuracy of
KNN algorithm was 85.32% which was better than the percentage of accuracy in case
of random forest (RF) classifier which was 79.56%. Estrada et al. [15] also performed
posture recognition using cameras and smartphones and had highest classification
accuracy with decision trees classifier as 89.83% (for spinal posture) and 95.35%
(for head and shoulder posture) as compared to KNN and SVM classifier algorithms.
C. Feature Selection and Extraction
Tian et al. [16] in their research work proposed a system which used three types
of feature extraction namely original features, linear discriminant analysis (LDA)
88 R. S. Chowdhary and M. Basu
features, and kernel discriminant analysis (KDA) features. In this research work
several features were extracted from accelerometer data in all three axis (viz. x-axis,
y-axis, and z-axis). Features were also extracted which was composed of data from
all three dimensions of accelerometer sensor. Detailed explanation of the extracted
features is given below where Axi,Ayi, and Azi were the acceleration values in all
three axes and the total number of samples in each axis is denoted by N.
•Mean in x-axis for sitting upright position is defined as the summation of accel-
eration values in x-axis while sitting upright divided by number of samples, i.e.,
N.
µ(Ax)sit =1
N
N
i=1
Axi (1)
•Mean in y-axis for sitting upright position is defined as the summation of accel-
eration values in y-axis while sitting upright divided by number of samples, i.e.,
N.
µ(Ay)sit =1
N
N
i=1
Ayi (2)
•Mean in z-axis for sitting upright position is defined as the summation of accel-
eration values in z-axis while sitting upright divided by number of samples, i.e.,
N.
µ(Az)sit =1
N
N
i=1
Azi (3)
•Variance in x-axis, VAx is given as the spread of the accelerometer data around
the mean in x-axis,
VAx =1
N
N
i=1
(Axi −µ(Ax)lean)2(4)
•Variance in y-axis, VAy is given as the spread of the accelerometer data around the
mean in y-axis,
VAy =1
N
N
i=1Ayi −µ(Ay)lean 2(5)
Discomfort Analysis at Lower Back and Classification … 89
•Variance in z-axis, VAz is given as the spread of the accelerometer data around the
mean in z-axis,
VAz =1
N
N
i=1
(Azi −µ(Az)lean)2(6)
Similarly, mean acceleration values (µ(Ax)lean ,µ(Ay)lean , and µ(Az)lean ) for lean
forward position was also calculated as was done in case of mean acceleration values
for sitting upright position in Eqs. 1,2, and 3.
5 Experimental Setup
The experiments were designed in such a manner that it will be able to capture the
accelerometer data from the sensor attached at the lower back of the subjects. A
written consent was taken from all subjects before performing experiments. The
subjects were addressed to perform predefined experiments calmly and without
feeling any pressure. All subjects were counseled about the execution of the exper-
iment in order to collect neutral data from all experiments. Experimental data was
collected from 12 subjects (2 females and 10 males) with ages from 24 to 40 years.
The device was fastened by a belt at the middle lower back to capture the data. The
subjects were asked to sit stationary in an upright position on a chair for 3 s and
then lean forward in the sagittal plane and stay there in that position for 10 s. The
normal vibrational values of lower back at sitting upright position were collected
and stored for analysis. Further the vibrational values from the accelerometer sensor
were collected when the lower back is at discomfort position as the spine was under
stress in the forward lean position. Figure 3shows the 2 positions used for getting
accelerometer data. The data was collected and transmitted wirelessly to the hub
where logs were created.
A. Experiment with Subjects
To perform experiments the subjects were equipped with sensor node which was
installed at their lower back. Sensor position was selected in such a manner to avoid
thick muscles which could affect the accuracy of classification. Further, the subjects
were asked to sit on a chair in an upright position for calibration of the sensor. The
calibration process will take 5 s and after that a status LED available on the sensor
node blinks to indicate completion of calibration process. Next, the synchronous
signal is transmitted from the hub by pressing a synchronization switch. The subject
will remain in upright position for 5 s so that the lower back acceleration data can be
transmitted to the hub. Then the subject was moved in a forward lean position making
a posture angle of less than 70° as shown in Fig. 3. Ikegami et al. [17] developed a
90 R. S. Chowdhary and M. Basu
Fig. 3 a Sitting upright
(Position: P1) and blean
forward (Position: P2) while
performing experiments
chair which prevents lower back pain while prolonged sitting and doing handwork
at the same time. To minimize the effect of sitting for long hours on the lower back,
the experiments in the present research work was designed for short duration.
6 Experimental Results
Table 1shows data values for different features used for discomfort analysis at lower
back and classification of subjects. In the experiment conducted, 10 male and 2
female subjects participated who performed an experiment at sitting upright and
lean forward position. It was found that mean acceleration value in sitting position
around x-axis was 9.893 and in lean forward position was 9.737. This shows that
subjects had more vibrations at the lower back while they were in sitting position
as compared to lean forward position for a duration of 10 s. This happened because
Tabl e 1 Details of features
used for discomfort analysis
and classification by different
classifiers
Feature Val u e Remark
µ(Ax)Sit 9.893 Mean of Ax at sitting upright position
µ(Ay)Sit 0.581 Mean of Ay at sitting upright position
µ(Az)Sit 1.457 Mean of Az at sitting upright position
µ(Ax)lean 9.737 Mean of Ax at lean forward position
µ(Ay)lean 0.445 Mean of Ay at lean forward position
µ(Az)lean 2.494 Mean of Az at lean forward position
VAx 0.047 Gives spread of data around µ(Ax)lean
VAy 0.036 Gives spread of data around µ(Ay)lean
VAz 0.089 Gives spread of data around µ(Az)lean
Discomfort Analysis at Lower Back and Classification … 91
x-axis is the axis which is passing through the lower back of the subjects in coronal
plane. On the other hand, mean acceleration values in sitting position around z-axis
was 1.457 and in lean forward position was 2.494. This shows that subjects had
more vibrations at the lower back while they were in position P2 (i.e., lean forward
position) as compared to position P1 (i.e., sitting upright position). This happened
because z-axis is the axis which is passing through the lower back of the subjects
in the sagittal plane. Hence more vibrational unrest is there in lean forward position
as compared to sitting upright position in the z-axis direction. Mean of acceleration
values in sitting and lean forward positions in all three axes were denoted as µ(Ax)sit,
µ(Ay)sit ,µ(Az)sit ,µ(Ax)lean ,µ(Ay)lean, and µ(Az)lean. All these values are shown in
Table 1along with variance in all three directions in lean forward position denoted
by VAx,VAy , and VAz.
Figure 4shows the accelerometer data for lower back at sitting upright and lean
forward position in y-axis direction. It showed the vibrational acceleration data for
approximately first 2500 samples in sitting upright position and the vibrational accel-
eration data for approximately 2000 samples (from sample no. 3000 to sample no.
5000) in lean forward position. The performed experiment showed that the average
acceleration had different values and is easily recognizable from Fig. 4. The measured
values also indicate that there is more vibrational unrest in the lean forward posi-
tion as µ(Az)lean >µ(Az)sit . This type of study can be used as meaningful data for
predicting posture correction techniques while playing and recovering from back
pain problems.
Figure 5shows the acceleration data (in all three directions) of lower back of
the subject while performing experiment. The shift in the acceleration values was
observed in the graphs which shows the movement of the subject while sitting and
performing experiment in the sagittal plane. The subject was sitting stationary for
the initial phase of the experiment and is denoted by the first 2700 sample values
in the graph. The subject then moved from position P1 (Sitting upright position) to
position P2 (Lean forward position). Position P2 is denoted by sample starting from
sample number 4000 and going up to sample number 6500.
Fig. 4 Acceleration data in
Y-axis direction for sitting
upright and lean forward
position
92 R. S. Chowdhary and M. Basu
Fig. 5 Acceleration data in Y-axis direction for sitting upright and lean forward position
Table 2shows the values of different features which were used with various clas-
sification algorithms as discussed in Sect. 4part B. AxS,AyS, and AzSdenote accel-
eration values in sitting upright position for x-axis, y-axis, and z-axis, respectively.
Similarly, AxL,AyL, and AzLdenote acceleration values in lean forward position
for x-axis, y-axis, and z-axis, respectively. VAx,VAy , and VAz denote variance for
Tabl e 2 Experimental values of classification features for sitting upright and lean forward position
AxSAySAzSAxLAy LAzLVAx VAy VAz
9.67 −0.83 2.98 9.57 −0.56 3.34 0.05 0.04 0.08
9.63 −0.92 3.06 9.61 −0.11 3.23 0.05 0.03 0.11
9.45 −0.93 3.57 9.47 −0.76 3.55 0.06 0.04 0.09
9.44 −0.94 3.55 9.46 −0.75 3.53 0.06 0.04 0.09
10.07 −0.80 1.04 9.48 −0.45 3.56 0.04 0.03 0.06
10.04 −0.66 1.38 9.71 −1.18 2.73 0.03 0.04 0.04
10.11 0.01 0.87 10.06 −0.24 1.35 0.03 0.05 0.06
10.10 −1.06 −0.41 9.96 −0.13 1.91 0.05 0.05 0.11
10.09 0.19 1.00 9.90 −0.37 2.22 0.04 0.03 0.06
10.14 0.06 −0.39 10.15 −0.10 −0.04 0.03 0.03 0.05
9.90 −0.26 1.83 9.40 −0.25 3.89 0.08 0.03 0.18
10.07 −0.84 −1.02 10.07 −0.43 0.73 0.04 0.02 0.14
Discomfort Analysis at Lower Back and Classification … 93
lean forward position in x-axis, y-axis, and z-axis, respectively. It was observed from
the table that there were increased values of feature AzLas compared to feature AzS
hence there is more vibrational unrest in Z-axis direction. A higher vibrational unrest
shows more discomfort at lower back due to unstable body posture. This data was
analyzed and matched with the subjective discomfort data collected from subjects
while performing experiments. The comparison of experimental data with the subjec-
tive data helps in assigning target class to the data set. This enables the classifiers
to classify the test data into binary class having value as ClassCand ClassDwhere
ClassCsignifies the target class which will have subjects with no or less discomfort
in lean forward position and ClassDsignifies the target class which will have subjects
with low or high discomfort in lean forward position.
Figure 6shows the confusion matrix for predicted classes using random forest
(RF) classification algorithm. In this the target class was set to “1” (for ClassD, i.e.,
subjects with discomfort at lower back in forward lean position) and it was set to “0”
(for ClassC, i.e., subjects with no discomfort at lower back in forward lean position).
There were true labels (0 and 1) for actual class and predicted labels (0 and 1) for
predicted classes. It was observed that random forest classification algorithm was
having a classification accuracy of 80% which was highest as compared to decision
trees, naïve Bayes, and support vector machine. Figure 6also shows that there was
one false prediction when the target class was falsely predicted as “1”. The complete
set of data was divided into two sets namely training set and test set. The prediction
was carried out the testing set after training the classification algorithms. Two more
performance metrics were calculated for all classifiers, i.e., F-score and computation
time and was found that FScore is 0.66 and computation time was 0.028 s for random
forest classifier.
Fig. 6 Classification result
in form of confusion matrix
for random forest classifier
94 R. S. Chowdhary and M. Basu
The classifiers were evaluated by three different performance measures viz. preci-
sion, recall, and FScore. The values of various performance metrics can be calculated
by using Eqs. 7,8, and 9.
Precision =True P
True P+FalseP
(7)
Recall =True P
True P+FalseN
(8)
FScore =2∗Precision ∗Recall
Precision +Recall (9)
Equation 7,8, and 9were used to calculate values of respective performance
metrics and the results were calculated as Precision =0.5, Recall =1 and FScore =
0.66.
7 Conclusion
The research work resulted in designing of a new device and presented a new way of
classification of subjects based on discomfort analysis at lower back using accelerom-
eter data. A wireless sensor node-hub architecture was used for data acquisition and
creating logs. This research proved that classification of person having discomfort at
lower back could be done from accelerometer data from the sensor mounted at the
lower back. It showed that the classification accuracy with random forest classifier
was highest with 80% accuracy as compared to naive Bayes, decision trees, and
support vector machine classifiers. It was also found that there is higher vibrational
unrest around z-axis in lean forward position as compared to sitting upright position.
The mean acceleration value in z-axis was 1.457 in sitting upright position and it was
2.494 in z-axis in lean forward position.
References
1. V. Oliverio, O.B. Poli-Neto, Case study: classification algorithms comparison for the multi-
label problem of chronic pelvic pain diagnosing, in IEEE 33rd International Conference on
Data Engineering, pp. 15071509 (2017)
2. S. Taghvaei, Y. Hirata, K. Kosuge, Visual human action classification for control of a
passive walker, in IEEE 7th International Conference on Modeling, Simulation and Applied
Optimization (ICMSAO) (2017), p. 15
3. D. Urukalo, P. Blazevic, S. Charles, J.-P. Carta, The TeachWear healthcare wearable device, in
27th IEEE International Symposium on Robot and Human Interactive Communication (2018),
pp. 638–643
Discomfort Analysis at Lower Back and Classification … 95
4. W. Du, O.M. Omisore, H. Li, K. Ivanov, S. Han, L. Wang, Recognition of chronic low back
pain during lumbar spine movements based on surface electromyography signals. IEEE Access
6, 65027–65042 (2018)
5. M. Molnar, M. Kok, T. Engel, H. Kaplick, F. Mayer, T. Seel, A method for lower back motion
assessment using wearable 6D inertial sensors, in 21st IEEE International Conference on
Information Fusion (FUSION) (2018), p. 799806
6. W. Xu, C. Ortega-Sanchez, I. Murray, Measuring human joint movement with IMUs, in 15th
IEEE Student Conference on Research and Development (2018), pp. 172–177
7. O. Chutatape, K. Naonueng, R. Phoophuangpairoj, Detection of improper postures leading to
dislocation of hip prosthesis by a smartphone, in 14th IEEE International Conference on Elec-
trical Engineering/Electronics, Computer, Telecommunications and Information Technology
(ECTI-CON) (2017), pp. 415–418
8. W. Kam, K. O’Sullivan, W.S. Mohammed, E. Lewis, Low cost portable sensor for real-time
monitoring of lower back bending, in 25th IEEE International Conference on Optical Fiber
Sensors (2017)
9. D.-M. Dobrea, M.-C. Dobrea, A warning wearable system used to identify poor body postures,
in IEEE International Conference on Advances in Wireless and Optical Communications
(2018), pp. 55–60
10. G.A. Sandag, N.E. Tedry, S. Lolong, Classification of lower back pain using K-Nearest
Neighbor algorithm, in 6th IEEE International Conference on Cyber and IT Service Manage-
ment (CITSM 2018) (2018), pp. 1–5
11. F. Galbusera, G. Casaroli, T. Bassani, Artificial intelligence and machine learning in spine
research. JOR Spine, Jan 2019
12. G. Shabrina, B.M. Iqbal, D.H. Syaifullah, Effect of shoes on lower extremity pain and low
back pain during prolonged standing on a sloping medium, in IEEE International Conference
on Intelligent Informatics and Biomedical Sciences (ICIIBMS), vol. 3 (2018), pp. 181–187
13. T. Nezam, R. Boostani, V. Abootalebi, K. Rastegar, A novel classification strategyto distinguish
five levels of pain using the EEG signal features. IEEE Trans. Affect. Comput. 1–9 (2018)
14. A.A. Abdullah, A. Yaakob, Z. Ibrahim, Prediction of spinal abnormalities using machine
learning techniques, in IEEE International Conference on Computational Approach in Smart
Systems Design and Applications (ICASSDA) (2018), pp. 1–6
15. J. Estrada, L. Vea, Sitting posture recognition for computer users using smartphones and a web
camera, in IEEE Region 10 Conference (TENCON), pp. 1520–1525 (2017)
16. Y. Tian, J. Zhang, L. Chen, Y. Geng, X. Wang, Single wearable accelerometer-based human
activity recognition via kernel discriminant analysis and QPSO-KELM classifier. IEEE Access
7, 109216–109227 (2019)
17. K. Ikegami, H. Hirata, H. Ishihara, S. Guo, Development of a chair preventing low back pain
with sitting person doing hand working at the same time, in IEEE international conference on
mechatronics and automation (2018), pp. 1760–1764
Design of Lyapunov-Based Discrete-Time
Adaptive Sliding Mode Control for Slip
Control of Hybrid Electric Vehicle
Khushal Chaudhari and Ramesh Ch. Khamari
Abstract This paper has developed discrete-time Fuzzy Adaptive sliding mode
control algorithm for controlling the slip ratio of a hybrid electric vehicle. Fuzzy
logic algorithm is used to develop controller for controlling slip ratio so as to handle
different road conditions. A discrete-time Sliding Mode Observer is designed to
observe the vehicle velocity. Furthermore, an adaptive SMC has been designed
by employing Lyapunov theory in order to adapt with slip dynamic change for
varying or changing road conditions. The performances of designed controller such
as ASMC, SMO, FLC, and Fuzzy PID for controlling slip ratio are compared using
MATLAB simulation and it is proved that the discrete-time fuzzy ASMC perform
most impressively and effectively.
Keywords Discrete-time sliding mode control ·SMO ·FLC ·HEV ·Slip ratio ·
FSMC
1 Introduction
Hybrid electric vehicles have good energy efficiency and reduce emissions. Hybrid
electric vehicles (HEVs) are more comfortable and preferable over conventional
vehicles (ICVs) [1]. Furthermore, multiple sources that provide power are used in
HEVs for driving. Further, with the help of control performance such as TCS and
ABS [2], it is easier to get and achieve useful, human favorable, and advanced
driving performance. Actuation system (electric motor) of HEV always produces
uncertainty error or variation in driving or braking torque as it is not present in
K. Chaudhari
Department of Electronics and Telecommunication, Government Polytechnic, Jintur, Maharashtra
431509, India
e-mail: khushalchaudhari@gmail.com
R. Ch. Khamari (B)
Department of Electrical Engineering, Government College of Engineering, Keonjhar, Odisha
758002, India
e-mail: rameshkhamari_fee@gcekjr.ac.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_9
97
98 K. Chaudhari and R. Ch. Khamari
that of hydraulic braking system and combustion engine system. Corresponding to
the main function of propulsion, because of torque performance and regeneration
capability of motors, electric motor will alternatively use the braking option. The
main problem during braking on roads according to different road conditions like
slippery road, road covered with ice, etc., is that locking condition of the wheels of
the vehicle may get occurred which implies that wheel speed nearly tends to approach
toward zero and which may result in that vehicle leaving the road. Hence, it creates the
possibility of overstepping, collision, or accident of the vehicle. The goal of Antilock
Braking System is to maximize wheel traction of the vehicle by preventing the wheels
from locking as wheel speed approaches to zero during braking while maintaining
vehicle stability on the road. Due to nonlinearity present in system dynamic and
time-variant unknown parameters related with wheel characteristic, controlling of a
braking system of electric and hybrid electric vehicle are difficult.
A numerous control strategy likes Iterative Learning Control [2], sliding mode
control (SMC) [3], fuzzy logic [4], neural networks [5], etc., have been used for
controlling the slip ratio of vehicles. Also, advanced controlling strategies such as
neural network [6], fuzzy logic control [7], adaptive control [8], hybrid control [9],
and PWM technique [10] have been developed for vehicles with IC engine to achieve
effective antilock braking performance. Undesirable phenomenon of oscillations
having finite frequency and amplitude is known as chattering [11]. Chattering is
a problem as it leads to less accuracy and more power losses.
The objective is to develop a controller for achieving the required value of slip
ratio such that locking of wheel is avoided and furthermore the controller must handle
the nonlinearity and time variation present in HEV dynamics (wheel and vehicle
dynamic). Also, observer design is required as vehicle speed is not measurable.
Even though in the recent literature, SMC and observer were used for designing
for slip ratio controlling of HEV earlier but discrete-time controller is essential for
practical implementation of controller on system. Hence unlike [2,4,10], this work
is focused on development of discrete-time sliding mode observer, discrete-time
Adaptive Sliding mode control, Fuzzy logic control, and Fuzzy PID control for slip
control of HEV.
The remaining structure of the paper is described as follows. Development of
discrete-time Slip Control model and actuator dynamic presented in Sect. 2, Sect. 3
describes the problem development, Sect. 4describes design of SMO, Sect. 5
describes design of adaptive SMC, design of FLC for SRC is described in Sect. 6,
Sect. 7describes fuzzy PID controller design, Sect. 8presents simulation results,
and Sect. 9concludes the paper.
Design of Lyapunov-Based Discrete-Time Adaptive Sliding … 99
2 Slip Control Model and Actuator Dynamic
The dynamics of vehicle in discrete domain is given by [12]
x1(k+1)−x1(k)
T=f1(x1(k))+d1μ(λ(k))(1)
x2(k+1)−x2(k)
T=f2(x2(k))+d2μ(λ(k))+d3τm(k)(2)
where Tis sampling time, x1=ωvand x2=ωw,
f1(x1)=−
cvrw
mx2
1,f2(x2)=−
rwfw(x2)
Iw
d1=nwg
rw
,d2=−
rwmg
Iw
,d3=1
Iw
Variables information are rwis radius of wheel, mis vehicle mass, Iwis moment
of inertia of wheel, nwis number of wheel, gis acceleration due to gravity, fwis
viscous wheel friction force, τmis braking torque, vis vehicle linear velocity, ωvis
angular velocity of vehicle, ωwis angular velocity of wheel, cvis aerodynamic drag
coefficient, μis adhesive coefficient.
The actuator consists of a dc motor for HEV given in [13]. The torque is negative
while brake applied on wheel of vehicle. Voltage required to produce braking torque
in discrete domain [12]is
e(k)=L
Km
τm(k+1)−τm(k)
T+R
Km
τm(k)+Kbωw(k)(3)
3 Problem Formulation
Slip ratio is defined as [12]
λ=ωw−ωv
max(ωv,ω
w)(4)
Equation (4) is rewritten as
λ(k)=x2(k)−x1(k)
max(x1(k), x2(k))(5)
Deceleration [12]:
For deceleration, x1>x2and hence
100 K. Chaudhari and R. Ch. Khamari
λ(k+1)−λ(k)
T=f(λ, x)+bu (6)
where f(λ, x)=f2(x2)−(1+λ)f1(x1)+[d2−(1+λ)d1]μ(λ)
b=d3,u=τm
x1
x=[x1,x2]T
Acceleration [12]:
For acceleration, x2>x1therefore,
λ(k+1)−λ(k)
T=f(λ, x)+bu (7)
where f(λ, x)=(1−λ)f2(x2)−f1(x1)+[d2(1−λ)−d1]μ(λ)
b=(1−λ)d3,u=τm
x2
Our main solution is for achieving effective braking and for these, goal of the
paper is to get control input uusing Eq. (6) as it is known that Iwis directly related
with bwhich is unknown constant gain. Estimated value of bis represented by ˆ
b.
μ−λcharacteristics of surface vehicle [14] and μis given by
μ(λ) =2μpλpλ
λ2
p+λ2(8)
where λpis optimal slip and μpis optimal adhesive coefficient. Now main goal of
the paper is to find the control input uso as to achieve tracking of desired slip ratio to
desired value for the Hybrid Electric Vehicle in presence of nonlinearity in f(λ, x)
due to adhesive coefficient and slip relation.
4 Design of Discrete-Time Sliding Mode Observer
Design of Fuzzy SMC is done in ref. [12]. While driving the vehicle, it is finding
to measure vehicle velocity online very difficult. Hence, we suggested the design of
a discrete-time SMO observer for estimating or observing vehicle angular velocity.
So, observer dynamic chosen in following form
ˆx1(k+1)=ˆx1(k)−Tc
vrwˆx2
1
m−TK
vsgn(˜x1)+c1μT(9)
where ˆx1is estimated vehicle velocity, ˜x1is measurement error and Kvis observer
gain. Now ˜x1is given as
Design of Lyapunov-Based Discrete-Time Adaptive Sliding … 101
˜x1(k)=x1(k)−ˆx1(k)(10)
Equation (10) rewritten as
˜x1(k+1)=x1(k+1)−ˆx1(k+1)(11)
Substituting x1(k+1)and ˆx1(k+1)from Eqs. (1) and (9)in(11) and solving
for ˜x1(k+1)we get
˜x1(k+1)=˜x1(k)−Tc
vrw˜x2
1
m+TK
vsgn(˜x1)(12)
where ˜x2
1=x2
1−ˆx2
1SMO observer dynamic represented in Eq. (9) is asymptotically
stable if observer gain chosen as
Kv≤
cvrw˜x2
1
m(13)
So as prove that, we have chosen the Lyapunov candidate function as
V=1
2˜x2
1(14)
and
ΔV=V(k+1)−V(k)=˜x1(k)˜x1(k+1)−˜x1(k)≤0 (15)
Putting Eqs. (9) and (11)in(15), we get
T˜x1(k)−cvrw˜x2
1
m+Kvsgn(˜x1)≤0 (16)
Under the assumption, wheel speed and vehicle speed are positive gives the
condition as
Kv≤
cvrw˜x2
1
m(17)
5 Design of Discrete-Time Adaptive Sliding Mode Control
Equation (6) can also be written as
102 K. Chaudhari and R. Ch. Khamari
λ(k+1)−λ(k)
T=fa(λ, x)+θh(λ, x)+bu (18)
where
fa(λ, x)=f2(x2)−(1+λ)f1(x1)
x1
(19)
h(λ, x)=1
x1
2[c2−(1+λ)c1]λpλ
λ2
p+λ2(20)
θ=μp(21)
Choosing of sliding surface is the same as given in the ref [12].
Choosing Lyapunov candidate function as
V=1
2s2+1
2(θ −θ
)2(22)
So,
ΔV=V(k+1)−V(k)=s(k)[s(k+1)−s(k)]−
θθ
(k+1)−θ
(k)≤0
(23)
where
θ=θ−θ
(24)
Substituting Eq. (18)inEq.(23) leads to
ΔV=s(k)Tfa+θh+bˆ
b−1−ˆ
fa−θ
ˆ
h+λd+(1−qT)s(k)+bεTsgn(s(k))
−
θθ
(k+1)−θ
(k)(25)
Now, rearranging different terms of Eq. (25), we get
ΔV=s(k)Tfa+θh−bˆ
b−1ˆ
fa−bˆ
b−1θ
ˆ
h+bˆ
b−1λd+bˆ
b−1(1−qT)s(k)
+bεTsgn(s(k))]−
θθ
(k+1)−θ
(k)(26)
Equation (26) can be rewritten as
ΔV=s(k)T(fa−ˆ
fa+(1−bˆ
b−1)ˆ
fa+(1−bˆ
b−1)θ
ˆ
h+bˆ
b−1λd
+θh−θ
ˆ
h+bˆ
b−1(1−qT)s(k)+bεTsgn(s(k)))
Design of Lyapunov-Based Discrete-Time Adaptive Sliding … 103
−
θθ
(k+1)−θ
(k)(27)
Following assumptions are considered for bounds in function faand h.
fa−ˆ
fa≤Faand h−ˆ
h≤H
Using above assumption in Eq. (27) leads to
ΔV=s(k)T(fa−ˆ
fa+(1−bˆ
b−1)ˆ
fa+(1−bˆ
b−1)θ
ˆ
h+bˆ
b−1λd+θˆ
h
+θH−θ
ˆ
h+bˆ
b−1(1−qT)s(k)+bεTsgn(s(k)))
−
θθ
(k+1)−θ
(k)≤0 (28)
Now, rearranging different terms, we get
ΔV=s(k)T(fa−ˆ
fa+(1−bˆ
b−1)ˆ
fa+(1−bˆ
b−1)θ
ˆ
h+bˆ
b−1λd
+θH+bˆ
b−1(1−qT)s(k)+bεTsgn(s(k)))
−
θθ
(k+1)−θ
(k)+s(k)T
θˆ
h≤0 (29)
For asymptotically stable,
ΔV≤0 and hence
s(k)Tfa−ˆ
fa+(1−bˆ
b−1)ˆ
fa+(1−bˆ
b−1)θ
ˆ
h+bˆ
b−1λd+θH
+bˆ
b−1(1−qT)s(k)+bεTsgn(s(k))=0(30)
and
−
θθ
(k+1)−θ
(k)+s(k)T
θˆ
h=0 (31)
From Eq. (30), we get
ε≥fa−ˆ
fa+1−bˆ
b−1ˆ
fa+1−bˆ
b−1θ
ˆ
h+bˆ
b−1λd+θH
+bˆ
b−1(1−qT)s(k)(bT sgn(s(k)))−1(32)
Equation (31) provide dynamic for θ
given by
θ
(k+1)=θ
(k)+s(k)Tˆ
h(33)
Equation (32) provide the value of εand Eq. (33) provide the estimated value of
optimal adhesive coefficient.
104 K. Chaudhari and R. Ch. Khamari
Tabl e 1 Rule base for computing control action
λeλe
NB NM NS ZPS PM PB
NB NB NB NM NM NS NS Z
NM NB NM NM NS NS ZPS
NS NM NM NS NS ZPS PS
ZNM NS NS ZPS PS PM
PS NS NS ZPS PS PM PM
PM NS ZPS PS PM PM PB
PB ZPS PS PM PM PB PB
6 Design of FLC for SRC
The design of an FLC for controlling of the desired value of slip ratio of HEV
considers the appropriate membership Functions (MFs). Appropriate MFs will be
selected on the basis that chosen MFs will occupy the whole universe of discourse
and also the selected membership functions overlap each other. This is to be done
so that any kind of discontinuity can be avoided. We have developed Fuzzy Logic
controller consisting of two inputs and one output for slip ratio tracking problem of
HEV. FLC inputs are λeand Δλeand output control action is u. We have chosen 7
numbers of triangular membership function for input and output variable of FLC.
FLC for slip ratio of HEV has the following rules formulated in Table 1.
Rule i: If λeis NB and Δλeis NB (negative big), then control action uis NB
(negative big).
where i=1, 2, 3,…, nand nis equal to 49 where nis the number of rules.
7 Fuzzy PID Controller Design
This section provides design of fuzzy PID controller. Two Inputs for Fuzzy PID are
λeand Δλeand Three outputs of Fuzzy PID are Kp,Ki,Kd. 7 triangular membership
functions will be chosen for λe,Δλe,Kp,Ki,Kd. Rules shown in Table 2.
Rule i: If λeis NB and Δλeis NB, then Kpis BS and Kiis BS and Kdis BS.
where i=1, 2, 3,…, nand nare the number of rules and it is equal to 49.
Design of Lyapunov-Based Discrete-Time Adaptive Sliding … 105
Tabl e 2 Rule base for computing Kp,Kd,Ki
λeλe
PB PM PS ZNS NM NB
NB MMS MS S S BS BS
NM M M MS MS S S BS
NS MB M M MS MS S S
Z B MB MB MMS MS S
PS B B MB MB M M MS
PM BB B B MB MB M M
PB BB BB B B MB MB M
8 Simulation Results and Discussion
All the developed controllers namely SMC, ASMC, FLC, and F-PID are simulated
in MATLAB considering discrete-time dynamic of HEV [Eqs. (14), (15)]. For simu-
lating purpose, the chosen value is the initial value of wheel speed 88 km/h, vehicle
speed 90 km/h. Values of parameter used for simulating the work designed are shown
in Table 3.
8.1 Analysis of Designed Fuzzy Sliding Mode Control
and Fuzzy Sliding Mode Control with Observer
Simulation result of fuzzy SMC is incorporated from ref paper [12]. Figures show
performing result for fuzzy SMC and fuzzy SMC with observer. Value of desired slip
ratio is selected as λdis −0.6, εis 0.05, qis 300. Necessity of SMO is that observe
vehicle speed online because vehicle speed is not measurable practically. Figure 1
shows that the value of slip ratio to controlled has been achieved and settling time
of response is nearly 0.18 s. Uncertainty estimation error is shown in Fig. 1for
FSMC and FMSO with observer is zero. Observed vehicle speed (FSMO) closely
Tabl e 3 Vehicle and
Actuator parameter Parameter Val u e Parameter Va l u e
Iw0.65 km m2λp−0.17
rw0.31 m cv0.595 N/m2/s2
m1400 kg T0.001 s
g9.8 m/s2Km0.2073 mkg/h
nw4Kb2.2 V/rad/s
fw3500 N R0.125 Ω
μp−0.8 L0.0028 H
106 K. Chaudhari and R. Ch. Khamari
Fig. 1 Slip ratio with FSMC
and FSMO
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
Time(sec)
Slip
FSMC
FSMO
Fig. 2 Estimation error with
FSMC and FSMO
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-1
0
1
2
3
4
5
Time(sec)
Uncertainty Est. Error
FSMC
FSMO
resemble vehicle speed as shown in Figs. 3and 4provide vehicle speed estimation
error approaches to zero. The breaking torque shown in Fig. 5and voltage required
for the actuator to carry out the process shown in Fig. 6.
8.2 Analysis of Designed Sliding Mode Control, Fuzzy Logic
Control, and Fuzzy PID Control
Simulation result of sliding mode control is incorporated from ref paper [12].
Performing results of SMC, FLC, and fuzzy PID control are shown in figures. Desired
slip ratio is selected as λdis −0.6, qis 300, εis 0.05. Controllers such as SMC, FLC,
and F-PID are used for tracking the desired value of slip ratio and compared perfor-
mance are shown in Table 4. Figure 7shows that the desired slip ratio is being
Design of Lyapunov-Based Discrete-Time Adaptive Sliding … 107
Fig. 3 Vehicle speed with
FSMC and FSMO
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
87
87.5
88
88.5
89
89.5
90
Time(sec)
Vehicle Speed(Km/hr)
FSMC
FSMO
Fig. 4 Vehicle speed
estimation error with FSMO
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Time(sec)
Velocity Estimation Error
Fig. 5 Braking torque with
FSMC and FSMO
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
200
400
600
800
1000
1200
1400
1600
1800
Time(sec)
Torque(Nm/rad)
FSMC
FSMO
108 K. Chaudhari and R. Ch. Khamari
Fig. 6 Required
voltage with FSMC and
FSMO
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
100
150
200
250
300
350
Time(sec)
Voltage(volts)
FSMC
FSMO
Tabl e 4 Tabular comparison Controller Chattering Rise time (s) Setting time (s)
SMC Zero 0.0699 0.12
FLC Zero 0.0066 0.0042
F-PID Zero 0.0243 0.048
Fig. 7 Slip ratio with FLC,
SMC and F-PID
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-0.8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
Time(sec)
Slip
FLC
SMC
F-PID
achieved by suggested controllers. Comparison proved that, as compared to SMC
and F-PID, response of FLC is impressive and as compared to SMC, the response of
F-PID is good. Braking torque required is shown in Fig. 8. By comparing response
of braking torque of all suggested controllers, initial value of torque in FLC is higher.
Vehicle speed and wheel speed performance shown in Figs. 9and 10, respectively,
and it is given in literature that for deceleration mode both speeds should decrease
so that slip ratio is maintained at desired value and it is happening in our figure
response also. Voltage excitation required for producing torque is shown in Fig. 11.
Design of Lyapunov-Based Discrete-Time Adaptive Sliding … 109
Fig. 8 Braking torque with
FLC, SMC and F-PID
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-1
0
1
2
3
4
5x 10 4
Time(sec)
Torque(Nm/rad)
FLC
SMC
F-PID
Fig. 9 Vehicle speed with
FLC, SMC and F-PID
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
87
87.5
88
88.5
89
89.5
90
Time(sec)
Vehicle Speed(Km/hr)
FLC
SMC
F-PID
Fig. 10 Wheel speed with
FLC, SMC and F-PID
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
20
30
40
50
60
70
80
90
Time(sec)
Wheel Speed(Km/hr)
FLC
SMC
F-PID
110 K. Chaudhari and R. Ch. Khamari
Fig. 11 Required
voltage with FLC, SMC and
F-PID
00.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
-1000
0
1000
2000
3000
4000
Time(sec)
Voltgae(volts)
FLC
SMC
F-PID
In Table 4, if we compare all responses of the controller, then F-PID gives impressive
performance in all respects as compared to FLC and SMC.
8.3 Analysis of Designed Adaptive Sliding Mode Control,
Fuzzy Logic Control and Fuzzy PID Control
Figures show results of adaptive SMC, FLC, and fuzzy PID control. For controlling
purpose, range 0.8 to −0.4 is selected for λd. Adaptive SMC, FLC, and fuzzy PID are
used for slip ratio (λd)tracking and performance of these controllers are observed.
Figure 12 shows that λdis being achieved by suggested controllers. So, as compared
to adaptive SMC and F-PID, response of FLC is impressive and as compared to
adaptive SMC, the response of F-PID is good. Comparison provides that initial value
of braking torque is higher when FLC is used for controlling slip ratio required is
Fig. 12 Slip ratio with FLC,
ASMC and F-PID
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-0.9
-0.8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
Time(sec)
Slip
FLC
ASMC
F-PID
Design of Lyapunov-Based Discrete-Time Adaptive Sliding … 111
Fig. 13 Braking torque with
FLC, ASMC and F-PID
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-3
-2
-1
0
1
2
3
4
5x 10 4
Time(sec)
Torque(Nm/rad)
FLC
ASMC
F-PID
Fig. 14 Vehicle speed with
FLC, ASMC and F-PID
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
20
30
40
50
60
70
80
90
Time(sec)
Vehicle Speed(Km/hr)
FLC
ASMC
F-PID
shown in Fig. 13. Figures 14 and 15 show vehicle speed and wheel speed profile and
it is given in literature that for deceleration mode both speeds should decrease so that
slip ratio is maintained at desired value and it is happening in our figure response also.
Figure 16 shows the voltage required for actuation purpose which indeed produced
torque and response is just following the braking torque response.
9 Conclusion
This paper presented the design of control algorithm schemes like discrete-time
Sliding Mode Control, FLC, and Fuzzy PID for achieving control over the desired
value of slip ratio. Also, observed control response output of slip ratio of an HEV
impressively speeds up using FLC and F-PID. The effectiveness of above-mentioned
112 K. Chaudhari and R. Ch. Khamari
Fig. 15 Wheel
speed with FLC, ASMC and
F-PID
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
10
20
30
40
50
60
70
80
90
Time(sec)
Wheel Speed(km/hr)
FLC
ASMC
F-PID
Fig. 16 Required
voltage with FLC, ASMC
and F-PID
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-3000
-2000
-1000
0
1000
2000
3000
4000
5000
Time(sec)
Voltage(volts)
FLC
ASMC
F-PID
controllers will be observed through MATLAB Simulation results. Problem of uncer-
tainty in the dynamic of slip ratio is addressed using fuzzy logic. Thereafter, designed
sliding mode observer using Lyapunov theory successfully estimated the vehicle
velocity online. The problems related to the tire, such as road dynamic, slip changes,
road changes, etc., are overcomed by designed discrete-time ASMC. Overall, chat-
tering problem is vanished by all designed controllers and hence actuation system is
avoided from damage.
References
1. P. Khatun, C.M. Bingam, N. Schofield, P.H. Mellor, Application of fuzzy control algorithms
for electric vehicle antilock braking/traction control systems. IEEE Trans. Veh. Technol. 52(5),
1356–1364 (2003)
Design of Lyapunov-Based Discrete-Time Adaptive Sliding … 113
2. C. Mi, H. Lin, Y. Zhang, Iterative learning control of antilock braking of electric and hybrid
vehicles. IEEE Trans. Veh. Technol. 54(2), 486–494 (2005)
3. C. Unsal, P. Kachroo, Sliding mode measurement feedback control for antilock braking systems.
IEEE Trans. Control Syst. Technol. 7(2), 271–281 (1999)
4. G.F. Mauer, A fuzzy logic controller for an ABS braking system. IEEE Trans. Fuzzy Syst.
Technol. 3(4), 381–388 (1995)
5. O. Hema Kesavulu, S. Sekhar Dash, N. Chellammal, M. Padma Lalitha, A novel control
approach for damping of resonance in a grid interfaced inverter system-fuzzy control approach.
Int. J. Ambient Energy, pp. 1–8
6. C.M. Lin, C.F. Hsu, Neural-network hybrid control for antilock braking systems. IEEE Trans.
Neural Netw. 14, 351–359 (2003)
7. C.M. Lin, C.F. Hsu, Self learning fuzzy sliding mode control for antilock braking systems.
IEEE Trans. Contr. Syst. Technol. 11, 273–278 (2003)
8. J.S. Yu, A robust adaptive wheel-slip controller for antilock brake system, in Proceedings of
36th IEEE Conference of Decision Control, vol 3 (1997), pp. 2545–2546
9. T.A. Johansen, J. Kalkkuhl, J. Ludemann, I. Petersen, Hybrid control strategies in ABS, in
Proceedings of 2001 American Control Conference, vol 2 (2001), pp. 1704–1705
10. M. Yoshimura, H. Fujimoto, Slip ratio control of electric vehicle with single-rate PWM consid-
ering driving force, in IEEE International Workshop on Advanced Motion Control (2012),
pp. 738–7432012
11. V. Utkin, H. Lee, Chattering problem in sliding mode control systems, in International
Workshop on Variable Structure Systems (2006)
12. K. Chaudhari, R. Khamari, Design and comparison of discrete-time sliding mode control and
discrete-time fuzzy sliding mode control for slip control of a hybrid electric vehicle. Int. J.
Manage. Technol. Eng. 9, 4199–4203 (2019)
13. B. Subudhi, S.S. Ge, Sliding mode observer based adaptive slip ratio control for electric and
hybrid vehicles. IEEE Trans. Intell. Transp. 13(4), 1617–1627 (2012)
14. F.L. Lewis, A. Yesildirek, K. Liu, Multilayer neural-net robot controller with guaranteed
tracking performance. IEEE Trans. Neural Netw. 7(2), 388–399 (1996)
An Improved Scheme for Organizing
E-Commerce-Based Websites Using
Semantic Web Mining
S. Vinoth Kumar, H. Shaheen, and T. Sreenivasulu
Abstract In the running of the Internet world, E- commerce industry has its own
benchmark in terms of its rapid growth and has made itself an established sector that
is indispensable for every industry to trade and do transactions online. As the world
is rushing in a rapid manner, India is slogging in the improvisation of the online
market, leading to the lack of customized needs of the customers. Bigger companies
are trying to put in a different strategic approach taking that into consideration an
approach of blended e-mining along with e-commerce has been devised. It would
be a design of the semantic- and neural-based page ranking algorithm [2]. This tool
upon launching would be a well-defined approach for e-commerce website ranking
[1]. It would also facilitate the customers to find the relevant websites on the top
of the page during their search for any particular product or business. It would be
further customized with all the relevant comparison of the other websites in terms of
the product quality and price.
Keywords Neural-based page ranking ·Website ranking ·E-mining
1 Introduction
One of the fastest growing businesses for the past decade is e-commerce [1]. The
customer’s needs have taken up the next level for satisfying the demands comparing
the competitors and to fetch revenue to the company. Researchers made an analysis
S. Vinoth Kumar (B)
Department of Computer Science and Engineering, SNS College of Technology, Coimbatore,
India
e-mail: profsvinoth@gmail.com
H. Shaheen ·T. Sreenivasulu
Department of Computer Science and Engineering, St. Peters Engineering College, Hyderabad,
India
e-mail: shaheen66@gmail.com
T. Sreenivasulu
e-mail: sreesmart@hotmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_10
115
116 S. Vinoth Kumar et al.
that in India is also rapidly growing among other countries. The reason behind this
sudden growth is due to the increase in the awareness of the Internet-based appli-
cations, computer literacy, change of lifestyle, and high rated income [6]. Apart
from that are the policies from the companies like trial and exchange feedback of
the product, cash upon delivery, and reviews about the product. One of the asso-
ciate research areas is intelligent neural web based mining would be taking up these
features of adaptability and would be able to transit the same to practical application
stage where the errors would be minimized to a larger extent and help the customer
to extract useful and abstract information from his surfing web pattern [3]. These
patterns would help us to do the necessary improvements on the website structure
enhancement of the website content making it user-friendly. This paper is going to
describe the web mining process deployment to get the maximum benefit of the e-
commerce area. It may not only be a user-friendly approach for the customer but also
for the data analysts to take a decision on the organizational needs. We have divided
the paper into various categories. The phase 2 category is task related. Phase 3 is
the identification of the query and analysis. Phase 4 illustrates the objectives of the
analysis. Phase 5 details about the analysis method and innovations taken for the page
ranking algorithm and website listing tool along with the graphical representation of
data [4].
2 Task Related
The semantic- and neural-based web mining technology leads to a better sophisticated
ranking of the e-commerce websites. The main objective for taking up such a blend
is that it might be able to assist the customer as well as the organizations to take
a more clear approach in terms of the transaction as well as on the data-oriented
decision-making. Since the old data mining area was not that successful it has ended
up in inventing new areas for progressing further [9]. The intention is to perform
a discussion of the quality of the e-commerce websites using the data envelopment
analysis model (DEA).
Eventually, they also compared the same with the other models like CCR, BCC,
and KH. A large amount of data was taken and was automated for obtaining results
on e-commerce websites. Various ideas regarding applications are implied in e-
commerce [7]. It also emphasized on the handling of customer behavior pattern and
feedback reciprocation. This contributed largely on the optimization of the websites.
While handling a large volume of data, most of the important information was hidden
during retrieving the data. This hidden information would be of great use for the struc-
turing of the web page and to enhance a better ranking [4]. They also added further
that the constant watch of the users, their methodology of watch search patterns, etc.
would also be helpful for the website optimization. The primary objective of this
research is to improvise the search engine algorithm and to minimize the complexi-
ties faced by the user. The semantic and neural network methodology is used for an
An Improved Scheme for Organizing E-Commerce-Based Websites … 117
unbiased ranking model of the e-commerce websites. This model has a wider perspec-
tive toward the e-commerce industry through an easier way of user navigation and
retrieval of specific information [8].
3 Identification of the Query and Analysis
The rapid growth of the e-commerce industry remains untold [5]. Every customer is
idealized to use only a search engine instead of the web catalog. Search engine may
not be able to fetch the exact requirement, as it is a syntactic-based query. It might
match and fetch the results based on the frequency count and proximity fetching the
data based on search query and web page. This syntactic match may lack semantics
producing wrong results which might fetch either number of unwanted pages or no
result at all. Apart from this, the result-producing pages are powered by the search
engines which make a very good revenue on the companies listing irrespective of
their content, reliability, and relevance to the customer. Few e-commerce companies
may not have got the authorization to sell the product. But even then they would
have published the same on their websites leading to a confused state of mind for
the customer to suffer from not knowing the product in detail along with the other
details of warranty, replacement, etc. The reason behind is that the search engines
failed to design their structure with reference to the customer’s queries and intention
of the customer. The other reason is the backpropagation of the errors and retrieval
algorithms lead to the biased ranking leading to only the top-based rankings popular.
4 Objectives of the Analysis
The whole aim of the analysis is to enhance the e-commerce website to be ranked in a
more better and efficient way using the ranking algorithm through the SNEC process
to assist the customer while carrying out online transactions in a more authentic
and rational manner. The research implies a semantic- and neural-based approach
to deal with the ranking problems. This would optimize the use of web dictionary
backpropagation and unbiased ranking process.
4.1 Analysis Method and Innovations
The research comprises backprocessing the retrieved company information using a
profiling and dictionary implementation module to improvise the incomplete entries
and data cleaning. The dictionary and the candidate web page are then analyzed by
another module called as the content priority module. The primary objective of this
module is to check for relevance and to remove unwanted data. Then the query is
118 S. Vinoth Kumar et al.
Fig. 1 Analysis method
passed on to the priority module which would check for the priority of the web page
based on the customers search and also on the previous searches for the same product
by other people. Taking these data the web page is now sent to the next phase of
the module called as the semantic module which identifies the user session from
other external sources using its algorithm and determines the search to avoid wrong
interpretation. One of the most popular NN algorithms is backpropagation algorithm.
The BB routine can be fragmented into four key phases followed by which loads
of the networks are chosen arbitrarily and the back broadcast routine is employed
for estimating the needed modifications. The routine could be disintegrated into the
stated four phases as feed onward estimation, back broadcast for the result layers,
back broadcast for the concealed layer, and load-based revision. The scheme is halted
on experiencing a fault function which is suitably small. It is very irregular and the
basic formula for the back broadcast holds some dissimilarity designed by others but
it must be precise and much easier for usage. The concluding phase is load revision
which occurs for the overall scheme (Fig. 1).
4.2 Nomenclature
Semantic- and Neural-based Algorithm
XiUser search product.
Min Minimum length of the keyword Xi.
Max Maximum length of the keyword Xi.
YKeyword search specific.
ST The web-based e-commerce document to be scanned.
PD Dictionary with reference to the Tth document.
TXT Words of the document.
A1 Time spent of the browsing by the other visitors.
A2 Time spent on the web page creation.
An Improved Scheme for Organizing E-Commerce-Based Websites … 119
FFrequency of the number of keywords found.
NF Not found keywords.
tanθLinear activation function for the training of the neural network.
MT Mass of the input.
4.3 Module
Module 1:
Step 1: Input from the user.
Step 2: Filtering of unwanted terms from the user.
Step 3: Track the movement of the pattern sequence of the user data.
Step 4: Track the web pages through the search engine.
Step 5: Divide the strings into various words like: Y1,Y2,…,Yn.
Step 6: Determination of minimum and the maximum length of the words
Min :=StrLen (Y1), Max : =StrLen (Y1)
For k=1tondo
Initialize F:=0 and NF:=0
If ST found in PD then
F:=F+1
Else F:=NF +1.
Step 11: Remove those web pages where NF > F.
Module 2
Step 12: To evaluate the timestamp A2 for the creation of web page.
Step 13: On the beginning of the user session, determine a1 which is session
duration of current page and determine new value of A1 as follows:
If A1=0 then A1=a1
Else A1 =(A1+a1)/2.
Step 14: Assign a higher priority to web page if A2 is low and A1 is high.
Step 15: Update the time database of tool with keywords, page address, and A1.
Module 3
Step 16: Identify navigation session by comparing user search query with each
of the search query present in user profile database as
LCS [c,d]=0ifc=0, or d=0OR
LCS [c,d]=LCS [c−1, d−1] +1, if c,d<>0 and X1c=X2dOR
LCS [c,d]=max(LCS[c−1, d], LCS[c,d−1], if c,d> 0 and X1c<>X2d.
Step 17: Class generation using Web Ontology Language.
120 S. Vinoth Kumar et al.
Module 4
Step 18: Normalize all the priority inputs from module 2, 3, and 4.
Step 19: Train the network using various sets of inputs and outputs with linear
activation function as
{O} =tan θ{I}.
Step 20: Use sigmoidal function for output evaluation in hidden and output layers
as {O}=[(1/1 +e−1)] and summation function as (C1MT1 +C2MT2 +
C3MT3 +C4MT4 +C5MT5 +B).
Step 21: The error rate is determined by adjusting the weight age of the synapses.
Step 22: At last the web pages are displayed in a decreasing manner in order of
the ranking priority.
4.4 Priority on Time Spent
SBPP algorithm mentioned in this exploration research implies the usage of the
importance to website priority under ASP.NET framework. Through this research,
we would be able to explore more than five e-commerce websites using the design
done. The tool will allow the number of entries based on the design. After entering
the data the tool would search in accordance with the content and the statistical data
(like the number of times the page has been visited, the product specification, etc.)
(Fig. 2).
Fig. 2 Priority on time spent
An Improved Scheme for Organizing E-Commerce-Based Websites … 121
4.5 A Statistical Approach Toward the Research
This segment shows the comparative outcomes of parameters precision and effi-
ciency, respectively. Outcomes were obtained while acting of weighted page rank
algorithm on the dataset of Internet pages. Diverse iterations have been completed
to test the consistency of the version. Precision: it approaches how properly the
website precedence tool (WPT) is working. Internet site precedence device allows
evaluation of websites, the use of drop-down container, and seek box to specify a
string of unique product. The drop-down field provides as many URL’s (uniform
resource locator) of the website and after contrast, WPT tool assigns precedence to
every candidate website based on the calculation of content priority module, time
spent priority module, advice module, and neural priority module. Subsequently,
precision is used to measure the consistency of the consequences for each and every
time the system runs. Greater the relevancy of the fetched web pages better might
be the consistency of the machine. The better consistency of the results implies that
the website priority tool is operating correctly. As an end result, better accuracy of
net site priority tool leads to higher precision. Relevancy is calculated with the aid of
measuring the space of the statistics. Statistics has been stored in array/matrix shape.
Distance will be calculated for every row by comparing it with all different rows. For
each row, lesser the gap among rows more relevant will be records and vice versa.
Precision values of the proposed machine had been received by way of making use
of more than one testing rounds (iterations) about 25 on the statistics set (Fig. 3).
Precision primarily based evaluation of the designed version. The foreseen graph
represents the correct values for Internet site precedence tool, Google, and the
proposed WPR. The line graph here truly shows that the proposed weighted page
rank set of rules has high precision values for all the iterations. The graphical layout
of both the model’s page rank and stepped forward weighted page rank showcases the
Fig. 3 Precision-based evaluation
122 S. Vinoth Kumar et al.
comparative evaluation which has been evaluated on the premise of the precision of
the simulated outcomes. The upgrades are stated which validates the better conduct
of the designed WPR scheme than the prevailing schemes.
5 Conclusion and Future Work
We see a developing interest in the use of the semantic and neural community for
solving net programs inside the approaching years. The inherent capability of neuro-
semantic techniques in managing indistinct, big-scale, and unstructured informa-
tion is a great fit for Internet-related problems. Earlier studies have a tendency to
consciousness on a way to extract wanted facts from unstructured net facts. Recently,
we have seen the use of neural methodologies in building a based web [10]. The
semantic Internet is one instance of such. The perception of a structured web can be
made extra practical while the concept is employed due to the fact that net records
have a tendency to be unpredictable in nature. We count on to peer an integration
of gentle computing techniques in semantic Internet methodologies within the near
destiny. A genetic set of rules for net software ought to also turn out to be extra
famous as Internet applications get large in scale.
References
1. N. Verma, Improved web mining for e-commerce website restructuring, in IEEE Interna-
tional Conference Computational Intelligence & Communication Technology, UP, India (2015),
pp. 155–160
2. N. Verma, D. Malhotra, M. Malhotra, J. Singh, E-commerce website ranking using semantic
web mining and neural computing, in Proceedings of International Conference on Advanced
Computing Technologies and Applications, Mumbai, India (2015). Procedia Comput. Sci.,
Science Direct. Elsevier, 45, 42–51. https://doi.org/10.1016/j.procs.2015.03.080
3. D. Malhotra, N. Verma, An ingenious pattern matching approach to ameliorate web page rank.
Int. J. Comput. Appl. 65(24), 33–39 (2013). https://doi.org/10.5120/11235-6543
4. D. Malhotra, Intelligent web mining to ameliorate web page rank using back propagation neural
network, in Proceedings of 5th International Conference, Confluence: The Next Generation
Information Technology Summit, Noida, India (IEEE, 2014), pp 77–81. https://doi.org/10.1109/
confluence.2014.6949254
5. S. Sharma, P. Sharma, Use of data mining in various fields. IOSR J. Comput. Eng. IOSR-JCE
16(3) (2014)
6. N. Padhy, P. Mishra, R. Panigrahi, The survey of data mining applications and feature scope.
Int. J. Comput. Sci. Eng. Inf. Technol. IJCSET 2(3) (2012)
7. H. Kaur, A review of applications of data mining in the field of education. Int. J. Adv. Res.
Comput. Commun. Eng. 4(4), 2319–5940 (2015)
8. S. Bagga, G.N. Singh, Applications of data mining. Int. J. Sci. Emerg. Technol. Latest Trends
(2012)
An Improved Scheme for Organizing E-Commerce-Based Websites … 123
9. N. Jain, V. Srivastava, Data mining techniques. Int. J. Res. Eng. Technol. eISSN: 2319-1163 |
pISSN: 2321-7308
10. D. Kumar, D. Bhardwaj, Rise of data mining: current and future application areas. Int. J.
Comput. Sci. Issues 8(5) (2011). ISSN (Online): 1694-0814
Performance Estimation and Analysis
Over the Supervised Learning
Approaches for Motor Imagery EEG
Signals Classification
Gopal Chandra Jana , Shivam Shukla, Divyansh Srivastava,
and Anupam Agrawal
Abstract In this paper, a comparative analysis has been done to estimate a robust
classifier to classify motor imagery EEG data. First, segment detection and feature
extraction have been done over the raw EEG data. Then the frequency domain features
have been extracted using FFT. Six classifiers DNN, SVM, KNN, Naïve Bayes,’
Random Forest, and Decision Tree have been considered for this study. The DNN
model configured with four layers and used the binary cross-entropy loss function
and sigmoid activation function for all layers. The optimizer used is “Adam” having
the default-learning rate of 0.001. In this experiment, for the purpose of the estimation
of the performance of various classifiers, the experiment used dataset IVa from BCI
Competition III, which consisted of EEG signal data for five subjects, namely ‘aa,’
‘al,’ ‘av,’ ‘aw,’ and ‘ay.’ The highest average accuracy of 70.32% achieved by the
DNN model, whereas the model achieved an accuracy of 80.39% over the subject
‘aw.’ The objective of this experiment encompasses the different models for the
classification of various motor tasks from EEG signals.
Keywords Electroencephalogram (EEG) ·Brain–computer interface (BCI) ·
Motor imagery ·Deep neural network (DNN) ·SVM ·KNN ·Naive Bayes ·
Random forest ·Decision tree
G. C. Jana (B)·A. Agrawal
Interactive Technologies and Multimedia Research Lab, Department of Information Technology,
Indian Institute of Information Technology-Allahabad (IIITA), Allahabad, India
e-mail: go.gopal.ch.jana@gmail.com
A. Agrawal
e-mail: anupam@iiita.ac.in
S. Shukla ·D. Srivastava
Research Intern, Interactive Technologies & Multimedia Research Lab, Department of
Information Technology, Indian Institute of Information Technology-Allahabad (IIITA),
Allahabad, India
e-mail: create.shivamshukla@gmail.com
D. Srivastava
e-mail: srivastavadivyansh98@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_12
125
126 G. C. Jana et al.
1 Introduction
Motor Imagery based on Brain–Computer Interface (BCI) technologies is a
constantly evolving field due to the immense researches being conducted and novel
methods being constantly brought out and improved. Motor Imagery (MI) is a
type of mental process which involves motor actions. In this particular process,
the person simulates a given action mentally rather than performing them. It may
involve different limb movements or body part movements, but all these tasks are
accomplished only mentally, using imagination to do that particular task. Exam-
ples include the right hand or foot movement. The development in this field may
bring about significant changes in the lives of persons with neurological disorders.
The BCI employs pathways between the brain and any other external actuators or
devices for performing the action without the actual limb movement. These systems
can be extremely beneficial for neural rehabilitation for the persons with neurolog-
ical and neuromuscular disorders. MI-based BCI is now evolving as a domain due
to the improvement in the performance of different classification techniques and
the introduction of non-invasive BCI techniques like EEG, MEG, etc., which are
easy and risk-free in terms of internal organ damage and cost. The electroencephalo-
graph (EEG)-based signals depend upon the neuron’s activity inside the brain of the
subjects. The main challenges in the field are due to signal temporal and spatial reso-
lutions and considering an efficient algorithm to work on EEG signal features with
its extraction strategies that are available. Other than this, the data collection using
invasive methods are less cost-effective and accurate in comparison to non-invasive
ones, while risky than non-invasive. This work has been done in two phases, one
is data preprocessing and another one is classification. Data preprocessing phase
have been done in parts; one is segment detection and another one is feature extrac-
tion. To estimate a robust classifier for this proposed approach, we consider different
classifiers for getting the best performance for the EEG signal-based motor imagery
classification for the different subjects. The extensive analysis for different classi-
fiers—Deep Neural Networks (DNN), Support Vector Machines (SVM), K-Nearest
Neighbor (KNN), Naïve Bayes algorithm, and Random Forests have been applied
to the EEG signal-based motor imagery dataset, which consists the signal data of
five different subjects. These models estimate the prediction accuracy of different
motor imagery tasks. This paper focusses on the performance analysis of the different
classifiers with a supervised approach for the classification of EEG signals for either
right hand or foot which are taken from BCI Competition III dataset IVa. The further
improvement in the techniques to classify the EEG-based signals will bring more
accuracy in the analysis of various disorders and the development of many different
devices and applications for the same.
Organization of paper is done in various sections: Next Section (Sect. 2) discusses
the related works and literature review, Furthermore, Sect. 3describes the dataset
used for classification models. Sect. 4discusses methodologies and classifiers model.
Sec. 5discusses the results and analysis of classification results. In Sect. 6,the
conclusions are provided and different references for the paper are shown.
Performance Estimation and Analysis Over … 127
2 Related Works
This approach is about MI classification among two classes of EEG signals generated
due to the right hand and right foot activity of the subjects. There has been contin-
uous work going in this field with the intent of making the classifiers more and more
accurate with different combinations of the feature extraction strategies and different
classifiers too. In [1], the motor imagery classification has been done using tradi-
tional classifier models like SVM and MLP are employed on handcrafted features
extracted from EEG signal data and both of classifier model achieved a considerable
accuracy with of 85% and 85.71%, respectively. With the purpose of designing, a
reliable and robust model of BCI, an enhanced method using A DNN-based approach
had been proposed in [2] for motor imagery classification. Similarly, an improved
approach has been proposed in [3], where the authors use SVM for classification
and Particle Swarm Optimizations (PSO) for the selection of kernels and penalty
parameters for motor imagery classification based on EEG signals. Different other
classifiers like KNN have also been used in the work of [4] for motor imagery EEG
signal classification, and the proposed model achieved a maximum classification
accuracy of 91% for left and right hand movements and extraction of features have
been done using the wavelet-based techniques. A comparative strategy for finding
the effectiveness of the different feature extraction techniques has been proposed
in [5], where the raw data samples are compared with DWT-based feature extrac-
tion. It demonstrated that the improvement with their strategy came out significantly
in terms of different performance measures. With the same consideration, a hybrid
method based on Discrete Wavelet Transform (DWT) and Artificial Neural Network
(ANN) [6,7] has been proposed in [8,9] for the classification of physical action,
while some other motor imagery classification technique has been proposed in [10]
using Gaussian Naive Bayes algorithm on the feature of EEG signal data extracted
using the Common Spatial Pattern (CSP) technique. Other than this some more
powerful models like CNN (Convolutional Neural Networks) are also being consid-
ered to make the generalization procedure even better. In [11], the paper uses the
five-layer CNN for the classification of different motor imagery tasks, while there is
also a comparative analysis that is performed with SVMs and other composite feature
extraction techniques. The CNNs have made the accuracy level even higher for motor
imagery data. In [12], the CNNs were used along with the time–frequency methods
using STFT (Short-Term Fourier Transform) and CWT (Continuous Wavelet Trans-
form) were employed as preprocessing stage. Also, the most popular Neural Network
(NN) techniques were used by the authors of [13] over the BCI Competition Data.
NN techniques provide an efficient and less cost-effective classification approach,
and they are continuously being applied on the different classification problems by
tuned the hyperparameters. This literature review show that still there is a scope of
analyzing the performance of an approach using DNN and FFT with respect to other
supervised classification approaches for motor imagery EEG signal classification for
a significant breakthrough in the area of BCI.
128 G. C. Jana et al.
3 Experimental Data Description
The dataset used for this experiment has been taken from BCI III Competition.
Dataset IVa [14] has been used for the estimation of the performance of different
models in the proposed methodology. As mentioned in the BCI III Competition
weblink, the data were acquired using a non-invasive technique from five healthy
participants or subjects whose data were labeled as: ‘aa,’ ‘al,’ ‘av,’ ‘aw,’ and ‘ay.’ An
EEG cap of 128 Ag/AgCl electrodes is used, out of which only 118 electrodes were
considered for data acquisition purposes. In the duration of 3.5 s, which is the duration
of the visual cue for each subject, the subjects have to perform one of imagery, namely
Left Hand (L), Right Hand (R), and Right Foot (F). In this experiment, we consider
two classes: one is right hand and another one is right foot. The data were available
in two different formats, it will be either in zipped ASC II format or mat format.
In this experiment, 100-Hz ASCII format of data has been used which contains two
different txt files, one for EEG values and another one is for the corresponding class
label.
4 Methodologies
Experiment of this proposed approach has been done in several stages. An illustration
of the proposed approach is shown in Fig. 1.
Experimental Environment Setup: Different environmental parameters are
considered and setup for this experiment. The experimental procedure assumes the
usage of the following libraries and tools
1. Anaconda (Python 3.6 or above)
2. Jupyter notebook (for implementation)
3. Virtual Environment (for Python) is enabled and workable with following
libraries: sklearn, sklearn.metrics (for accuracy_score, confusion_matrix),
sklearn.svm (for svm implementation), sklearn.datasets, sklearn.naive_bayes(for
Naïve Bayes implementation), sklearn.neighbors (for K-nearest Neighbour
implementation), sklearn.ensemble (for Random Forest implementation),
Fig. 1 Illustration of the
proposed approach
Performance Estimation and Analysis Over … 129
sklearn.tree (for Decision Tree implementation), Numpy (basic numeric oper-
ations on matrix, vectors, load txt file), scipy.fftpack, keras.utils (up_utils),
matplotlib.pyplot (for visualization purposes), keras.layers (dense), keras.models
(sequential).
The methodology used in this paper encompasses some important key components
which are as follows:
A. Segment Detection
The input raw EEG signals data are provided by the BCI Competition III—Dataset
IVa. This dataset has few segments and these are belonging to class label NaN
which are not important for our classification process. So, in this phase, the segment
detection has been performed to extract the correct segment of signal from the original
signal. The extracted segment is now passed for the feature extraction stage.
B. Feature Extraction
In the feature extraction phase, transform EEG values from the segmented signal are
taken as representative for that original signal. Fast Fourier Transform has been used
for this phase which is mentioned below
Fast Fourier Transform: The Fast Fourier Transform has been used in this
phase for the transformation time-domain EEG signal into the frequency domain.
This conversion is crucial for this experiment to extract signal features in terms of
frequency from the time-domain segmented samples. The FFT transform [15]theN
points time-domain signal samples into the separated Ndifferent frequency-domain
signals points. Finally, from these Npoint spectra, the Nfrequency spectrum is
extracted. This Npoint frequency spectrum is meant to be extracted into a single
spectrum for the synthesis of the overall signals into the spectrum. FFT is one of
the prominent ways to apply Discrete Fourier Transform (DFT). The basic equation
governing the DFT (which is the main process behind DFT for efficient working) is:
F(n)=
N−1
0
x(k)e−j2πkn
N.
The value of F(n)is the amplitude of the signal in the frequency domain at value
n, while Nis the total discrete samples of signals taken.
In this phase, we have applied FFT on all segmented signal samples for extracting
frequency-domain features. Extracted frequency-domain features pass into the
classifier for the classification of right hand or foot.
C. Classification
Deep Neural Networks (DNN): DNN [16,17] is one of the classifiers that have
been used for this experiment. The DNN configured with four layers having input
layer, hidden layer-1, hidden layer-2, and output layer. The hidden layer-1 has seven
130 G. C. Jana et al.
Tabl e 1 Layer configuration
of DNN Layers Input shape Output shape
Dense (118, 1) (None, 12)
Dropout (None, 12) (None, 12)
Dense (None, 12) (None, 8)
Dropout (None, 8) (None, 8)
Dense (None, 8) (None, 6)
Dense (None, 6) (None, 3)
Tabl e 2 Subject-wise
training epochs and batch size Subject Epochs Batch size
aa 150 32
al 100 32
av 115 16
aw 150 32
ay 25 4
neurons, while the hidden layer-2 has six neurons. The final output layer is provided
with the two output neurons. The classification into two classes is accomplished by
this network.
The DNN model uses the loss function as binary cross-entropy, activation function
as sigmoid for all layers. The optimizer used is “Adam” having the default learning
rate of 0.001. The input layer takes an input of 118 frequency-domain features.
Layer-wise details of DNN have been mentioned in Table 1.
The following Table 2describes the number of epochs and batch size of each
subject data have been used in the training session for the DNN.
Support Vector Machines (SVM):SVM[1] are one of the popular classification
models which have a pretty simple algorithm. In this experiment, SVM have been
used for classification. Different kernel functions have been used to estimate the best
kernel function for which SVM achieves high classification accuracy over the motor
imagery EEG data. Two kernels have been considered which are mentioned below.
1. RBF Kernel (Radial Basis Function)
2. Sigmoid Kernel
The input data for the SVM is kept the same as that of other classifiers and no
change has been done. The results of this particular technique have been shown in
the result and analysis section.
Naive Bayes Classifier: Naive Bayes classifiers [10] are another category of clas-
sifiers that depends upon the probabilistic analysis for their classification algorithm.
The foundation of this algorithm is based on Bayes Theorem, these are very powerful
and faster models that perform better on some class of data. It has been used in this
Performance Estimation and Analysis Over … 131
experiment to compare the classification capability w.r.t other classifiers over the
same input data. The results of this particular technique have been shown in the
result and analysis section.
K-Nearest Neighbor (KNN): K nearest neighbor [4] is also one of the renowned
supervised algorithms in which the vicinity of a particular data point w.r.t to other
available points. It does not make up any structure of the overall data during training.
Rather all the training examples are kept at the same time for the test example for
the nearest neighbor.
In this experiment, KNN with three nearest neighbors has been taken to classify
the test data correctly. The results of this particular technique have been shown in
the result and analysis section.
Random Forest Method: RandomForest models [15] work by preparing multiple
decision trees at the time of their training. The output of particular test data is deter-
mined by either mode or mean of the different decision trees that are prepared. This
is a type of ensemble-learning algorithms. For this experiment, the number of esti-
mators for this method is ten for the overall implementation of this paper. The results
of this particular technique have been shown in the result and analysis section.
Decision Tree: Decision Tree (DT) [18] which is a powerful supervised learning
algorithm. DT classification techniques have the capability of representing the infor-
mation in terms of trees. These are able to provide decisions with the help of this
underlying structure. It works like a flowchart, a flow goes from one point to another
in flowchart w.r.t different conditions and parameters.
In this experiment, we have also considered decision trees as classifiers for
comparative analysis with different models that have been used over the same input
data. The maximum depth of decision trees for this experiment has been consid-
ered as 3. The results of this particular technique have been shown in the result and
analysis section.
5 Results and Analysis
With the intention of finding out the best classifiers from the six different types of
supervised classification approaches, we hereby estimated the performance of six
different kinds of classifiers to classify the motor imagery dataset. The performance
of the classifiers has been recorded subject-wise. Classification performance has been
shown in section in terms of confusion matrix and accuracy. In this section, we try
to show the results of all the considered classifiers separately, and this section ends
with the results of the DNN classifier which has been achieved the highest accuracy
for all subjects.
Support Vector Machines (SVM): SVM with Radial Basis kernel function (RBF)
and sigmoid kernels have been tested over the input data. These model performances
have been recorded subject-wise. Performance variation has been found w.r.t the
kernel and datasets, whereas both models achieved the highest accuracy of 57.14%
for the subject ‘ay’ while in case of other subjects’ performance is less than 57.14%.
132 G. C. Jana et al.
Fig. 2 Confusion matrix
over the training data of
subject ‘aa’ for SVM (RBF)
Fig. 3 Confusion matrix
over the testing data of
subject ‘aa’ for SVM (RBF)
Figures 2and 3depict the confusion matrix of the SVM classifier with RBF kernel
over the training data and testing data of subject ‘aa,’ whereas Fig. 4shows subject-
wise accuracy (%) achieved by the SVM with RBF and sigmoid kernels. It is pretty
evident that SVM with both kernels achieved similar accuracy for the individual
subjects.
Naive Bayes Classifier:
The NB classifier had shown the best accuracy over the subject ‘av’ of 52.17%,
which is pretty lower than the other classifiers’ performance. Figures 5and 6show
the depiction of the confusion matrix plotted the training data and the testing data of
subject ‘aa’ for NB, whereas Fig. 7shows the subject-wise performance (accuracy)
achieved by the classifier.
Evidently from Fig. 7, the NB classifier achieved an accuracy of 52.17% over
subject ‘av.’ Other than this, the NB classifier achieved even less accuracy than 50%
over the other subject’s data.
Performance Estimation and Analysis Over … 133
Fig. 4 Subject-wise
accuracy (%) achieved by the
SVM with RBF and sigmoid
kernels
Fig. 5 Confusion matrix
over the training data of
subject ‘aa’ for Naïve Bayes
classifier
Fig. 6 Confusion matrix
over the testing data of
subject ‘aa’ for Naïve Bayes
classifier
134 G. C. Jana et al.
Fig. 7 Subject-wise
accuracy (%) achieved by the
Naïve Bayes classifier
Fig. 8 Depicts the
confusion matrix over the
training data of subject ‘aa’
for KNN classifier
K-Nearest Neighbour (KNN):
KNN algorithm has been successfully applied over the motor imagery data and
performance has been estimated for all subjects. The overall performance of the KNN
is pretty good in this case. The best accuracy of 64.70% has been achieved over the
Subject ‘aw’ which has been shown in Fig. 10, where KNN achieved an accuracy of
42% over the Subject ‘aa.’ Figures 8and 9are the depiction of the confusion matrix
of the subject ‘aa’ for the KNN-based classification.
The performance of the KNN over all subjects has been shown in Fig. 10, which
indicated that the KNN cannot be granted as the robust model for the motor imagery
based on EEG data with achieved accuracies of 42%, 47.16%, 52.17%, 64.70%, and
57.14% over the subjects, namely ‘aa,’ ‘al,’ ‘av,’ ‘aw,’ and ‘ay,’ respectively.
Performance Estimation and Analysis Over … 135
Fig. 9 Depicts the
confusion matrix over the
testing data of subject ‘aa’
for KNN classifier
Fig. 10 Depicts
subject-wise accuracy (%)
achieved by the KNN
classifier
Random Forest Method:
In this experiment, Random Forest classification approach has been applied over
all subjects. Following Figs. 11 and 12 are the confusion matrixes over training data
and testing data of subject ‘aa’ of the subject ‘aa’ for classification using random
forest approach.
The accuracy of 50%, 24.52%, 56.52%, 23.52%, and 57.14% achieved by the
Random Forest approach over the subjects, namely ‘aa,’ ‘al,’ ‘av,’ ‘aw,’ and ‘ay,’
respectively. The random forest approach achieved at most 57.14% accuracy over
the subject ‘ay.’ While the accuracy was pretty lower with respect to other subjects.
Following Fig. 13 is the depiction of the same.
Figure 13 shows the inconsistent performance with respect to the different
subjects. Moreover, the performance was significantly less for the subject ‘al,’ and
‘aw,’ in comparison to the subject ‘aa’ and ‘ay.’ The overall performance of Random
136 G. C. Jana et al.
Fig. 11 Depicts the
confusion matrix over the
training data of subject ‘aa’
for random forest classifier
Fig. 12 Depicts the
confusion matrix over the
testing data of subject ‘aa’
for random forest classifier
Forest classifier is significantly less than that the KNN which indicated that the
Random Forest classifier cannot be granted as the robust model for the motor imagery
EEG data classification.
Decision Tree:
In this experiment, DT approach has been used over all subjects. The performance
has been estimated for all subjects. Following Figs. 14 and 15 are the confusion matrix
over the training data and testing data of subject ‘aa.’
For DT, the performance of classifier attained a maximum accuracy of 71.42% for
the subject ‘ay,’ while for others, it achieved an accuracy of 44.0%, 11.32%, 69.56%,
and 11.76% for the subjects aa, al, av, aw, respectively, which has been shown in
Figure 16. The classifier performed very poorly with an accuracy of 11.32%.
Performance Estimation and Analysis Over … 137
Fig. 13 Shows subject-wise
accuracy (%) achieved by the
random forest classifier
Fig. 14 Depicting the
confusion matrix over the
training data of subject ‘aa’
for the decision tree classifier
It is evident that DTs although performed well over the subjects ‘aa,’ ‘av,’ and
‘ay,’ while for other two subjects, DT classifier achieved very less accuracy. This is
pretty evident from the overall data that decision trees cannot perform very well with
the present combination of feature extraction techniques.
Deep Neural Networks:
For this experiment, DNN and their different architectures have been applied over
the pre-processed motor imagery data. In this approach, the Fast Fourier Transform
technique over the segmented data and then the DNN has been applied to classify.
The performance of DNN has been estimated over all subjects which have been
shown in Figure 19. It shows subject-wise performance in terms of accuracy of
63.99%, 67.30%, 63.77%, 80.39%, and 76.19% achieved by the DNN classifier over
the subjects, namely ‘aa,’ ‘al,’ ‘av,’ ‘aw,’ and ‘ay,’ respectively. The DNN model
138 G. C. Jana et al.
Fig. 15 Depicting the
confusion matrix over the
testing data of subject ‘aa’
for decision tree
Fig. 16 Shows subject-wise Accuracy (%) achieved by decision tree classifier
Fig. 17 Trend of training and testing accuracy w.r.t epochs over the subject ‘aa’
Performance Estimation and Analysis Over … 139
Fig. 18 Loss trend of the DNN model over the training and testing of subject ‘aa’
Fig. 19 Shows subject-wise
accuracy (%) achieved by the
DNN Classifier
has achieved better accuracy than the SVM, Naïve Bayes, KNN, Random Forest,
and Decision Tree classifiers over all subjects. The variation of DNN accuracy with
respect to the number of epochs over the subjects ‘aa’ has been shown in Fig. 17.
Also, the loss trend of the DNN model has been estimated with respect to the number
of epochs over all subjects. Figure 18 shows the loss trend of the DNN model over
the subjects ‘aa.’
Finally, an average accuracy has been calculated for the all considered classifier
over all subjects, and it has been found that average accuracy of 70.32%, 52.63%,
47.90%, 47.03%, 43.58%, 42.34%, and 41.61% achieved by DNN, KNN, SVM with
RBF, SVM with sigmoid, Naive Bayes, Random Forest, Decision Tree classifier,
respectively. It indicates the robustness of the DNN to classify the motor imagery
EEG over all subjects. This comparative analysis shows the considered classifiers
performed well at some of the subjects with the motor imagery data. Thus, the
considered DNN architecture with feature extraction techniques is one of the robust
approaches for motor imagery EEG data classification. The further analysis can be
140 G. C. Jana et al.
done with a different set of hyperparameters to enhance the overall classification
accuracy.
6 Conclusion with Future Scope
In this paper, a comparative analysis has been done upon the performance of some
selective classifiers using a supervised learning technique over the motor imagery
EEG data. The classification approaches have been applied to the frequency-domain
features that have been extracted using FFT. The performance of the seven classifiers
has been estimated over the five subjects of motor imagery EEG data. The estimated
average accuracy indicates the considered DNN architecture achieved the highest
classification performance over the extracted features. Thus, the estimated result
shows DNN and further architecture is preferably one of the suitable approaches
for classification of the motor imagery based EEG data. Several feature extraction
and selection strategies can be considered which will enhance the robustness of the
classification models which will serve as the future work.
Source code and experimental results of this paper can be found from the
author’s website (https://sites.google.com/site/gcjanahomepage/publications/Public
ations-Source-Codes).
Acknowledgments This work was supported by the Indian Institute of Information Technology
Allahabad, UP, India. The authors are grateful for this support.
References
1. R. Chatterjee, T. Bandyopadhyay, EEG based motor imagery classification using SVM and
MLP, in 2016 2nd International Conference on Computational Intelligence and Networks
(CINE), Bhubaneswar (2006), pp. 84–89
2. G.C. Jana, A. Swetapadma, P.K. Pattnaik, Enhancing the performance of motor imagery classi-
fication to design a robust brain computer interface using feed forward back-propagation neural
network. Ain Shams Eng. J. 9(4), 2871–2878 (2018)
3. Y. Ma, X. Ding, Q. She, Z. Luo, T. Potter, Y. Zhang, Classification of motor imagery EEG signals
with support vector machines and particle swarm optimization. Comput. Math. Methods Med.
2016, Article ID 4941235 (2016)
4. R. Aldea, M. Fira, A. Laz˘ar, Classifications of motor imagery tasks using k-nearest neighbors,
in 12th Symposium on Neural Network Applications in Electrical Engineering (NEUREL),
Belgrade (2014), pp. 115–120
5. L. Vega-Escobar, A.E. Castro-Ospina, L. Duque-Munoz, DWT-based feature extraction for
motor imagery classification, in 6th Latin-American Conference on Networked and Electronic
Media (LACNEM 2015), Medellin (2015), pp. 1–6
6. M. Wairagkar, Motor imagery based brain computer interface (bci) using artificial neural
network classifiers, in Proceedings of the British Conference of Undergraduate Research,vol
1 (2014)
Performance Estimation and Analysis Over … 141
7. H. Yang, S. Sakhavi, K.K. Ang, C. Guan, On the use of convolutional neural networks
and augmented CSP features for multi-class motor imagery of EEG signals classification,
in Conference Proceedings of IEEE Engineering in Medicine and Biology Society (2015)
8. G.C. Jana, A. Swetapadma, P.K. Pattnaik, An intelligent method for classification of normal
and aggressive actions from electromyography signals, in 1st International Conference on
Electronics, Materials Engineering and Nano-Technology (IEMENTech) (2017), pp. 1–5
9. G.C. Jana, A. Swetapadma, P.K. Pattnaik, A hybrid method for classification of physical action
using discrete wavelet transform and artificial neural network. Int. J. Bioinf. Res. Appl. (IJBRA)
17(1), xx–xx (2021). (In press)
10. S.S.R.J. Rabha, K.Y. Nagarjuna, D. Samanta, P. Mitra, M. Sarma, Motor imagery EEG
signal processing and classification using machine learning approach, in 2017 International
Conference on New Trends in Computing Sciences (ICTCS) (2017), pp. 61–66
11. Zhichuan Tang, Chao Li, Shouqian Sun, Single-trial EEG classification of motor imagery using
deep convolutional neural networks. Optik 130, 11–18 (2017)
12. S. Chaudhary, S. Taran, V. Bajaj, A. Sengur, Convolutional neural network based approach
towards motor imagery tasks EEG signals classification. IEEE Sens. J. 19(12), 4494–4500
(2019)
13. L. Ming-Ai, W. Rui, H. Dong-Mei, Y. Jin-Fu, Feature extraction and classification of mental
EEG for motor imagery, in 2009 Fifth International Conference on Natural Computation
(2019), pp. 139–143
14. [DataSet] http://www.bbci.de/competition/iii/desc_IVa.html
15. D. Steyrl, R Scherer, G.R. Müller-Putz, Using random forests for classifying motor imagery
EEG, in Proceedings of TOBI Workshop IV (2013), pp. 89–90
16. N. Lu, T. Li, X. Ren, H. Miao, A deep learning scheme for motor imagery classification based
on restricted Boltzmann Machines. IEEE Trans. Neural Syst. Rehabil. Eng. 25(6), 566–576
(2017)
17. S. Kumar, A. Sharma, K. Mamun, T. Tsunoda, A deep learning approach for motor imagery
EEG signal classification, in 2016 3rd Asia-Pacific World Congress on Computer Science and
Engineering (APWC on CSE) (2016), pp. 34–39
18. S. Guan, K. Zhao, S. Yang,Motor imagery EEG classification based on decision tree framework
and Riemannian geometry. Comput. Intell. Neurosci. 2019, Article ID 5627156 (2019)
Fully Automated Digital Mammogram
Segmentation
Karuna Sharma and Saurabh Mukherjee
Abstract The “Computer-Aided Detection and Diagnosis” (CADx) system plays a
vital role as second look to identify and analyze the breast carcinoma. The functioning
of CADx can be degraded due to some factors like the appearance of impulse and
speckle noise, artifacts, and low contrast both in CC and MLO views and pectoral
muscles appears in mammogram’s MLO view. For this reason, noise elimination,
artifacts and pectoral muscles, mammogram image enhancement, and breast profile
extraction are significant prior process stages in the CADx system for breast carci-
noma analysis. The research is aimed to propose a precise and effective method
for completely automated mammogram image segmentation which includes CC and
MLO views. In this paper, median filter and Wiener filter are used for noise removal,
threshold based on otsu’s method for artifact removal, “Contrast limited histogram
equalization” method for mammogram enrichment in both views and multilevel
threshold with canny edge detection for pectoral muscle segmentation and breast
parenchyma extraction. The proposed work was examined on mammographic images
containing both views from CBIS-DDSM and MIAS databases. Jaccard similarity
index, Dice similarity index, and Score Matching methods have been employed to
estimate the performance of segmentation results that represents the proposed work’s
effectivity and usability.
Keywords Computer-aided detection and diagnosis ·Noise ·Image
enhancement ·Thresholding ·Otsu’s method ·Edge detection ·Segmentation
K. Sharma (B)·S. Mukherjee
Department of Computer Science, Faculty of Mathematics and Computing, Banasthali Vidyapith,
Banasthali, Tonk 304022, Rajasthan, India
e-mail: karuna4frenz@gmail.com
S. Mukherjee
e-mail: mukherjee.saurabh@rediffmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_13
143
144 K. Sharma and S. Mukherjee
1 Introduction
The breast carcinoma utmost continual ill health spotted among women, in both
emerging and established countries’ death. As stated by “World Health Organization”
(WHO) and “National Carcinoma Institute” [1] in 2004, breast carcinoma reported
as the cause of 13% of total demises in the globe [2] and one among eight women
faces breast carcinoma in some phase throughout her lifetime in the “United States”
(US). The radiologists use “Computer-Aided Detection and Diagnosis” (CADx)
system extensively as an investigative and assessment tool for breast carcinoma
detection at the primary phase. It is an extremely trustworthy method for the primary
detection of breast carcinoma, decreasing life-threatening rates up to 25%. But the
performance of CADx can be degraded due to some factors like the appearance of
impulse and speckle noise, artifacts, and low contrast both in CC and MLO views
and pectoral muscles appear in mammogram’s MLO view. For this reason, noise
elimination, artifacts and pectoral muscles, mammogram image enhancement, and
breast profile extraction are significant preprocessing stages in the CADx system to
analyze breast carcinoma. [3] proposed an algorithm to remove artifacts and pectoral
muscle by using a region description, split and merge method to extract pectoral
muscle. In [4], authors studied the several filters as mean, median, and Wiener filters
by applying various window sizes using DDSM (“Digital Database for Screening
Mammography”) database, evaluated with PSNR (“Peak Signal to Noise Ratio’). [5]
proposed an approach to remove pectoral muscles by using ROR (“Robust Outlying
Ratio”) method and to remove Gaussian and impulse noise by using DCT (“Discrete
Cosine Transform”) filter ROR-NLM. [6] performed histogram equalization [7],
and used histogram equalization with Grey relational analysis for the enhancement
of mammogram. Agarwal et al. [8] used homomorphic filtering (MH-FIL) method
to modify the histogram-based contrast enhancement method. CLAHE enhances
smaller regions in mammograms better. [9,10] proposed CLAHE (“Contrast Limited
Adaptive Histogram Equalization”)-based mammogram image enhancement, math-
ematical morphology, and multiscale laplacian Gaussian pyramid transform. [11]
proposed Otsu threshold and multiple regression analysis-based pectoral muscle
segmentation. Shape-based mask with morphological operators is employed to the
mammographic image by [12] and fitting function with cubic polynomial for the
segmentation of pectoral muscle region [13]. In this paper, the clustering tech-
nique based on K-Means to eliminate pectoral muscle morphology-based opera-
tions, and “seeded region growing” (SRG) techniques to remove noise and artifacts
are proposed. Kwok et al. [14] used iterative “cliff detection” to detect the pectoral
muscle region. The pectoral muscle detection with morphology-based operations and
the “random sample consensus” method named “RANSAC” are proposed in [15].
The region of pectoral muscle is segmented by using geometrical shapes and CLAHE
for mammogram image enhancement is proposed in [16]. In [17], histogram-based 8-
neighborhood connected component labeling to remove pectoral muscle is proposed.
[18] proposed binary threshold-based pectoral muscle segmentation. In [19], a hybrid
approach to delineate the pectoral region border by applying Hough transform and
Fully Automated Digital Mammogram Segmentation 145
to segment pectoral muscle active contour is proposed. In [20], threshold with an
active contour model for breast boundary extraction and canny edge detection for
pectoral muscle removal are used. In [21], the authors proposed a two-phase novel
“Margin Setting Algorithm” (MSA)-based breast and pectoral region segmentation.
[22] proposed a low-contrast pectoral region segmentation by using local histogram
equalization and estimation of polynomial curvature for the selected areas. In [23], an
adaptive gamma correction-based pectoral muscle detection with 98.45% accuracy is
proposed. [24] proposed linear enhancement mask and global threshold-based tech-
nique to detect the pectoral region boundary. In [25], the author presented a review of
the segmentation method for pectoral muscle, breast boundary, micro-calcification,
and mass lesions. [26] proposed analysis on particle swarm optimization and Ant
bee colony optimization methods to search optimal multilevel threshold value for
segmentation.
2 Research Gaps
After going through a literature review, we found the following research gaps:
1. There is little research work available for segmenting mammographic images
including both CC and MLO views using a single system.
2. There is only a little research work available for integration of artifact removal,
noise suppression, pectoral muscle segmentation, breast boundary extraction,
and mammogram image enhancement.
3 Objectives
1. To develop a comprehensive algorithm for completely automated segmentation
of mammograms on CC and MLO views.
2. To create a GUI for the implementation of a fully automated segmentation of
mammograms on CC and MLO views.
4 Research Methodology
a. Proposed Research Methodology
See Fig. 1.
b. Tools and Techniques
Image Database: The secondary data are used for this study from the “Mammo-
graphic Image Analysis Society” (MIAS) and “Curated Breast Imaging Subset”
of DDSM (“CBIS-DDSM”). The MIAS database includes 320 films of only
mediolateral oblique (MLO) views with 1024 ×1024 pixels and investigated and
146 K. Sharma and S. Mukherjee
Mammogram
Image (Input)
Resize
Mammogram
Image (256*256)
ArƟfact and
Label removal Noise Removal
Pectoral Muscle
SegmentaƟon
Mammogram
Image
Enhancement
Breast
profile
ExtracƟon
ImplementaƟon
of proposed
system using
MATLAB
Fig. 1 Sequence of methods to be used in this research
labeled by specialized radiologists. The DDSM consists of 2620 mammograms
of normal, benign, and malignant cases with verified pathology information.
Techniques used in proposed Algorithm
Normalization of Grayscale Image: Normalization is a technique to transform
a gray image with intensities in the (Min, Max) range, into a new image with
intensities in the (IMin, IMax) range. The normalization of a digital grayscale
image is performed according to the formula equation:
INormalized =((I−Min)∗((IMax −IMin)÷(Max −Min))) +IMin (1)
Global Image Thresholding: Otsu developed a discriminate analysis-based
method to determine the optimum thresholds to separate the level of the classes in
grayscale image at its maximum [27]. Otsu suggested the between-class variance
which is the sum of sigma functions of each class is given by Eq. (2):
f(t)=σ0+σ1(2)
σ0=ω0(μ0−μt)2,σ
1=ω1(μ1−μt)2(3)
Here, (μT)is the gray image’s mean intensity. The mean level (μ1) of two classes
for the case of bi-level threshold is given by Eq. (4)[28]:
σ0=
t−1
i=0
iP
i
ω0
σ1=
t−1
i=0
iP
i
ω1
(4)
The optimal value of threshold to maximize the function for a between-class
variance can be specified using Eq. (5)[28]:
σ2
ω(t)=ω0(t)σ 2
0(t)+ω1(t)σ 2
1(t)(5)
Fully Automated Digital Mammogram Segmentation 147
Optimal Multilevel Threshold: The multilevel threshold method subdivides
the image pixels into several distinct groups having similar gray levels within
a specific range. The Otsu’s method can be applied for multiple threshold [29]
segmentation. The optimum threshold values t∗
1and t∗
2can be computed as Eq. (5).
f(x,y)=⎧
⎨
⎩
aifg(x,y)>t2
bift
1<g(x,y)≤t2
cifg(x,y)≤t1
⎫
⎬
⎭
(6)
“Canny Edge Detection”: Canny method for detection of an edge is the summa-
tion of four complex exponentials calculated by Gaussian’s first derivative. As
stated by Canny, detection of edges can be done by convolution of the noisy
image with a function f(x) and denoting the edges in the output at the maxima
of convolution. The procedure to compute edges using a Canny edge detector is
given as follows:
•Compute the image gradient f(a,b) by convolution of image with Gaussian’s
first derivative in xand ydirections which is expressed by Eqs. (7) and (8):
fa(a,b)=f(a,b)∗−a
σ2e−a2+b2
2σ2(7)
fb(a,b)=f(a,b)∗−b
σ2e−(a2+b2)/2σ2(8)
•Apply Non-Maxima Suppression on the result of above step.
•Perform hysteresis threshold on the resultant gradient of above step.
Canny edge detection method works on two thresholds by scanning image form
leftward to rightward and upward to downward. If the non-maxima-suppressed
gradient magnitude at the pixel is larger than the highest threshold, then it is
declared as an edge point. On the other hand, if the non-maxima-suppressed
gradient magnitude at the pixel is larger than the lowest threshold, then it is also
declared as an edge point.
“Contrast-Limited Adaptive Histogram Equalization”: The “Contrast-
Limited Adaptive Histogram Equalization” computes the contrast-limited trans-
formation function for each small region and, as a result of this operation, each
small area’s contrast is enhanced as the ‘Distribution’ value histogram matches
approximately with the histogram of the outcome region. The neighborhood
regions are then combined to remove falsely inferred boundaries using bilinear
interpolation. The contrast is limited to avoid amplify any noise in homogeneous
areas that may exhibit in the input image.
Performance evaluation of segmentation result: Jaccard similarity index, Dice
similarity index, and Score Matching methods are employed to estimate the
effectiveness of segmented region results.
148 K. Sharma and S. Mukherjee
(a) The Jaccard Similarity index Jcomputes the likeness of real (“ground truth”)
image and segmented image as follows:
J(Ig,Is)=
Ig∩Is
Ig∪Is
(9)
where Igis the real (“ground truth”) image segmented manually and Isis
the segmented image by proposed work.
(b) The Dice Similarity index Dcomputes the likeness of real (“ground truth”)
image and segmented image as follows:
D(Ig,Is)=2∗
Ig∩Is
Ig
+|Is|(10)
where Igis the real (“ground truth”) image segmented manually and Is
represents the segmented image by proposed work.
(c) The Score matching method measures the closeness of the segmented image
boundary and the ground truth image boundary as follows:
sm =2∗p∗r/(r+p)(11)
where pis the precision and r is the recall value and the score match is the
harmonic mean of pand rwith a distance error tolerance.
c. Implementation of Proposed Algorithm
See Fig. 2.
Algorithm 1. Pseudocode for the Proposed Algorithm START
1. Read the mammogram image
Segmented Breast ROI (Output Image for Further Analysis)
Breast Profile ExtracƟon
Canny Edge DetecƟon for Breast Boundary ExtracƟon Breast ROI enhancement using CLAHE
Pectoral Muscle Segmentaion
OpƟmal Otsu's Method based MulƟlevel Thresholding Using Median filter to sharpen the image
Labels and other ArƟfact Removal
OpƟmal Otsu's Method based global Image Threshold Using Gaussian filter and Weiner Filter to remove noise
Mammogram Image[Input]
Resize the Image
Fig. 2 Steps for implementation of proposed work
Fully Automated Digital Mammogram Segmentation 149
2. Resize the grayscale image to 256*256 pixels
3. IF orientation is right
a. Flip the image left to right
End IF
4. Apply “Contrast-Limited Adaptive Histogram Equalization” to enhance the
breast ROI.
5. Apply optimal Otsu’s method-based global image threshold
6. Filter the mammogram image using a Gaussian filter and Wiener filter to remove
impulse and speckle noise.
7. IF MLO view image then
a. Apply optimal Otsu’s method-based multilevel Threshold
b. Filter the image using the median filter to sharpen the image.
End IF
8. Detect the edges of breast boundary using edge detection method developed by
Canny.
9. Filtering the mammogram using Gaussian filter to smoothen the edges of image.
END
5 Result and Discussion
We have implemented the proposed approach using GUI created in Matlab2018b.
A result of the complete algorithm is presented with a number of input and output
mammogram images of both MLO and CC views with both breasts. In our work, we
used a global threshold technique based on Otsu’s method to eliminate artifacts and
labels. Level five Multi-threshold, morphological structuring, and region fill are used
for pectoral muscle segmentation and extraction. Breast structure excluding pectoral
region is extracted using an edge detector developed by Canny and Gaussian filtering.
The proposed work’s result exhibited the capability to eliminate artifacts and label,
remove and extract pectoral muscle, and extract breast profile and to enhance the
contrast of image without losing any information. The outcome exhibit reducing the
size of image will reduce the computational time of proposed work (Figs. 3,4,5,6,
7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24, and 25).
6 Conclusion
This research paper dealing with the identification and segmenting of the pectoral
muscle region and breast profile from both MLO and CC view of mammogram
150 K. Sharma and S. Mukherjee
Fig. 3 Mdb004.pgm
Fig. 4 Mdb009.pgm
Fig. 5 mdb012.pgm
Fully Automated Digital Mammogram Segmentation 151
Fig. 6 Mdb014.pgm
Fig. 7 Mdb020.pgm
Fig. 8 Mdb022.pgm
152 K. Sharma and S. Mukherjee
Fig. 9 Mdb024.pgm
Fig. 10 Mdb030.pgm
Fig. 11 Mdb032.pgm
Fully Automated Digital Mammogram Segmentation 153
Fig. 12 Mdb034.pgm
Fig. 13 Mdb036.pgm
Fig. 14 Mdb040.pgm
154 K. Sharma and S. Mukherjee
Fig. 15 Mdb042.pgm
Fig. 16 Mdb046.pgm
Fig. 17 Mdb047.pgm
Fully Automated Digital Mammogram Segmentation 155
Fig. 18 Mdb048.pgm
Fig. 19 Mdb050.pgm
Fig. 20 Mdb058.pgm
156 K. Sharma and S. Mukherjee
Fig. 21 Mdb060.pgm
Fig. 22 Mdb064.pgm
Fig. 23 Mdb068.pgm
Fully Automated Digital Mammogram Segmentation 157
Fig. 24 Mdb069.pgm
Fig. 25 Mdb070.pgm
contains the left as well right breast. The proposed approach performs Level Five
Multilevel threshold to segment pectoral muscle which does not deal with the detec-
tion of a straightline. The local histogram equalization performs enhancement of
pectoral muscles. Then the pectoral region border is found with morphological struc-
turing and breast region, receptively, using Canny edge detection and threshold tech-
nique. The investigated outcomes obtained for the proposed work are tested on 100
mammograms, 50 from the MIAS database mammograms and 50 from the CBIS-
DDSM database consisting of both views of breasts. The future work of this research
will involve a comparative study of the proposed approach with other approaches.
The results obtained on 100 images of CBIS-DDSM and MIAS database have shown
an excellent output. The outcome mammogram images can be used further for the
next CADx stage to assist for discrimination, identification, and diagnosis of breast
carcinoma.
158 K. Sharma and S. Mukherjee
References
1. J. Ferlay, I. Soerjomataram, R. Dikshit, S. Eser, C. Mathers, M. Rebelo, D.M. Parkin, D.
Forman, F. Bray, Carcinoma incidence and mortality worldwide: sources, methods and major
patterns in GLOBOCAN 2012. Int. J. Carcinoma 136(5), E359–E386 (2015)
2. P.K. Dhillon, Breast carcinoma fact sheet. Mortality 11(5), 10–18 (2009)
3. J.A. Ojo, T.M. Adepoju, E. Omdiora, O. Olabiyisi, O. Bello, Pre-processing method for extrac-
tion of pectoral muscle and removal of artefacts in mammogram. IOSR J. Comput. Eng.
(IOSR-JCE) e 16(3), 6–9
4. S. Kannan, N.P. Subiramaniyam, A.T. Rajamanickam, A. Balamurugan, Performance compar-
ison of noise reduction in mammogram images. Image, 1(3 x 3), 5 x 5
5. S. Sreedevi, E. Sherly,A novel approach for removal of pectoral muscles in digital mammogram.
Procedia Comput. Sci 46, 1724–1731 (2015)
6. E.D. Pisano, S. Zong, B.M. Hemminger, M. DeLuca, R.E. Johnston, K. Muller, M.P. Braeuning,
S.M. Pizer, Contrast limited adaptive histogram equalization image processing to improve the
detection of simulated spiculations in dense mammograms. J. Digit. Imaging 11(4), 193–200
(1998)
7. B. Gupta, M. Tiwari, A tool supported approach for brightness preserving contrast enhance-
ment and mass segmentation of mammogram images using histogram modified grey relational
analysis. Multidimension. Syst. Signal Process. 28(4), 1549–1567 (2017)
8. T.K. Agarwal, M. Tiwari, S.S. Lamba, Modified histogram-based contrast enhancement using
homomorphic filtering for medical images, in 2014 IEEE International on Advance Computing
Conference (IACC) (IEEE, 2014), pp. 964–968
9. E.D. Pisano, E.B. Cole, B.M. Hemminger, M.J. Yaffe, S.R. Aylward, A.D. Maidment,
R.E. Johnston, M.B. Williams, L.T. Niklason, E.F. Conant, L.L. Fajardo, Image processing
algorithms for digital mammography: a pictorial essay 1. Radiographics 20(5), 1479–1491
(2000)
10. I.K. Maitra, S. Nag, S.K. Bandyopadhyay, Technique for preprocessing of digital mammogram.
Comput. Methods Programs Biomed. 107(2), 175–188 (2012)
11. C.C. Liu, C.Y. Tsai, J. Liu, C.Y. Yu, S.S. Yu, A pectoral muscle segmentation algorithm for
digital mammograms using Otsu thresholding and multiple regression analysis. Comput. Math
Appl. 65(5), 1100–1107 (2012)
12. C. Chen, G. Liu, J. Wang, G. Sudlow, Shape-based automatic detection of pectoral muscle
boundary in mammograms. J. Med. Biol. Eng. 35(3), 315–322 (2015)
13. N. Alam, M.J. Islam, Pectoral muscle elimination on mammogram using K-means clustering
approach. Int. J. Comput. Vis. Sign. Process. 4(1), 11–21 (2014)
14. S.M. Kwok, R. Chandrasekhar, Y. Attikiouzel, M.T. Rickard, Automatic pectoral muscle
segmentation on mediolateral oblique view mammograms. IEEE Trans. Med. Imaging 23(9),
1129–1140 (2004)
15. W.B. Yoon, J.E. Oh, E.Y. Chae, H.H. Kim, S.Y. Lee, K.G. Kim, Automatic detection of pectoral
muscle region for computer-aided diagnosis using MIAS mammograms. BioMed Res. Int.
(2016)
16. S.A. Taghanaki, Y. Liu, B. Miles, G. Hamarneh, Geometry-based pectoral muscle segmentation
from mlo mammogram views. IEEE Trans. Biomed. Eng. 64(11), 2662–2671 (2017)
17. R. Boss, K. Thangavel, D. Daniel, Automatic Mammogram Image Breast Region Extraction
and Removal of Pectoral Muscle (2013). arXiv preprint arXiv:1307.7474
18. A.K. Mohideen, K. Thangavel, Removal of pectoral muscle region in digital mammograms
using binary thresholding. Int. J. Comput. Vis. Image Process. (IJCVIP) 2(3), 21–29 (2012)
19. A.L. Pavan, A. Vacavant, A.F. Alves, A.P. Trindade, D.R. de Pina, Automatic identification and
extraction of pectoral muscle in digital mammography, in World Congress on Medical Physics
and Biomedical Engineering 2018 (Springer, Singapore, 2019), pp. 151–154
20. A. Rampun, P.J. Morrow, B.W. Scotney, J. Winder, Fully automated breast boundary and
pectoral muscle segmentation in mammograms. Artif. Intell. Med. 79, 28–41 (2017)
Fully Automated Digital Mammogram Segmentation 159
21. H. Al-Ghaib, Y. Wang, R. Adhami, A new machine learning algorithm for breast and pectoral
muscle segmentation. Eur. J. Adv. Eng. Technol. 2, 21–29 (2015)
22. M. Mustra, M. Grgic, Robust automatic breast and pectoral muscle segmentation from scanned
mammograms. Sig. Process. 93(10), 2817–2827 (2013)
23. S.J.S. Gardezi, F. Adjed, I. Faye, N. Kamel, M.M. Eltoukhy, Segmentation of pectoral muscle
using the adaptive gamma corrections. Multimedia Tools Appl. 77(3), 3919–3940 (2018)
24. P.S. Vikhe, V.R. Thool, Detection and segmentation of pectoral muscle on MLO-view
mammogram using enhancement filter. J. Med. Syst. 41(12), 190 (2017)
25. J. Dabass, S. Arora, R. Vig, M. Hanmandlu, Segmentation techniques for breast carcinoma
imaging modalities-a review, in 9th International Conference on Cloud Computing, Data
Science & Engineering (2019)
26. B. Akay, A study on particle swarm optimization and artificial bee colony algorithms for
multilevel thresholding. Appl. Soft Comput. 13(6), 3066–3091 (2013)
27. T. Kurban, P. Civicioglu, R. Kurban, E. Besdok, Comparison of evolutionary and swarm based
computational techniques for multilevel color image thresholding. Appl. Soft Comput. 23,
128–143 (2014)
28. A.K. Bhandari, A. Kumar, G.K. Singh, Modified artificial bee colony based computationally
efficient multilevel thresholding for satellite image segmentation using Kapur’s, Otsu and
Tsallis functions. Expert Syst. Appl. 42(3), 1573–1601 (2015)
29. ChH Bindu, K.S. Prasad, An efficient medical image segmentation using conventional OTSU
method. Int. J. Adv. Sci. Technol. 38, 67–74 (2012)
Empirical Study of Computational
Intelligence Approaches for the Early
Detection of Autism Spectrum Disorder
Mst. Arifa Khatun, Md. Asraf Ali, Md. Razu Ahmed,
Sheak Rashed Haider Noori, and Arun Sahayadhas
Abstract The objective of the research is to develop a predictive model that can
significantly enhance the detection and monitoring performance of Autism Spec-
trum Disorder (ASD) using four supervised learning techniques. In this study, we
applied four supervised-based classification techniques to the clinical ASD data
obtained from 704 patients. Then, we compared the four machine learning (ML)
algorithms performance across tenfold cross-validation, ROC curve, classification
accuracy, F1 measure, precision, recall, and specificity. The analysis findings indi-
cate that Support Vector Machine (SVM) achieved the uppermost performance than
the other classifiers in terms of accuracy (85%), f1 measure (87%), precision (87%),
and recall (88%). Our work presents a significant predictive model for ASD that can
effectively help the ASD patients and medical practitioners.
Keywords Autism spectrum disorder (ASD) ·Machine learning ·Classification ·
Screening tools ·SVM
1 Introduction
Autism spectrum disorder (ASD) is a complex neurobehavioral syndrome that refers
to impairments in different social skills and developmental, repetitive activities
combined with nonverbal communication. In 2018 alone, according to the Autism
and Developmental Disabilities Monitoring (ADDM) Network, a division of Centers
Mst. A. Khatun ·S. R. H. Noori
Department of Computer Science and Engineering, Daffodil International University, Dhaka,
Bangladesh
Md. A. Ali ·Md. R. Ahmed (B)
Department of Software Engineering, Daffodil International University, Dhaka, Bangladesh
e-mail: razu35-1072@diu.edu.bd
A. Sahayadhas
Artificial Intelligence Research Lab, Vels Institute of Science, Technology and Advanced Studies,
Chennai, India
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_14
161
162 M. A. Khatun et al.
for Disease Control and Prevention (CDC), 1 in 59 is recognized with ASD [1]. More-
over, males are four times more likely to be diagnosed with ASD than females [1].
Studies have shown that 31% of adolescents affected with ASD have a severe psycho-
logical impairment (i.e., intelligence quotient less than 70), another 25% adolescents
are in the marginal range (i.e., IQ: 71–85), and only 44% adolescents have IQ in
the normal range (i.e., IQ greater than 85) [2]. The existing scientific evidence indi-
cates that the possible factors causing ASD in children may be due to ecological
and genomic factors [3]. These possible factors indicate that the primary signs of
autism usually appear at early age (i.e., 2 or 3 years). Consequently, a study shows
that early diagnosis within the beginning stage of life can lead to positive outcomes
of autism-affected people [4].
Researchers have presented different autism diagnosis tools and techniques which
are based on some scripted questionnaires with scoresheets that yield results. Based
on the inference, the clinical practitioner can take their decision. Example of different
autism screening tools and techniques are the “Modified Checklist for Autism in
Toddlers (M-CHAT) [5], Autism Diagnostic Interview (ADI) [6], Autism Diag-
nostic Observation Schedule Revised (ADOS-R) [7], and Child Behavior Check-
list (CBCL)” [8], etc. The existing ADI-R and ADOS have shown that competitive
performance as inputs to train autism screening algorithms was clinically confirmed
in several experimental research findings [9–12]. Cognoa [13] is another data-driven
machine learning (ML)-based tool and it has been validated by multiple research
studies [12,14,15]. The Cognoa ASD screening tool has been developed with the
aim of helping the practitioners to diagnose autism in children between the age of
18 and 72 months [16].
ML algorithms have been used to improve the diagnosis process of ASD as well
as to obtain information to deliver faster access to clinical services for the medical
practitioner so that an effective decision based on the ASD diagnosis can be made
[17,18]. Moreover, ML classifiers are the most operational and effective methods for
low-cost screening ASD tools including different clinical contexts. Most of the time,
clinical data are generated from different clinical environments but the data consists
of an imbalance, unstructured format making it difficult to be applied to the screening
tools. The consistency and reliability of ML models may vary conditionally based
on the real-time trained data. In this work, we highlight these challenges and current
practical tools and strategies for early diagnosis using several ML algorithms. The
main aspect of this work is to examine different classifiers’ performance through
different performance measurement techniques and attain more effective decisions
from clinical data. Many of the studies focused on only accuracy and classification
based on ASD data. Therefore, the outperform classification techniques have been
considered for the predictive model.
The rest of the paper is presented as follows, the materials and methods of this
study are presented in Sect. 2. In Sect. 3, the experimental results are described.
Finally, conclusions and future recommendations of this work are illustrated in
Sect. 4.
Empirical Study of Computational Intelligence … 163
2 Materials and Methods
A. Dataset Collection
In this work, we used the ASD dataset created by Fadi Fayez Thabtah and which
is provided by the “UCI Machine Learning Repository” [19,20]. We considered
704 instances including 21 attributes for our predictive model. Table 1shows the
original attributes from the dataset which is collected from the UCI Machine Learning
Repository.
B. Classification Techniques for clinical tools
In the last decades, ML algorithms have been performing a significant role and it is
recognized to solve the medical problem and clinical diagnosis [21,22]. Studies show
that ML approaches can perform a significant role in improving the consistency and
effectiveness of clinical ASD screeners [23–25]. In this work, four ML classifiers
have been considered to perform the ASD screening model: “k-nearest neighbors
(KNN), Logistics Regression (LR), Random Forest (RF) and Support Vector Machine
(SVM)”.
Tabl e 1 Dataset attribute
from UCI machine learning
repository
#ID Feature Data type
1Age of patients Numerical
2 Gender of patients string_value
3Ethnicity of patients string_value
4Born alongside jaundice Ye s / no
5 Family_member has a PDD Yes / n o
6 Completed test (i.e. doctors/parents) String_value
7Residence String_value
8Used_the_screening app previous yes/no
9Screening techniques Numerical
10 The answer to Q1 binary_value (0/1)
11 The answer to Q2 binary_value (0/1)
12 The answer to Q3 binary_value (0/1)
13 The answer to Q4 binary_value (0/1)
14 The answer to Q5 binary_value (0/1)
15 The answer to Q6 binary_value (0/1)
16 The answer to Q7 binary_value (0/1)
17 The answer to Q8 binary_value (0/1)
18 The answer to Q9 binary_value (0/1)
19 The answer to Q10 binary_value (0/1)
20 The screening test results Numerical
21 Target class Yes/ n o
164 M. A. Khatun et al.
Tabl e 2 The confusion
matrix of actual and projected
class
Projected class
Actual class True_positive
(TPsample)
False_positive
(FPsample)
False_negative
(FNsample)
True_negative
(TNsample)
TPs: the amount of true_positive samples; FPs: the amount of
false_positive samples
FNs: the amount of false_negative samples; TNs: the amount of
true_negative samples
Tabl e 3 Performance
measurement benchmark Performance metrics Mathematical equation
Accuracy (TP+TN)
(TP+FP+TN+FN)
Precision TP
(TP+FP)
Recall TP
(TP+FN)
F1 measure 2∗(Recall*Precision)
(Recall+Precision)
Specificity TN
(TN+FP)
C. Performance Measurement
In this paper, performances have been validated using the technique of tenfold cross-
validation [26] and different performance measurement approaches. The confusion
matrix is an operational tool for examining how good a ML classification technique
can recognize of dissimilar classes. Table 2shows a confusion matrix for actual and
projected classes.
In order to evaluate and compare these supervised ML techniques for ASD
screening, we used different performance measurement techniques as shown in
Table 3.
D. Experimental Setup
In this experiment, the ASD dataset has been considered to develop the ASD predic-
tive model. Then we performed different data pre-processing methods on the ASD
dataset to make a tidy dataset, such as correlation analysis to find the redundant values,
missing values analysis, feature selection analysis, etc. The detailed workflow of the
ASD predictive model is presented in Fig. 1.
Empirical Study of Computational Intelligence … 165
Data
Processing
Raw Data
Data Split
(Training 80%,
Testing 10%,
Validation 10%)
Applying 4
Classifiers Analyze Model Approve the
Outperform model
Feature
Selection
Train Data
Validation
Data
Test Data
Fig. 1 Workflow of ASD screening model
3 Result and Discussions
A. Data Preprocessing
In this section, we considered various evaluations to investigate the ASD dataset.
Figure 2shows that the white and European country-wise distribution of ASD and it
is looking likely to similar for topmost five countries such as the United States, the
United Kingdom, Australia, New Zealand, and Canada as top providers of positive
ASD. Moreover, Fig. 3the observation shows that the adults and toddlers have
an extreme risk of being an ASD patient in which people are based on white and
Europeans Ethnicities. The Black and Asian people are following the next risk factor
with a smaller ratio. We conclude that our study has presented a possible genetic link
for ASD positive.
Here, Fig. 4shows that the ASD positive ratio is more familiar among boys than
girls. But our study presents a different scenario of adults. Whereas in adults, it is
Fig. 2 The ratio of ASD based on white and European country
166 M. A. Khatun et al.
Fig. 3 ASD positive relatives with autism distribution for different ethnic peoples
Fig. 4 ASD positive with jaundice based on gender
lower than the female gender but, considering toddlers, it is likely four times higher
ratio than girls. From the ASD dataset, 704 samples and 21 attributes were taken into
the analysis of the predictive ASD model. We splitted the ASD dataset into three
chunks, whereas the training set comprises 80%, the test set comprises 10%, and
another 10% of the splitted data for validation test. Hence, the ASD dataset was as
well examined to validate the redundant values. we have used heatmap to find the
redundant and correlated columns in the ASD dataset. Our analysis shows that there
are no columns correlated with one to one in Fig. 5.
B. Analysis of the performance
The prediction performance of four ML classifiers was examined for the classification
of ASD. Figure 6shows the performance of four supervision-based classification
techniques. With respect to precision, recall, and f1 measure, SVM achieved the
highest performance than other classifiers. Moreover, when considering the accuracy
(Fig. 7), SVM obtained the highest (i.e., 85%). In addition, LR achieved the lowest
performance in terms of accuracy, precision, recall, and f1 measure. By looking at
KNN and RF classification techniques, we can observe that their performance was
almost similar. However, the performance results suggesting that the SVM is more
effective and reliable in the prediction of ASD models.
Another measurement for classifiers is ROC (receiver operating curve) [27], which
is based on “false positive rate (x-axis) and true positive rate (y-axis)”. However, the
Empirical Study of Computational Intelligence … 167
Fig. 5 Heat map for checking correlated columns
Fig. 6 The figure shows the
performance of four
classifiers
168 M. A. Khatun et al.
Fig. 7 Classification accuracy for ASD models
Fig. 8 Receiver operating curve for the classification techniques
ROC curve is unbiased of both classes for measuring the capability of the predictive
classifier. Figure 8shows that SVM performs better than other classifiers to predict
ASD.
4 Conclusion
Early diagnosis of ASD can prevent the occurrences of ASD patients and can have
a significant influence on its clinical treatment. This work presents a workflow that
is based on computational intelligence techniques for the forecast and diagnosis of
ASD. This work used four classification techniques in the early identification of
ASD patients. These four ML methods were validated with tenfold cross-validation
techniques including different statistical measurement techniques. The performance
result shows that the SVM achieved the highest performance (i.e., 85%). Therefore,
Empirical Study of Computational Intelligence … 169
this ML-based predictive application can be used for early diagnosis of ASD patients
and which will be helpful for clinical practitioners and the health-care research
community.
Conflict of Interest None.
References
1. J. Baio, Prevalence of Autism spectrum disorder among children aged 8 years—autism
and developmental disabilities monitoring network, 11 Sites, United States, 2014. MMWR.
Surveill. Summ. 67(6), 1–23 (2018)
2. Autism and Health: A Special Report by Autism Speaks | Autism Speaks. [Online].
Available: https://www.autismspeaks.org/science-news/autism-and-health-special-report-aut
ism-speaks. Accessed 7 Sept 2019
3. What Is Autism? | Autism Speaks. [Online]. Available: https://www.autismspeaks.org/what-
autism. Accessed 9 Sept 2019
4. M.S. Durkin et al., Socioeconomic inequality in the prevalence of autism spectrum disorder:
evidence from a US cross-sectional study. PloS one 5(7), e11551 (2010)
5. C. Chlebowski, D.L. Robins, M.L. Barton, D. Fein, Large-scale use of the modified checklist
for autism in low-risk toddlers. Pediatrics 131(4), e1121–e1127 (2013)
6. C. Gillberg, M. Cederlund, K. Lamberg, L. Zeijlon, Brief report: ‘The autism epidemic’. The
registered prevalenceof autism in a Swedish Urban Area. J. Autism Dev. Disord. 36(3), 429–435
(2006)
7. C. Lord, S. Risi, L. Lambrecht, E.H. Cook, B.L. Leventhal, P.C. DiLavore, A. Pickles, M.
Rutter, The Autism diagnostic observation schedule—generic: a standard measure of social
and communication deficits associated with the spectrum of Autism. J. Autism Dev. Disord.
30(3), 205–223 (2000)
8. T.M. Achenbach, L. Rescorla, Manual for the ASEBA school-age forms & profiles: an
integrated system of multi-informant assessment. ASEBA (2001)
9. D.P. Wall, R. Dally, R. Luyster, J.-Y. Jung, T.F. DeLuca, Use of artificial intelligence to shorten
the behavioral diagnosis of autism. PLoS ONE 7(8), e43855 (2012)
10. E. Ruzich et al., Measuring autistic traits in the general population: a systematic review of
the Autism-Spectrum Quotient (AQ) in a nonclinical population sample of 6,900 typical adult
males and females. Mol. Autism 6(1), 2 (2015)
11. D.P. Wall, J. Kosmicki, T.F. DeLuca, E. Harstad, V.A. Fusaro, Use of machine learning to
shorten observation-based screening and diagnosis of autism. Transl. Psychiatry 2(4), e100–
e100 (2012)
12. M. Duda, J. Daniels, D.P. Wall,Clinical evaluation of a novel and mobile autism risk assessment.
J. Autism Dev. Disord. 46(6), 1953–1961 (2016)
13. Cognoa | Home. [Online]. Available: https://www.cognoa.com/. Accessed 9 Sept 2019
14. S.M. Kanne, L.A. Carpenter, Z. Warren, Screening in toddlers and preschoolers at risk for
autism spectrum disorder: evaluating a novel mobile-health screening tool. Autism Res. 11(7),
1038–1049 (2018)
15. A. Sarkar, J. Wade, A. Swanson, A. Weitlauf, Z. Warren, N. Sarkar, A Data-Driven Mobile
Application for Efficient, Engaging, and Accurate Screening of ASD in Toddlers (Springer,
Cham, 2018), pp. 560–570
16. Cognoa autism devices obtain FDA breakthrough status. [Online]. Available: https://www.med
icaldevice-network.com/news/cognoa-autism-devices/. Accessed 9 Sept 2019
17. F. Thabtah, Machine learning in autistic spectrum disorder behavioral research: A review and
ways forward. Inf. Heal. Soc. Care 44(3), 278–297 (2019)
170 M. A. Khatun et al.
18. J.L. Lopez Marcano, Classification of ADHD and non-ADHD Using AR Models and Machine
Learning Algorithms (2016)
19. F. Thabtah, Autism spectrum disorder screening, in Proceedings of the 1st International
Conference on Medical and Health Informatics 2017 - ICMHI ’17 (2017), pp. 1–6
20. “UCI Machine Learning Repository: Autism Screening Adult Data Set.” [Online]. Available:
https://archive.ics.uci.edu/ml/datasets/Autism+Screening+Adult. Accessed 9 Sept 2019
21. M.R. Ahmed, S.M. Hasan Mahmud, M.A. Hossin, H. Jahan, S.R. Haider Noori, A cloud based
four-tier architecture for early detection of heart disease with machine learning algorithms, in
2018 IEEE 4th International Conference on Computer and Communications (ICCC) (2018),
pp. 1951–1955
22. S.M.H. Mahmud, M.A. Hossin, M.R. Ahmed, S.R.H. Noori, M.N.I. Sarkar, Machine learning
based unified framework for diabetes prediction, in Proceedings of the 2018 International
Conference on Big Data Engineering and Technology - BDET 2018 (2018), pp. 46–50
23. H. Abbas, F. Garberson, E. Glover, D.P. Wall, Machine learning approach for early detection of
autism by combining questionnaire and home video screening. J. Am. Med. Inf. Assoc. 25(8),
1000–1007 (2018)
24. K.K. Hyde et al., Applications of supervised machine learning in autism spectrum disorder
research: a review. Rev. J. Autism Dev. Disord. 6(2), 128–146 (2019)
25. E. Stevens, D.R. Dixon, M.N. Novack, D. Granpeesheh, T. Smith, E. Linstead, Identification
and analysis of behavioral phenotypes in autism spectrum disorder via unsupervised machine
learning. Int. J. Med. Inform. 129, 29–36 (2019)
26. M. Duda, R. Ma, N. Haber, D.P. Wall, Use of machine learning for behavioral distinction of
autism and ADHD. Transl. Psychiatry 6(2), e732–e732 (2016)
27. C.E. Metz, Basic principles of ROC analysis. Semin. Nucl. Med. 8(4), 283–298 (1978)
Intelligent Monitoring of Bearings Using
Node MCU Module
Saroj Kumar, Shankar Sehgal, Harmesh Kumar, and Sarbjeet Singh
Abstract This paper discusses the application of NodeMCU to intelligent moni-
toring of bearings via an online method using an accelerometer to detect the vibra-
tion level. An accelerometer was used to detect the vibration level and NodeMCU
module for sending a message to the end-user regarding excessive vibration levels.
NodeMCU module serves as a low-cost industrial-internet-of-things setup for online
monitoring of bearings. In the experiment, the set-up had a motor (to provide torque
to the shaft), two ball bearings set, a shaft coupling (to connect main shaft to motor
shaft), a NodeMCU (for sending a warning message), an accelerometer (to detect the
vibration level), and Blynk app (to control the NodeMCU). The experimental setup
was designed to detect the vibration level in time domain as well as in frequency
domain and the setup was able to send the warning message in both the cases. By
using this type of experimental setup, the unwanted breakdown and uncertain failure
of machines due to bearing failure can be avoided. The setup helped in alerting the
user about any failure in real time whenever the magnitude of vibrations exceeded its
predetermined threshold limit. This experimental setup is found to be very relevant
for applications in small- and medium-scale industries due to its low-cost, ease of
operation, and good accuracy.
Keywords Accelerometer ·Bearings ·Blynk app ·Industrial-internet-of-things ·
NodeMCU
1 Introduction
In the current scenario, the unwanted shutdown and breakdown problems in industries
as well as other organizations is most common. Due to the ever-changing nature of
S. Kumar ·S. Sehgal (B)·H. Kumar
Mechanical Engineering Department, UIET, Punjab University, Chandigarh 160014, India
e-mail: sehgals@pu.ac.in
S. Singh
Mechanical Engineering Department, GCET, Jammu 181122, India
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_15
171
172 S. Kumar et al.
the working environment, it is desired to put into practice an effective e-maintenance
strategy for evolving proper usage of existing assets and reliability and safety [1,2].
The control of these types of failures in real-time can be possible by the application
of industrial-internet-of-things (IIoT). IIoT can be used in real-time monitoring by
integrating different components like Arduino UNO, GSM SIM900A module (for
sending a message to the user), and also Fast Fourier Transformation (FFT) setup. In
the current paper, the NodeMCU module has been used to send the warning message
in real time and this module is also cheaper than Arduino UNO and SIM GSM900A
developed earlier.
Hoshi, in 2006, developed a simple test setup for monitoring of damages produced
on rolling surfaces in a ball bearing-based setup of a machine tool spindle [3]. This
setup was able to monitor the initiation as well as the progress of the damage occurring
on rolling surfaces of the bearing and could also be used to predict the life of a bearing
assembly. In 2008, Cao and Jiang [4] developed a service-oriented architecture-based
system for supporting the decision-making process in the maintenance of machines.
Yang et al. [5], in 2009 used the vibration signal produced by a motor along with phase
currents for detecting the presence of a fault in the system and for further diagnosis
purposes. Later in 2010, Zhao et al. [6] developed service-oriented architecture-based
remote machine health-monitoring system which can be used to diagnose the faults
present in industrial machinery located at remote places by using the concepts of web
services, smart client, extensible mark-up language technologies, and visual studio.
Lazar et al. [7] in 2014 used vision-based robot predictive control in flexible automatic
manufacturing systems. In 2019, Goyal et al. [8,9] developed a laser-based non-
contact vibration measurement system for monitoring the condition of the machine
in rea -time. It verified the effectiveness and practicality of the system. Shelkeet al.
[10] used the time domain as well as frequency-domain methods for monitoring the
health of ball bearing by measuring their high-amplitude dynamic responses arising
out of wear, corrosion, fatigue or crack, etc., and also by extracting their features
to reduce the downtime of machines. In 2017, Hassin [11] used modulation signal
bispectrum method to analyze the dynamic responses. Recently, in 2019, Pesch and
Scavelli [12] performed the health monitoring of active magnetic bearings using an
online monitoring method.
Although several methods have been proposed and implemented for condition
monitoring of various types of industrial machines. But it is observed that NodeMCU
cum Blynk mobile app-based online condition monitoring of bearings assembly is
not used earlier. The proposed NodeMCU-based method is low cost and easy to use
due to its linking with the android-based mobile app Blynk. This paper deals with
the development of such low-cost user-friendly online condition monitoring system
that can be adopted easily by small- and medium-scale industries.
Intelligent Monitoring of Bearings Using Node MCU Module 173
2 Materials and Methods
An experimental system as shown in Fig. 1was developed which consisted of a
steel table containing a motor, two bearing-set housings, and a shaft connected to a
coupling. On the left-hand side bearing housing, an ADXL335 accelerometer was
mounted for detection of vibration level of that bearing. The experimental setup
also contained NodeMCU microcontroller, a low-cost module for sending real-time
notifications. NodeMCU system can transfer real-time data to user mobile through
Blynk app-based IIoT system. Initially, the experimental data were collected for
correct as well as faulty bearings. The collected data were then analyzed and the
program coding was done for NodeMCU and Blynk app. The entire work could be
divided into two prototypes. First prototype comprised of collection and analysis of
data in the time domain, while the second prototype was based on collection and
interpretation of data in the frequency domain.
The whole procedure of the experiment using NodeMCU can be summarized
as follows. First of all, vibration data were collected in the time domain by the
accelerometer. An Arduino sketch (program) was made to convert real time data
from time-domain format into the frequency domain format. Analyzed data points
(time-domain format in the first prototype; frequency-domain format in the second
prototype) were processed statistically to find the mean and standard deviation. A
threshold value was set to keep a check for faulty bearing readings. Blynk app was
developed to send a warning message to the end user in case the vibration signal
exceeds the threshold value. Analyzed values were compared with a threshold value
Fig. 1 NodeMCU-based system for online monitoring of vibration responses
174 S. Kumar et al.
and appropriate warning notifications were sent to the end-user through the Blynk
mobile app.
3 Results and Discussion
The time-domain signals for correct and faulty bearings signals were captured
consecutively by using the ADXL335 accelerometer. Blynk app was linked with
the NodeMCU microcontroller and was used to send the warning messages to the
end-user as shown in Figs. 2and 3for time domain and frequency domain, respec-
tively. Blynk app was also used to send the warning message on the email id of the
end-user as shown in Fig. 4.
Fig. 2 Blynk App graphical
user interface in time domain
Fig. 3 Blynk App graphical
user interface in frequency
domain
Intelligent Monitoring of Bearings Using Node MCU Module 175
Fig. 4 Warning message of Blynk app in time domain
Fig. 5 Acceleration signal for correct bearing
After the study of time-domain signal, FFT function was applied to the time
domain signals to convert it into frequency-domain format. The experimental plot for
the FFT of a healthy bearing is shown in Fig. 5and for the faulty bearing is shown in
Fig. 6. The peak results obtained in the frequency domain for healthy bearing signals
were observed to be 0.7 g and, in case of faulty bearing, it was 0.85 g. Based on
these results, the threshold limit for Blynk app was set accordingly to the maximum
peak signal of the correct bearing. Figure 7shows the corresponding experimental
warning message received in the inbox email of end-user.
4 Conclusion
The IIoT has significantly attracted the attention of researchers throughout the most
recent couple of years. With the progress in sensor hardware technology and cost-
effective materials, sensors are expected to be attached to all the items around us,
so that these can interact with each other with minimal human intervention. The
results of NodeMCU and Blynk app-based setup show that the proposed low-cost,
176 S. Kumar et al.
Fig. 6 Acceleration signal for faulty bearing
Fig. 7 Warning message of
Blynk app in frequency
domain
user-friendly system can be successfully utilized in online condition monitoring of
the bearings for sending a warning message to the end-user regarding any unwanted
high-vibration signals arising due to fault in the bearing setup. The proposed setup has
the potential for small- and medium-scale enterprises to avoid unplanned shutdowns
due to unexpected failure of bearings.
Acknowledgments Authors are thankful to the team members Pankaj Tiwari, Prachi Kaura,
Harshita Sharma, Muskan Goyal, Raunak Sharma, and Anamika Saini, all from UIET, Panjab
University, Chandigarh, India, for their support during the execution of this project.
References
1. X. Cao, P. Jiang, Development of SOA based equipments maintenance decision support system,
ed. by C. Xiong, et al., in International Conference on Intelligent Robotics and Applications.
Lecture Notes in Computer Science (Springer, Berlin, Heidelberg, 2008), pp. 576–582. https://
doi.org/10.1007/978-3-540-88518-4_62
2. D. Goyal et al., Non-contact sensor placement strategy for condition monitoring of rotating
machine-elements. Eng. Sci. Technol. an Int. J. 22(2), 489–501 (2019). https://doi.org/10.1016/
J.JESTCH.2018.12.006
3. D. Goyal et al., Optimization of condition-based maintenance using soft computing. Neural
Comput. Appl. 28(Suppl 1), S829–S844 (2017). https://doi.org/10.1007/s00521-016-2377-6
Intelligent Monitoring of Bearings Using Node MCU Module 177
4. D. Goyal, B.S. Pabla, Development of non-contact structural health monitoring system for
machine tools. J. Appl. Res. Technol. 14(4), 245–258 (2016). https://doi.org/10.1016/J.JART.
2016.06.003
5. D. Goyal, B.S. Pabla, The vibration monitoring methods and signal processing techniques for
structural health monitoring: a review. Arch. Comput. Methods Eng. 23(4), 585–594 (2016).
https://doi.org/10.1007/s11831-015-9145-0
6. O. Hassin, et al., Monitoring mis-operating conditions of journal bearings based on modulation
signal bispectrum analysis of vibration signals, in First Conference on Engineering Sciences
and Technology (Libya, 2018), pp. 509–517. https://doi.org/10.21467/proceedings.4.18
7. T. Hoshi, Damage monitoring of ball bearing. CIRP Ann. 55(1), 427–430 (2006). https://doi.
org/10.1016/S0007-8506(07)60451-X
8. C. Lazar, et al., Vision-guided robot manipulation predictive control for automating manu-
facturing, ed. by T. Borangiu, et al., in Service Orientation in Holonic and Multi-Agent
Manufacturing and Robotics. Studies in Computational Intelligence (Springer, Cham, 2014),
pp. 313–328. https://doi.org/10.1007/978-3-319-04735-5_21
9. A.H. Pesch, P.N. Scavelli, Condition monitoring of active magnetic bearings on the internet of
things. Actuators 8(17), 1–13 (2019). https://doi.org/10.3390/act8010017
10. S.V. Shelke et al., Condition monitoring of ball bearing using vibration analysis and feature
extraction. Int. Res. J. Eng. Technol. 3(2), 361–365 (2016)
11. Z. Yang, et al., A study of rolling-element bearing fault diagnosis using motor’s vibration
and current signatures, in 7th IFAC Symposium on Fault Detection, Supervision and Safety of
Technical Process (IFAC, Barcelona, Spain, 2009), pp. 354–359. https://doi.org/10.3182/200
90630-4-ES-2003.0307
12. F. Zhao et al., SOA-based remote condition monitoring and fault diagnosis system. Int. J. Adv.
Manuf. Technol. 46(9–12), 1191–1200 (2010). https://doi.org/10.1007/s00170-009-2178-5
Image Denoising Using Various Image
Enhancement Techniques
S. P. Premnath and J. Arokia Renjith
Abstract The main aim of the image enhancement technique is to process any image
given as input and to obtain the resultant outcome more accurately than the existing
image. The level of accuracy of an image can be restored in different forms using
image enhancement techniques. The choices of choosing different image enhance-
ment techniques may vary depending upon the quality of the picture, task, and atmo-
spheric conditions. The different algorithms used for enhancement and the concepts
are discussed in this paper. The frequency and spatial domain of the image can be
enhanced by using these processes.
Keywords Image enhancement ·Spatial domain enhancement ·Frequency
domain-based enhancement ·Histogram equalization
1 Introduction
Image enhancement is the most predominant and not easy technique in image
research. The image enhancement technique intends to provide a high-quality clear
image such as satellite images, real-life photographs, medical images experiencing
low disparity, and unpleasant noise. To increase the quality of an image, it is important
to enhance the contrast and remove the noise (Fig. 1).
Different techniques are available to refine the standard of the digital image
without diminishing it. The two broad categories, which are utilized to intensify
the clarity of the image are as follows:
(a) Spatial Domain
S. P. Premnath (B)
Department of Electronics and Communication Engineering, Sri Krishna College of Engineering
and Technology, Coimbatore, India
e-mail: premnathsp@yahoo.com
J. Arokia Renjith
Department of Computer Science and Engineering, Jeppiaar Engineering College, Chennai, India
e-mail: arokiarenjith@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_16
179
180 S. P. Premnath and J. Arokia Renjith
Fig. 1 Enhancing technique
(b) Frequency Domain.
The spatial domain technique defines that the picture or an image is composed
of the grouping of pixels. In this method, the quality of an image can be enriched
by directly operating on the pixels of an image [1]. The expression for the spatial
domain process can be given as follows:
g(x,y)=T[f(x,y)]
where g(x, y) denotes the given input image, in this gindicates the gray-level value
and the coordinates are denoted by (x,y).Tbe the transformation applied on the
input image, to produce a newly improved image f(x,y). The two different ways in
which the image can be spatially enhanced are as follows:
(a) Point Processing
(b) Neighborhood Processing.
The next approach to improve the quality of the image is based on the frequency-
domain method. In this approach when an image is given as input, the first stage
is to convert the input image into a frequency domain, during the conversion of
the image into the frequency domain, Fourier transform of the image is quantified
first [2]. In order to retrieve the resultant image, the inverse Fourier transform is
performed over Fourier transform. In order to enrich the clarity of an image, the
gain frequency-domain method is categorized into three types. They are as follows:
Image smoothing, Image Sharpening, and Periodic noise reduction (Fig. 2).
To enrich the quality of an image (i.e., based on image brightness, contrast, etc.),
the above-said image enhancing techniques are performed. The resultant image will
Fig. 2 Effect of image
enhancement
Image Denoising Using Various Image Enhancement Techniques 181
be more improved in quality due to the application of transformation function in the
given input values.
2 Point Processing Operations
Point processing operation is one of the simplest spatial domain operations that
deals with the individual pixel intensity values. The intensity values are altered using
transformation techniques as per the requirement [3]. Gray level at any point of the
image plays an important role in magnification of the image at any point (Fig. 3).
Point processing can be represented as follows:
S=T(r)
where Srepresents the pixel value after processing and rdenotes the original pixel
value of an image.
(A) Negative Transformation
The primary and simple operations in image processing are to compute the negative
image. The Gray (or) White features deeply in a surrounding mass of an image in the
darkish point can be magnified through negative images. The negative transformation
can be defined by
S=(L−1)−r
Fig. 3 Some basic
gray-level transformation
function
182 S. P. Premnath and J. Arokia Renjith
Fig. 4 Effect of negative transformation
Fig. 5 Identity
transformation
representation
Image intensity level in negative transform lies between in the range of [0, L−
1]. L −1 denotes the maximum pixel values; ris the pixel value of an image [4]
(Fig. 4).
(B) Identity Transformation
In identity transformation, the point of the pre-image and the image pixel is same,
the point is united (Fig. 5).
If all the points of pre-image and image are same, the entire figure is united.
(C) Log Transformation
In certain cases, the effective variation between upper and lower limits of an image
may be greater in size than the desired device capacity. Due to the variation in the
image, low-pixel values get masked. To overcome this, effective variation needs to
be compressed, i.e., log transformation method enlarges the values of darkish pixels
and compresses the bright pixel [5]. The high quality of an image compression ratio
can be obtained by the log operator. So, this log operator is used to compress the
effective range of an image. Log transformation can be given as follows:
S=C.log(1+|r|)
Normalization constant can be denoted by Cand rdenotes the input intensity
(Fig. 6).
Image Denoising Using Various Image Enhancement Techniques 183
Fig. 6 Log transformation
representation
The shape of log curve indicates the variation of low gray-level pixel values in
the input image to a bright value in the outcome and vice versa.
(D) Contrast Stretching
Due to a lack in lighting effect, effective range (or) improper line aperture conse-
quence in the formation of low-contrast images. In order to magnify the image, the
effective range of the gray level needs to be enhanced. So, the contrast stretching
technique is used to enrich the quality of an image [6]. Contrast stretching can be
represented as follows:
k=1
1+(p+r)z
where rrepresents the values of a given input image, kdenotes the value of the
resultant image, and pbe the thresholding value and be the slope (Fig. 7).
The above diagram shows the outcome of the variable z. If z=1, then stretching
becomes edge transformation. If z>1, then the range of intensity value of an image
is defined by the curve which is smoother, and when z<1, the transformation makes
the negative and also stretching.
Fig. 7 Contrast stretching
Bright Dim
n n
Bright Dim
Bright DimBright Dim
184 S. P. Premnath and J. Arokia Renjith
(E) Intensity-Level Slicing
Intensity-level slicing method is used to represent the specific range of the image.
For example, if the figure does not contain the values below 40 (or) above 255, then
the contrast of the image decreases [7]. If it is remapped by increasing the brightness
of the image in the range of [0, 255], then the contrast stretching of the image can
be increased.
Figure 8a shows the variation features a range of gray levels and makes other
range levels as a less desirable state. Figure 8b shows the variation features range
and it maintains a further level of an image.
v=L,a≤u≤b
0,otherwise v=L,a≤u≤b
u,otherwise
(F) Bit Plane Slicing
This transformation is useful in determining the number of visually significant bits
in an image. Consider a 256*256 image with 256 gray levels. Suppose each pixel is
represented by 8 bit, it is desired to extract the nth most significant bit (Fig. 9).
In this transformation, higher order bits contain visually sufficient data and low-
order bits contain suitable details of an image.
Fig. 8 Specific range of gray level
Image Denoising Using Various Image Enhancement Techniques 185
Fig. 9 Bit plane slicing
(G) Power Law Transformation
The quality of any image can be retained by applying power law transformation.
Power law transformation can also be called as Gamma Correction. It can be used
to rectify the errors that occurred while taking the photo and processing the image
further.
The power law transformation can be given as follows:
f(x,y)=c.g(x,y)µ
s=c.rμ
where cand μare the non-negative constants. Figure 10a is the real image, for a
positive constant c=1, then the positive γis 3, 4, 5, respectively, for the image B,
C, and D.
186 S. P. Premnath and J. Arokia Renjith
Fig. 10 Gamma correction
3 Neighborhood Pixel Processing
Neighborhood pixel processing technique is one of the spatial domain approaches
to enrich the quality of an image. In this technique, one pixel is considered at a
time and it is modified accordingly. The adjacent pixel values are also taken into
consideration. Pixel value can be changed based on 8 neighbors [1]. The quality of
the image can be increased by using neighborhood pixel processing when compared
with point processing.
Figure 11a shows the original image and Fig. 11b after filtering the image.
(A) Spatial Averaging And Spatial Low-Pass Filtering
Spatial averaging method is used here to reduce the noise produced by the isolated
and random pixels. The pixel value by range from small to high range value. In order
to retain the image, the small range values are weighted to the average of adjacent
pixel values.
Image Denoising Using Various Image Enhancement Techniques 187
Fig. 11 a,b
Ts(s,z)=
(i,j)∈l
a(i,j)b(s−i,z−j)
where b(s, j) represents the output image and T(s, z) represents the input image. W
is the windowing technique, a (i, j) is the filter weights. Low-pass filtering technique
is used to reduce the noise present in the pixels.
(B) Directional Smoothing
To protect edges from blurring while smoothing, spatial averages are calculated in
several directions, and the direction giving the smallest changes before and after
filtering is selected.
s(k,r:θ)=1
Nθ
(U,V)∈Zθn(k−u,r−v)
The direction θis found in a manner such that |n(k,r)−s(k,r:θ|is minimum.
(C) Directional Smoothing
Generally, there is a desire to enlarge the specified location of the image [8]. This
requires taking an image and displaying it as a larger image.
•Zero-Order Hold: Zero order is performed by considering the pixel values of
an image, i.e., the previous pixel values are also used repeatedly to perform this
function. In this, (n*n) image size can be varied to (2n)*(2n). Then convolve with
188 S. P. Premnath and J. Arokia Renjith
H=11
11
K (i, j) =c (x, y) x and yare given as int m
2, respectively.
The straightline is fitted between pixels along rows and columns. It can be obtained
by interlacing the image by rows and columns of zero and convolve with
M=⎡
⎣
1
4
1
2
1
4
1
211
2
1
4
1
2
1
4
⎤
⎦
Image Denoising Using Various Image Enhancement Techniques 189
4 Histogram Processing
Histogram processing is the act of reconstruction of an image by modifying its
histogram [9]. The gray of a histogram digital image lies in the range of [0, L−1].
The discrete function can be given as follows:
h(γk)=nk
where rkdenotes the gray level of an image and nkdenotes the pixels of an gray-level
image rk. Histograms regularly produce the new pixel values of an image by shifting
the values of the existing image. The normalized histogram is given by, p(γk)=nk
n,
where K=0,1,…,L−1.
(A) Histogram Equalization
Histogram equalization is a simple method used in enhancing the image. In the
enhancing processes, histogram equalization is used to reduce the gray level pixel
values, so that the brightness level of an image can be enhanced. The resultant image
will be more accurate.
(B) Frequency-Domain Techniques
In frequency-domain approach, Fourier transform of an image can be determined first
by converting an image to a frequency domain. In order to retrieve the desired quality
of an image in the resultant, the intensity pixel values of the image are modified by
taking inverse Fourier transform (Fig. 12).
Fig. 12 Frequency domain filtering operation
190 S. P. Premnath and J. Arokia Renjith
The transfer function can be given as follows:
I(a,b)=T(a,b)F(a,b)
where I(a, b) is the enhance image, T(a, b) represents the filter function, and F(a, b)
is the original image before processing. Since it depends on the frequency-domain
approach, parameters such as high-frequency components can be enhanced easily.
(C) Low-Pass Filtering
The simple filtering techniques used in image enhancement are low-pass filtering.
The low-pass filter reduces noise content present in the image. Noise present in each
pixel may vary and it affects the quality of the image. Enabling low-pass filter, it
allows selectively smooth image. Due to this, resultant image obtained is a quality
image [10]. The cutoff frequency of the low-pass filter can be represented as follows:
H(i,j)=⎧
⎨
⎩
1,ifi2+j2≤r0
0,ifi2+j2≥γ0
The cutoff frequency γ0of the ideal low-pass filter denotes the amount of noise
gets suppressed. Contrast and blurring of the image can be reduced when the value
of the γ0tends to reduce.
(D) High-Pass Filtering
Image enhancement can be performed by various image processing tasks. Among the
various filtering techniques, high-pass filtering makes a more smoothened image on
resultant by allowing the high frequency to pass and attenuates the low frequencies.
H(i,j)=⎧
⎨
⎩
1,ifi2+j2≤r0
0,ifi2+j2≥γ0
5 Conclusion
Image enhancement algorithms offer various enhancing approaches for retrieving the
images to obtain quality as well as clarity of an image. These image enhancement
techniques in spatial domain and frequency domain have been successfully discussed.
Depending on the quality of the image and noise with which it is corrupted by varying
the methods, visual quality can be improved. The computational cost of enhancement
algorithms is not discussed here. It plays a major responsibility in picking accurate
Image Denoising Using Various Image Enhancement Techniques 191
algorithms for different conditions. The combination of the above algorithms can be
used to produce a more enhanced image.
References
1. R. Maini, H. Aggarwal, A Comprehensive Review of Image Enhancement Techniques
2. A.K. Jain, Fundamentals of Digital Image Processing (Prentice Hall, Englewood Cliffs, NJ,
1989)
3. S.M. Pizer et al., Adaptive histogram equalization and its variations. Comput. Vis. Graph.
Image Process. 39, 355–368 (1987)
4. R.C. Gonzalez, R.E. Woods, Digital Image Processing (Prentice Hall, Upper Saddle River,
New Jersey, 2002)
5. R. Jain, R. Kasturi, B.G. Schunck, Machine Vision (McGraw-Hill International Edition, 1995)
6. H. Lidong, Z. Wei, W. Jun, S. Zebin, Combination of Contrast Limited Adaptive Histogram
Equalisation and Discrete Wavelet Transform for Image Enhancement (2014)
7. H.C. Andrews, Digital image restoration: a survey. Computer 7(5), 36–45 (1974)
8. Y. Yao, B. Abidi, M. Abidi, Digital Imaging with extreme Zoom: System Design and Image
Restoration
9. R. Hummel, Histogram modification techniques. Comput. Graph. Image Process. 4, 209–224
(1975)
10. G.L. Anderson, A.N. Netravah, Image restoration based on a subjective criterion. IEEE Trans.
Syst. Man Cybern. SMC-6, 845 (1976)
Energy Consumption Analysis
and Proposed Power-Aware Scheduling
Algorithm in Cloud Computing
Juhi Singh
Abstract Cloud computing is a large-scale paradigm using remote servers hosted
on the internet. Servers over the data centers provide services in the form of SaaS,
PaaS, and IaaS. The concept of cloud computing is the virtualization of resources
that provides services. Cloud is an infinite resource pool along with the availability of
these resources to multiple users. The energy consumption in a cloud environment is
due to the number of parameters that have fixed values or it is variable. In the paper,
we had rectified a model for energy consumption and an algorithm is proposed for the
power awareness scheduling algorithm. Also, the paper shows a relationship between
workload, utility, power, and energy consumption is presented as a model.
Keywords Cloud computing ·Cost ·Energy consumption ·Green computing ·
Power ·Resource management ·Scheduling ·Utility ·Workload
1 Introduction
Cloud Computing is an environment that provides services on users’ requests as
per demand. From the consumer’s point of view, the cloud is a scalable service and
consumers can get as many as resources as per the usage. So, from the provider’s
point of view, it is important to manage resources with both points that it can serve
the consumer by providing resources in a better way and gets profit without compro-
mising the quality of services or violating the SLA. There are a number of parameters
taken into consideration to provide better services. Complications of these param-
eters may lead to a penalty, time limitation, unavailability, less utilization, etc. The
provider needs to first keep the track with the limitations of time and energy resources
availability. Also, services should be optimized with minimum cost.
Various resources like compute, infrastructure, applications, storage, network,
platform, and database are used in the cloud environment. The resources can be
classified into two categories.
J. Singh (B)
Faculty of Computer Science and Engineering, SRMU, Lucknow, UP, India
e-mail: juhisingh.srmcem@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_17
193
194 J. Singh
Physical Resources →Computer, Disk, Database, Network, Scientific Instru-
ments.
Logical Resources →Execution, Monitoring, Communicate applications.
These resources need to manage efficiently and in an optimized way over the
cloud environment. Cloud computing has concepts of virtualized resources. Resource
management is done with the objectives to reduce overhead, to reduce latency, and to
improve throughput. Resource management also includes scalability, optimal utility,
specialized environment, and cost-effectiveness. Several challenges of resource
management can be categorized into broad categories of hardware and software.
1. Resource Management—Challenges (Hardware resources)
•CPU (central processing unit)
•Memory
•Storage
•Workstations
•Network elements
•Sensors/actuators
2. Resource Management—Challenges (Logical resources)
•Operating system
•Network throughput/bandwidth
•Energy Consumption
•Load balancing mechanisms
•Protocols
•Delays
•Security of Information
•Application Programming Interfaces
Taking into consideration these points, the term resource management indicates
the methods or steps that control the capabilities provided by cloud resources.
Resource management also includes how services can be made available in the
best way to other entities; these entities could be users, applications, services in
an optimized and efficient manner [1]. It can be efficient in energy optimization,
maximizing in profit, efficient in providing the best quality or efficient with respect
to SLA. Cloud service providers have huge data centers such as Google, Microsoft,
Yahoo, Amazon, and IBM. These resources are outsourced from these providers. It
is calculated that the server consumes 0.5% of total electricity usage over the world
[2]. There are different approaches and methods done by researchers for resource
management.
Resource provisioning →Resources are assigned to the workload given.
Resource allocation →Resources are distributed among the workload
Resource Adaptation →Resources adjust themselves automatically to fulfill the
user’s workload requirements.
Energy Consumption Analysis and Proposed … 195
Resource Mapping →Resources that are required by the workload and resources
that are provided by the cloud infrastructure are co-related correspondingly.
Resource Modeling →A structure is designed to find the resource requirements
of a workload with attributes.
Resource Estimation →Approximate prediction of resources required is done for
the execution of workload.
Resource Discovery →Resources identified for execution that are available.
Resource Brokering →A third party agent negotiate the resources for the
assurance of resource availability.
Resource Scheduling →A proper schedule of resources and events is determined
to depend on the number of factors, i.e., its duration, resources allocated, predecessor
activities, and predecessor relationships [3].
Green Computing can be taken as advanced scheduling schemas to reduce energy
consumption, power-aware, thermally aware, data center designs to reduce Power
Usage Effectiveness, cooling systems rack design. There are various reasons to intro-
duce Green Data Centers for economic research with the aim to reduce costs, since
many facilities are at their peak operating stage and we need to reduce cost at
feasible availability of resources without expanding new power sources. Also, in
terms of energy, the majority of energy sources are fossil fuels and a huge volume of
CO2emitted each year from power plants. Thus, green computing is the advanced
technology that works on power awareness methodology (Fig. 1).
Fig. 1 Green cloud framework
196 J. Singh
2 Related Work
Management of resources over the cloud environmentaffects a lot on power consump-
tion so we need to use the resources more efficiently. Various scheduling algorithm
in the cloud environment is done for a heterogeneous system to optimize the cost by
using the resources more efficiently [4–6]. Data center plays a vital role in power
consumption. Dynamic methods are used for managing the utilization of energy and
resources over data centers [7–9]. Initially, the work was done for a single user to opti-
mize energy consumption in the research paper [7]. Further, the research continued
for power control by adjusting the frequency at the processing level [8]. Later, the
research work was done to find relation power and frequency for optimal power
allocation [9].
Further to make data centers eco-friendly and to save energy, a new technique
came into role, i.e., Green Computing. Moving toward optimal power consumption as
introduced in green computing advance research, work is proposed as Powernap, i.e.,
an approach of energy conservation, where the migration of servers is done between
active state and a near-zero power idle state [10]. A unique state of server called
“Somniloquy” is introduced to save power more effectively [11]. Li et al. proposed a
cost model based on total cost and utilization cost [12]. There was a limitation in the
work that the calculation was based on a single hardware component. So, on further
research work, Jung et al. focused on physical host power consumption [13]. The
work–energy consumption had no effect on workload and hardware specifically. Li
et al. introduced a new infrastructure for CPU utilization based on dynamic console
VMs and Verma et al. enhanced the work, where the characteristics of VMs were
used [14–16]. Various works have been done in terms of energy optimization and
power consumption [17–19] (Fig. 2).
Data Centre
Environmental Logistic (UPS,
AC, Setups)
Compute (Server, Storage,
Networks)
Fig. 2 Datacenter resources
Energy Consumption Analysis and Proposed … 197
3 Energy Consumption Analysis Model
In other research papers [20], it is discussed that total energy consumption is taken
as a summation of all fixed energy consumed and variable energy consumed referred
to as Efix and Evar . The total consumption defined as Etotal is shown by the formula
given below:
Etotal =Efix +Evar (1)
where, in the cloud environment, variable energy consumption can be classified in
storage, computation, communication resources, and other resources.
1. Estorage referred to as the energy consumption of storage resources or memory.
2. Ecomp referred to as the energy consumption of computation resources or servers.
3. Ecomm referred to as the energy consumption of communication resources or
networks.
4. Eother referred to as other resources as environmental logistic or UPS, AC, Setups.
Etotal =Efix +Estorage +Ecomp +Ecomm +Eother(2)
For task scheduling algorithm, we can count total energy consumption as energy
consumed by all tasks. And, energy consumed by all tasks is not just the addition
of all tasks. Scheduling overhead will generate energy consumption referred to as
Esche. So further for scheduled task total energy consumed can be taken as follows:
Etotal =Efix +Estorage +Ecomp +Ecomm +Eother+Esche (3)
where represents the summation of each task 1 to a number of task n. For each task,
energy consumption is tightly bound with workload, so we say energy consumption
is directly proportional to the workload of each task, and the energy consumption
increases with the number of processes in execution with the task [21].
4 Power Model
In concern of power consumption, data center are classified into two categories:
1. Compute
2. Environmental Logistic
To calculate energy, the information requires power efficiency and task resource
demands of each VM, as in general, power is defined as the energy consumed per
unit time. And, resource demand is taken static parameters to compile a task, i.e.,
198 J. Singh
size of data through a disk, number of instructions, size of data through network,
and job id where it is generated from. Contrast to this dynamic parameter is taken as
energy consumption and execution time of the task [22]. Power and energy of CPU
in the cloud are modeled as follows:
P(u)=k×Pmax +(1−k)×Pmax ×u(4)
where Pmax is the maximum power consumed when the server is fully utilized, u
is the CPU utilization, and k is the fraction of power consumed by the idle server.
The utilization of CPU changes with time since workload changes. Thus, the CPU
utilization is taken as a function of time which is represented by u(t). Therefore,
the total energy consumed by a physical host is taken as an integral of the power
consumption function over a period of time:
E=P(u(t)) (5)
Energy consumption by hosts in data centers includes that of disk storage, CPU,
and network interfaces. The dynamic power consumption of cloud data centers is
mainly generated by the workload on each running server, while the main sources
for power consumption are resource demands of tasks that drive server workloads.
From the above-derived formulas, we find that there is a tight bonding between CPU
utilization and power consumption of system [22].
5 Proposed Algorithm
Power-Aware Scheduling of VM
1. Initialize Buffer
2. Start the loop till resource pool and assign
processing elements.
for i =1 to i <=|pool| do
pei=num cores in pooli
end for
3. While queue is not empty calculate power
and energy
while(true)
for i =1to i <=|queue| do
vm =queuei
for j =1 to j <=|pool| do
if pej>=1 then
if check capacity vm on pejthen
Energy Consumption Analysis and Proposed … 199
schedule vm on pej
P(u) =k×Pmax +(1 −k) ×Pmax ×u
E=P(u(t))
pej–1
end if
end if
end for
end for
wait for internal t
end while
6 Cost and Penalty Cost
In the virtualization technique, physical resource is assigned a logical name, and
when there is a request made physical resource is provided a pointer. Resource
management helps to maintain SLA i.e., a contract between service provider and
service consumer. Resources are provided on availability. KPI of SLA are:
•Uptime
•Downtime
•Inbytes.
Cloud computing finds the availability on two parameters, namely service uptime
and service downtime with a formulation
Availability =1−(service downtime/service uptime)(6)
For example
Service uptime =128 min
Service downtime =32 min
Availability =1 – (32/128) =0.75
Cloud is cheaper when utility premium is less than the ratio of peak demand and
average demand. Since demand is an essential parameter related to utility, we can
calculate utility pricing details as follows:
CT =U×B×Dtdt
=A×U×T0B×TBT
=P×B×T(7)
While owning resources, there is a penalty whenever your resources do not match
the instantaneous demand, which leads to pay for unused resources, or suffer the
penalty of missing service delivery that is formulated as follows:
200 J. Singh
D(t)—Instantaneous Demand at time t
R(t)—Resources at time t
Penalty Cost α|D(t)−R(t)|dt
If demand is flat, penalty =0
If demand is linear periodic provisioning is acceptable.
7 Conclusions and Future Work
Energy consumption in the cloud computing model is analyzed for static and dynamic
consumption in the cloud environment. Dynamic consumption of energy at data
centers is optimized. The optimized energy consumption improves energy efficiency
according to workload at datacenters. The energy consumption model is analyzed
and derived the relationship between workload, utility, power, and energy consumed.
In this paper, the proposed power scheduling algorithm for a virtual machine in the
cloud environment is elaborated. Also, the papers represent the derivation of the
resource management parameters, i.e., availability and utility. Resource manage-
ment maintains the service-level agreement to reduce the penalty cost. Although a
set of parameters is considered to model energy consumption, there are a number of
environmental parameters that affect and can be taken to rectify energy consump-
tion. Further, the research directions include how to conserve energy within a cloud
environment more accurately.
References
1. http://www.cse.hcmut.edu.vn/~ptvu/gc/2012/GC-pp.pdf
2. A.J. Younge, G. von Laszewski, L. Wang, S. Lopez-Alarcon, W. Carithers, Efficient resource
management for cloud computing environments
3. N.M. Gonzalez, T.C.M. de Brito Carvalho, C.C. Gonzalez et al, Review open access cloud
resource management: towards efficient execution of large-scale scientific applications and
workflows on complex infrastructures. J Cloud Comput. Adv. Syst. Appl. (2017)
4. S. Abrishami, M. Naghibzadeh, Epema, Deadline-constrained workflow scheduling algorithms
for infrastructure as a service clouds. Futur. Gener. Comput. Syst. 29(1), 158–169 (2013)
5. A. Bala, I. Chana, A survey of various workflow scheduling algorithms in cloud environ-
ment, in IJCA Proceedings on 2nd National Conference on Information and Communication
Technology, Nagpur vol. NCICT (2011), pp 26–30
6. H. Arabnejad, J.G. Barbosa, List scheduling algorithm for heterogeneous systems by an
optimistic cost table. IEEE Trans. Parallel Distrib. Syst. 25(3), 682–694 (2014)
7. V. Sharma, A. Thomas, T. Abdelzaher, K. Skadron, Power-aware QoS management in web
servers, in Proceedings of the Real-Time Systems Symposium, 2003
8. X. Wang, M. Chen, Cluster-level feedback power control for performance optimization, in
Proceedings of the IEEE 14th International Symposium on High Performance Computer
Architecture, 2008
9. A. Gandhi, M. Harchol-Balter, R. Das, and C. Lefurgy, “Optimal Power Allocation in Server
Farms”, Proc. Of SIGMETRICS Conference on Measurement and Modeling of Computer
Systems, 2009
Energy Consumption Analysis and Proposed … 201
10. D. Meisner, B.T. Gold, T.F. Wenisch, PowerNap: eliminating server idle power. in Proceedings
of ASPLOS, 2009
11. Y. Agarwal et al., Somniloquy: augmenting network interfaces to reduce PC energy usage, in
Proceedings of NSDI, 2009
12. X. Li, Y. Li, T. Liu, J. Qiu, F. Wang, The method and tool of cost analysis for cloud computing,
in The IEEE International Conference on Cloud Computing (CLOUD 2009), Bangalore, India,
2009, pp. 93–100
13. G. Jung, M.A. Hiltunen, K.R. Joshi, Mistral: dynamically managing power, performance,
and adaptation cost in cloud infrastructures, in The International Conference on Distributed
Computing Systems, Genova, Italy, 2010, pp. 62–73
14. L. Liu, H. Wang, X. Liu, X. Jin, W. He, Q. Wang, Y. Chen, Greencloud: a new architecture for
green data center, in IEEE Conference on Autonomic Computing (ICAC), Barcelona, Spain,
2009, pp. 29–38
15. A. Verma, P. Ahuja, A. Neogi, Power-aware dynamic placement of HPC applications, in The
22nd Annual International Conference on Supercomputing (ICS 2008), Island of Kos, Aegean
Sea, Greece, 2008, pp. 175–184
16. R. Nathuji, K. Schwan, Virtualpower: coordinated power management in virtualized enter-
prise systems, in The 21st ACM Symposium on Operating Systems Principles (SOSP 2007),
Stevenson, Washington, USA, 2007, pp. 265–278
17. J. Stoess, C. Lang, F. Bellosa, Energy management for hypervisor-based virtual machines, in
The USENIX Annual Technical Conference (USENIX) 2007, pp. 1–14
18. Q. Chen, P. Grosso, K.v.d. Veldt, C.d. Laat, R. Hofman, H. Bal, Profiling energy consumption
of VMs for green cloud computing, in The 9th IEEE International Conference on Dependable,
Autonomic and Secure Computing (DASC), Sydney, Australia, 2011, pp. 768–775
19. A. Kansal, F. Zhao, N. Kothari, A.A. Bhattacharya, Virtual machine power metering and provi-
sioning, in The 1st ACM Symposium on Cloud Computing (SoCC 2010), Indianapolis, Indiana,
USA, 2011, pp. 39–50
20. F.F. Chen, J.-G. Schneider, Y. Yang, J. Grundy, Q. He, An energy consumption model and anal-
ysis tool for cloud computing environments, in 1st ICSE Workshop on Green and Sustainable
Software Engineering, Zurich, Switzerland, 3rd June 2012
21. Z. Zhang, S. Fu, Characterizing power and energy usage in cloud computing systems, in The
3rd IEEE International Conference on Cloud Computing Technology and Science (CloudCom
2011), Athens, Greece, 2011, pp. 146–153
22. P. Mathiyalagan, M. Rajshree, S. Shobika, M. Sowmiya, An efficient dynamic scheduling based
on greedy algorithm for heterogeneity clouds, Int. J.Adv. Res. Comput. Commun. Eng. 6(3),
2017. ISO 3297:2007 Certified
Relationship Between Community
Structure and Clustering Coefficient
Himansu Sekhar Pattanayak, Harsh K. Verma, and A. L. Sangal
Abstract Overlapping community is a phenomenon often observed in numerous
real-world networks. Fire-spread (Pattanayak et al. in Swarm Evol Comput. 44: 1–48
(2019) [11]) community detection algorithm is an efficient algorithm to detect over-
lapping community structures. In this work, the Fire-spread algorithm is modified
to establish a relationship between community structure and clustering coefficient.
By using different networks and executing the modified Fire-spread algorithm, it is
found that the clustering coefficient is highly correlated with community structure.
Finally, a simpler community detection algorithm, derived from the fire-spread algo-
rithm, is proposed, where the clustering coefficient is used as a threshold value. To
validate the proposed algorithm, it is compared with some state of art community
detection algorithms based on the NMI score.
Keywords Community ·Detection ·Complex network ·Probabilistic computing ·
Clustering ·NMI score ·Fire-spread algorithm
1 Introduction
Real-world networks vary in sizes and structures. These networks include both natu-
rally occurring networks and human-made networks. In a network, there are few
actors (participants) and the interaction between those actors [1]. The actors are
generally represented as nodes of the network and interactions are represented as
edges. The edges can be directional or non-directional. A social network such as a
friendship network, the edges are non-directional whereas an email network [2]
H. S. Pattanayak (B)·H. K. Verma ·A. L. Sangal
Department of CSE, NIT Jalandhr, Jalandhr, India
e-mail: himansusekharpattanayak@gmail.com
H. K. Verma
e-mail: vermah@nitj.ac.in
A. L. Sangal
e-mail: sangalal@nitj.ac.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_18
203
204 H. S. Pattanayak et al.
representing email communication between the actors is necessarily directional.
Similarly, edges of a network can either be weighted or unweighted. Edges of a
friendship network can be unweighted, whereas edges of a co-purchase network
[2] are weighted. Networks include social groups, families, villages [3,4], World
Wide Web (web pages containing similar topics are connected more densely among
themselves) [4], biological networks [5].
The real-world networks have been studied extensively by the researchers. These
networks exhibit scale-free [6] and community structure [7] properties. Commu-
nity structure is an organization of nodes where some of the nodes are tightly
connected with each other compared to the outside nodes [7,8]. Community detec-
tion techniques are used for the detection of suspicious events in the telecom network,
detection of terrorist networks, refactoring of software, lung cancer detection [9].
Community detection is the process of identifying the functional groupings
without any other information except for the structural information. In community
detection, usually, the Modularity [9] and Conductance [10] scores are used. As
described in [10], the clustering coefficient is not used as a parameter in commu-
nity detection. However, the clustering coefficient has a very strong co-relation with
community structure, which will be discussed in the following sections.
The organization of the paper is given in the following order. A literature survey
related to the subject is presented in Sect. 2. The proposed algorithm, which is a
simplified version of Fire-spread community detection algorithm [11], is discussed
in Sect. 3. Implementation of the proposed algorithm and datasets are described in
Sect. 4, a comparison of the proposed algorithm with other community detection
algorithms is presented in Sect. 5and the paper is concluded in Sect. 6.
2 Related Work
A brief review of the related research and literature is presented in this section.
Some earliest efficient community detection algorithms are based on optimization
of modularity score [12,13]. The algorithm proposed in [14] is based on a similar
technique. However, it is capable of detecting hierarchical communities. Similar
hierarchical community detection algorithms, which use an incremental method by
combining sub-graphs, are presented in [15–19].
The authors of [20] have the set of weights for each node of the network and
have applied the K-means clustering algorithm. However, the value of k, i.e., the
number of communities needed to be predefined. The proposed algorithm in [21], is
an enhancement over the previous work, in which the authors have used a nonlinear
Lagrangian programming model that reduces the complexity of the previous algo-
rithm. The authors have used the K-means technique to get an approximate solution
and then refined the solution using simulated annealing in [22]. The community
detection algorithm in [23] is also relied on K-means clustering.
Relationship Between Community Structure and Clustering … 205
The work described in [24] has applied a supervised learning technique. The
algorithm needs prior knowledge about the network. The authors in [25]haveused
unsupervised division algorithm and verified their technique using 408 networks.
The authors of [26] have used multimodal optimization based on the firefly algo-
rithm. The work described in [27], is an enhancement of the previous work by adding
a multi-objective Pareto-optimal solution to the original. The authors of [28]have
used PSO (particle swarm optimization technique) for community discovery which
is efficient even for isolated nodes. There are some other community detection algo-
rithms based on swarm intelligence described in [29,30]. The authors of [31,32]have
used genetic algorithms with multi-objective solutions for community detection. The
research presented in [33], uses ant colony optimization technique by defining each
node as an ant. The algorithm proposed in [11], is a novel nature-inspired community
detection algorithm that relies on local community detection using the Fire-spread
algorithm.
The authors of [34] have proposed a community detection algorithm based on
Integer programming and computation of the capacity of different communities. The
author of [35] has designed a community detection algorithm that can able to work
in a noisy environment and responsive to the dynamic behaviour of the community
structures.
3 Proposed Work
The research work presented in this paper is divided into two parts; (1) study on the
relationship between community structure and clustering-coefficient, (2) designing
a simple and efficient community detection algorithm derived from the Fire-spread
algorithm [11]. The original Fire-spread algorithm is a community detection algo-
rithm, based on the propagation of fire in real scenarios. It takes its inspiration from
naturally occurring fire. Neighbouring points of the source of fire are always at a
higher risk of catching fire. We have modified the fire-spread community detec-
tion algorithm to simplify its application and investigate the relationship between
community structure and clustering coefficient [10].
(A) Fire-spread Community Detection Algorithm
The probability to catch fire (Pi) at a node iis proportional to fire transmission
coefficient pi. The fire propagation probability between node i and jis denoted as
Pi|j and is defined as follows:
pi|j=1
G·degree(j)(1)
where jis a neighbour node of iin graph G.
206 H. S. Pattanayak et al.
The total probability of catching fire at node ican be defined in terms of sum of
the fire transmission coefficient of all the neighbours.
Pi=P1∗pi|1+P2∗pi|2+P3∗pi|3... +Pj∗pi|j+··· (2)
where P1,P2,P3… and Pjare the probabilities of catching fire at 1st, 2nd, 3rd and
jth neighbours, respectively.
The algorithm starts with a seed node. An R-radius subgraph (g) containing all
the neighbours of seed which are less than or equal to R-edge distance from seed
are selected to localize the search of community nodes. Initially, the fire catching
probability of seed is set to 1.0, and for other nodes, it is set to 0. In each step, the
fire catching probability (Pj) of any node j, with an edge (i,j) form node i, is updated
as Pj=Pj+Pi∗pi|j. The number of steps required for the algorithm is R,asthe
maximum distance from the seed node to any other node is R.
After the completion of the updating of the fire catching probability for all the
nodes, the nodes are compared with threshold values. Nodes having greater fire
catching probability than the threshold are selected as member nodes along with the
seed node. The value of threshold varies for each node depending on the distance
from seed node, average node density, degree of the node and number of nodes of
the subgraph g. This process is continued repeatedly until no node is left without
being a part of any community. The seed nodes are the nodes that are not part
of any communities discovered so far. This algorithm is efficient for overlapping
community detection as it allows a node to be part of more than community does.
The disadvantage of the algorithm is the complexity in calculating the threshold
values for each node. Therefore, a simpler version of the algorithm is proposed, in
which the threshold values are calculated, based on the clustering coefficient.
(B) The R-radius Neighbourhood
Let, two nodes Aand Bare part of community Cwhich has nnodes. Let, Ahas
node degree d1and Bhas node degree d2. Since both Aand Bbelong to the same
community C, the probability that both nodes connected to each other by an edge
can be calculated.
Probability of immediate neighbour =
n−2Cd1−1−n−2Cd2−1/n−1Cd1+n−1Cd2(3)
Probability that both nodes are connected through a single intermediate node=
(d1+d2)
(n−2)(4)
Probability that both nodes are connected with 2-radius neighbourhood (P)=
Probability of immediate neighbour +Probability that both nodes are connected by
single intermediate node =
Relationship Between Community Structure and Clustering … 207
n−2Cd1−1+n−2Cd2−1/n−1Cd1+n−1Cd2+(d1+d2)
(n−2)(5)
Without losing generality, it is assumed that d1=d2=d=average node degree
inside community C. So, replacing din Eq. (5)
P=n−2Cd−1+n−1Cd−1/n−1Cd+n−1Cd+2d
(n−2)(6)
=(n−2)!
(d−1)!∗(n−d−1)!∗(n−d−1)!∗d!
(n−1)!+2d
n−2=d
(n−1)+2d
n−2(7)
P=d
(n−1)+2d
n−2≥d
(n−1)+2d
n−1=3d
(n−1)≥3d
n(8)
If d≥n
3then P≥1
Therefore, it can be concluded that it is highly probable that any two nodes of
the community Cwill be in 2-radius neighbourhood, if the average degrees of nodes
inside Care greater than or equal to n/ 3. Therefore, for dense graphs, the R-value
can be set 2. However, depending on the density of the graph the value of Rmay be
higher.
C. The Proposed Algorithm
In the original FS algorithm, the fire propagation probability and threshold values are
variable depending upon nodes and their edges. However, in the proposed algorithm
the fire propagation probability is kept constant for all the edges. The threshold
value is also kept constant for each node. By changing the threshold values from
0.0 to 1.0, with a fixed fire propagation probability value, the recall and precision
are calculated. The best values of recall and precision corresponding to the chosen
threshold value are recorded and compared with the proposed algorithm with the
clustering coefficient as a threshold. For this experiment, the network with a known
community structure is available.
4 Implementation
The Fire-spread working of algorithm is consisting of two parts. The first algorithm
is a local fire-spread algorithm that discovers a community around a seed node.
The second part of the algorithm is the community detection algorithm which selects
random seed nodes, calls the local Fire-spread algorithm and divides the network into
overlapping communities. For all the networks, the fire propagation probability is set
to 0.2. In this section, the original Fire-spread algorithm analysed by modifying it to
work with fixed propagation probability and a varying threshold value. Community
membership of all the datasets are already available, hence we are able to calculate
208 H. S. Pattanayak et al.
the recall and precision values by executing local community detection algorithm
with specific seed nodes. The best values of recall and precision will be used to
investigate the effect of community structure on the clustering coefficient, which
will be subsequently used to set a fixed threshold value for the proposed modified
algorithm.
(A) Dataset:
The datasets that are used for the experiments are Zackary’s Karate Club network
[36] (Fig. 1). Bottlenose Dolphins network [37] (Fig. 2), American Political Book
network [38] (Fig. 3), Adjective–Noun Network [39] (Fig. 4), Relaxed Caveman
Network [40] (8 clusters containing 10 nodes each with 0. 1 being the probability of
inter-cluster edge, Fig. 5) and Shared Community Network (A synthetic network with
5 communities with 20 nodes in each community, 25% nodes of each community
are shared with other communities, Fig. 6).
Fig. 1 Books on American
politics network graph
Fig. 2 Adjective–noun
network graph
Relationship Between Community Structure and Clustering … 209
Fig. 3 Dolphin network
graph
Fig. 4 Karate club network
graph
Fig. 5 Relaxed Caveman
N/W graph
210 H. S. Pattanayak et al.
Fig. 6 Shared community
graph
(B) Recall and Precision
The following formulas are used for calculations of precision and recall.
Recall(C)=No.of Nodes of Community Cfound by Algorithm
Actual no of nodes of commmunity (9)
Precision(C)=No.of Nodes of Community Cfound by Algorithm
No of nodes found as community by the Algorithm (10)
The probability value is fixed at 0.2 for all the networks and we calculate the value
of precision and recall for all threshold ranging from 0.0 to 1.0 with 20 different
values. The precision and recall values against the different threshold values are
plotted in the following graphs. The six plots are recall and precision values for six
datasets. The aim is to find the optimal threshold value which will maximize both
recall and precision. The American political book graph is shown in Fig. 7. The best
value of threshold for this network is 0.15 with recall =0.93 and precision =0.95.
Figure 8represents the adjective–noun network. The best value of recall =0.97
and precision =1.0 is found at a threshold value of 0.30. For the Dolphin network
plot given in Fig. 9, the performance is best with recall =0.94 and precision =0.94
at threshold =0.05. Figure 10 is a plot for Karate club network. The FS algorithm
performs best when threshold =0.2. Recall and precision values are found to be
0.89 and 0.85, respectively. As it can be seen in Fig. 11, FS algorithm performs best
with recall =1.0 and precision =1.0 when the threshold value is set to 0.35 for the
Relaxed Caveman network. The shared network plot is shown in Fig. 12. The FS
algorithm has best recall =1.0 and precision =1.0 when the threshold issetat0:2.
For most of the networks, best values of recall and precision are obtained with
threshold =0.2 expect for the Dolphin network where the best threshold is found to
be 0.05. The best values of thresholds found are used for further experimental works.
By observing all the six networks, it can be concluded that with fixed probability
Relationship Between Community Structure and Clustering … 211
Fig. 7 Books politics
network
Fig. 8 Adjective–noun
network
Fig. 9 Dolphin network
graph
212 H. S. Pattanayak et al.
Fig. 10 Zachary’s Karate
club graph
Fig. 11 Relaxed Caveman
graph
Fig. 12 Shared community
network
Relationship Between Community Structure and Clustering … 213
Tabl e 1 Threshold value of networks using clustering coefficient
Network graph Avg node degree Clustering coefficient 2.0Clustering Coefficient
Avg Node degree
Political book 8.204 0.515 0.1254
Noun–adjective 8.0625 0.195 0.0482
Dolphin 4.18 0.325 0.1554
Karate 4.75 0.529 0.2228
Relaxed Caveman 90.71 0.1578
Shared 4.222 0.783 0.3704
value; an increase in threshold results in a decrease in recall but an increase in
precision. Setting higher threshold value results in the detection of high-resolution
hierarchical structures. Lower threshold values result in the detection of communities
with a higher degree of sharing of nodes.
For calculating threshold, we have used the local clustering coefficient which is
a measure to quantify how close a node along with its neighbour to be a clique. To
calculate the local clustering coefficient for a node ithe following formula is used
Clustering Coefficient =2∗
ejk
|Ni|∗(|Ni|−1)(11)
where |Ni|=number of neighbour nodes of node iand ejk is set of all possible
pairs of nodes such that j,kare neighbours of i. The subgraph, which has a higher
clustering coefficient with a lesser average node degree, has a higher threshold value
than a subgraph with a lesser clustering coefficient and higher average node degree.
Table-1shows the threshold value for the networks based on clustering coefficients.
Threshold =constant ∗Clustering Coefficient
Avg Node degree (12)
The constant =2.0 is set for all the networks. We have experimentally set the
value after repeated experiments.
The recall and precision of the proposed algorithm are compared with the best
recall and precision of the original algorithm.
C. Local Fire-Spread Algorithm:
Input: Graph G,fire propagation probability p,radius R
For each node in G
If node does not belong to any community
Use node as seed node
Find Ego-Graph for the node with radius R
Find threshold =(2*Clustering Coefficient)/Average node degree
214 H. S. Pattanayak et al.
Call local Fire spread Algorithm with input g,p&t
Store the list of nodes returned by the Local Fire Spread algorithm.
return list of overlapping communities
Fire-Spread Algorithm:
Input: R-radius subgraph g,fire propagation probability p,Threshold t, seed s
Initialize each node in g
Initialize seed Probability to catch fire Ps=1.0
Initialize nodes other than seed with Pi=0.0
For each node in g
For i∈Neighbor (node)
Update Pi=Pi+Pnode∗p
Update Pnode =Pnode +Pi∗p
For each node in g
Check if a node belongs to the community using threshold t
return list of community nodes
D. Comparison of Proposed Algorithm with Fire-Spread Algorithm
The proposed algorithm is compared with the Fire-spread algorithm in terms of
precision and recall and presented in Table 2. The best precision and recall values
are taken for the Fire-spread algorithm.
As it can be seen from Table 3, the performance of the proposed algorithm and the
original algorithm are same for the synthetic networks with recall and precision values
being 1.0. For the real-world networks, both the algorithms perform very similarly
baring a few cases. The above experimental analysis is used to set a simpler threshold
value for the network based on local clustering coefficient.
Tabl e 2 Comparison between proposed algorithm and fire-spread algorithm
Network graph Fire-spread algorithm Proposed algorithm
Recall Precision Recall Precision
Political book 0.9268 0.95 0.8536 1.0
Noun–adjective 0.9726 1.0 1.0 0.9634
Dolphin 0.9473 0.9473 0.9473 1.0
Karate 0.8889 0.8421 1.0 0.8571
Relaxed Caveman 1.0 1.0 1.0 1.0
Shared 1.0 1.0 1.0 1.0
Relationship Between Community Structure and Clustering … 215
Tabl e 3 NMI score: books on American politics
Algorithms Max Mean Min Standard deviation
Proposed algorithm 0.6005 0.5996 0.5982 0.0005
Enhanced firefly 0.5953 0.5927 0.5915 0.0004
Firefly 0.5829 0.5819 0.5803 0.0007
Blondel 0.5954 0.5937 0.5924 0.0012
Ronhovde-Nussinov 0.5948 0.5941 0.5878 0.0029
Clauset-Newman 0.5427 0.5386 0.5304 0.00327
5 Results and Discussion
In this section, NMI (Normalized Mutual Information) score is used for all the
networks and the proposed algorithm is compared with some well-known algo-
rithms, such as Enhanced Firefly algorithm [26], Firefly algorithm [27], Blondel
algorithm [12], Ronhovde–Nussinov [14] algorithm and Clauset–Newman [13]. All
the experiments are repeated ten times for accuracy.
For any two clusters C1and C2,NMI (C1,C2) is a measure to find the amount
of mutual information, the two clusters share with each other for the networks,
where communities or clusters may overlap. The version of NMI we used for our
comparisons is based on [41]: Given two clusters C1and C2
NMI(C1,C2)=H(C1)+H(C1)−H(C1,C2)
((H(C1)+H(C1))/2)(13)
where H(C1)=entropy of cluster C1,H(C2)=entropy of cluster C2and H(C1,C2)
joint entropy of cluster C1and C2.H(C) for any cluster C,H(C)=−
c∈Cp(c)∗
lg(p(c)) (14)
H(C1,C2)=
c1∈C1
c2∈C2
P(c1,c2)∗lgP(c1,c2)
p(c1)∗p(c2)(15)
where P(c) for any node c is the probability of and node c.∈community C.
To calculate probability P(c) for any node in community C, we have used the
formula
P(c)=No.of edges of c connected to nodes inside community C
Degree(c)(16)
Table 3represents the NMI score for Books on American politics network. We
have calculated the maximum, minimum, average NMI score for each algorithm
along with standard deviation with 10 runs of the algorithms. The proposed algorithm
gives the best maximum =0.6005, minimum =0.5982 and average =0.5996 NMI
scores among all the algorithms. Enhanced firefly algorithm gives second best result
216 H. S. Pattanayak et al.
with maximum =0.5953, minimum =0.5915 and average =0.5927 NMI score.
Clauset-Newman algorithm performs worst with maximum =0.5427, minimum =
0.5304 and average =0.5386 NMI score. Enhanced firefly algorithm gives the least
S.D. (standard deviation) with S.D. value =0.0004. However, the proposed algorithm
give second best result with S.D. value =0.0005.
Table 4contains the NMI score for Adjective Noun network. The proposed algo-
rithm performs best with maximum =0.8635, average =0.8619, minimum =0.8584
NMI score. Enhanced firefly algorithm performs second best in terms of NMI score.
In this network, proposed algorithms perform best in terms of S.D. with S.D. value
=0.0012. Blondel algorithm performs second best in terms of S.D. value with S.D.
=0.0013.
Table 5represents the Dolphin network. Among all the algorithms our proposed
performs best with 0.9907, 0.9898, and 0.9884 as maximum, average and minimum
NMI scores, respectively. In this network Firefly and Ronhovde–Nussinov algorithms
perform best in terms of S.D. with S.D. value =0.0001. The proposed algorithm
along with the Enhanced Firefly algorithm performs second best in terms of S.D.
value with S.D =0.0002.
Table 6represents the famous Karate club network data. In this network, the
proposed algorithm performs best in terms of maximum, minimum and average
NMI score and Enhanced firefly algorithm performs second best. In terms of SD, the
proposed algorithm performs best with S.D =0.0001.
Relaxed Caveman network data is stored in Table 7. The proposed algorithm
has the best maximum =0.3803, average =0.3791 and minimum =0.3786 NMI
Tabl e 4 NMI score: adjective noun network
Algorithms Max Mean Min Standard deviation
Proposed algorithm 0.8635 0.8619 0.8584 0.0012
Enhanced firefly 0.8592 0.8566 0.8512 0.0021
Firefly 0.8498 0.8466 0.8396 0.0025
Blondel 0.8535 0.8517 0.8487 0.0013
Ronhovde–Nussinov 0.8526 0.8488 0.8410 0.0028
Clauset–Newman 0.8321 0.8276 0.8189 0.0034
Tabl e 5 NMI score: Dolphin network
Algorithms Max Mean Min Standard deviation
Fire-spread 0.9907 0.9898 0.9884 0.0002
Enhanced firefly 0.9857 0.9838 0.9825 0.0002
Firefly 0.9832 0.9826 0.9817 0.0001
Blondel 0.9837 0.9822 0.9809 0.0003
Ronhovde–Nussinov 0.9868 0.9863 0.9855 0.0001
Clauset–Newman 0.5910 0.5881 0.5807 0.0025
Relationship Between Community Structure and Clustering … 217
Tabl e 6 NMI score: Karate club network
Algorithms Max Mean Min Standard deviation
Fire-spread 0.9975 0.9969 0.9964 0.0001
Enhanced firefly 0.9969 0.9966 0.9961 0.0001
Firefly 0.9792 0.9753 0.9683 0.0021
Blondel 0.9897 0.9847 0.9812 0.0015
Ronhovde–Nussinov 0.9868 0.9860 0.9856 0.0002
Clauset–Newman 0.7019 0.6941 0.6883 0.0261
Tabl e 7 NMI score: Caveman network
Algorithms Max Mean Min Standard deviation
Fire-spread 0.3803 0.3791 0.3786 0.0004
Enhanced firefly 0.3765 0.3741 0.3723 0.0009
Firefly 0.3739 0.3691 0.3658 0.0018
Blondel 0.3743 0.3712 0.3687 0.0009
Ronhovde–Nussinov 0.3716 0.3675 0.3624 0.0023
Clauset–Newman 0.3484 0.3409 0.3357 0.0092
score and Enhanced Firefly has the second best NMI scores. In terms of SD, the
Fire-spread algorithm performs best with value 0.0004. Both Enhanced Firefly and
Blondel algorithm have second best S.D. with value 0.0009.
Table 8represents data for Shared Community Network. This network was
specially designed; with 25% nodes are shared between communities, to evaluate
the proposed algorithm for overlapping communities. As it is given in the table,
proposed algorithm performs best in terms of maximum, minimum, average NMI
score along with S.D. with values 0.6165, 0.6127, 0.6145 and 0.0007, respectively.
Tabl e 8 Shared community network
Algorithms Max Mean Min Standard deviation
Fire-spread 0.6165 0.6145 0.6127 0.0007
Enhanced Firefly 0.6153 0.6138 0.6104 0.0008
Firefly 0.6127 0.6086 0.6049 0.0016
Blondel 0.6119 0.6095 0.6072 0.0009
Ronhovde–Nussinov 0.6111 0.6063 0.6016 0.0019
Clauset–Newman 0.5732 0.5574 0.5421 0.0057
218 H. S. Pattanayak et al.
6 Conclusion
From the data presented in the result section, it can be concluded that the proposed
algorithm has the best performance with both real networks and synthetic networks.
The algorithm is capable of detection overlapping community structure as it permits
any node to be part of more than community does.
Since the proposed algorithm is derived from the Fire-spread community detection
algorithm, it inherits some of the best features of its parent algorithm. The proposed
algorithm can be executed at local levels for finding the community around a seed
node and can able to assign a single node to multiple communities. However, the
algorithm has a simpler implementation in comparison with its parent algorithm, as
the threshold value can be calculated from the clustering coefficient.
From the experiments presented in Sect. 4, it is also clear that the clustering index
is co-related to community structure, hence can be used as a threshold value. In this
work, we also analysed the behaviour of Fire-spread algorithm in great detail with
different propagation probabilities and threshold values.
The proposed algorithm is used only on an undirected and unweighted network.
The algorithm can be applied for a directed and weighted network as a future work.
References
1. S. Wasserman, K. Faust, Social Network Analysis: Methods and Applications (Cambridge
University Press, 1994)
2. J. Leskovec, A. Krevl, https://snap.stanford.edu/data/
3. M.S. Granovetter, The strength of weak ties. Am. J. Sociol. 78, 1360–1380 (1973)
4. G. Flake, S. Lawrence, C. Giles, Efficient identification of web communities, in KDD’00:
Proceedings of the 6th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining (2000), pp 150–160
5. G. Palla, I. Derényi, I. Farkas, T. Vicsek, Uncovering the overlapping community structure of
complex networks in nature and society. Nature 435(7043), 814–818 (2005)
6. R. Pastor-Satorras, A. Vespignani, Epidemic spreading in scale-free networks. Phys. Rev. Lett.
86(14), 3200 (2001)
7. M. Girvan, M.E. Newman, Community structure in social and biological networks. Proc. Natl.
Acad. Sci. 99(12), 7821–7826 (2002)
8. A. Gog, D. Dumitrescu, B. Hirsbrunner, Community detection in complex networks using
collaborative evolutionary algorithms. Adv. Artifi. Life 886–894 (2007)
9. M.S. Ahuja, J. Singh, Neha practical applications of community detection. Int. J. Adv. Res.
Comput. Sci. Softw. Eng. 6(4) (2016)
10. H.S. Pattanayak, H.K. Verma, A.L. Sangal, Community detection metrics and algorithms in
social networks, in 2018 First International Conference on Secure Cyber Computing and
Communication (ICSCCC). IEEE (2018, December), (pp. 483–489)
11. H.S. Pattanayak, A.L. Sangal, H.K. Verma, Community detection in social networks based on
fire propagation. Swarm Evol. Comput. 44, 31–48 (2019)
12. V.D. Blondel, J.L. Guillaume, R. Lambiotte, E. Lefebvre, Fast unfolding of communities in
large networks. J. Stat. Mech: Theory Exp. 2008, P10008 (2008)
13. A. Clauset, M.E.J. Newman, C. Moore, Finding community structure in very large networks.
Phys. Rev. E 70, 066111 (2004)
Relationship Between Community Structure and Clustering … 219
14. P. Ronhovde, Z. Nussinov, Multiresolution community detection for mega scale networks by
information-based replica correlations. Phys. Rev. E 80, 16109 (2009)
15. S. Boettcher, A.G. Percus, Extremal optimization for graph partitioning. Phys. Rev. E 64 (2001).
(Article ID: 026114)
16. D. Fisher, J. Artif, Iterative optimization and simplification of hierarchical clusterings. J. Artif.
Intell. Res. 4, 147–180 (1996)
17. D. He, J. Liu, D. Liu, D. Jin, Z. Jia, Ant colony optimization for community detection in
large-scale complex networks, in Proceedings of 7th International Conference on Natural
Computation (ICNC 2011), vol. 2 (IEEE Press, 2011), pp. 1151–1155
18. T. Herlau, M. Mørup, M.N. Schmidt, L.K. Hansen, Detecting hierarchical structure in networks,
in Proceedings of 3rd International Workshop on Cognitive Information Processing (CIP 2012)
(IEEE Press, 2012), pp. 1–6
19. Z. Zhao, C. Li, X. Zhang, F. Chiclana, E.H. Viedma, An incremental method to detect
communities in dynamic evolving social networks. Knowl. Based Syst. 163, 404–415 (2019)
20. Z. Gao, N. Jin, Detecting community structure in complex networks based on k-means clus-
tering and data field theory, in Proceedings of 20th Chines Control and Decision Conference
(CCDC 2008) (IEEE Press, 2008), pp. 4411–4416
21. W. Li, 229 (2013) Revealing network communities with a nonlinear programming method,
Inform. Sci. 18–28
22. J. Liu, T. Liu, Detecting community structure in complex networks using simulated annealing
with k-means algorithms. Phys. A: Stat. Mech. Appl. 389(11), 2300–2309 (2010)
23. J.B. MacQueen, Some methods for classification and analysis of multivariate observations, in
Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability,vol.
1 (1967), pp. 281–297
24. Z.Y. Zhang, Community structure detection in complex networks with partial background
information, EPL 101 (4), 6 (2013). Article No. 48005
25. A. Ghasemian, H. Hosseinmardi, A. Clauset, Evaluating overfit and underfit in models of
network community structure. IEEE Trans. Knowl. Data Eng. (2019)
26. B. Amiri et al., Community detection in complex networks: multi–objective enhanced firefly
algorithm. Knowl. Based Syst. 46, 1–11 (2013)
27. X.S. Yang, Firefly Algorithms for Multimodal Optimization (Foundations and Applications,
Stochastic Algorithms, 2009), pp. 169–178
28. X. Duan, C. Wang, X. Liu, Y. Lin, Web community detection model using particle swarm
optimization, in Proceedings of IEEE Congress on Evolutionary Computation (CEC 2008)
(IEEE Press, 2008), pp. 1074–1079
29. R.J. Kuo, Y.D. Huang, C.C. Lin, Y.H. Wu, F.E. Zulvia, Automatic kernel clustering with bee
colony optimization algorithm. Inform. Sci. 283, 107–122 (2014)
30. Q. Cai, M. Gong, L. Ma, S. Ruan, F. Yuan, L. Jiao, Greedy discrete particle swarm optimization
for large-scale social network clustering. Inform. Sci. 2014 (2014). Available online 7 October
31. C. Pizzuti, GA-Net: a genetic algorithm for community detection in social networks, in Proceed-
ings of 10th International Conference on Parallel Problem Solving from Nature (PPSN 2008),
Lecture Notes in Computer Science, vol. 5199 (2008) pp. 1081–1090
32. A.I. Hafez, N.I. Ghali, A.E. Hassanien, A.A. Fahmy, Genetic algorithms for community detec-
tion in social networks, in Proceedings of 12th International Conference on Intelligent Systems
Design and Applications (ISDA 2012) (IEEE Press, 2012), pp. 460–465
33. V. Estivill-Castro, Why so many clustering algorithms: a position paper. ACM SIGKDD Explor.
Newslett. 4(1), 65–75 (2002)
34. D.J. Watts, P.S. Dodds, M.E.J. Newman, Identity and search in social networks. Science
296(5571), 1302–1305 (2002)
35. L. Wang, J. Wang, Y. Bi, W. Wu, W. Xu, B. Lian, Noise-tolerance community detection and
evolution in dynamic social networks. J. Comb. Optimiz. 28(3), 600–612 (2014)
36. W.W. Zachary, An information flow model for conflict and fission in small groups. J. Anthropol.
Res. 452–473 (1977)
220 H. S. Pattanayak et al.
37. D. Lusseau, K. Schneider, O.J. Boisseau, P. Haase, E. Slooten, S.M. Dawson, The bottlenose
dolphin community of Doubtful Sound features a large proportion of long-lasting associations.
Behav. Ecol. Sociobiol. 54, 396–405 (2003)
38. V. Krebs. http://www.orgnet.com/
39. M.E.J. Newman, Phys. Rev. E 74, 036104 (2006)
40. D.J. Watts, Small Worlds (Princeton University Press, 1999)
41. L. Danon, A. Diaz-Guilera, J. Duch, A. Arenas, Comparing community structure identification.
J. Stat. Mech.: Theory Exp. 2005, P09008 (2005)
Digital Image Forensics-Image
Verification Techniques
Anuj Rani and Ajit Jain
Abstract In the present scenario of the digital era, we are on the brink of marginal
transformation in digital imaging technology. Developments of high computational
and artificial intelligence techniques have produced wonderful image editing tech-
niques that create imposing results in stint frames. And, doing this will affect the
original artifacts present in digital images. This refers to the manipulation or forgery
of digital data with the help of image editing applications and this also erodes our
trust in digital images. To rely on the semantics of digital images, there is significant
research contribution in terms of various image forgery detection techniques. These
techniques help to re-evaluate the authenticity of digital images. In this article, an
examination of various research contributions is conducted. The primary goal of this
work is to give a glimpse of various current existing techniques related to digital image
forensics. These techniques are assessed as per their proficiencies and boundaries.
Whenever possible, a comparison between similar domain techniques is presented.
The main focus is on pixel and physics-based techniques. Our analysis discusses the
challenges that are to be addressed in this area of forensic science and provides insight
into various datasets available for researchers to develop and test new approaches.
Finally, the work recommends possible future research directions.
Keywords Content authentication ·Forgery detection ·Image forensics ·Passive
techniques
A. Rani (B)
Department of Computer Science, G L Bajaj Institute of Technology and Management, Greater
Noida, Utter Pradesh, India
e-mail: anuanuj1989@gmail.com
A. Jain
Department of Computer Science, Banasthali Vidyapith, Tonk, Rajasthan 304022, India
e-mail: ajitjain_2k@yahoo.co.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_19
221
222 A. Rani and A. Jain
1 Introduction
Today, there barely exists in any field where digital images are not used. Some of the
popular areas, where digital images are used, are digital media, e-media, military,
law and jurisdiction, science and technology, medical imaging science, newspapers,
social media, and many more. As per the usage of different types of digital images
in various fields, the importance of their originality is a major concern. We believe
in what we see rather than verbal content. Therefore, give our consent to this digital
data. The widespread use of this data is commonly tampered with digital imaging
techniques and hence misrepresents the meaning of information to us [1]. Image
tampering changes the meaning of image data for a kind of malevolent purpose.
On the other side, these manipulations are detected using digital image forensics.
Tampering with images is not a new concept even first forgery is reported in the
1860s. Many types of forgery making techniques are available in the literature. An
example of image forgery is depicted in Fig. 1. In this figure, the photomontage
forgery is shown in which different images are joined to create a new image using
pasting, overlapping, and rendering operations. These operations are carried out
using freely available and accessible image editing applications such as Photoshop,
GIMP, Paintshop, and many more. Under such conditions of image manipulations,
it becomes essential to find whether the produced image is real or fake because
of their usage in many secret domains. Therefore, this brings the importance or
motivation to design new image counterfeit recognition methods that can identify
and prove the authenticity of digital images. Many researchers are currently working
in this cutting-edge domain of digital forensics to develop new techniques to combat
image forgeries. And, here it becomes important to state that many research papers
are published by IEEE and Elsevier from the year 2000 to 2017 which shows the
increased pattern of publications in this domain. An analysis is shown in Fig. 2which
confirms that digital image forensic techniques are growing at a high pace. To find
the image vulnerabilities, forgery detection methods are majorly divided into two
types, namely active and passive/blind approaches. In an active approach, a kind
Fig. 1 A forged image
created using kiwi fruit and
lemons using GIMP
application software [2]
Digital Image Forensics-Image Verification Techniques 223
Fig. 2 A pattern of published papers by IEEE and Elsevier from 2000 to 2017. The analysis is
done using query “Image forgery Detection”
of preprocessing over data is required and examples of this approach are signature
generation and watermarking (hiding of data in host signal).
On the other side, passive techniques do not have such steps and, therefore, more
vigorous compares to active methods. In these techniques, a significant challenge is
to detect the image tempering because there exists no concealed data.
Many approaches are used to recognize the traces of manipulations. The passive
methods are divided into five types as explained below [3].
•Pixel-Based Techniques:
These techniques are used to detect pixel-level geometric anomalies during the
tempering process. Inconsistencies are identified at pixel levels. Techniques included
in this are mainly copy-move and splicing-based forgery detection approaches.
•Format (Compression)-Based Techniques:
Format-based methods use the anomalies generated at the time of compression of
the image. The generally used methods in this category are double JPEG.
•Camera or source-based Technique:
This technique changes the facts by sensor noise, camera lens properties, and on-chip
processing in the cameras.
•Physics-based Techniques:
Such approaches detect manipulations in 2D or 3D lighting parameters with
reflections and shadows of intrinsic lighting features.
224 A. Rani and A. Jain
Fig. 3 Classification of various image forgery detection techniques
•Geometric Techniques:
These approaches implement the concept of geometrical analysis of object location
related to camera location for forgery detection. A detailed description of these
approaches is shown in Fig. 3. We attempted the following research questions in this
study.
RQ1: How to analyze data to find image forgeries.
RQ2: Provide availability of popular image forensics techniques for researchers.
RQ3: A comparative study for various image forensics techniques.
This paper focuses on passive forensic approaches. The paper is segregated into
three sections. Sec. 2is about the related comprehensive study of various forgery
detection techniques followed by Sect. 3about the conclusion and future aspects of
this domain.
2 Critical Review of Various Forgery Detection Techniques
A critical review of various good quality publications in the domain of image foren-
sics is presented in this section. Our major target is to bring various types of current
image manipulation-detecting techniques to the researchers, which can help them to
advance in this field.
Digital Image Forensics-Image Verification Techniques 225
(A) Pixel-Based Forgery Detection Techniques
In pixel-based techniques, the copy-move-based forgery detection methods are
mainly discussed. The copy-move-based approaches are alienated into keypoint-
based, block-based, hybrid-based, and splicing-based approaches. Copy-move
forgery is the most common type of manipulation where a portion from the image is
copied and inserted on other regions of a similar image. This will hide the existing
object from the image and create a duplicate forged version of the image [4]. There
are many ways to make forgery using copy-move. Some of these techniques are
known as keypoint- and block-based copy-move forgeries. In the keypoint-based
forgeries, keypoints or features are extracted using feature extraction algorithms.
Distinct features such as corners, blobs, and contours/edges/boundaries are estimated
form images, and later descriptors and features are matched to identify the forged
copied regions. These algorithms are mainly implemented using three categories
like SIFT (scale-invariant feature transform), SURF (speed up robust features), and
Harris Corner points.
Silva et al. [5] proposed a copy-move-based image tampering detection approach
using a multi-scale analysis and voting process of the image. They extracted interest
points and clustered them into regions based on geometrical constraints. These
features are robust against operations like rotation and scaling. Using such points
finally, a multi-scale image representation is generated and the final decision of the
process is taken by a voting process for various image regions. They compared their
approach to 15 other techniques and produced batter results. Cozzolino et al. [6]
implemented a copy-move-based forgery algorithm for 44 precisely detection and
localization of forgery using rotation-invariant features. They applied the fast approx-
imate nearest-neighbor search algorithm and PatchMatch. Their experimental results
are robust and quicker than other state-of-art techniques. Yu et al. [7] suggested
a method to solve the problem of redundant feature points and feature fusion to
detect image forgery. The two-stage process enhances the matching performance
by merging MROGH and HH descriptors. Their technique is effective in detecting
highly precise copy-move forgeries. Zandi et al. [8] suggested a high computation
cost-efficient technique to identify copy-move forgery. A new nearest point detector
is recommended using the block-based and keypoint-based approaches. The esti-
mated keypoint from the image is considered even from the low-contrast areas of the
image so that a unique matrix is computed. A filtering algorithm is also used to filter
the wrongly matched areas. Their process is iterated with keypoint and, finally, the
copied areas are estimated. Their results are satisfactory. Wang et al. [9] suggested
a keypoint-based copy-move forgery technique for small even image portions. In
the first step of their approach, the tampered part of the image is segmented into
non-overlapping and superpixels. Using the local entropy information, these super-
pixels are classified into smooth, textured, and strong textures. Then, further, in the
second step, the stable image keypoint is obtained from each superpixel by using
their content-based adaptive feature point detector. Further, the local visual features
such as exponent moments magnitudes are generated for every key point and the best
is generalized using 2NN (nearest-neighbor) algorithms for matching keypoints. In
226 A. Rani and A. Jain
the last step of the algorithm, the false matched points are removed and duplicated
areas are estimated using zero-mean normalized cross-correlation measure. These
techniques produce better results against geometric transformations, compression,
additive Gaussian noise, etc.
On the contrary, the block-based copy-move forgeries extract the features imple-
mented on segmented images. These segmented images can have many overlap-
ping and non-overlapping image blocks for which feature vectors are estimated and
matched using the block-based procedures. There exist many block-based matching
algorithms in previous works; some of them are discussed here. Lee et al. [10]
implemented a blind forensics technique to detect block-based copy-move forg-
eries. In their work, the input image is isolated into various overlapping blocks,
and HoG gradients are used for every block. Similarities of these blocks are esti-
mated using the statistical features. In the last step of their algorithm, the feature
vectors are lexicographically arranged and cloned image areas are identified. This
technique has the capability to detect copy-move forgeries and also deals with
various image distortion features. Liang et al. [11] presented an approach for object
removal by exemplar-based inpainting integrates principal pixel mapping, greatest
zero-connectivity component labeling, and fragment splicing detection. Central pixel
mapping speeds up the copied areas by matching such blocks with similar hash values.
The zero-connectivity component is basically required to identify the manipulated
pixels in the blocks estimated using labeling. Finally, fragment splicing detection is
used to discretize the forged regions from the best match image areas. This approach
is capable of producing 85% forgery detection accuracy. Huang et al. [12] proposed
a method to find copy-move forgery using JPEG compression and Gaussian noise ith
blurring attacks. Their technique involves steps such as feature extraction, matching,
and copied block identification. For feature extraction, they used Fast Fourier Trans-
formation (FFT), SVD, and Principal Component Analysis (PCA). The combination
of FFT, SVD, and PCA is used to extract similar features. In the last step of their
approach, pixels on the upper left corners of duplicated regions are used for visual
examination. Their results are satisfactory. Zhou et al. [13] suggested a method based
on color moments and the other five descriptors. Forged image is divided into fixed-
size overlapped blocks, and clustering shares whole search space into regions of
alike color distribution. Tampered blocks reside in the same cluster in which copied
and similar blocks have the same color distribution. A group of deep compositional
(PNN) pattern neural network is trained with estimated features. Finally, the feature
vector from various clusters indicated possible forged areas. Their accuracy for image
forgery detection is good even against various blurring attacks.
In splicing-based image forgery detection, two or more number of images are
compounded to make a new copy of the image. Splicing operation is somehow close
to copy-move forgery. Various authors already published good research papers based
on splicing-based approaches. Zhou et al. [14] proposed a 2D noncasual Markov
model for detecting splicing-based image forgeries. They estimated discriminative
model parameters to differentiate the spliced regions. The model is applied in block
DCT domain and cross-domain features are picked as final features for the classi-
fication purpose. They also used SVM to classify the spliced images on the public
Digital Image Forensics-Image Verification Techniques 227
image splicing dataset. The results are superior compared with other techniques in
their domain. Bahrami et al. [15] proposed a splicing detection based approach using
image blur estimation. In their work, after block-based image partitioning, the local
blur approximation is applied to find the features from the local blur kernel. Image
blocks are categorized in the motion blur-based features to produce the invariant blur
forgeries. Finally, a fine splicing localization is performed to enhance the accuracy
of region edges. And, the differences in these image areas are detected as a sign of
forgeries. Their results are able to estimate blur-type image splicing manipulations.
Zhan et al. [16] suggested a splicing detection technique that is based on irrele-
vant areas between spliced and original image patches. These areas are computed
by minimum eigenvalues estimated using PCA without knowing any prior knowl-
edge. The similarity matrix for PCA is estimated using a similar pixel strategy so
that local minimum eigenvalue is obtained for the separation of spliced portions of
images. Cozzolino and Verdoliva [17] implemented a method in which expressive
local features are obtained from noise residue of image and provide to the autoen-
coder as input to generate the implicit model of the data. This autoencoder finds
the spliced region as an anomaly. Their demonstrated results are able to identify
spliced-based forgery detection. Abrahim et al. [18] proposed a technique to find
spliced regions using image texture properties. They used various texture features to
capture the edge of objects and color features. Using these features, a feature vector
is generated to find the spliced object. Three features are obtained and trained using
ANN. In the second model, the majority voting for three features is performed. By
combining different features and information the NN classifies the spliced image.
Song et al. [19] also proposed motion blur-based forgery detection. Their approach is
divided into three parts. In the first step, a CNN-based motion blur kernel estimation
method is implemented to find image patches available in the image. In the second
step, a shared motion blur kernel-based image manipulation detection technique is
suggested to detect whether a cluster of motion blur kernels is estimated from the
same 3D camera trajectory. In the last part, a consistency propagation approach is
implemented to specify the forged image portions. Their results are compelling and
detect motion blur based-image splicing.
There exist some hybrid approaches to detect pixel-based forgeries. Li et al. [20]
proposed a method to enhance the functioning of forgery localization using tampering
possibility maps. They estimated statistical feature-based identifier and copy-move
forgery detector and, further, adjusted these results to find tampered possibility maps.
Afterward, exploring these maps and comparing them with many fusion schemes,
they finally localized their results. Their findings are improved over other tech-
niques. Korus and Huang [21] investigated a multi-scale analysis technique that uses
many candidates tapering maps to analyze the windows which are more reliable and
produce better localization resolution of images. They combined three techniques
for multi-scale fusion and verified their results ith other state-of-art techniques. This
hybrid technique can detect image forgeries. For more work related to pixel-based
techniques can be referenced from [22–26]. A comparative study between various
pixel-based approaches is shown in Table 1.
228 A. Rani and A. Jain
Tabl e 1 Comparative study between various pixels based approaches
Author citation Features estimated Classifier used Performance measure
[5]Speed up robust features Nearest neighbour
distance ratio
85.35%
[6] Zernike moments,
principal components
transformations, Fourier
Mellin transform
Dense linear fitting 95.92%
[7]MROGH and hybrid using
MROGH and HH
G2NN 94%
[8]Polar cosine transform RANSAC 90%
[9]Speed up robust features Rg2NN 93.3%
[11] Hash values GZCL 96%
[12]Cellular automata and
local binary pattern
FL ANN Not defined
[13]Color moments, color
layout features, color and
edge directive descriptive
Neural network Not defined
[15]Local blur kernals EM algorithm 92.8%
[16] Local minimum
eigenvalues
Threshold-based
clustering
94%
[17] Image residuals Auto encoder
network
Not defined
[26]Speed up robust features Threshold-based
clustering
98%
(B) Source And JPEG Compression-Based Forgery Detection
Identification of camera source using which photos are captured has a significant
importance in digital image forensics. Using source-based techniques, the identifi-
cation of the source camera can be identified. These techniques identify and extract
the features of image acquisition devices using the source processing elements such as
lens aberration, sensor deficiencies, and color filter array (CFA), etc. By analyzing
the irregularities in these parameters, source-based forgeries can be detected. On
the other side, the compression-based techniques find the irregularities in the JPEG
compression format. The primary purpose of these techniques is to find the area of
images where the differences of JPEG compression are detected. These approaches
for JPEG-based methods use block-based artifacts caused due to JPEG compression.
These types of inconsistencies are used to locate the compression- and resampling-
based forgeries in images. The techniques related to source and compressions are
discussed in this section.
Choi et al. [27] proposed a source identification approach using intrinsic lens aber-
ration. They focused on the features if lens radial distortions and obtain such features
for all the pixels of images. They used a classifier to classify the source camera. Their
Digital Image Forensics-Image Verification Techniques 229
results are producing 87–91% of source identification accuracy. Lukas et al. [28]
implemented a novel method to detect the source using the sensor’s pattern noise.
For each camera, they estimated the reference pattern noise as a sign of uniqueness of
the camera. They referred to the reference noise as a spectrum of identity for multiple
images. Their experiment shows that their technique is able to detect the source of
the image effectively. Chan et al. [29] proposed a confidence map and pixel-weighted
correlation method to recognize the camera model. A denoizing approach is imple-
mented to find the photo response non-uniformity (PRNU) between real and fake
images. Non-linear regression model is used to test the noise in images and, subse-
quently, a confidence map for pixels is estimated. This identifies the PRNU noise
and therefore the camera model. In a similar kind of work, Tapinar et al. [30]show
that for a given numerous seam carved images of a camera, the source identification
is possible using PRNU. Source identification is possible for uncarved blocks of 50
by 50. They estimated PRNU from the images using these blocks and identification
of the camera model is to be carried out. A recent work by Yang et al. [31] suggested
a solution to identify the camera using content-adaptive fusion residual networks.
This image is divided into three sub-sets, namely saturation, smoothness, and others
to find image content. Then, they trained the fusion network for saturated images
through transfer learning. The fusion network is formed with three parallel residual
networks, and the differences in these networks are recorded. These are obtained
as per the presented feature data. The convolution is performed in the initial stage
to improve the features. Finally, these features find the differences in image data
to justify the source-based forgery. There are other approaches also available in the
literature which identified CFA-based inconsistencies and, therefore, camera models.
Nguyen and Katzenbeisser [32] proposed a resampling detection-based tech-
nique. They combined various resampling techniques and hybrid median filtering
for covering the hints of resampling. But, their technique is robust to identify resam-
pling artifacts to find the traces of forgery. Their results show that their results are
better even against various attacks on images. Kirchner and Gloe [33] find resam-
pling detection using re-compression in images. The blocking artifacts rise during the
compression and the previous compression helps to locate the forgery. They applied
affine transformation on compressed images to identify the forgeries. Su et al. [34]
derived a method based on image resampling detection algorithm for blind deconvo-
lution by proposing an inverse filtering process. Their method is effectively avoiding
the interfaces that arise due to JPEG block artifacts. The results of the technique are
convincing. Zach et al. [35] suggested an approach for powering the identification
of JPEG ghosts. These ghosts are expended to differentiate the single and twofold
JPEG compression and this could be a common clue for detection image forgeries.
They estimated jpeg ghosts automatically and, finally, manipulation in images. The
outcomes of this technique are comparatively better to other such methods. Wang
and Zhang [36] propose a double JPEG compression detection approach using CNN.
CNN is used to design and classifies the histograms of DCT coefficients. These
coefficients differ as per single and double compression. The localization results are
obtained according to classification and tampered areas are identified. Their results
are able to find tampering in double compression and forgery localization. Pasquini
230 A. Rani and A. Jain
et al. [37] proposed a technique for detecting JPEG compression artifacts stored
in various image formats. They computed DCT coefficients from 8 by 8 blocks
and applied a hypothesis of no compression and compression with specific quality
features. Further to classification, tests are conducted using statistics relying on the
discriminative threshold. This statistical analysis is based on the Gaussian distribu-
tion of the coefficients which provides better detectors for any uncompressed image.
These identify the differences in image data find the manipulations. Other relevant
studies in these areas can be referred to [1,2,38–40].
(C) Physics- and Geometric-Based Forgery Detection
Physics-based approaches included the analysis of inconsistencies in lighting envi-
ronments. These methods further include 2D and 3D lighting and lighting envi-
ronments. These methods believe in the supposition that any change in the image
will change the geometrical and physical characteristics of images and the changed
aspects can be used to find the traces of forgeries. In lighting-based techniques, the
direction of light sources are estimated to locate the forged regions or objects in the
images for various kind of lighting environments including from simple to complex.
Very few researchers explored this area of forgery detection and very limited liter-
ature is available based on this type of forgery detection techniques. Some of the
excellent works in this area are discussed here in this paper.
Johnson and Farid [41] proposed a technique to detect forgeries using 2D and 3D
lighting. They estimated intensity parameters and respective lighting differences in
image areas to localize the forgeries. Their results can find the directions for infi-
nite light sources. The same researchers in [42] described a technique for finding
forgeries by detecting lighting inconsistencies in the complex lighting environment.
They show a technique to estimate the complex lighting with low-dimensional model
and estimated the parameters of this model using spherical coefficients. The irregu-
larities in these lighting coefficients are used as an indication of image manipulation.
Kee and Farid [43] proposed a technique to estimate 3D lighting using Lambertian
reflectance model. Their technique is based on the estimation of spherical harmonics
and 3D is estimated for the face profile of the objects. The limitation of the tech-
nique is that it works only for the face profiles and respective 3D models but not
for other generalized objects. 3D surface normals are estimated from this model to
identify the image forgery. Peng et al. [44] implemented an improved 3D lighting
environment based approach using a general surface reflection model. The consid-
ered local geometry and texture information for their reflection model. Their method
is more applicable for objects like human faces which are non-convex and bumpy.
Their results of the technique are effective for finding forgery detection. Kumar and
Srivastava [45] designed a lighting based technique to detect light source direction.
They estimated lighting features from various objects of the image using the surface
normal and intensity estimation. They selected patches areas from the same intensity
portion and estimated light source directions. For a change in the intensity, their
technique performs poorly. Their result detects forgeries for similar kinds of image
Digital Image Forensics-Image Verification Techniques 231
intensity patches using lighting estimation. Riess [46] proposed a method to opti-
mize the constraints of surface albedo and shape from shading. Riess detected image
forgery using illumination estimation from an image. The author considered objects
boundary to estimate the direction of the light environment. Their results are able
to detect lighting using illumination profiles. Kumar and Srivastava [47] proposed a
technique to identify the inconsistencies in lighting direction. They estimated angle
of incidence for each object present in the image and further calculations for surface
normal are done. They considered the RED band to access the surface properties for
illumination estimation. Their results are satisfactory and detect forgery for various
types of objects available in images. Same authors also proposed lighting-based
forgery detection in [48,49].
Complex lighting-based work is very limited in this domain and very few authors
explored this area. The complex lighting is defined when the lighting environment
is lighted by more than one light source. That is if more than one light source is
available in the scene then that scene environment is considered a complex lighting
environment. Some of the works which can be refereed in this direction are [50–54].
A comparison between various physics-based approaches is shown in Table 2.
Tabl e 2 Comparison between various physics-based approaches
Citation Features used Classifier used Performance/error measure
[41]Localization of occluding
boundaries
Kind of least square
estimation
Error percentage 18%
[44] Surface reflectance model
for illumination estimation
and spherical coefficients
Position-dependent transfer
function
89.2%
[45]Surface geometry –Angle error threshold
8°–10°
[46]Illumination and color
features
– –
[47]Surface texture and color
features
Least square estimation 95%
[50] Intensity, curve, and
recursive refinement
fittings
Approximation technique Mean error of 8.55° and
8.84°
[52]Calibration object
parameters and critical
point estimation
–Error threshold 5°
[53]Modified critical point
estimation and intensity
–Threshold for an error of
2.5°–4°
[54]PCA –AUC of 98.1% and a DR of
97.4%
232 A. Rani and A. Jain
3 Conclusion and Future Direction
A thoughtful speculative study is presented in this paper, which highlights image
forgery detection techniques. The paper aims to provide a classification on various
state-of-art methods so that researchers can explore more and gain a better under-
standing of these techniques in this area. Some of the valuable and recent researches
in this direction are discussed in this paper. Various works state that knowledge of
detection features and approaches, modeling techniques, and tools are highly required
to work in this domain. Even there is no technique that can identify all types of image
forgeries, which is a challenging task. In the future direction of this work, we need
to design more powerful approaches that can estimate forgeries in complex lighting
environments and a technique that can provide a higher percentage of successful
detection of various types of image manipulations. Also, the requirement of a dataset
is much needed as most of the researcher establishes their work using either a limited
number of forged images or small datasets for a kind of forgeries.
References
1. M. Kumar, S. Srivastava, Image forgery detection based on physics and pixels: a study. Aust.
J. Forensic Sci. 51(2), 119–134 (2019)
2. S. Walia, K. Kumar, Digital image forgery detection: a systematic scrutiny. Aust. J. Forensic
Sci. (2018). https://doi.org/10.1080/00450618.2018.1424241
3. H. Farid, Image forgery detection. IEEE Signal Process Mag. 26(2), 16–25 (2009)
4. M. Dilshad, S. Ghrera, V. Tyagi, Pixel based image forgery detection: a review. IETE J. Educ.
55(1), 40–44 (2014)
5. E. Silva, T. Carvalho, A. Ferreira, A. Rocha, Going deeper into copy-move forgery detection:
exploring image telltales via multi-scale analysis and voting processes. J. Vis. Commun. Image
Represent. 26, 16–32 (2015)
6. D. Cozzolino, G. Poggi, L. Verdoliva, Efficient dense-field copy-move forgery detection. IEEE
Trans. Inf. Forensics Secur. 10(11), 2284–2297 (2015)
7. L. Yu, Q. Han, X. Niu, Feature point-based copy-move forgery detection: covering the non-
textured areas. Multimedia Tools Appl. 75(2), 1159–1176 (2016)
8. M. Zandi, A. Mahmoudi-Aznaveh, A. Talebpour, Iterative copy-move forgery detection based
on a new interest point detector. IEEE Trans. Inf. Forensics Secur. 11(11), 2499–2512 (2016)
9. X.-Y. Wang, S.L.Y.-N. Liu, Y. Niu, H.-Y. Yang, Z.-L. Zhou, A new keypoint-based copy-
move forgery detection for small smooth regions. Multimedia Tools and Applications 76(22),
23353–23382 (2017)
10. J.-C. Lee, C.-P. Chang, W.-K. Chen, Detection of copy–move image forgery using histogram
of orientated gradients. Inf. Sci. 321, 250–262 (2015)
11. Z. Liang, G. Yang,X. Ding, L. Li, An efficient forgery detection algorithm for object removal by
exemplar-based image inpainting. J. Vis. Commun. Image Represent. 30(2015), 75–85 (2015)
12. D.-Y. Huang, C.-N. Huang, W.-C. Hu, C.-H. Chou, Robustness of copy-move forgery detection
under high JPEG compression artifacts. Multimedia Tools Appl. 76(1), 1509–1530 (2017)
13. H. Zhou, Y. Shen, X. Zhu, B. Liu, Z.F.N. Fan, Digital image modification detection using color
information and its histograms. Forensic Sci. Int. 266, 379–388 (2016)
14. X. Zhao, S. Wang, S. Li, J. Li, Passive image-splicing detection by a 2-D noncausal markov
model. IEEE Trans. Circuits Syst. Video Technol. 25(2), 185–199 (2015)
Digital Image Forensics-Image Verification Techniques 233
15. K. Bahrami, A.C. Kot, L. Li, H. Li, Blurred image splicing localization by exposing blur type
inconsistency. IEEE Trans. Inf. Forensics Secur. 10(5), 999–1009 (2015)
16. L. Zhan, Y. Zhu, Z. Mo, An image splicing detection method based on PCA minimum
eigenvalues. J. Inf. Hiding Multimedia Signal Process. 7(3), 610–619 (2016)
17. D. Cozzolino, L. Verdoliva, Single-image splicing localization through autoencoder-based
anomaly detection, in IEEE International Workshop on Information Forensics and Security
(WIFS), Abu Dhabi, United Arab Emirates, 2016
18. A.R. Abrahim, M.S.M. Rahim, G.B. Sulong, Splicing image forgery identification based on
artificial neural network approach and texture features. Cluster Comput. 22(S1), 647–660
(2019)
19. C. Song, P. Zeng, Z. Wang, T. Li, L. Qiao, L. Shen, Image forgery detection based on motion
blur estimated using convolutional neural network. IEEE Sens. J. (2019). https://doi.org/10.
1109/jsen.2019.2928480
20. H. Li, W.Luo, X. Qiu, J. Huang, Image forgery localization via integrating tampering possibility
maps. IEEE Trans. Inf. Forensics Secur. 12(5), 1240–1252 (2017)
21. P. Korus, J. Huang, Multi-scale fusion for improved localization of malicious tampering in
digital images. IEEE Trans. Image Process. 25(3), 1312–1326 (2016)
22. F. Akhtar, H. Qayyum, Two fold image forgery detection system using combined keypoint
based method and block based method. J. Inf. Commun. Technol. Robotic Appl. 9(2), 62–70
(2018)
23. X.-Y. Wang, L.-X. Jiao, X.-B. Wang, H.-Y. Yang, P.-P. Niu, A new keypoint-based copy-move
forgery detection for color image. Appl. Intell. 48(10), 3630–3652 (2018)
24. X.-Y. Wang, Y.-N. Liu, H. Xu, P. Wang, H.-Y. Yang, Robust copy–move forgery detection using
quaternion exponent moments. Pattern Anal. Appl. 21(2), 451–467 (2018)
25. H.-Y. Yang, S.-R. Qi, Y. Niu, P.-P. Niu, X.-Y. Wang, Copy-move forgery detection based on
adaptive keypoints extraction and matching. Multimedia Tools Appl. 1–28 (2019). https://doi.
org/10.1007/s11042-019-08169-w
26. P. Srivastava, M. Kumar, V. Deep, P. Sharma, A technique to detect copy-move forgery using
enhanced SURF. Int. J. Eng. Adv. Technol. 8(6S), 676–680 (2019)
27. K.S. Choi, E.Y. Lam, K.K. Wong, Source camera identification using footprints from lens
aberration, in Proceedings of SPIE-IS&T Electronic Imaging, SPIE, vol. 6069, 2006
28. J. Lukas, J. Fridrich, M. Goljan, Digital camera identification from sensor pattern noise. IEEE
Trans. Inf. Forensics Secur. 1(2), 205–214 (2006)
29. L.-H. Chan, N.-F. Law, W.-C. Siu, A confidence map and pixel-based weighted correlation for
PRNU-based camera identification. Digit. Invest. 10(2013), 215–225 (2013)
30. S. Taspinar, M. Mohanty, N. Memon, PRNU-based camera attribution from multiple seam-
carved images. IEEE Trans. Inf. Forensics Secur. 12(12), 3065–3080 (2017)
31. P. Yang, R. Ni, Y. Zhao, W. Zhao, Source camera identification based on content-adaptive
fusion residual networks. Pattern Recogn. Lett. 119, 195–204 (2019)
32. H. C. Nguyen, S. Katzenbeisser, in Robust resampling detection in digital images, ed. by B.
De Decker, D.W. Chadwick. Communications and multimedia security. CMS 2012. Lecture
Notes in Computer Science, vol. 7394 (Springer, Berlin, Heidelberg, 2012)
33. M. Kirchner, T. Gloe, On resampling detection in re-compressed images, in First IEEE
International Workshop on Information Forensics and Security (WIFS), London, UK, 2009
34. Y. Su, X. Jin, C. Zhang, Y. Chen, Hierarchical image resampling detection based on blind
deconvolution. J. Vis. Commun. Image Represent. 48, 480–490 (2017)
35. F. Zach, C. Riess, E. Angelopoulou, in Automated image forgery detection through classi-
fication of JPEG ghosts, ed. by A. Pinz, T. Pock, H. Bischof, F. Leberl. Pattern Recogni-
tion. DAGM/OAGM 2012. Lecture Notes in Computer Science, vol. 7476 (Springer, Berlin,
Heidelberg, 2012)
36. Q. Wang, R. Zhang, Double JPEG compression forensics based on a convolutional neural
network. EURASIP J. Inf. Secur. 23(2016), 1–12 (2016)
37. C. Pasquini, G. Boato, F. Pérez-González, Statistical detectionof JPEG traces in digital images
in uncompressed formats. IEEE Trans. Inf. Forensics Secur. 12(12), 2890–2905 (2017)
234 A. Rani and A. Jain
38. P. Korus, Digital image integrity—a survey of protection and verification techniques. Digit.
Signal Proc. 71(2017), 1–26 (2017)
39. S. Mandelli, P. Bestagini, S. Tubaro, D. Cozzolino, L. Verdoliva, Blind detection and local-
ization of video temporal splicing exploiting sensor-based footprints, in 26th European Signal
Processing Conference (EUSIPCO), Rome, Italy, 2018
40. B. Gupta, M. Tiwari, Improving source camera identification performance using DCT based
image frequency components dependent sensor pattern noise extraction method. Digit. Invest.
24(2018), 121–127 (2018)
41. M. K. Johnson, H. Farid, Exposing digital forgeries by detecting inconsistencies in lighting, in
ACM Multimedia and Security Workshop, New York USA, 2005
42. M. K. Johnson, H. Farid, Exposing digital forgeries in complex lighting environments, IEEE
Trans. Inf. Forensics Secur. 2(3), 450–461 (2007)
43. E. Kee, H. Farid, Exposing digital forgeries from 3-D lighting environments, in IEEE
international workshop on information forensics and security, Seattle, WA, USA, 2010
44. B. Peng, W. Wang, J. Dong, T. Tan, Improved 3D lighting environment estimation for image
forgery detection, in IEEE International Workshop on Information Forensics and Security
(WIFS), 2015
45. M. Kumar, S. Srivastava, Identifying photo forgery using lighting elements. Indian J. Sci.
Technol. 9(48), 1–5 (2016)
46. C. Riess, in Illumination analysis in physics-based image forensics: a joint discussion of illu-
mination direction and color, ed. by A. Piva, I. Tinnirello, S. Morosi. Digital Communication.
Towards a Smart and Secure Future Internet. TIWDC 2017. Communications in Computer and
Information Science, vol. 766 (Springer, Cham, 2017)
47. M. Kumar, S. Srivastava, Image authentication by assessing manipulations using illumination.
Multimedia Tools Appl. 78(9), 12451–12463 (2019)
48. M. Kumar, S. Srivastava, in Image tampering detection based on inherent lighting fingerprints,
ed. by D. Hemanth, S. Smys. Computational Vision and Bio Inspired Computing. Lecture Notes
in Computational Vision and Biomechanics, vol. 28 (Springer, Cham, 2018)
49. M. Kumar, A. Rani, S. Srivastava, Image forensics based on lighting estimation. Int. J. Image
Graph. 19(3):195004:1–195004:14 (2019)
50. J. Lopez-Moreno, E. Garces, S. Hadap, E. Reinhard, D. Gutierre, Multiple light source
estimation in a single image. Comput. Graph Forum 32(8), 170–182 (2013)
51. S. Xu, A. Wallace, Recovering Surface reflectance and multiple light locations and intensities
from image data. Pattern Recogn. Lett. 29(11), 1639–1647 (2008)
52. Y. Wang, D. Samaras, Estimation of multiple directional light sources for synthesis of mixed
reality images, in 10th Pacific Conference on Computer Graphics and Applications, Beijing,
China, 2002
53. M. Kumar, S. Srivastava, N. Uddin, Forgery detection using multiple light sources for synthetic
images. Aust. J.Forensic Sci. 51(3), 243–250 (2019)
54. A. Mazumdar, P.K. Bora, Estimation of lighting environment for exposing image splicing
forgeries. Multimedia Tools Appl. 78(14), 19839–19860 (2019)
Using Automated Predictive Analytics
in an Online Shopping Ecosystem
Ruchi Mittal
Abstract This paper is based on a UCI dataset analysis, which includes data gener-
ated through Google analytics. The data attributes describe the shopping intent of
consumers and their likelihood of exiting the e-commerce website during browsing.
This study has selected five attributes out of a total of 18 that were generated and seeks
to capture the shopping intent based on the website browsing behavior of shoppers.
All selected attributes are in numeric form, and the dependent variable is pagevalue
which represents the average value for the user-visited webpage before the comple-
tion of a transaction. The proximity of a forthcoming special day is checked for its
mediation effect on other independent variables based on web browsing behavior.
The findings give some very interesting and relevant insights into the use of analytics
on human online behavior.
Keywords Predictive analytics ·Structural model ·Shopping intent ·Browsing
behavior
1 Introduction
The online shopping industry is growing rapidly and customers are increasingly
preferring to buy online because of varied factors such as ease, variety, schemes,
and saving their time and cost. It is thus a promising domain for the researchers to
make use of the technology to understand the changing requirements of the online
shoppers to help online retailers to make appropriate decisions to meet the compe-
tition and to have an edge over their competitors. For this purpose, many different
types of demographical, behavioral, and other information can be collected from the
customers, who are making online purchases that are quite significant to be analyzed
critically in order to provide solutions to many business problems. Many studies have
been conducted to analyze the impact of demographic factors of customers on online
purchases and one of such study [1] indicated that in context to online shopping, only
R. Mittal (B)
Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, India
e-mail: ruchi.mittal@chitkara.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_20
235
236 R. Mittal
a few of the demographic factors like age, gender, income, and education are signifi-
cant for the researchers to analyze the predictive power of the online purchase corre-
sponding to particular products. Previous research extended the study to understand
the important factors contributing to the success of the online purchase experience
other than the basic demographics and concluded that social connections and prior
online shopping experience also attribute to various online shopping categories [2].
Reference [3] recognized the problem faced by the online customers while making
a decision for the product just on the basis of a given image on the website, and
thus some of the other associated information such as price, brand, and previous
customers’ reviews on price, reliability, durability, and quality and shopping expe-
rience can be vital in building the faith in the new customer and in increasing their
online shopping convenience. They proposed a ranking allocation mechanism to the
previous customers’ reviews as it is difficult for the new shoppers to get through all
the reviews [3].
Data mining is the inclusion of many disciplines such as database technologies,
statistics, artificial intelligence, machine learning, etc., that advance the analytical
capabilities of the managers in uncovering the hidden facts, important for the decision
making of an organization. Data mining is quite popular in all the business sectors
(online and onsite) to optimize their cost and benefits. Business analysts and data
scientists can choose among various functionalities of data mining such as classifi-
cation and prediction, clustering, association analysis, class description, outlier or
time-series analysis, etc., based on the type of data and the and the requirement of
the application and can suggest intelligent insights for the growth and sustainability
of the organization.
Also with the availability of high-speed and low-cost internet facility, selling
and buying online is widely becoming popular and convenient. Incorporation of the
appropriate data mining tools and technologies can help online retailers to collect,
store, process and analyze the large amount customers’ data and can exploit the bene-
fits in important decision-making such as identifying the potential customers; market
basket analysis; future sales prediction; product planning; segmenting and profiling
customers for target marketing; and improved customer relationship management
for online shopping organizations [4].
This study takes a different and unique approach to online shopping. An existing
open-source dataset was selected from the UCI repository and the following research
question was framed based on the web browsing behavior of customers: How can
we predict the “pagevalue” of an e-commerce site based on select attributes and
how do certain attributes mediate the relation among these different attributes? Here
“pagevalue” refers to the average value for a webpage that a shopper visited before
completing an e-commerce transaction.
To achieve the objectives the following attributes were selected as predictors of
“pagevalue”: “product-related information”, “bounce rate”, “exit rate”. An upcoming
“special day” was used as a mediator in the hypothesized relationships between the
predictors and the criterion attribute (or independent and dependent attributes).
The paper is organized as follows; in Sect. 2the review of the literature is
discussed. This is followed by research methodology in Sect. 3. Section 4deals
Using Automated Predictive Analytics in an Online Shopping … 237
with the results and analysis Sect. 5deals with discussions and interpretation. The
final section is the conclusion and future research.
2 Review of the Literature
A prior study emphasized the need for various data mining techniques such as
segmentation, association, and classification on the vast amount of data collected
from the emerging style of online shopping instead of the conventional shopping
approach and also described that how such mining and analysis can help to have
better customer relationship management (CRM) and to further optimize the profits
of an organization [5]. In order to analyze the customers’ requirements and prefer-
ences, a novel method based on online customers’ reviews was examined by first
clustering the customers and then by identifying the relevant and salient attributes
of a product in these clusters based on the customer choice criteria [6]. Looking at
customer feedback as an integral part of CRM, a previous study, after identifying
the importance of feedback system, not only to analyze the customer satisfaction
post-purchase/usage, but also to help the app developers in future developments and
modifications, the authors presented an ordered review of different studies in the
domain of online product reviews and customers’ feedback on mobile app store and
also acknowledged the effectiveness of applying data mining techniques for the same
[7]. In a separate study, the authors tried to aid the business decision-making by using
data mining functionalities such as decision trees to discover the type of buyers in
the times when discounts were offered by the stores and also, to devise the strategies
to increase the customer base for the organization [8]. In another interesting study,
a prototype of an application was discussed which provides instant product sugges-
tions to potential shoppers based on even a few preferred product attributes. Based
on footwear retailing, the researchers collected the data related to the parameters of
foot and shows description of the ease on feet and then applied data mining function-
alities such as decision tree, neural network and KNN to come up with a predictive
model giving the relevant parameters such as the length of the shoe and also stated
that decision tree algorithm has given the best predictive accuracy [9].
Another study on e-commerce recognized the potential of an automated and intelli-
gent system to extract knowledge from the tremendously growing data in e-commerce
and applied and compared different data mining classification techniques using Weka
machine learning software on a credit-card dataset for identifying possible frauds
during online payments to sustain the possibility of paying online which is very
important in e-commerce systems. The authors also described that J48 outperforms
other classifiers on the mentioned dataset [10]. Another study categorized online
shoppers based on their demographics and consumption patterns to further under-
stand the behavioral aspects of shoppers belonging to these identified segments to
help the retailers to devise appropriate strategies to cater to the needs of different
types of customers. The study used a soft-clustering method for segmentation as it
gives better cluster quality over other approaches [11]. In another study, researchers
238 R. Mittal
identified the importance of customers’ reviews for the products purchased online
to not only help the retailers to make more accurate business decisions but also
the new buyers to decide on the purchase of a particular product. The authors used
Weka machine learning tool to mine such reviews generated from online purchases
and conducted a sentiment analysis (SA) followed by classification of these reviews
using different classification algorithms [12]. In a different study, the authors used
TANGARA machine learning tool in order to investigate the reviews of buyers on
different products and brands and for this purpose, the authors applied various mining
techniques of classification, clustering, association and correlation to get online busi-
ness insights like the relationship between customers and sellers and also to classify
customers based on their remarks on different products [13]. In a separate study, the
authors provided a comparative analysis of 11 data mining techniques of clustering
and classification on the online customers’ shopping dataset to identify the most
appropriate classifier and proposed that decision tree classification scheme gives
better results over clustering for customer segmentation on e-commerce websites
[14].
Many studies have been carried out emphasizing the use of tools and technologies
to analyze the vast amount of data being generated in various domains of business
organizations existing online and offline, so based on the review of previous studies,
this study aims to get the insights for online shopping behavior of customers based
on select attributes drawn from the web browsing behavior of customers.
3 Research Methodology
3.1 Dataset
The dataset has been adopted from the UCI machine learning repository. The dataset
comprises a total of 12,330 records of transactions and deals with the intention of
shoppers towards purchases while shopping online [15]. The original dataset includes
18 attributes (10 numerical and 8 categorical attributes) extracted threw aggregated
page views, which were collated by Google Analytics. Select attributes have been
shortlisted for the purpose of this study, namely: (1) product-related duration, (2)
bounce rate, (3) exit rate, (4) pagevalue, and (5) special day.
3.2 Description of Attributes
Product-related duration refers to the duration of time spent by the potential
consumer on the product-related page of the e-commerce portal. Bounce rate refers
to the feature for a web page that calculates the percentage of visitors, who enter the
site from that page and then leave without triggering any other requests to the Google
Using Automated Predictive Analytics in an Online Shopping … 239
Analytics server during that session. The value of Exit Rate feature for a specific web
page is calculated as for all page views to the page, the percentage that was the last
in the session. The Pagevalue feature represents the average value for a web page
that a user visited before completing an e-commerce transaction. The Special Day
feature indicates the closeness of the site visiting time to a specific special day (e.g.,
Mother’s Day, Valentine’s Day) in which the sessions are more likely to be finalized
with transaction.
3.3 Predictive Model
The data analysis was done using AMOS version 18—structural equation modeling
(SEM) and tested a total of seven hypotheses drawn from the conceptual model as
per Fig. 1.
“Structural equation modeling (SEM) is a collection of data analyses techniques
that allow a set of relationships between one or more independent variables (IVs),
either continuous or discrete, and one or more dependent variables (DVs), either
continuous or discrete, to be examined. Both IVs and DVs can be either factors
or measured variables. Structural equation modeling is also referred to as causal
modeling, causal analysis, simultaneous equation modeling, analysis of covariance
structures, path analysis, or confirmatory factor analysis” [16]. SEM has been very
effectively used for prediction and model testing in earlier studies [17–20]. Based on
the model as per Fig. 1, eight hypotheses have been formulated. The SEM analysis
performs the hypothesis testing and the results are presented in the next section.
Fig. 1 Conceptual model
240 R. Mittal
4 Results and Analysis
See Tables 1,2,3,4,5and 6.
Tabl e 1 Regression weights: (group number 1—default model)
Estimate S.E. C.R. PLabel
Special day <— Product related 0.000 0.000 1.623 0.105
Special day <— Bounce rate −0.533 0.091 −5.844 ***
Special day <— Exit rate 0.922 0.093 9.906 ***
Pagevalue <— Special day −3.840 0.828 −4.636 ***
Pagevalue <— Product related −0.004 0.004 −1.105 0.269
Pagevalue <— Bounce rate 91.415 8.395 10.889 ***
Pagevalue <— Exit rate −149.497 8.597 −17.390 ***
Tabl e 2 Standardized regression weights: (group number 1—default model)
Estimate
Special day <— Product related 0.015
Special day <— Bounce rate −0.130
Special day <— Exit rate 0.225
Pagevalue <— Special day −0.041
Pagevalue <— Product related −0.010
Pagevalue <— bounce rate 0.239
Pagevalue <— Exit rates −0.391
Tabl e 3 Standardized indirect effects (group number 1—default model)
Product related Exit rate Bounce rate Special day
Special day 0.000 0.000 0.000 0.000
Pagevalue −0.001 −0.009 0.005 0.000
Tabl e 4 Standardized indirect effects—two-tailed significance (BC) (group number 1—default
model)
Product related Exit rates Bounce rates Special day
Special day … … … …
Pagevalue 0.035 0.001 0.001 …
Using Automated Predictive Analytics in an Online Shopping … 241
Tabl e 5 Standardized direct effects (group number 1—default model)
Product related Exit rate Bounce rate Special day
Special day 0.015 0.225 −0.130 0.000
Pagevalue −0.010 −0.391 0.239 −0.041
Tabl e 6 Standardized direct effects—two-tailed significance (BC) (group number 1—default
model)
Product related Exit rate Bounce rate Special day
Special day 0.039 0.001 0.001 …
Pagevalue 0.211 0.001 0.001 0.001
5 Discussions and Interpretation
H1 There is a significant relationship between special day and product-related
duration.
Table 1shows the regression weight in the relationship between special day and
product-related duration. The regression coefficient value of 0.000, standard error
value of 0.000, critical ratio 1.623 which is below 1.96, and p value greater than <0.01
show that there is no significant relationship between special day and product-related
duration.
H2 There is a significant relationship between special day and bounce rate.
Table 1shows the regression weight in the relationship between special day and
bounce rate. The regression coefficient value of −0.533, standard error value of
0.091, critical ratio −5.844 which is below 1.96, and p value <0.01 show that special
day has a significant relationship with bounce rate. Table 2also indicates a negative −
0.130 standardized beta value of between the relationship of special day and bounce
rate. This indicates that the bounce rate is a negative significant predictor of special
day. Hence, it is concluded that bounce rate significantly affects the special day in a
negative manner, which means that more the bounce rate lower the probability of a
forthcoming special day.
H3 There is a significant relationship between special day and exit rate.
Table 1shows the regression weight in the relationship between special day and
exit rate. The regression coefficient value of 0.922, standard error value of 0.093,
critical ratio 9.906 which is above 1.96, and p value<0.01 show that special day has
a significant relationship with exit rate. Table 2also indicates a positive standardized
beta value of 0.225 between the relationship of special day and exit rate. This indicates
that exit rate is a positive significant predictor of special day. Hence, it is concluded
that exit rate significantly affects the special day in a positive manner which means
that more the exit rate higher the probability of a forthcoming special day.
242 R. Mittal
H4 There is a significant relationship between pagevalue and special day.
Table 1shows the regression weight in the relationship between pagevalue and special
day. The regression coefficient value of −3.840, standard error value of 0.828, critical
ratio −4.636, and pvalue <0.01 show that special day has a significant relationship
with pagevalue. Table 2also indicates a negative standardized beta value of −0.041
between the relationship of special day and pagevalue. This indicates that special day
is a negative significant predictor of pagevalue. Hence, it is concluded that special day
significantly affects the special day in a negative manner, which means that higher
the probability of a forthcoming special day lower the pagevalue.
H5 There is a significant relationship between pagevalue and product related
duration.
Table 1shows the regression weight in the relationship between pagevalue and
product-related duration. The regression coefficient value of −0.004, standard error
value of 0.004, critical ratio −1.105, and pvalue >0.01 show that product-related
duration has a nonsignificant relationship with pagevalue.
H6 There is a significant relationship between pagevalue and bounce rate.
Table 1shows the regression weight in the relationship between pagevalue and bounce
rate. The regression coefficient value of 91.415, standard error value of 8.395, critical
ratio 10.889 which is above the threshold of 1.96, and pvalue <0.01 show that
bounce rate has a significant relationship with pagevalue. Table 2also indicates a
positive standardized beta value of 0.239 between the relationship of bounce rate
and pagevalue. This indicates that bounce rate is a positive significant predictor of
pagevalue. Hence, it is concluded that bounce rate significantly affects the pagevalue
in a positive manner, which means that more the bounce rate higher the pagevalue.
H7 There is a significant relationship between pagevalue and exit rate.
Table 1shows the regression weight in the relationship between pagevalue and exit
rate. The regression coefficient value of −149.497, standard error value of 8.597,
critical ratio of −17.390 which is above the threshold of 1.96, and pvalue <0.01
show that exit rate has a significant relationship with pagevalue. Table 2also indicates
a negative standardized beta value of −0.391 between the relationship of exit rate and
pagevalue. This indicates that exit rate is a negative significant predictor of pagevalue.
Hence, it is concluded that exit rate significantly affects the pagevalue in a negative
manner which means that more the exit rate higher the pagevalue.
H8 An upcoming special day mediates the relationship between (a) product-related
duration and pagevalue, (b) bounce rate and pagevalue, (c) exit rate and pagevalue.
Results of the mediation analysis show that special day mediates the relationship
between product-related duration and pagevalue, bounce rate and pagevalue, and
exit rate and pagevalue (please refer to Table 3,4,5, and 6).
Using Automated Predictive Analytics in an Online Shopping … 243
6 Conclusion and Future Research
The following select attributes were shortlisted for the purpose of this study, namely,
product-related duration, bounce rate, exit rate, pagevalue and special day. The data
were subjected to structural equation modeling (SEM). As per the data analyses,
there is no significant relationship between special day and product-related duration.
Bounce rate significantly affects the special day in a negative manner which means
that more the bounce rate lower the probability of a forthcoming special day. Exit
rate significantly affects the special day in a positive manner which means that
more the exit rate higher the probability of a forthcoming special day. Special day
significantly affects the special day in a negative manner which means that higher
the probability of a forthcoming special day lower the pagevalue. Product-related
duration has a nonsignificant relationship with pagevalue. Bounce rate significantly
affects the pagevalue in a positive manner, which means that more the bounce rate
higher the pagevalue. Exit rate significantly affectsthe pagevalue in a negative manner
which means that more the exit rate higher the pagevalue. Special day shows an
effective mediation effect as hypothesized in the conceptual model.
In a future study, the balance 13 attributes may be included and different structural
models may be created and tested. Different attributes may be checked for mediation
effect and some attributes may also be evaluated for moderation effect.
References
1. M.K. Chang, W. Cheung, V.S. Lai, Literature derived reference models for the adoption of
online shopping. Inf. Manag. 42(4), 543–559 (2005)
2. M.B. Naseri, G. Elliott, Role of demographics, social connectedness and prior internet experi-
ence in adoption of online shopping: applications for direct marketing. J. Target. Meas. Anal.
Mark. 19(2), 69–84 (2011)
3. K. Zhang, R. Narayanan, A.N. Choudhary, Voice of the customers: mining online customer
reviews for product feature-based ranking. WOSN,10,11
4. M. Ismail, M.M. Ibrahim, Z.M. Sanusi, M. Nat, Data mining in electronic commerce: benefits
and challenges. Int. J. Commun. Netw. Syst. Sci. 8(12), 501 (2015)
5. M. Riaz, A. Arooj, M.T. Hassan, J.B. Kim, Clustering based association rule mining on online
stores for optimized cross product recommendation, in The 2014 International Conference on
Control, Automation and Information Sciences (ICCAIS 2014), 2014 (pp. 176–181). IEEE
6. R. Rai, Identifying key product attributes and their importance levels from online customer
reviews, in ASME 2012 international design engineering technical conferences and computers
and information in engineering conference, 2013 (pp. 533–540). American Society of
Mechanical Engineers Digital Collection
7. N. Genc-Nayebi, A. Abran, A systematic literature review: opinion mining studies from mobile
app store user reviews. J. Syst. Softw. 125, 207–219 (2017)
8. S. Kumar, R.K. Arora, Analyzing customer behaviour through data mining. IJCATR (Online)
4, 12 (2015)
9. P. Sudta, K. Kanchan, C. Chantrapornchai, Children shoes suggestion system using data mining.
Int. J. Database Theory Appl. 5(1), 21–36 (2012)
10. F. Alam, S. Pachauri, Detection using weka. Adv. Comput. Sci. Technol. 10(6), 1731–1743
(2017)
244 R. Mittal
11. R.S. Wu, P.H. Chou, Customer segmentation of multiple category data in e-commerce using a
soft-clustering approach. Electron. Commer. Res. Appl. 10(3), 331–341 (2011)
12. A.A. Kshirsagar, P.A. Deshkar,. Review analyzer analysis of product reviews on WEKA
classifiers, in 2015 International Conference on Innovations in Information, Embedded and
Communication Systems (ICIIECS), 2013 (pp. 1–5). IEEE
13. V. Jain, Narula, G.S., M. Singh, Implementation of data mining in online shopping system
using tanagra tool. Int. J. Comput. Sci. Eng. (2013)
14. R.A.E.D. Ahmeda, M.E. Shehaba, S. Morsya, N. Mekawiea, Performance study of classifi-
cation algorithms for consumer online shopping attitudes and behavior using data mining, in
2015 Fifth International Conference on Communication Systems and Network Technologies,
2015 (pp. 1344–1349). IEEE
15. C.O. Sakar, S.O. Polat, M. Katircioglu, Y. Kastro, Real-time prediction of online shoppers’
purchasing intention using multilayer perceptron and LSTM recurrent neural networks. Neural
Comput. Appl. 1–16
16. J.B. Ullman, P.M. Bentler, Structural equation modeling. Handb. Psychol. 607–634 (2003)
17. R. Mittal, V. Rattan, Evaluating rule based machine learning classifiers for customer spending
in a shopping mall. J. Adv. Res. Dyn. Control Syst. 11(09), 716–719 (2019)
18. A.S. Kumar, P.K. Chand, A. Mittal, Emotional intelligence and its relationship to employability
skills and employer satisfaction with fresh engineering graduates. Int. J. Qual. Res. 13(3),
735–752 (2019)
19. Vikas Uppal, Amit Mittal, Antecedents of post purchase consumer dissatisfaction in household
white goods. J. Adv. Res. Dyn. Control Syst. 11(5), 1958–1966 (2019)
20. Ruchi Mittal, Predictive modeling for bank direct marketing: a sem based approach on select
attributes. Int. J. Adv. Sci. Technol. 27(1), 283–288 (2019)
Design and Development of Home
Automation System
Anjali Pandey, Yagyiyaraj Singh, and Medhavi Malik
Abstract Internet of things is based on the integration of various real-world objects
by interconnecting them using the Internet. This concept can be applied to build an
automated smart home. A phone application will be used to control various elec-
trical and electronic appliances such as fans, lights, fire alarms, automatic doors,
smoke detectors, etc., through various sensors and actuators. The microcontroller
used in the proposed prototype is node MCU. It is Wi-Fi equipped and can be
programmed using Arduino IDE. The working principle of the proposed system
is based on different types of wireless communication techniques such as ZigBee,
Wi-Fi, Bluetooth, etc. This enables us to remotely control manage and monitor home
appliances. An automated home is smarter, safer, more convenient, time saving as
well as energy efficient.
Keywords Home automation ·Internet of Things ·Server ·Mobile devices
1 Introduction
Internet of things is a technology that connects devices, machines, and tools to the
Internet by means of wireless technologies. Over 9 billion things are connected to the
Internet as of now. It also involves the unification of technologies such as low-power
embedded systems, cloud computing, big data, machine learning, and networking.
Internet of things contains a network of physical objects which are interconnected
to one another through the Internet. Some areas identified as IoT enablers are RFID,
nanotechnology, sensors, and smart networks [1].
A. Pandey ·Y. Singh ·M. Malik (B)
SRM IST Delhi NCR Campus Ghaziabad, Ghaziabad, India
e-mail: medhavimalik28@gmail.com
A. Pandey
e-mail: Anjalipandey868@gmail.com
Y. Singh
e-mail: yagyiyaraj@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_21
245
246 A. Pandey et al.
Benefits of Using Internet of Things
•Optimization: It is mainly required in industries, for example, strict quality control,
supply chain management, inventory management.
•Increase in productivity: With the help of IoT, employees can have improved
decision-making and efficient communication.
•Better business models: IoT enables companies to provide service based on real-
time sensor data and information [2].
Modern Day IoT Applications
•Traffic congestion,
•Smart home,
•Medical applications,
•Smart city,
•Weather prediction,
•Smart parking, and
•Soil moisture monitoring.
IoT Market Share
•Business manufacturing,
•Health care,
•Retail, and
•Security [3].
Home Automation
A smart home includes the interconnection of various devices using a variety of
wired and wireless technologies. “It is a home that is equipped with lighting, heating,
electronic devices that can be controlled remotely by smart phone or computer”. For
example, you can switch on the air condition of your room while sitting in any other
part of the home or installing an alarm system that sends a notification to your phone
the moment alarm rings. An automated home can provide a safe home environment
and personalization according to our needs. A smart home optimizes structures,
systems, service, and management [4].
Home area network (HAN): It can be referred to as a network within a home that
allows controlling the devices remotely. HAN elements are as follows:
•Internet protocol—It is a set of protocols and rules for exchanging messages and
information among devices in a network connected through the Internet [5].
•Wired HAN—In this, we can use various types of wires like optical fiber, power
lines, coaxial cables, and telephone lines. It is a low-cost method and provides
efficient use of already existing house infrastructure.
•Wireless HAN—It is less complex as compared to wired HAN. It uses
technologies such as Bluetooth, ZigBee, Wi-Fi, etc.
Design and Development of Home Automation System 247
SMART HOME
Communication
network
infrastructure
Intelligent
control and
management
Sensor
networks
Smart feature
and automatic
response
Advantages of Home Automation System
The ability to manage your home’s electronic system from one main control system
can make your household run smoother, feel better, and save energy. The trick is to
find a system that will meet all the demands of your household now and in the future.
1. Energy Saving: It is one of the most important advantages of home automation
and can be done in various ways such as automatically switching off the air condi-
tioner when the temperature is optimum or generating energy using photovoltaic
cells (solar cells) installed on the rooftop of the house [6].
2. Remote access: This feature will allow you to monitor your home’s safety and
alter the appliance’s settings remotely using your phone.
3. Interoperability: The beauty of a home automation system lies in its ability to tie
diverse electronics together so that they can work as one unified system.
4. Variety of user interface: There are many different ways you can control your
automated home for example by touching the colorful icons on a portable touch
panel of sliding fingers across the screen of your smartphone [7].
2 Proposed Work
The fundamental purpose of this paper is to develop an environment in which one
can control home appliances through a mobile application on a smartphone device.
All the electronic devices connected to the Internet in the system will run without
any human interference.
For building a smart home, the following are the basic requirements:
1. A microcontroller such as Node MCU, Arduino UNO, etc.
2. Various Sensors such as light sensor, temperature sensor, gas sensor, water flow,
motion sensor, and many other sensors.
3. Electronics appliances such as lights, fans, AC, television, speakers, surveillance
cameras, etc.
248 A. Pandey et al.
Microcontroller (Node MCU):
The information about various sensors is gained by using a microcontroller.
Node MCU, technically isn’t a device but it is firmware. There are two types of
Node MCU boards, version 0.9 and 1.0.
1. Version 0.9: This board is blue and loaded with ESP-12 chip and it comes with
16 pins
2. Version 1.0: This board is black and comes with ESP-12E (enhanced). And this
version comes with 22 pins [8].
ESP8266 is more recent release than the Arduino. There is a 32-bit RISC processor
clocked at 80 MHz, along with generous RAM complement and storage for up to
16 Mb of external flash storage.
It has a USB connection for easy interface with the computer.
“Node MCU is based on ESP8266-12E Wi-Fi system on chip loaded with an open
source firmware. It is perfect for IoT applications and various other situations where
wireless connection is required” [9].
Sensors:
A sensor senses some kind of change in a system and forwards or processes this
gathered information in a certain manner.
“A device which detects or measures a physical property and records, indicates, otherwise
responds to it.”
—Oxford dictionary [10]
When a sensor senses some physical change in a system then it performs input
functions and responds to stimuli.
Operation of the sensor is very similar to the sensors in the human body. For
example, skin is one of the sensors of the human body. When it senses or receives a
pain trigger, then it causes the organism to “be aware”’. Similarly, sensors sense the
changes and respond to these changes accordingly.
There are two categories of sensors: Analog and Digital.
1. Analog Sensors: Analog sensors produce a continuous output signal or voltage,
which is generally proportional to the quantity being measured.
Physical quantity such as temperature, speed, pressure, displacement, strain, etc.
are all analog quantities and they tend to be continuous in nature.
2. Digital Sensors: Digital sensors produce discrete digital output signals and
voltage that are in digital representation of the quantity being measured [3].
Digital sensors produce a binary output signal in the form of logic “1” or logic
“0”, (“ON” or “OFF”)
Sensor Types:
1. Light sensor: It converts light energyinto electrical signals. Example: photodiode,
light-dependent sensor.
Design and Development of Home Automation System 249
2. Temperature sensor: It is used to measure temperature. It uses a converter to
convert the temperature value to an electrical value. Example: thermocouple,
thermistor.
3. Sound sensor: It uses diaphragm which vibrates when hit by sound and this
vibration is converted into an electrical signal by the sensor. Example: carbon
microphone and piezoelectric crystal [3].
There are a number of such examples of sensors.
Actuators:
An actuator is a component of a machine or a system that moves or controls the
mechanism of the system.
An actuator requires a control signal and a source of energy.
Upon receiving a control signal, the actuator responds by converting the energy
into mechanical motion [3].
Planning:
There are various technologies that can be incorporated into a Smart Home:
1. Smart lights:
Light is a very critical component of any home. It affects almost every part of
the household. So replacing lights with smart lights in homes can save both our
time and money.
Smart lights are efficient as they can be automatically switched ON and OFF
according to the usage.
2. Smart fans:
Smart fans can be automatically turned ON and OFF after a certain temperature
has reached.
3. Smart security system:
A security system can detect when someone has entered your property or home
and sound an alarm. The automation system can send images captured by
surveillance cameras to your smartphone.
4. Heating and cooling system:
It detects the temperature of a room. If the temperature is high then it turns ON
the cooling system of the room.
Implementation:
Software Architecture:
To implement home automation, two software are used that are Arduino IDE and
mobile app (Blynk app).
Arduino IDE:
The Arduino IDE is an open-source software that is mainly used for writing and
compiling the code into the Arduino module (board).
A range of modules available are Arduino UNO, Arduino Mega, Node MCU, and
many more.
250 A. Pandey et al.
Each of these modules consists of a microcontroller that is programmed and
accepts the information in the form of code.
The IDE environment mainly contains two basic parts: Editor and Compiler.
Editor: It is used for writing the required code.
Compiler: It is used for compiling and uploading the code into the given board.
It supports both C and C++ languages [11].
Mobile Application:
Home automation demands a proper home automation mobile app to coordinate all
the system components and create an efficient infrastructure. All the home appliances
can be controlled using a mobile application.
The operation becomes very easy by using a mobile application.
You can control each and every appliance in the system by just one tap on your
mobile screen.
We can create our own mobile application for our system or we can use different
pre-made mobile applications.
There are different mobile applications for home automation available on the play
store. We just need to download it and set it up according to our requirements. Blynk
app is one of the pre-made mobile applications for home automation. In this paper,
we will be focusing on building a smart home using Blynk app [12].
Blynk App:
Blynk app is a new platform that allows you to quickly build interface for controlling and
monitoring your hardware projects from your iOS and Android devices. After downloading
the Blynk app, you can create a project dashboard and arrange buttons, sliders, graphs and
other widgets onto the screen.
Blynk Architecture:
The Blynk platform includes the following components:
Design and Development of Home Automation System 251
1. Blynk app builder: Allows you to build apps for your project using various
widgets. It is available for android and iOS platforms.
2. Blynk Server: Responsible for all the communications between your mobile
(device) that is running the Blynk app and the hardware.
3. Blynk libraries: Enables communication with the server and processes all the
incoming and outgoing commands from your Blynk app and the hardware.
4. Connection types: Blynk app supports the following connection types to connect
your microcontroller board (hardware) with Blynk cloud and Blynk’sserver:
Ethernet, Wi-Fi, Bluetooth, Cellular, Serial [13].
Steps to Setup Blynk App:
Blynk app is used to visualize data, control equipment remotely, set notifications
and rules, manage multiple devices, etc. Since app development is an expensive
pre-made module and the feature reduces cost.
Step 1: Select the project name and choose the device.
Step 2: After selecting the project name and board, the token is sent to your
registered id by the Blynk app.
Step 3: Now select the widgets from the widget box. You can select buttons,
sliders, timer, vertical slider, and various other widgets.
Step 4: Select the required type and number of pins.
252 A. Pandey et al.
Step 5: Six buttons have been selected for controlling lights and fans and gauge
that will indicate temperature. You can add widgets according to your requirement.
Step 6: The status is displayed on the top, when you are online you can control
devices using the buttons.
Design:
The system proposed above is designed using an Android-based smartphone that
has touch screen operation and an application for controlling the elements of an
automated home. This application will use the MQTT protocol for the transmission
of the signal to node MCU or any other microcontroller that has been used in the
system. MQTT stands for message queuing telemetry transport which uses publish-
subscribe messaging pattern. Thus by using our smartphone, we can send the ON/OFF
signal to the microcontroller which is in turn connected to various sensors and home
appliances that we need to monitor, control, and manage. Hence a home automation
system can be designed with the help of an Android application that can be controlled
Design and Development of Home Automation System 253
Fig. 1 Design of a basic home automation system
remotely by touching a particular button on the smartphone which will turn on/off
an appliance or sensor through a wireless network [14] (Fig. 1).
3 Scope of Home Automation System
1. Security: A smart home can have various devices such as security alarm, smoke
detectors, and security cameras. These devices can be controlled easily with the
help of a home automation system and hence improving the security of the house.
2. Saving time: An efficient home automation system can simplify our daily routines
by providing the feature to easily control various elements of our home [3].
3. Comfort and remote access: By providing the ability to remotely manage and
monitor various home appliances, a home automation system provides a great
deal of comfort and peace of mind to people living in that house as it can easily
adapt and fit according to the lifestyle of people and enable easy customization
options.
4. Boosting efficiency: The efficiency of a household can be increased in various
ways by using a home automation system as it provides interoperability, energy
saving, cost saving, and adding to the convenience of user [15].
254 A. Pandey et al.
4 Conclusion
Using node MCU is cost-effective as it will decrease the overall cost of the automa-
tion system since it as cheaper than other microcontrollers. Node MCU has many
things in common with Arduino like both of them are microcontroller equipped
prototyping board which can be programmed using Arduino IDE. Hence node MCU
is more compact and Wi-Fi-equipped alternative of Arduino. By using Wi-Fi for
connectivity rather than using other technologies such as Bluetooth and ZigBee we
can overcome the problem of short-range connectivity. Since Wi-Fi will provide
wireless connectivity it will be more efficient than wired connections because if
devices are connected through wires then data will be transferred at comparatively
low speed and finding out any fault in wiring is not easy.
References
1. https://www.sap.com/india/trends/internet-of-things.html
2. https://readwrite.com/2019/03/07/how-iot-is-optimizing-costs-for-industrial-sectors/
3. NPTEL online course- introduction to internet of things by Prof. Sudip misra
4. https://www.xfinity.com/hub/smart-home/home-automation
5. https://www.techopedia.com/definition/5366/internet-protocol-ip
6. https://www.wespeakiot.com/how-smart-homes-help-saving-energy/
7. https://www.smarthouseintegration.com/smarthouse-blog/10-key-features-of-home-automa
tion-systems/
8. https://www.instructables.com
9. https://www.electronicwings.com/nodemcu/basics
10. Oxford dictionary
11. https://www.theengineeringprojects.com/2018/10/introduction-to-arduino-ide.html
12. https://www.sam-solutions.com/blog/why-is-a-mobile-app-important-for-an-efficient-smart-
home-solution/
13. https://subscription.packtpub.com/book/application_development/9781788995061/1/ch01lv
l1sec10/what-is-blynk
14. Kodali RK, Jain V, Bose S, Boppana (2016) IoT based smart security and home automa-
tion system. In: International conference on computing, communication and automation
(ICCCA2016)
15. https://www.slideshare.net/DanKinsella/smart-home-systems-52056045
Performance Analysis Study
of Stochastic Computing Based Neuron
A. Dinesh Babu and C. Gomathy
Abstract Deep learning systems are computationally more expensive and their
performance depends strongly on the underlying hardware platform. The basic
requirements of a deep neural network (DNN) are simple and high-precision compu-
tations that can be performed faster. A deep neural network involves enormous data
manipulations, of which basic arithmetic operations such as addition and multipli-
cation form the building blocks. But implementing these operations in DNN is a
challenging task with less hardware requirements. In this paper, a study has been
made on how DNN effectively processes the diverse multi-sensor data. All these can
be taken into consideration for the applications which work on constrained resource
environment, e.g., IoT devices.
Keywords Stochastic computing ·DNN ·Stochastic neuron ·Deep learning
1 Introduction
With the advent of smart mobile phones in the past decade, our day-to-day life has
become easier. Security has become mandatory for these devices. Speech recognition,
visual recognition are playing a key role. For all these, a large amount of sensory
inputs are taken and processed to give the final expected output. Moreover, a major
part of these processes takes place in the warehouse computer which is communicated
through the cloud. Moving certain processes from cloud into the device would save
a lot of power and make the system an intelligent one. All these require enhancing
the efficiency of the algorithms used and also the efficiency of the mobile devices
which perform these functions.
A. Dinesh Babu (B)·C. Gomathy
Department of ECE, SRM Institute of Science and Technology, Vadapalani, Chennai, India
e-mail: dineshba@srmist.edu.in
C. Gomathy
e-mail: hod.ece.vdp@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_22
255
256 A. Dinesh Babu and C. Gomathy
In recent years, many improvements have been made in the algorithmic predic-
tions which are more accurate. This is mainly because of the introduction of deep
neural networks for these applications. These networks are nowadays used in object
classification, detections. Past few years, various researches have been made in the
deep neural network which is inspired from biological systems [1]. Restricted Boltz-
mann machine in one such deep learning system, which has created a great impact
because of its higher performance in solving complex function [2].
It is classified into two major domains; learning phase and inference phase. The
learning phase comprises mapping the input data to their respective desired outputs,
while the inference phase makes use of this configuration to calculate the outputs for
the data. Deep belief networks have produced desirable results for vision recognition
processes. DBN are made by peeling the RBM one above the other to form a deep
neural network [3]. However, power consumption is very high for all these processes.
With the advent of the IoT for mobile devices have pushed the researchers to build
lower power machine learning pervasive devices and also to make the maximum use
of the onboard power supply [4].
Stacked layers of RBMs are used to construct the Deep belief network, with
the output layer at the final level as seen in Fig. 1. These networks comprise of
various matrix multiplications in every levels. These multiplications are very costly
in terms of hardware implementation since they require a large silicon portion and
high-power consumption. Moreover, the nonlinearity functions are performed with
the help of lookup tables, which require a large amount of memory [5]. Thus DBN
implementation using VLSI efficiently is a big task [6].
Stochastic computing, nowadays has been seen as an alternative solution for the
above-stated problem. Stochastic computing (SC) methods may result in fast compu-
tation, low power consumption during hardware implementation. With the help of
Fig. 1 DBN with N-layers
where weights are denoted
using Wand number of
layers are denoted by N
Performance Analysis Study of Stochastic Computing Based Neuron 257
SC, various complex computational units can be easily implemented. Various multi-
plication circuits and adder circuits can be implemented with the help of unipolar
SC, i.e. using AND gate and multiplexer.
Since the DNN consist of a numerous number of multipliers, various approaches
have been taken into consideration for minimizing the computational speed, power,
and area [7]. All these are required for end IoT devices since they have a limited
onboard power supply, as well as the size of these devices is very small. Literature
survey has also been done, and a study has made on the possible solutions to overcome
the previously existing problems.
2 Literature Survey and Related Works
2.1 Structure of a Neuron
The scope of DNN is to find a solution to complex problems. DNN is a group of
neurons with processing capability and connection-related algorithms. The neuron
acts as a processing unit for information [8]. The neurons, are similar to the biological
neurons, since both of them have inputs, output, and hidden layers, which produce
an output with respect to the given inputs. The artificial neurons (Fig. 2) are formed
by [9]:
•Inputs: Direct inputs or the outputs from the other neuron’s activation unit.
•Weights and Bias: Can be either positive or negative. These play an important
role in determining the function of the neural network.
•Summation Unit: It is used for adding up the input signals, which were previously
multiplied with their corresponding weights. Finally, bias is added to it.
•Activation Unit: It controls the previous stage output and normalizes it between
0 and 1 or −1, 1.
•Output: The output of the neuron which drives the next neuron.
Fig. 2 Basic neuron architecture
258 A. Dinesh Babu and C. Gomathy
2.2 Neural Network and its Types
The neural network comprises a number of neurons that are arranged in a series of
sequential layers [10,11]. The neural network (Fig. 3) comprises different layers:
•Input layer—This layer receives real-time inputs, and starts processing the input.
The input layer is a visible layer.
•Hidden layer—This layer is sandwiched between the input and output layers. The
role of this layer is to convert the inputs by computing it with their respective
weights and bias and then process them to the output layer
•Output layer—This layer computes the output of the neural network.
Some of the neural network types includes the following:
•Single-layer perceptron: Network with two inputs and one output without any
hidden layers [12]. It is also known as the Basic Perceptron model (Fig. 4a).
•Multilayer perceptron: Network contains many hidden layers between the input
and the output layers. It is also known as Feedforward Deep Neural Network
(Fig. 4b) [13].
•Restricted Boltzmann machine neural network: The nodes in the network are with
the next layers but not within the same layer. The neurons in this network make
stochastic decisions, i.e., to turn on or turn off. (Fig. 5)
Fig. 3 Basic neural network
architecture
Fig. 4 a Structure of
single-layer perceptron
bstructure of multilayer
perceptron
Performance Analysis Study of Stochastic Computing Based Neuron 259
Fig. 5 Structure of restricted
boltzmann machine
3 Stochastic Computing (Sc) for DNN
The computational functionality of the neuron mainly depends on the weights and the
bias, which are multiplied and added, respectively, with the given inputs. Moreover,
the implementation of the deep belief networks in VLSI is more expensive in terms
of computation because it needs a large number of matrix multiplications and there
is no direct representation for activation unit (sigmoid function, tanh function, etc.)
[14]. It is generally generated with the help of Lookup Tables which in turn occupy
some more additional memory. In order to overcome this, a better functionality is
needed. The following are the challenges and ways to overcome them.
3.1 Computions in Each Neuron
Each neuron performs some computational functions in which the input value of
input nodes or the visible nodes are multiplied with their corresponding respective
weighs. The weights are assigned to each neuron in the first layer. These weights are
multiplied with their respective activations from the input layer and their weighted
sum is computed [15]. After this, bias is added to the weighted sum.
Xi=
N
k=1
WikVk+Bi
Gi=1=σ(Xi)1+exp(−Xj)
260 A. Dinesh Babu and C. Gomathy
where Nrepresents the number of input visible nodes, Vrepresents the activation
value of the input nodes, Brepresents the bias, Wrepresents the weights, Xrepresents
the temporary midway value, and Grepresents the output value of the neuron [16].
Similarly, for a neural network, many such computations take place which results
in the use of a large number of multipliers and adder circuits. In order to minimize
such computational load, stochastic computing can be seen as a better alternative
[17].
3.2 Computation Using Stochastic Computing
•One of the major advantage of using stochastic computing is that the multiplication
of stochastic numbers needs only a single AND gate (Fig. 6). These numbers are
a sequence of random bits. For stochastic multiplication, bit-wise AND operation
is performed between the two stochastic streams.
•X=AND(C,D)=C·D
•Similarly, addition operations can be performed with the help of OR gate or scaled
adder (Figs. 7and 8).
•Z=C·S+D·(1−S)for MUX and
•Y=A+B−A·BforORgate
Fig. 6 Structure of
stochastic AND gate used for
multiplication
Fig. 7 Structure of
stochastic OR gate used for
addition
Fig. 8 Structure of
stochastic MUX used for
addition
Performance Analysis Study of Stochastic Computing Based Neuron 261
•Using stochastic computing, the number of multiplication and addition processes
in the neural network gets reduced to a large extent. Since in stochastic computing
these multiplication and addition operations can be performed by using bitwise
AND gates in unipolar format and addition operations can be performed with the
help of scaled adders or an OR gate.
4 Neuron Design Using Stochastic Computing
Neuron can be built using stochastic computing [18]. For this integer, stochastic
stream is used. It is built by using the summation of the two binary to stochastic
(B2S) stream inputs (Fig. 9). The binary to stochastic unit consists of LFSR, constant
number generator (Fig. 10), comparator (Fig. 11), and a counter unit (Fig. 12).
Fig. 9 Structure of binary to stochastic unit (B2S)
Fig. 10 Structure of LFSR to generate random numbers
Fig. 11 Structure of
comparator to produces 1’s
and 0’s binary output
COUNTER
262 A. Dinesh Babu and C. Gomathy
Fig. 12 Structure of counter
to count the number of 1’s
and 0’s
Binary output
Stochastic input Binary output
Clock
COUNTER
The LFSR unit is used to generate the random numbers, these generated random
numbers are then compared with the constant number. The constant number is gener-
ally denoted by the count of LFSR numbers which are taken into consideration. These
two values are then compared using a comparator. Such that if the value of the LFSR
number is less than the constant number then a high output (1’s) is produced by the
comparator and if the value of the constant number is lesser than the value of the
LFSR number then a low output (0’s) is produced.
The counter is used to count the number of 1’s and 0’s which are produced by
the comparator unit. The final output from the counter is the output of the binary to
the stochastic unit. The next stage is to construct the binary to integer stochastic unit
(B2IS) (Fig. 13). This is constructed with the help of two binary to stochastic units
[19].
Integer addition is performed between the two B2S, and the output of this unit
produces the output of the B2IS unit. This output acts as one of the inputs for the
bitwise AND operator, which performs the multiplication function in the stochastic
Fig. 13 Structure of binary
to integer stochastic unit
(B2IS)
Performance Analysis Study of Stochastic Computing Based Neuron 263
Fig. 14 Structure of stochastic multiplier using bitwise AND operator
computing process. The other input for the bitwise AND operator is the corresponding
weighs (Fig. 14).
Similarly, many such inputs node which performs bitwise operators are fed into a
binary adder, where the summation of the values from various activation units takes
place. Along with that, the bias value is added to it. The final output from the adder
unit is fed into an activation unit. The activation unit can be a sigmoid function or
tanh function [20]. The activation unit determines the on or off state of each node
(Fig. 15).
Thus, when a neuron is built using stochastic computing logic, the number of
multiplier and adder used is replaced with the help of bitwise AND operator and
scaled adder or an OR gate, respectively.
Fig. 15 Structure of neuron using stochastic computing
264 A. Dinesh Babu and C. Gomathy
Tabl e 1 Comparison
between the normal neural
network and stochastic
computing based neural
network
No of multiplication process
No. of multiplier used in
normalneuralnetwork
540 +120 +100 +10 =780
780 multipliers contains 1520
multiplication process
No. of used bitwise AND
operators in SC
780 stochastic multiplication
processes
5 Results and Discussions
With the advent of the stochastic computing process, the number of multiplication
processes and addition processes reduces to a large extent during hardware implan-
tation. Consider a neural network with 540 visible nodes, 120 hidden nodes in the
first hidden layer, 100 hidden nodes in the second layer, and 10 visible nodes at the
output. Therefore to compute the output of the first layer, 540 such multiplication
operations are required, along that their respective weighs occupies a large memory
space. Moreover simultaneous implementation of this network requires a large area
in the silicon wafer.
All these can be reduced with the help stochastic computing logic. Instead of
performing these many numbers of multiplications, in stochastic computing, these
multiplication operations can be computed using stochastic multiplication where
bitwise AND operations are used. And the summation of the multiplication process
is done with the help of an OR gate (Table 1).
6 Conclusion
Thus, implementation of a deep neural network using stochastic computing will
reduce more than half of the area which was previously required for the conventional
neural network [21]. Moreover when it is implemented for a complex function,
precision is more since stochastic computing uses the random probabilistic value
to determine the output. Since the area consumption is reduced their corresponding
power consumption is also reduced to a large extent. Such a type of neural network
using stochastic computing can be used for end IoT devices, which result in better
performance under a constrained resource environment. The processing capability
of the DNN for diverse sets of data can be increased efficiently using the stochastic
computing process. Thus, stochastic computing based DNN performs better than
conventional DNN models.
Performance Analysis Study of Stochastic Computing Based Neuron 265
References
1. J.A. Bullinaria, Introduction to neural networks, Lectures, University of Birmingham, UK,
2004
2. R.F. Molanes, K. Amarasinghe, J.J. Rodriguez-adina, M. Manic, Deep learning and reconfig-
urable platforms in the internet of things, 2018
3. Y. Chen, T. Luo, S. Liu, S. Zhang, L, He, J. Wang, L. Li, T. Chen, Z. Xu, N. Sun, O. Temam,
DaDianNao: a machine-learning supercomputer, 2014
4. H. Li, K. Ota, M. Dong, Learning IOT in edge: deep learning for the internet of things with
edge computing, 2018
5. S. Han, H. Mao, W.J.Dally, Deep Compression: compressing deep neural network with pruning,
trained quantization and huffman coding (2015)
6. P. Maechler, C. Studer, D.E. Bellasi, A. Maleki, A. Burg, N. Felber, H. Kaeslin, R.G. Baraniuk,
VLSI design of approximate message passing for signal restoration and compressive sensing,
2012
7. T. Rzayev, S. Moradi, D.H. Albonesi, R, Manohar, Deeprecon: dynamically reconfigurable
architecture for accelerating deep neural networks, 2015
8. L. Ranganath, J. Kumar, P.S.N. Reddy, Design of MAC unit in artificial neural network
architecture using verilog HDL, 2013
9. C. Sheng, Brodersen, Weste & Harris, 3rd ed. Hardware Multiply/Accumulate (MAC) Unit,
Motorola, 1992
10. L. Cavigelli, L. Benini, A 803 GOp/s/W convolutional network accelerator. IEEE Trans. Circ.
Syst. Video Technol 99, 1–1 (2016)
11. N.K. Bose, P. Liang, Neural network fundamentals with graphs, algorithms, and applications,
1996
12. T. Theocharides, G. Link, N. Vijaykrishnan, M.J. Irwin, V. Srikantam, A generic reconfigurable
neural network architecture implemented as a network on Chip, 2004
13. S. Haykin, Neural networks a comprehensive foundation, 2nd edn. (Prentice Hall Publishing,
New Jersey, VoU, 1999), pp 6–7
14. Y.H. Chen, T. Krishna, J. Emer, V. Sze, Eyeriss: an energy-efficient reconfigurable accelerator
for deep convolutional neural networks, 2016
15. Y.S. Anshika, N. Goee, S. Indu, Neuronal logic gates realization using vedic mathematics,
2015
16. L.A. Basheer, M. Hajmeer, Artificial neural networks: fundamentals, computing, design and
application. J. Microbiol. Methods 43, 3–31 (2000)
17. O.C., Yolcu, E. Bas, E. Egrioglu, U. Yolcu, Single multiplicative neuron model artificial neural
network with autoregressive coefficient for time series modelling, 2018
18. E. Nurvitadhi, J. Sim, D. Sheffield, A. Mishra, S. Krishnan, D. Marr, Accelerating recurrent
neural networks in analytics servers, 2016
19. I. Kouretas C. Basetas, V. Paliouras, IEEE, Low-power logarithmic number system addi-
tion/subtraction and their impact on digital filters, 2012
20. A. Amaricai, O. Boncalo, C.-E. Gavriliu, Low-precision DSP-based floating-point multiply-add
fused for field programmable gate arrays (2013)
21. G. James, D. Witten, T. Hastie, R. Tibshirani, An introduction to statistical learning: with
applications, in R. Springer Texts in Statistics, 2014
Scheduling of Parallel Tasks in Cloud
Environment Using DAG MODEL
Sakshi Kapoor and Surya Narayan Panda
Abstract The scheduling of tasks in a heterogeneous multiprocessor system in the
cloud is still a demanding problem that is being explored by many researchers.
Parallel computing which itself is a research area is now integrated with cloud
computing. In this paper, we present a systematic study of a directed acyclic
graph (DAG) model in a parallel multiprocessor system. Basic concepts of parallel
computing along with issues like task scheduling have been discussed in detail.
Computational solutions can be demonstrated as directed acyclic graphs (DAG)
model having edges and nodes with weights. Each task or job in the DAG has its
own execution time, which incorporates into various processors. This paper gives the
basic idea of parallel processing using DAG. The main objective behind scheduling
using the DAG model is to reduce the finish time or completion time of parallel appli-
cations through proper assignment of tasks among multiprocessors. Most of these
schemes consider the precedence constraints among tasks. Up to now, Direct Acyclic
Graph (DAG) is considered a prominent approach used for modeling the precedence
constraints between the nodes or tasks. There are various scheduling algorithms that
use DAG for scheduling jobs. The most popular heterogeneous earliest finish time
(HEFT) is one of them. A literature review on various task scheduling schemes in
combination with artificial intelligence concepts is also presented.
Keywords Task scheduling ·Cloud computing ·Parallel computing ·Directed
acyclic graph ·Multiprocessor ·Heterogeneous environment
1 Introduction
Cloud Computing is the delivery of services over the Internet. It is a kind of structure
that provides software, hardware, and platform in the form of SaaS (Software-as-
a-service), IaaS (Infrastructure-as-a-service), PaaS (Platform-as-a-service), known
as service models of cloud. The service provider is the one, who provides all these
S. Kapoor (B)·S. N. Panda
Chitkara Institute of Engineering and Technology, Chitkara University, Punjab, India
e-mail: sakshi.kapoor@chitkara.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_23
267
268 S. Kapoor and S. N. Panda
services to the clients via the Internet [1–3]. The parallel execution of jobs is needed
to get the work done more speedily. Parallel computing is basically an architecture
in which various processors execute applications or tasks simultaneously. Parallel
computing can be called parallel processing [2,3]. It divides the workload of large
computations between multiple processors, at the same time. Parallel computing
principles are followed by most of the supercomputers in order to operate properly
and efficiently. Parallel computing makes the execution of various portions of tasks
or applications possible simultaneously in a heterogeneous multiprocessor environ-
ment. A high degree of parallelism enhances the high conduct of existed systems
[4]. Not only parallel, but serial or sequential execution of the jobs or tasks and auto-
mated interprocess synchronization become efficient through parallel processing.
Scientific problems such as modeling of human bones, environmental modeling,
statistical mechanics as well as genetic engineering are very complicated if solved
through simulation as it needs powerful computers. Parallel computing architectures
solve these problems with high efficiency. The division of parallel tasks is analytical
for the indication of distributed computing systems [5–7].
The main purpose behind task scheduling is proper assignment of jobs or tasks to
multiprocessors in order to reduce the finish time of parallel applications [6]. There
are two forms of scheduling problems: one is static and the other is dynamic. Static
scheduling is the one in which characteristics or properties of a parallel program like
task communication, synchronization requirements as well as data dependency are
known before the actual execution of the program. Therefore, a parallel program can
be symbolized by a weighted directed acyclic graph (DAG) consisting of weighted
edges and nodes [7].
Nowadays, scheduling in the multiprocessor system or environment is an active
research area, in which various assumptions are freely recommended. Shockingly,
some assumptions are neither obviously expressed nor reliably utilized by the vast
majority of the analysts. Therefore, it is hard to welcome the benefits of different
scheduling algorithms and quantitatively to assess their efficiency and performance.
To keep away from this issue, we present directed acyclic graph (DAG), understood
portrayal of parallel applications is a DAG (Directed Acyclic Graph) in which jobs or
tasks are symbolized by nodes in the DAG. Inter-task dependencies are represented
by arcs. The actual finish time of the last job or node in the DAG gives the schedule
length or makespan [3,6]. DAG is a task precedence model, where directed edges
give the communication time between the nodes or the tasks and is popular for static
scheduling of tightly coupled tasks or jobs on multiprocessors in a heterogeneous
environment. A task precedence graph can model the program more efficiently in
most of the parallel applications as it conquers all the temporal dependencies between
the jobs or tasks.
When a system is demonstrated as a group of homogeneous processors, where one
task is executed or computed at a time, the goal is to limit the schedule length or finish
time or makespan. Sometimes, even the communication cost is explicitly modeled.
Subsequently, on assigning two neighboring jobs directly to different processors,
the subsequent one is postponed by a measure of time relative to the weight of
the directed edge between them, yet if allocated to a similar processor, no such
Scheduling of Parallel Tasks in Cloud Environment Using DAG Model 269
delay in task happens. Many limited instances of scheduling using DAG have been
demonstrated to be NP-complete, implying that it is unrealistic to search for an ideal
schedule. This is genuinely notwithstanding for very limited exceptional cases, for
example, at the point when communication is free and all the jobs have unit execution
times [8].
The paper is systematized in this sequence: Firstly, the introduction of Parallel
Computing in the cloud with a brief idea about DAG is given. In Sect. 2, we focus
on the need and issues of parallel computing like task scheduling along with their
solutions like DAG. In Sect. 3, the DAG Model along with an explanation about DAG
scheduling and performance parameters are discussed. Lastly, a literature review
on DAG scheduling using different artificial intelligence techniques is presented,
followed by the heterogeneous earliest finish time (HEFT) algorithm for efficient
task scheduling is discussed. The conclusion is discussed in Sect. 4.
2 Parallel Computing in Cloud
It is the set of multi-processors or multiple processing elements that solves problems
concurrently. Problems are decomposed into subproblems which are to be distributed
among multiprocessors and to be solved in parallel or simultaneously [9].
2.1 Need for Parallel Computing
•Larger and complex problems are harder to solve through serial computing.
•The advantage of all nonlocal resources can be taken in case local resources are
finite [10].
•In serial computing, there is a huge wastage of potential computing power which
makes a parallel computing environment utilize better work of the hardware.
•The entire real world is dynamic in nature like numerous things occur at a specific
time, however at different places simultaneously. This information is broadly
immense to handle [7].
•There is a need for dynamic simulation and modeling in real-world data and for
accomplishing the equivalent, parallel processing is the key.
•Complex and enormous datasets and their management can be sorted out by
utilizing parallel processing methodology [11].
•With parallel computing, there will be effective and efficient utilization of the
assets. Hardware is destined to be utilized adequately though in serial computing,
less portion of the hardware was utilized and the rest rendered inactive.
270 S. Kapoor and S. N. Panda
2.2 Issues in Parallel Computing
•Task scheduling of Parallel Tasks: Task scheduling in a heterogeneous multi-
processor environment is found to be a difficult issue [6,8,10,12]. Achievement
of high performance in a multiprocessor system is a key factor in scheduling
parallel tasks. The main objective of task scheduling is to map tasks in parallel on
multiprocessors and order their execution in such a way that a reduced schedule
length is given under the point of confinement of undertaking priority necessities.
Scheduling is considered to be a highly important issue since improper planning
of jobs can neglect to abuse the genuine capability of the framework and can
balance the increase from parallelization [12]. The goal of scheduling is to limit
the completion or finish time of a parallel application by appropriately assigning
the jobs to the processors. There are so many scheduling schemes that are found to
be efficient in parallel task scheduling in a multiprocessor environment. The vast
majority of these plans consider the priority limitations among assignments. Up
to now, Direct Acyclic Graph (DAG) is the popular and significant model utilized
as displaying the precedence constraints among the jobs [13].
•Load Balancing: Partitioning of work correspondingly between several proces-
sors, so that no processor becomes under-loaded or overloaded.
•Reducing Execution time: Noted executed time of last executed job should be
reduced in order to overcome high energy consumption.
Computation time—It is associated with the resources used by the nodes in the
DAG.
Idle time—It is the waiting time for the data from other processors.
Communication time—It is time taken by the processors to transmit communi-
cations among them (Fig. 1).
3 Task Scheduling
Task scheduling can be characterized as relegating the job or task to a processor
for executing at a specific time. The process of task scheduling relies on the
accompanying components:
•Total number of processors.
•Performance or efficiency of processors.
•Mapping of tasks to processors.
•Arrangement of tasks in order to execute on a specific processor.
The above components profoundly rely upon one another and figure out the
advanced outcomes. They all cooperate together and nobody considered the indi-
vidual. Static scheduling is generally done at compile time, in which the qualities
of a parallel program like communication, synchronization prerequisites are known
before the execution of the program [7]. In contrast to static scheduling, dynamic
scheduling takes a few assumptions about the parallel program before the execution
Scheduling of Parallel Tasks in Cloud Environment Using DAG Model 271
Fig. 1 Working of parallel processing [3]
takes place and thus, scheduling choices must be made on-the-fly. The objective of
dynamic scheduling not only includes a reduction in the program finish time yet, also,
the minimization of the planning overhead which comprises a noteworthy segment of
the cost paid for running the scheduler. Parallel computing is an enthusiastic type of
data processing, in which assignments or tasks are separated into components, which
are executed at the same time each on its individual processor [9,11]. The processors
might be orchestrated in homogeneous or heterogeneous conditions. Representation
of task scheduling is efficiently done by Directed Acyclic Graphs (DAG).
3.1 Model for Task Scheduling Problem
Task scheduling helps to limit the overall or total finish time by properly assigning
tasks to the processors and provisioning of execution sequencing of all the given
tasks. Due to efficient scheduling, the precedence constraints among the tasks are
saved. The general completion time of a parallel program is called the makespan.
The objective of efficient scheduling is to reduce or lessen the makespan.
272 S. Kapoor and S. N. Panda
Fig. 2 DAG model [5]
1
25
6
4
3
78
9
10
3.2 DAG Model
The Direct Acyclic Graph (DAG) Model Direct Acyclic Graph is the authentic model
of a parallel framework or parallel system [7]. Parallel functions, as well as parallel
programs, can be described in the form of DAG, where V and E are set of vertices
and edges. Tasks are represented by the nodes which are basically set of instructions
that must be executed sequentially without preemption in the same processor [14].
Every node contains some weight known as computation cost whereas the edge in
the graph gives communication cost between tasks and nodes. The starting node is
the source node or the parent node while the sink node is known as the child node.
A node having no parent is the entry node and a node having no child is known as
an exit node of the DAG [15].
Rules to be followed in an efficient DAG Scheduling:
•Execution will only be as per the level.
•Next-level jobs cannot be executed if there is any job left at the previous level.
•No child can be executed before its parent.
•Level jobs will be ranked as per some rule architecture (Fig. 2).
3.3 Literature Review on Task Scheduling Using DAG
The recent trends in Cloud Computing involve efficient energy utilization for an
optimal number of jobs in the least span of time. Therefore, various researchers
have implemented in different algorithms and artificial intelligence concepts in order
to reduce makespan and energy consumption. A literature survey on various task
scheduling schemes is given in Table 1.
Through Literature Survey, it is cleared that with the use of artificial intelligence
algorithms and techniques with the DAG Model, the energy consumption, as well
Scheduling of Parallel Tasks in Cloud Environment Using DAG Model 273
Tabl e 1 Literature review on task scheduling using DAG
S. No. Year Paper title Scheduling
parameters
Tool used Improvement
1. 2016 Task scheduling
algorithms with
multiple factors in
cloud computing
environment [16]
Cost metrics,
load balancing
CloudSim Improving
performance in
comparison to
conventional
algorithms
2. 2016 Enhanced bee
colony algorithm
for efficient load
balancing and
scheduling in Cloud
[17]
Schedule length,
cost
MATLAB The decrease in
schedule length
which enhances the
overall
performance
3. 2014 A task scheduling
algorithm based on
genetic algorithm
and ant colony
optimization in
cloud computing
[18]
Execution time MATLAB Increase in the
efficiency of the
proposed algorithm
4. 2014 Improved ant
colony algorithm
basedonPSOand
its application on
cloud computing
resource scheduling
[19]
Execution time MATLAB Decrease in
schedule length
5. 2014 Task scheduling
algorithm based on
improved Min-Min
algorithm in a cloud
computing
environment [20]
Schedule length,
load balancing
CloudSim High performance
6. 2015 Credit-based
scheduling
algorithm in cloud
computing
environment [21]
Schedule length CloudSim Decrease in the
schedule length
7. 2012 Dynamic task
scheduling
algorithm with load
balancing for
heterogeneous
computing system
[22]
Resource
utilization,
schedule length
Distributed
algorithm
simulator
The concept of
clustering is used
for resolving load
balancing as well as
decreasing
schedule length
274 S. Kapoor and S. N. Panda
as schedule length in the DAG, will be minimized to a great extent. In parallel
computing, the task scheduling process becomes energy efficient in applying genetic
algorithms and artificial intelligence concepts.
3.4 Task Scheduling Algorithm Using DAG
Heterogeneous Earliest Finish Time (HEFT)
HEFT is an algorithm which considers dependent tasks onto a network of processors
in a heterogeneous environment and taking communication time or cost between the
tasks into consideration. HEFT takes a DAG consisting of jobs or tasks as input,
processors, execution time of tasks on each processor as well as the time to commu-
nicate results from each job to each of its children between each pair of processors
[23–25].
1. Creation of DAG consisting of tasks Tjin the Cloud.
2. Set computation cost (Tj) and communication cost (Tj) between the nodes or
tasks.
3. For j=1toTi
(Calculation of the order of execution of tasks)
4. if (Tj=last task), then
5. O(Tj)=Avg ( Tj) on all processors, where Ois the order of tasks and Avg is
the average of tasks.
6. else
7. O(Tj)=Avg ( Tj)+max (order of task value of predecessor task of current
task) +communication cost (Tj).
8. end if
9. end for
10. Arrangement of tasks in decreasing order in a list based on O(Tj).
11. for (Tin list)
12. Assign Tto the processor which has the min execution time.
13. end for
14. End.
4 Conclusion
Larger and complex problems are harder to solve through Serial computing. To
overcome the problems of serial computing, parallel computing comes into existence.
In parallel computing, task scheduling is a very difficult and challenging issue. DAG
Model can be used for easy and efficient scheduling of parallel jobs or tasks. The
goal of scheduling is to limit the completion or finish time of a parallel application
by appropriately assigning the jobs to the processors. A literature review on DAG
Scheduling of Parallel Tasks in Cloud Environment Using DAG Model 275
Scheduling using an artificial intelligence approach is also given. A popular DAG-
based algorithm HEFT is also discussed in the paper.
References
1. J. Gibson, R. Rondeau, D. Eveleigh, Q. Tan, Benefits and challenges of three cloud computing
service models, in 2012 Fourth International Conference on Computational Aspects of Social
Networks (CASoN). IEEE (2012, November), pp. 198–205
2. J. Ekanayake, G. Fox, High performance parallel computing with clouds and cloud technolo-
gies, in International Conference on Cloud Computing (Springer, Berlin, Heidelberg, 2009),
pp. 20–38
3. T. Hagras, J. Janecek, A high performance, low complexity algorithm for compile-time task
scheduling in heterogeneous systems, in Proceedings of 18th International Parallel and
Distributed Processing Symposium. IEEE (2004), p. 107)
4. C. Cai, L. Wang, S.U. Khan, J. Tao, Energy-aware high performance computing: a taxonomy
study, in 2011 IEEE 17th International Conference on Parallel and Distributed Systems. IEEE
(2011), pp. 953–958
5. G.P. Gibilisco, M. Li, L. Zhang, D. Ardagna, Stage aware performance modeling of DAG based
in memory analytic platforms, in 2016 IEEE 9th International Conference on Cloud Computing
(CLOUD). IEEE (2016), pp. 188–195
6. A.R. Arunarani, D. Manjula, V. Sugumaran, Task scheduling techniques in cloud computing:
a literature survey. Future Gener. Comput. Syst. 91, 407–415 (2019)
7. S. Baskiyar, K.K. Palli, Lowpower scheduling of dags to minimize finish times, in International
Conference on High-Performance Computing (Springer,Berlin, Heidelberg, (2006, December),
pp. 353–362
8. R. Ghosh, Y. Simmhan, Distributed scheduling of event analytics across edge and cloud. ACM
Trans. Cyber Phys. Syst. 2(4), 24 (2018)
9. F. Suter, F. Desprez, H. Casanova, From heterogeneous task scheduling to heterogeneous
mixed parallel scheduling, in European Conference on Parallel Processing (Springer, Berlin,
Heidelberg, 2004), pp. 230–237
10. F. Guo, L. Yu, S. Tian, J. Yu, A workflow task scheduling algorithm based on the resources’
fuzzy clustering in cloud computing environment. Int. J. Commun Syst 28(6), 1053–1067
(2015)
11. X. Zhou, G. Zhang, J. Sun, J. Zhou, T. Wei, S. Hu, Minimizing cost and makespan for workflow
scheduling in cloud using fuzzy dominance sort based HEFT. Future Gener. Comput. Syst. 93,
278–289 (2019)
12. Z. Tang, L. Qi, Z. Cheng, K. Li, S.U. Khan, K. Li, An energy-efficient task scheduling algorithm
in DVFS-enabled cloud environment. J. Grid Comput. 14(1), 55–74 (2016)
13. S. Yassa, R. Chelouah, H. Kadima, B. Granado, Multi-objective approach for energy-aware
workflow scheduling in cloud computing environments. Sci. World J. (2013)
14. M. Mezmaz, N. Melab, Y. Kessaci, Y.C. Lee, E.G. Talbi, A.Y. Zomaya, D. Tuyttens, A parallel
bi-objective hybrid metaheuristic for energy-aware scheduling for cloud computing systems.
J. Parallel Distrib. Comput. 71(11), 1497–1508 (2011)
15. S. Selvarani, G.S. Sadhasivam, Improved cost-based algorithm for task scheduling in cloud
computing, in 2010 IEEE International Conference on Computational Intelligence and
Computing Research. IEEE (2010, December), pp. 1–5
16. N. Bansal, A. Awasthi, S. Bansal, Task scheduling algorithms with multiple factor in cloud
computing environment, in Information Systems Design and Intelligent Applications (Springer,
New Delhi, 016), pp. 619–627
17. K.R. Babu, P. Samuel, Enhanced bee colony algorithm for efficient load.balancing and
scheduling in cloud, in Innovations in Bio-inspired Computing and Applications (Springer,
Cham, 2016), pp. 67–78
276 S. Kapoor and S. N. Panda
18. C.Y. Liu, C.M. Zou, P. Wu, A task scheduling algorithm based on genetic algorithm and ant
colony optimization in cloud computing, in 2014 13th International Symposium on Distributed
Computing and Applications to Business, Engineering and Science. IEEE (2014, November),
pp. 68–72
19. H. Yang, Improved ant colony algorithm based on PSO and its application on cloud computing
resource scheduling, in Advanced Materials Research, vol. 989. (Trans Tech Publications,
2014), pp. 2192–2195
20. G. Wang, H.C. Yu, Task scheduling algorithm based on improved Min-Min algorithm in
cloud computing environment, in Applied Mechanics and Materials, vol. 303 (Trans Tech
Publications, 2013), pp. 2429–2432
21. A. Thomas, G. Krishnalal, V.J. Raj, Credit based scheduling algorithm in cloud computing
environment. Procedia Comput. Sci. 46, 913–920 (2015)
22. D.M. Abdelkader, F. Omara, Dynamic task scheduling algorithm with load balancing for
heterogeneous computing system. Egypt. Inform. J. 13(2), 135–145 (2012)
23. H. Topcuoglu, S. Hariri, M.Y. Wu, Task scheduling algorithms for heterogeneous processors,
in Proceedings. Eighth Heterogeneous Computing Workshop (HCW’99). IEEE, (1999, April),
pp. 3–14
24. H. Topcuoglu, S. Hariri, M.Y. Wu, Performance-effective and low-complexity task scheduling
for heterogeneous computing. IEEE Trans. Parallel Distrib. Syst. 13(3), 260–274 (2002)
25. Z. Cao, L. Zhou, B. Hu, C. Lin, An adaptive scheduling algorithm for dynamic jobs for dealing
with the flexible job shop scheduling problem. Bus. Inf. Syst. Eng. 61(3), 299–309 (2019)
GREENIE—Smart Home with Smart
Power Transmission
N. Noor Alleema, S. Babeetha, V. L. Hariharan, S. Sangeetha,
and H. Shraddha
Abstract The energy consumption is increasing in day-to-day life. Nowadays, the
growth of smart devices has increased, which consumes more energy so people are
shifting to renewable resources for producing energy. Modern homes are shifting
from conventional switches to a centralized control system. Home automation refers
to handling and controlling home. This project mainly focusses on transmitting
energy wirelessly using GREENIE in home automation. The energy from various
sources and it is distributed to the devices through WPT, where these devices are
controlled using IoT, by using Ardunio the various sensor are connected to the devices
and controlled using a web server and smartphones.
Keywords WPT––Wireless Power Transmission ·IoT––Internet of things
1 Introduction
Nowadays, wires are playing an important role in the transmission of electricity
in many sectors such as homes, industry, medical sector, etc. Transmitting electric
current leads to various problems, for example, when the overload current is passed
through wires it leads to the short circuit and wire is easily set to fire, this shows
the highly inflammable nature of wires, secondly the skilled persons like electrician
N. Noor Alleema (B)·S. Babeetha ·V. L. Hariharan ·S. Sangeetha ·H. Shraddha
SRM Institute of Science and Technology, SRM University, Ramapuram, Chennai 603203, India
e-mail: noor25nrs@gmail.com
S. Babeetha
e-mail: babeeths14@gmail.com
V. L. Hariharan
e-mail: hariharanvl04@gmail.com
S. Sangeetha
e-mail: sankarmanickam1965@gmail.com
H. Shraddha
e-mail: shraddhamutha9940@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_24
277
278 N. Noor Alleema et al.
is required to work or else leads to shocks, leakage of current compared to all the
problems occurred due to the usage of wires the complexity of wires is a major one
in homes, industries. The idea is mainly focusing on the problem that is to reduce
the complexity of wire in home, to solve this problem a powerful technology is used
that is called WPT to transmit the current wirelessly.
2 Methodology
2.1 Wireless Power Transmission (WST)
It is the technology used to transfer electric current without the usage of wire by
the principle of magnetic induction. Here the primary coil is connected to a power
source, which produces the magnetic field that is received by the secondary coil in the
form of electric current by this way the transmission is done more efficient and low
cost, there are three ways to transmit which depend on the range of the transmission
such as Short range, Medium range, and Long range.
(a) Short-Range Wireless Power Transmission
The short-range wireless electric transmission works on the principle based on induc-
tive coupling, for example, transformer passed electricity in the form of electromag-
netic fields. When the current is passed to the primary coil which is connected to
the power source directly, it produces a magnetic field inside and outside the coil.
The field produced inside the coil is a strong magnetic field and outside is the weak
magnetic field. The secondary coil (as shown in Fig. 1) receives the fields produced
by the primary coil. The electromagnetic flux produced between two coils induces
the secondary coil.
(b) Medium-Range Wireless Powerless Transmission
Fig. 1 Construction of transformer
GREENIE—Smart Home with Smart Power Transmission 279
Fig. 2 Resonance inductive
circuit
Fig. 3 Microwave power
transmission
This approach works on the combination of resonance and inductive coupling (shown
in Fig. 2), where resonance is used to send the magnetic field from primary coil
to secondary coil more efficiently and inductive coupling is used to increase the
magnetic flux between two coils. The transmission of power in this method is possible
only if both the coils are having the same frequency. This type of power transmission
is mostly used in homes.
(c) Long-Range Wireless Power Transmission
In this transmission, a beam of microwaves or laser waves are used to accomplish the
wireless transmission of electric current. Antennas are used to transmit the electric
current from source to destination. First, the electric current is converted into a
microwave signal which is transmitted through an antenna, signals are received at
destination and it is converted back to the electric current. For the conversion of an
electric current into microwave signals, AC signal is converted into DC signal at the
source, at the receiver end the DC signal is converted to AC signal (Fig. 3).
3 Architechture of Smart Home with GREENIE
The renewable energy is collected from various resources then it is transmitted to
GREENIE Board wirelessly, which is also attached with the main power supply
separately shown in Fig. 3.1. From the GREENIE Board, the power is transmitted to
280 N. Noor Alleema et al.
Fig. 4 Architecture
devices in the Smart Home. The devices are controlled using smartphones with the
help of IoT (Fig. 4).
4 Implementation
4.1 GREENIE Board Architechture
It is divided into two segments: sender and receiver. The sender consists of the
source coil attached to the main power supply. LC circuit is used for resonance
frequency. The receiver side consists of a destination coil along with the LC circuit.
The sender board is connected with the main power supply and receiver board is
connected with the respective devices (as shown in Fig. 3.2) each board is tuned with
the resonance frequency. Once the power starts flowing through board it produces
magnetic field lines which induce the receiver board in this way the power supply
for the house is transmitted. The whole house is connected with various sensors
and they are controlled using a single application each device will be assigned with
the corresponding port number in Arduino board that can be operated using the
application Fig. 5.
4.2 Smart Home Construction
The construction of a smart home is divided into three segments:
GREENIE—Smart Home with Smart Power Transmission 281
Fig. 5 GREENIE Board
•Construction,
•Connectivity, and
•Application.
(a) Construction
The smart home consists of three rooms, where Arduino board controls each room
separately which acts as a slave. Raspberry Pi acts as a master. The I2C signal is
passed to the Arduino boards with specified slave addresses (Figs. 6and 7).
The devices are connected to the Arduino board with a respective sensor. For
example, the living room consists of devices like lights, fans, etc. which is connected
Fig. 6 I2C connection between Raspberry Pi and multiple Arduino boards
282 N. Noor Alleema et al.
Fig. 7 Room configuration
to the Arduino board using relay module. Passive IR sensor is used to detect the
motion. LDR sensor is used to maintain the intensity of light. LM35 sensor is used
to monitor the temperature and humidity of the room.
(b) Connectivity
The SDA and SLC pins in Raspberry Pi are connected to the A4 and A5 pins of the
Arduino board for the I2c connection. Sensor is connected to the digital I/O pins.
Here 2. 2-channel relay connections are used to connect the devices. When Raspberry
Pi requests the Arduino with help of slave address on the other side Arduino board
runs the slave, program, which fetches the inputs from the sensors and sends the
signal to the devices (Fig. 8).
GREENIE—Smart Home with Smart Power Transmission 283
Fig. 8 Overall connectivity of a room
(c) Application
The application is developed and deployed in the Raspberry Pi, which is used to
control the devices. Here we should enter the Arduino board slave address so that it
can be operated (Fig. 9).
Fig. 9 Application interface
284 N. Noor Alleema et al.
5 Conclusion
The Smart Home is built using wireless transmission of electric current, which
enhances the feature of a smart home. The usage of wires can be reduced and the
efficiency of power transmission can be increased. The maintenance cost of the smart
home can be reduced and the wires are replaced by coils, which helps in preventing
the wire’s hazards.
References
1. J. Han, C.-S. Choi, W.-K. Park, I. Lee, S.-H. Kim, Smart home energy management system
including renewable energy based on ZigBee and PLC. IEEE Trans. Consum. Electron. 60(2)
(2014)
2. A. Singh, H. Mehta, A. Nawal, O.V. Gnana Swathika, Arduino based home automation control
powered by photovoltaic cells, in Proceedings of the Second International Conference on
Computing Methodologies and Communication (ICCMC 2018) IEEE Conference Record #
42656, IEEE Xplore ISBN 978-1-5386-3452-3 (2018)
3. A. Sharma, Application of wireless power transfer for home appliances using inductive
resonance coupling. Int. J. Eng. Trends Technol. (IJETT) 16(4) (2014)
4. Y. Wang, J. Qiao, J. Du, F. Wang, W. Zhang, A view of research on wireless power transmission.
J. Phys. Conf. Ser. 1074, 012140 (2018)
5. B. Che, Research on Omni-Direction Wireless Power Transfer Based on Rotation Magnetic
Field (Harbin Institute of Technology, 2014)
6. A. Kurs, A. Karalis, R. Moffatt et al., Wireless power transfer via strongly coupled magnetic
resonances. Science 317(5834), 83–86 (2007)
7. T. Nikola, Apparatus for transmitting electrical energy US1119732, 1914
8. Z. Zhao, Y. Zhang, K. Chen, New progress of mangetically-coupled resonant wireless power
transfer technology. Proc. CSEE, 33(3), 1–13 +21 (2013)
9. Z. Zhao, Y. Zhang , K. Chen, New Progress of magnetically-coupled resonant wireless power
transfer technology. Proc. CSEE 33(3), 1–13 +21 (2013)
10. L. Chen, R. Zeng, Study of wireless power transmission system based on magnetic resonance
communications power supply technology 34(4), 10–12 (2017)
ANAVI: Advanced Navigation Assistance
for Visually Impaired
Arjun Sharma, Vivek Ram Vasan, and S. Prasanna Bharathi
Abstract The advancement in technology over the decades has provided for excep-
tional solutions and aid for the visually impaired. However, they are standalone solu-
tions that target specific problems encountered by visually impaired individuals. This
paper addresses some of the most common problems faced by the visually impaired
and to provide an innovative and inexpensive solution that collectively addresses
the problems. The solution proposed in this paper makes use of RF communication,
AWS (Amazon Web Services) IoT (Internet of Things) core, and Computer Vision to
aid visually impaired people to detect an object and use public transportation inde-
pendently. Computer vision labels identify the obstacles or objects in front to aid the
visually impaired individual to navigate through rough terrain. The problem faced
by the visually impaired is access to public transport, which a majority of them use
to commute daily as it is the only viable mobility option in our country. The visually
impaired are forced to always rely on other individuals to help them travel making
them feel dependent. The other problem is when navigating the streets or in their
home, it is hard to recognize the type of object and the distance at which it is in front
of them. Hence, there is a system that is required to allow them to travel indepen-
dently and help them feel empowered. The handheld unit will be connected to the
bus stop modules using RF module and Py-camera. The bus stop modules will have
an ESP 8266 which provides the Internet connection. The modules and the buses
will be connected using the services of AWS IoT core. Even though there are many
ways for the visually impaired individuals to navigate the streets, as the environment
and the objects around us change rapidly, it is close to impossible for them to spot
using conventional methods and that is where this system comes in handy. For one
A. Sharma (B)
ESP-AS13, Robert Bosch Engineering and Business Solutions, Adugodi, Bangalore, India
e-mail: arj.sharma1997@gmail.com
V. R. Vasan
ESP-AS2, Robert Bosch Engineering and Business Solutions, Adugodi, Bangalore, India
e-mail: vivek435@gmail.com
S. Prasanna Bharathi
Department of ECE, SRM Institute of Science and Technology, Vadapalani, Chennai, India
e-mail: prasannss@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_25
285
286 A. Sharma et al.
such instance, an open drainage, which is 1 by 1 foot wide, cannot be detected using
the tap and go method or sound sensing methods while the system has a pretrained
dataset which can send a voice alert saying there is an open drain at 20 m in front and
a vibration-based alert stimulation is given at the stick handle to help circumnavigate
the open drain.
Keywords RF ·AWS ·Computer vision ·Py-camera ·Visually impaired ·IoT ·
ESP8266
1 Introduction
Eye Blindness or the number of visually impaired people around the world stands at
about 1.3 billion. The studies have found that visually impaired people suffer from
secondary road traffic accidents and there are times when they get lost in a crowded
place, in some instances they fall due to objects placed in the middle of walk paths.
The main factor of such untoward incidents is the lack of guidance or help provided
by the general public. There have been a lot of emerging technologies but never has
there been a combination of such principles used for the lesser fortunate in our nation.
With emerging technologies like machine learning and artificial intelligence, there
is a huge scope to put such technology into work in a much crucial situation and give
the impaired with an eye of technology to assist and guide them [1]. With the advent
of IoT and the widespread availability of internet connectivity, we propose a system
that is easy to implement over the existing infrastructure with minimal changes. The
need of an established connected system is to connect the unconnected and offer
people with visual impairment a mean to see things [1], instead of walking towards
their destination they can afford a public transport with a click of a button [2]. The
system is ideated keeping in mind the development of technology, which caters to
the needy and gives a solution to a problem statement or a purpose.
The existing solutions for this particular problem include the usage of RFID tags
to identify the required bus or its user, a bus arrival based detection by the use
of RF waves by the visually impaired person [3]. Though this method has been
proved to work, it has a lot of disadvantages such as cost and change in existing
infrastructure to serve a select group of people. RFID is a complex but trivial method
to cater to the visually impaired person to move from point Ato point B[4]. The
use of sensor and artificial based system to provide assistance to visually impaired
people can be seen from the plethora of works in the reference section below [5–
11]. However, technology has not served large communities to solve the common
problem such as obstacle avoidance, safe and convenient transportation to the visually
impaired people. The common problems found by studies indicate that the price
of this life-changing technology is higher or the device gets bulky and difficult to
manage.
Public transportation is the only viable mode of transport for the majority of the
visually impaired. Hence, it is imperative that there exists a system that is inclusive
ANAVI: Advanced Navigation Assistance for Visually Impaired 287
of them as well. The current scenario forces the visually impaired to rely on another
individual for guidance. We offer a device and an infrastructure that will empower
them and make them feel independent. We propose to design a system that integrates
all the buses deployed and the bus stops through IoT. The visually impaired individual
will be provided with a Handheld Unit that allows them to enter the bus number that
they wish to board. All other solutions revolve around the similar concept of some
sort of usage in the RFID field of technology. Clearly, none of these measures have
been scaled or implemented on a national level to serve the users.
Our solution involves the use of a handheld device which will be given to the
visually impaired user to input the number of the bus through voice recognition
or manual input and can use the onboard Computer Vision based module to detect
whether there is an object in the front. The distance and the type of object will
be captured and analyzed to give a voice-based alert. When a bus-based request
is requested, a request is sent to the bus stop from which they wish to board the
bus through radio frequency. The modules placed at the bus stops (with internet
connectivity) will be connected to a gateway device (server). This module will publish
the message to the gateway, where the buses will be at the receiving end of the
gateway, enabled through an Android application at the disposal of every bus driver.
The bus driver will now know the location of the visually impaired individual and
will be given the choice to accept or clear the request and an auditory signal will be
heard to indicate the location of the door. The particular devices and technologies
used to provide the aforementioned system will be explained further.
2 Methodology
The method which is proposed deals with the identification of real-time objects when
a visually impaired person uses the walking stick. The walking stick is multifunc-
tional as it can be used to communicate with local transport to board the vehicle and
identify objects while walking to their destination.
The working of the product is split into three parts:
1. Handheld Unit—User to Bus stop.
2. Bus Stop Module—Bus stop to Amazon Server.
3. ANAVI Driver app—Amazon Server to Bus driver.
2.1 Handheld Unit
This subsystem consists of a Matrix Keypad, Ultrasonic Sensor, NRF module, and a
Py-camera to obtain the inputs and transfer the data to the Bus stop receiver as well
as to detect objects and measure the distance.
288 A. Sharma et al.
2.1.1 Matrix Keypad and NRF Module
The Matrix Keypad will be present on the top of the module. The connector lines
enable the keypad to be used directly. We check the logic levels of all the rows and
columns infinitely. Any changes in the state of the rows or columns will be detected
in the code. Each key has a unique combination of a row and a column. Hence
by obtaining the row and column whose states have changed, we can obtain the key
pressed by the user. The visually impaired individual can identify all the characters by
use. This is required as it act as an easy method of obtaining the required bus number
which the user wishes to board. This will be processed by the microcontroller and
the data will be transmitted to the bus stop utilizing RF communication.
We utilize a cost-effective NRF module to transmit the data (bus number) to the bus
stop receiver. The module is an 8-pin device, with each pin having a unique function
to implement the communication protocol. The module requires a constant supply
of 3.3 V. The module runs on a 2.4 GHz bandwidth and provides a range of 100 m.
It can be programmed as a transmitter or a receiver based on the requirements. The
module can open up to 6 different channels for communication. We are employing
this module at two locations. Firstly, it is programmed as a transmitter on the HHU
module. Secondly, it is programmed as a receiver at the bus stop module.
2.2 Py-camera and SENIC Ultrasonic Sensor
The Py-camera and the ultrasonic sensor module is connected to a Raspberry Pi
for Image and Distance Processing. Object detection is done using OpenCv and a
pretrained dataset from the open-source Tensorflow Detection Model Zoo. This has
some restrictions due to the speed and accuracy of the data detection. To counter such
discrepancies, SSDLite-MobileNet is used to train the model and to get fast image
recognition. To support high accuracy, any model with small and faster quantized
kernels that allows fixed-point matching is implemented and datasets are trained
in the PC and an export is used in the Pi to make the model optimized for faster
detection.
The Ultrasonic Sensor can measure a distance that has a range of 70 feet and an
accuracy of 0.833 inches. In this project, we use it to tag the image detected so that it
enables the visually impaired person to know the distance of objects, buildings, and
any other obstacle in front of them. An earphone can be connected to the onboard
Raspberry Pi module to get Audio feedback which the visually impaired person can
use to navigate. A vibrator is placed on the stick to send an alert if an object is
detected too close to the individual (Fig. 1).
ANAVI: Advanced Navigation Assistance for Visually Impaired 289
PyCamera
Raspberry Pi
Vibration Module
Ultrasonic Sensor
Earphones
Fig. 1 Raspberry pi interfaced with Py-camera for computer vision and ultrasonic sensor for
distance calculation
2.3 Bus Stop Module
The bus stop module acts both as a receiver and a transmitter. It receives data from
the HHU and it is sent through serial communication to the Wi-Fi module. This is
provided with an internet connection to transmit data through a wireless connection
to the server which will be accessed by the bus.
2.3.1 Wi-Fi Connectivity Board-ESP8266
The ESP8266 is a cost-efficient Wi-Fi-enabled microchip with a full TCP/IP stack
and computational capability, embedded onto the NodeMCU board. We require this
device to access the Amazon server to publish the data received from the previous
subsystem. As it has internet capability, it will wirelessly transmit data to AWS IoT
gateway. It has storage capability and 9 digital GPIOs. It is powered with a supply
voltage of 3.3 V. The code to identify, connect and transmit data to the gateway is
run on this device.
290 A. Sharma et al.
Fig. 2 Bus stop module block diagram
This module will be enabled with wireless communication capability. Hence, it
can transfer the data from the client to the gateway server endpoint provided by
Amazon. The bus stop receives and transmits data. Accommodations are made to
store data in a buffer in case of multiple entries simultaneously (Fig. 2).
2.4 ANAVI Driver App
The final subsystem will be on the mobile phone of every bus driver. An Android
application will be deployed on their devices. This app will run a Graphical User
Interface. It will also serve as a subscriber to the server gateway and will be run in the
background. The incoming data will be received from the server vis-à-vis AWS IoT
services. This data will present to the driver about the presence of a visually impaired
person along with the name of the bus stop. The driver needs to indicate his arrival
at the bus stop. This is enabled by providing the ACCEPT or CLEAR command to
the driver.
The ACCEPT push button on the GUI will result in giving an auditory output to
the user indicating the exact location of the entrance to the bus (Fig. 3).
2.5 IoT Enabler
Our IoT enabler provides these services through AWS. The details of AWS imple-
mentation will be mentioned ahead. AWS IoT cloud platform provides a secure,
two-directional communication between hardware objects (such as smart sensors,
ANAVI: Advanced Navigation Assistance for Visually Impaired 291
Fig. 3 ANAVIdriverapp
digital actuators, embedded devices with edge connectivity, and smart appliances),
the AWS Cloud has encryption of MQTT. We have implemented the data transfer
using the MQTT protocol.
Devices acknowledge the sent commands by sending ACK messages through a
JSON format when an MQTT protocol is used. There can be multiple MQTT-based
Connection Topics relating to a single group name that identifies the unique device
which is updated based on its state and it sends back an acknowledgment for the state
change. When a message is published, the message is dispatched to the message
broker (MQTT), which ensures sending the messages published to that particular
MQTT group to reach all the clients subscribed to that group. The IoT broker uses
topics that routes messages to receiving clients from the sender. The topic to publish
to or subscribe from can be entered in the dialog box as shown and the MQTT broker
can transfer the messages from end to end. This process of subscribing and publishing
is included in our code at every bus stop module exclusively (Fig. 4).
3 Implementation and Result Analysis
In the proposed work, the use of a handheld device was implemented which will
be given to the visually impaired user to input the number of the bus through voice
292 A. Sharma et al.
Fig. 4 Module block diagram
recognition or manual input. The use of an onboard Computer Vision based module
to detect whether there is an object in front is implemented. The distance and the
type of object will be captured and analyzed to give a voice-based alert. When a
bus-based request is processed, the request is sent to the bus stop from which they
wish to board the bus through radio-frequency communication. The modules placed
at the bus stops (with internet connectivity) will be connected to a gateway device
(server). This module will publish the message to the gateway, where the buses will
be at the receiving end of the gateway, enabled through an Android application at
the disposal of every bus driver. The bus driver will now know the location of the
visually impaired individual and will be given the choice to accept or clear the request
and an auditory signal will be heard to indicate the location of the door of the bus
once the bus has arrived at the corresponding bus stop. The particular devices and
technologies used to provide the aforementioned system will be explained further.
3.1 Design Verification
To verify the design of the blind stick and the camera position along with the ultrasonic
sensor placement for optimal performance and to avoid low lying objects and wrong
calibration, simulations were run to study various scenarios that could be encountered
by the visually impaired individuals. The developed prototype was also tested on the
field with visually impaired individuals handling the prototype equipment. Figure 5
shows the placement of the aspects being calibrated. The collective data from the
simulations as well as the user experience from the visually impaired individuals was
processed and the following results were tabulated for optimal performance (refer
Table 1).
The app is tested for the range of requests processed by the AWS IoT core system.
The driver’s app must come under a range of 10 km of the visually impaired person to
get a request to accept or clear. Figure 6shows the request from a visually impaired
individual being transmitted from the server to the bus driver’s app. The notification
is displayed on the app with the bus stop name details and the options to Accept or
Clear the request. The reliability of internet connection and the latency for request
processing were also tested which resulted in improving our Android app code to
handle unexpected exceptions to increase the performance. The design and architec-
ture were verified under various scenarios and the experiments concluded with the
ANAVI: Advanced Navigation Assistance for Visually Impaired 293
Fig. 5 Walking stick attached with the modules
Tabl e 1 Blind stick
calibration with mean stick
length 130 cm
Average
precision
Intersection over
union
Area Max detections
Average
precision
0.49:0.95 All 0.453
Average
precision
0.50 All 0.0668
Average
precision
0.75 All 0.4721
Average
precision
0.49:0.95 Small 0.267
Average
precision
0.49:0.95 Small 0.470
Average
precision
0.49:0.95 Medium 0.589
Average
recall
0.49:0.95 Large 0.337
Average
recall
0.49:0.95 All 0.531
Average
recall
0.49:0.95 All 0.577
Average
recall
0.49:0.95 Small 0.368
Average
recall
0.49:0.95 Large 0.702
Average
recall
0.49:0.95 Medium 0.604
294 A. Sharma et al.
Fig. 6 ANAVI interface
successful working of the driver app to process requests as well as provide auditory
signals at the acceptance of corresponding requests.
3.2 Result Analysis
SSD_Mobilenet_V2_fpn_Coco is fast compared to other models by giving about
525 ms. The drawback of Raspberry Pi running on such powerful datasets is it heats
the Raspberry Pi, though a heat shrink can solve this problem. Images trained and
tested were of the resolution 420 * 240 in RBG though the size is big, it would
improve the model accuracy and reduce the recall. The camera was running on a 4–7
FPS and this can simultaneously capture about 25 objects and label them with an
accuracy of 85.69% according to the testing datasets (Table 2).
ANAVI: Advanced Navigation Assistance for Visually Impaired 295
Tabl e 2 6 data object
experiment results Position calibration Remarks
Ultrasonic sensor 24.6 cm From below
Vibration sensor 4.5 cm From the handle
Camera AOV 62.2 ×48.8° Front angle
Camera position 14.8 cm From the top
Microcontroller module 30 cm From the top
4 Summary
ANAVI is a user-friendly device that works primarily on computer vision and RF
transmission to send crucial information to the user and cloud. This solution has
solved the most important aspect which is providing essential visual and navigation
assistance to the visually impaired individuals in a simple and cost-effective way.
Field trials show that the camera has a wide focus area, which leads to multiple
object detection which gives the user more data to process. For better user decision-
making, in future developments, we can narrow down the focus area of the camera.
This solution can also be scaled up by improving the infrastructure of the connectivity
of the device to the cloud and expanding its reach to other transport sectors. This
serves as a precursor to the Intelligent Transport System toward which the world is
progressing.
From the results of our experiment, we can observe that the accuracy of image
processing to detect obstacles is improved to 85.69% resulting in better performance
compared to other developments in the field. Further improvements can be made
by learning more image data sets to recognize real-life obstacles with improved
accuracy. With this solution, the objectives achieved include enabling the visually
impaired to access public transportation independently with minimal changes to the
existing infrastructure.
References
1. M. Hersh, M. Johnson, Assistive Technology for Visually Impaired and Blind People (Springer
Publishing Company, 2008)
2. D. Dakopoulos, N.G. Bourbakis, Wearable obstacle avoidance electronic travel aids for blind:
a survey. IEEE Trans. Syst. Man Cybern. Part C (Appl. Rev.) 40(1), 25–35 (2010)
3. M.Z.H. Noor, I. Ismail, M.F. Saaid, Bus Detection Device For The Blind Using RFID
Application (IEEE Xplore, 2009)
4. K.N. Sushma, B.C. Smitha, B.N. Kiran, Implementation of RFID for blind bus boarding system,
in A Treatise on Electricity and Magnetism, ed. by J.C. Maxwell, ICCSP, vol. 2, 3rd edn.
(Clarendon, Oxford, 1892), pp. 68–73
5. S. Sivan, G. Darsan, Computer vision based assistivetechnology for blind and visually impaired
people. https://doi.org/10.1145/2967878.2967923
296 A. Sharma et al.
6. S. Abbouda, S. Hanassya, S. Levy-Tzedek, S. Maidenbaum, A. Amedi, Eyemusic: introducing a
visual colorful experience for the blind using auditory sensory substitution. Restorative Neurol.
Neurosc. 32(2) (2014)
7. A. Geiger, M. Roser, R. Urtasun, Efficient large-scale stereo matching, in Asian Conference
on Computer Vision (ACCV) (2010)
8. R. Tapu, B. Mocanu, A. Bursuc, T. Zaharia, A smartphone-based obstacle detection and classi-
fication system for assisting visually impaired people, in 2013 IEEE International Conference
on Computer Vision Workshops (ICCVW) (2013), pp. 444–451
9. L. Chen, B.-L. Guo, W. Sun, Obstacle detection system for visually impaired people based on
stereo vision, in 2010 Fourth International Conference on Genetic and Evolutionary Computing
(ICGEC) (2010), pp. 723–726
10. F.L.M. Milottta, D. Allegra, F. Stanco, G.M. Farinella, An electronic travel aid to assist blind
and visually impaired people to avoid obstacles, in International Conference on Computer
Analysis of Images and Patterns (2015), pp. 604–615
11. A. Morar, F. Moldoveanu, L. Petrescu, A. Moldoveanu, Real time indoor 3D pipeline for
an advanced sensory substitution device, in International Conference on Image Analysis and
Processing (2017)
THIRD EYE—Shopfloor Data
Processing and Visualization Using
Image Recognition
S. Prasanna Bharathi, Vivek Ram Vasan, and Arjun Sharma
Abstract This paper describes how superimposed data using an Augmented Reality
based system is appropriate in manufacturing and warehousing locations to improve
operations. Access to Data is penultimate in the manufacturing industry, and giving
the required data in real-time based on the component or a machine can help in
increasing productivity with an added insight of using, a handheld mobile phone,
thus avoiding any interaction with computers. Production and machine details give
the shop floor managers the flexibility to analyze machines at the shop floor, which
aids on time defect resolution and machine status. This paper focuses on object
detection based data analytical application, which can differentiate between two or
more similar machines using feature extraction methods and texture mapping to
enhance object detection in the used game engine and also to develop the application
and map those to the database of that machine, hence provide an analytical infor-
mation in a hands-free manner. The effectiveness of the algorithm was tested on
machines that look similar or near to identical was differentiate using our method-
ical image processing algorithm and the data pertaining to the machine is processed
on the cloud. Improvements to process the data on the device to improve overlay
efficiency and speed was experimented using various studies. Compared to various
image processing software, THIRD EYE was rated 93.6% efficient in image recog-
nition and data overlay to the user was one new feature we provide as an application
to the user.
Keywords Augmented reality ·Object detection ·Feature extraction ·Texture
mapping
S. Prasanna Bharathi (B)
Department of ECE, SRM Institute of Science and Technology, Vadapalani, Chennai, India
e-mail: prasanns@srmist.edu.in
V. R. Vasan
ESP-AS2, Robert Bosch Engineering and Business Solutions, Adugodi, Bangalore, India
e-mail: vivek435@gmail.com
A. Sharma
ESP-AS13, Robert Bosch Engineering and Business Solutions, Adugodi, Bangalore, India
e-mail: arj.sharma1997@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_26
297
298 S. Prasanna Bharathi et al.
1 Introduction
Augmented Reality (AR) is a game-changing technology in the manufacturing
industry which are used as part of the Industry 4.0 solutions. This technology gels
with automation and digitizing the industries by making the process simpler and
easily understood. Mobile augmented reality can be accessed on smartphones as it
extends and enhances the user experience, thus making it more customer friendly.
The current use cases in the manufacturing industries are for process training and
shop floor employees are replacing maintenance manuals for service technicians.
This methodology is adopted by industries to ensure process training is appreciated
and effortless to explain, giving the operator a hands-on experience, which proves
to be more efficient than the current training practices. Service technicians exten-
sively use AR devices instead of manuals, as the steps are elementary, exclusively
because of the crunch of operating space and visual aid which helps in identification
of the correct part to be serviced, apart from the step by step explanation. This is a
game-changing technology since there’s no necessity for the operator to have prior
experience in doing such tasks and provides a visual guide. Augmented Reality also
holds several applications in other industries and functions. Product advertisements
can also adapt to this technology to provide the user an interactive platform. The user
can perceive the product in an uncomplicated virtual environment.
AR technology superimposing data onto the real world the application is yet to be
explored for manufacturing usage and needs extensive research [1]. Virtual Reality
uses CAD-based 3D objects that allow the operator to experience and interact with
virtual environments [2–4]. The operators will be able to interact with the virtual
world which could be a simulation of a complex process or the viewing of volumetric
data [5]. Augmented Reality (AR) aids the user to visualize the reality with virtual
objects superimposed on the live feed. Virtual data can provide significant informa-
tion that the user cannot perceive with their own processing capability. Augmented
Reality based applications are used in various areas such as training and entertainment
(Thomas et al. [6]), social topics (Dähne et al. [7]) and product assembly, machine
downtime and repair of manufacturing equipment (Curtis et al. [8]; Navab et al. [9]).
AR technologies have been implemented in a few scientific visualization tasks such
as HUD, ultrasound scans, and in the healthcare sector for the visualization of CT
scans or MRI. In the above-mentioned applications, Augmented Reality is a unique
platform for data visualization for enhancing the understandability of data among
the users, but it has not been employed to embed insightful information pertaining
to the real object itself [10]. From the above-cited paper, it can be observed that
AR visualization increases the understandability and interaction of the displayed
data, however, there is no implementation of the AR technology in the industrial and
manufacturing field [11], in which effective ways are there to view and analyze data.
However, the problem of image/object identification in real time has been exten-
sively studied. Although there are many platforms to do such processing, there is
limited or no special package for AR-based applications, which can do recognition
on a very precise scale. Nevertheless, applications have come out with superimposed
THIRD EYE—Shopfloor Data Processing and Visualization Using … 299
images on real-time video feed which is popular despite several restrictions in terms of
3D plane and the quality of image or object it tracks. Currently, there are no systems
which can give real-time information of a process in the manufacturing industry,
and simultaneously display data when the detected object or machine changes, even
though they look alike. The main motive of object detection and recognition is to
discover the scene and find the target. Firstly, for the classification of information,
we have to rely on the enhanced image information for better detection and ordering
of objects. Research shows there was a lack in parallelism strategy to develop a
mobile-based image processing algorithm to deal with limited CPU-GPU memory
utilization without compromise of memory [12], in this work the ability to parallelly
process data and perform image recognition on the target. Image processing using
OpenGL ES 1.1 is studied [13] to save computing space by the efficiency of such
platform for processing decreases the accuracy of recognition, and hence advanced
yet compact versions of OpenCV paired with Vuforia engine is studied for compute
efficiency.
In the present paper, a method is proposed that identifies similar machines and
gives the current production data of the machine which is crucial to line managers
at a particular interval of time. The recognition and detection of the image targets
are done on OpenCV, which runs in parallel to the Vuforia game engine to classify
the images. The image targets are matched with respect to the feature value. This
method is put into Microsoft HoloLens and an AR-enabled mobile phone which
the operator uses to get insights into the production data as he scans around the
production floor using his AR glass. Studies have shown that users are well adapted
when provided with enriching data which is spatially aligned thus not affecting the
view of the user [14]. The system exploits a cyber-physical system, which enables
the user to seamlessly be connected with the database in a basic way, to view data of
different machines instantly. AR devices are popular, yet expensive to be scaled to
such business verticals since it gives possibilities for exploring Augmented Reality
using a mobile device to guide and visualize data.
2 Methodology
The method which is proposed deals with image identification of similar objects,
predominantly used for industrial manufacture (focuses mainly on production
machines on the production floor). The application is designed in such a way that, it
can differentiate two similar objects depending upon various parameters which tend
to differ within a similar equipment/machine when placed in a dynamic environment
such as a production floor. An example of such similar machines is depicted in Fig. 1.
The first step for image processing is the application should be able to process
two resembling objects thereby and able to differentiate based on a few features.
The design should also be able to cater to fast-moving object/camera feed in order
to deliver accurate data based on the image detected. Various methods are incorpo-
rated into the design such as contrast-based feature extraction, contour detection and
300 S. Prasanna Bharathi et al.
Fig. 1 Similar datasets for image training
texture mapping. The algorithm is divided into 2 stages to simplify the detection of
similar objects. Contour detection and texture mapping are used to identify objects
based on the contour and the texture, which helps to narrows down the detection
process. Mobile platform architecture is shown in Fig. 2.
Fig. 2 Mobile app
architecture on iOS/Android
THIRD EYE—Shopfloor Data Processing and Visualization Using … 301
Fig. 3 Classification of feature points of similar images
Once the machine or object is detected the images with a similar algorithm is run
to check if there is any matching object and further it is narrowed down to one unique
object. Feature extraction uses a search engine that checks for unique elements in an
image, which can be based out of the high-contrast spots due to the lighting condition
in that area or certain stickers or markers for that particular machine. There can also
be some wiring density, which makes it different from a similar machine. These
feature points are stored once the engine is trained with the image sets and forms
a defined constellation based on the feature points it detects. Once the application
is narrowed down on the object, the maker is recognized by the constellation. The
above process is depicted in Fig. 3.
The detection algorithm and the application are tested against the conditions as
follows:
•The detection algorithm should be able to operate in real time.
•The algorithm must be able to classify two similar objects based on certain feature
points.
•The efficiency is tested against varying environmental conditions such as
background change, light variations and wiring density changes.
•The detection speed should operate within 800 ms.
•The algorithm should work efficiently with human or object disturbance at
permissible levels.
302 S. Prasanna Bharathi et al.
3 Proposed System Design
3.1 Unity 3D and OpenCV
Augmented reality requires a powerful and efficient game engine like unity 3D, to
put into use, the mobile’s webcam for real-time object detection using its inbuilt
computer vision-based libraries. Unity host flexibility of the environment where
the application is developed, every layer can have a script to dedicate a certain
functionality. Game objects are tagged for the text values which are going to be
displayed post-image detection. The application will have communication protocols
for simultaneous communication with the cloud database, which is Azure cloud. For
unity instance and the image detection algorithm which works on unity platform to
communicate with each other, UDP (User Datagram Protocol) is used over TCP as
it doesn’t expect any acknowledgment token before sending the next packet. This
method is opted to optimize speed for the application to identify the image within
the stipulated time frame. This offers lower bandwidth and latency (Figs. 4and 5).
OpenCV in the unity platform uses Python API to detect the image based on the
algorithm implemented. An asset identification will be sent in terms of image targets
that are compatible with Vuforia inbuilt along with a unity platform to identify the
object to a predefined port using sockets. This is used to trigger a game object, which
syncs with the cloud and displays information pertaining to that object (Fig. 6).
The detection and processing phases performed by OpenCV is as follows:
1. Frames are captured from the camera on board the device as images.
2. The image is stored and a Gaussian blur is added to reduce the noise of the image
and for further processing.
3. A color filter is added by replacing the image with black and white so that an
outline of the image is obtained. Contours are detected and are marked on the
image.
Fig. 4 Feature extraction and classification
THIRD EYE—Shopfloor Data Processing and Visualization Using … 303
Fig. 5 Software architecture
Fig. 6 Object detection––basic scheme
4. Image is then scanned for feature points and these points are later represented
and stored as a constellation for identification of the object.
5. Image target script is invoked to match the feature and contour detection points
to a specific image target stored in the application (Fig. 7).
6. Image target ID is communicated through the UDP protocol to unity platform.
304 S. Prasanna Bharathi et al.
Fig. 7 Probability of false detection feature and texture-based biasing––blue dotted line
3.2 Port Forwarding and Game Object Setup
Unity platform handles the inputs via UDP and creates a trigger to generate the
instances of the game object. A game object is a component that can consist of
text, 3D models, and other functions that can also create legacy components using
the Scripting API in unity. An outline marker is marked on the identified object to
give a visual representation, so that the object is detected. The function which is
optional can be implemented via script. To receive data from the UDP layer go to
the project hierarchy under managers tabs −> player controller, the script named
Player Controller. CS is modified to receive the commands (Script is shown below).
A Boolean variable is added to check if the variable is initialized on each frame.
Based on that trigger action, the flag will be set. The trigger enables a handshake
technique to receive data from the database, which uses the object ID to parse data
back to the text game object.
Commands to be Added in the Script File
Thread recieveThread;
UdpClient client;
Int port;
Public Gameobject Player;
InputFielddataViz;
bool data;
Back in the Text Game Object, a script file is added to invoke the script to receive
the trigger from the UDP script. Boolean variable is initialized to false, and it sets to
true when a trigger is received from the OpenCV Python code. A text game object is
THIRD EYE—Shopfloor Data Processing and Visualization Using … 305
attached to the player controller, which stores the data and is parsed back from the
database when the trigger is set.
Script File
The port is initialized in this script.
Void start()
{
port =5066;
data =false;
dataViz =gameObject.GetComponent < InputField > ();
InitUDP();
}
3.3 Scene Creation and Database Connection
The scene can be created using a real-time feed of the camera as well as for the
representation of the object, to sense if it has been detected or not. There are panels that
change color from red to green when the object is detected. Enabling the back camera
in the project properties’ settings so that the camera opens when the application starts.
Text game objects are mapped to the 3D plane accordingly. The text game object also
has a script file that parses data from the cloud database based on the objectID. For
example, JSON parsing is used in terms of object name/objectID, which is parsed
from DB to the application. Data is superimposed on the realtime feed from the
camera, giving it an augmented reality feel.
For the information to be displayed on the screen 2D or 3D, text objects can be
used based on the software complexity. The detection panels can be downloaded
from the unity asset store or can also be used from PNG images. The text object
and image target must be destroyed using command Destroy GameObject() so that
it saves space and avoids rendering of many image targets which is active in the
background and may slow down the application (Fig. 8).
The data parsed uses the code as follows:
The data parsed uses the code as follows:
void Start()
{
stringurl=“http://10.10.0.25:8080/TestRESTService/MachineData.svc/Yield/{NET05000206}";
WWWFormformDate = new WWWForm();
formDate.AddField ("name ", "product");
WWW www = new WWW(stringurl, formDate);
StartCoroutine(request(www));
}
IEnumerator request (WWW www)
{
yield return www;
displayText.text = www.text;
}
306 S. Prasanna Bharathi et al.
Fig. 8 Unity scene setup with script files in the background
4 Implementation and Results
The final application consists of three layers, the primary layer of image detection
followed by, the network layer which transports the data between both interfaces and
initiates a trigger to the third layer of cloud computing. The speed and efficiency are
tested layer-wise and, the object detection efficiency is marked to 94% while, at the
network layer there is 98% of packet delivery. In cloud computation, the later speed of
fetch and display is about 300 ms. Various environmental parameters are accounted
and tested during the application testing. From these tests, light and materials of
objects which cause reflection and color of the object whether it’s light or dark
doesn’t take the features of the object, which leaves it undetectable by the algorithm.
To counter these abnormalities, we have adopted the match criteria threshold based
on the number of feature points for feature matching to accommodate these changes.
The final application has to be compressed due to the heaviness of the package due
to OpenCV embedded into the application. Graphics and other display texture were
limited to few 3D objects and mostly 2D objects were applied.
The object detection algorithm consists of feature matching, followed by contour
and texture mapping to increase the match probability. However, rules are evaluated to
make the algorithm flexible in case of 90% match is not possible due to environmental
impacts such as light conditions that have the ability to change the image feature
extraction properties. Unity platform has also been tweaked not only to serve as a
game and AR platform but also to be used as a gateway to view vast operation critical
data in industries (Fig. 9).
THIRD EYE—Shopfloor Data Processing and Visualization Using … 307
Fig. 9 Scene after object detection and data is parsed
5 Conclusion
Finally, the basic goal of the design is achieved very closely. The application is able
to identify similar machines and is also able to display operation crucial information
on a superimposed way to the user. AR devices can be integrated with this solution
or it can also fit into the phone segment as an Android/iOS app. We have presented
a flexible and extendable AR-based environment for the implementation of complex
image classification environments in this application.
The system logs which were taken during the image processing when running on
the smartphone shows parallelism strategy was used to balance CPU–GPU computa-
tional needs and does not overload the device. The test was run on Samsung mid-range
smartphones. The application was found to be more efficient than other similar appli-
cations done by the referenced researchers and thus giving an operational efficiency
of the application stood at 92.3% for image recognition and data computing stood at
86%.
Due to the complexity of the image detection package in the application, it tends to
crash frequently when the background process runs simultaneously which the authors
are focusing on the further scope of improvement. The scope for implementation
is integrating single-shot detectors and Mobile Nets for fast and efficient object
detection. To achieve high detection rates, feature point extraction can be used for
better contrast and texture pattern differentiation to get significant feature points,
hence improving the image classification. To reduce software complexity, we can
also use single-shot detectors and send the captured image over the cloud to a cloud-
based GPU for image processing and parse the machine information back to the
scene.
308 S. Prasanna Bharathi et al.
In comparison to the object recognition, the game engine operates by using a
scanned 3D model to detect the object. The conventional combined platform to recog-
nize and classify images using a combination of features, contour and texture can
achieve higher than 94% efficiency under normal conditions, while we are extending
the capabilities to recognize objects under bright light using other techniques such
as Gaussian blur after grayscale image conversion. Post the erosion and dilation of
the image, the extract information has been applied to label the image sets.
References
1. R. Azuma, A Survey of Augmented Reality (ACM SIG-Graph, 1997), pp. 1–38
2. A. Van Dam, A. Forsberg, D. Laidlaw, J. LaViola, R. Simpson, Immersive VR for scientific
visualization: a progress report. IEEE Comput. Graph. Appl. 20(6), 26–52 (2000)
3. W. Sherman, A. Craig, Understanding Virtual Reality: Interface, Applications and Design
(MorganKaufmann Publishers, 2003)
4. G. Giraldi, R. Silva, J.C. Oliveira, Introduction to Virtual Reality (LNCC, Brazil, 2003).
Technical Report: 06/2003
5. A. Fuhrmann et al., Occlusion in collaborative augmented environments.Comput. Graph. 23(6),
809–819 (1999)
6. B. Thomas, B. Close, J. Donoghue, J. Squires, P.D. Bondi, M. Morris, W. Pieraski, ARQuake:
an outdoor/indoor augmented reality first person application, in Proceedings of International
Symposium on Wearable Computers (2000)
7. P. Dähne, J.N. Karigiannis, Archeoguide: system architecture of a mobile outdoor augmented
reality system, in Proceedings of IEEE and ACM International Symposium on Augmented
Reality (Darmstadt, Germany, 2002)
8. D. Curtis, D. Mizell, P. Gruenbaum, A. Janin, Several devils in the details: making an ar
application work in the airplane factory, in Proceedings of IEEE Workshop on Augmented
Reality (San Francisco, 1998–2011)
9. N. Navab, M. Mitschke, A. Bani-Hashemi, Merging visible and invisible: two camera-
augmented mobile C-arm (CAMC) applications, in Proc. IEEE and ACM Int. Workshop on
Augmented Reality. San Francisco, CA, pp. 134–141 (1999)
10. A. Fuhrmann, H. Löffelmann, D. Schmalstieg, M. Gervautz, Collaborative visualization in
augmented reality. IEEE Comput. Graphics Appl. 18–4, 54–59 (1998)
11. S. Boivin, A. Gagalowicz, Image-Based Rendering for Industrial Applications (ERCIM News,
2001)
12. R. Thabet, R. Mahmoudi, M.H. Bedoui, Image processing on mobile devices: an overview, in
IEEE IPAS’14: International Image Processing Applications And Systems Conference (2014).
978-1-4799-7069-8/14
13. C. Xiao, Z. Lifeng, Implementation of mobile augmented reality based on Vuforia and Rawajali,
in 2014 IEEE 5th International Conference on Software Engineering and Service Science,
978-1-4799-3279-5/14 (2014)
14. Z. Szalavari, D. Schmalstieg, A. Fuhrmann, M. Gervautz, Studierstube—an environment for
collaboration in augmented reality. Virtual Reality Syst. Dev. Appl. 3(1), 37–49 (1998)
A Hybrid Search Group Algorithm
and Pattern Search Optimized PIDA
Controller for Automatic Generation
Control of Interconnected Power System
Smrutiranjan Nayak, Sanjeeb Kar, and Subhransu Sekhar Dash
Abstract A hybrid Search Group Algorithm (SGA) and Pattern Search (hSGA-
PS) technique with the PIDA controller, to deal with Automatic Generation Control
(AGC) of power system is presented. In the first stage, three nonlinear power systems
with PID controller is considered and the controller parameters are tuned by SGA.
The supremacy of the SGA -tuned PID-controlled AGC system is demonstrated
by comparing the published Firefly Algorithm (FA) optimization procedure for the
same interconnected power system. Then in the second stage, the PID controller is
replaced with Proportional–Integral–Derivative and Acceleration (PIDA) controller
and the optimum gains of the PIDA controller are optimized employing the SGA
technique. It has been demonstrated that SGA-tuned PIDA controller improves the
performance significantly compared with the SGA -tuned PID controller. Pattern
Search (PS), a local optimization method is used in the third stage to fine-tune the
PIDA controller parameters delivered by the SGA. The advantage of the hSGA-
PS-tuned PIDA controller over the SGA-tuned PIDA controller, SGA-tuned PID
controller, FA-tuned PID controller is demonstrated. Furthermore, in the sensitivity
analysis, the system parameters, operation load conditions, and the location of distur-
bance are changed and the results are analyzed. The performance and results from the
sensitivity analysis reveal the effectiveness of the hSGA-PS-tuned PIDA controller
aimed at AGC of the power system.
S. Nayak ·S. Kar
Department of Electrical Engineering, ITER, Sikha ‘O’ Anusandhan University, Bhubaneswar,
Odisha, India
e-mail: smrutikiit40@gmail.com
S. Kar
e-mail: sanjeebkar@soa.ac.in
S. S. Dash (B)
Department of Electrical Engineering, Government College of Engineering Keonjhar, Keonjhar,
Odisha, India
e-mail: subhransudash_fee@gcekjr.ac.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_27
309
310 S. Nayak et al.
Keywords Automatic generation control (AGC) ·Governor dead band (GDB) ·
Generation rate constraint (GRC) ·PIDA controller ·Search group algorithm
(SGA) ·Pattern search (PS)
1 Introduction
In the multi-area power system, units of individual areas are interconnected with
another area via transmission lines. If real power load changes abruptly, the frequency
of the power system and the power flow in tie line changes from its nominal value.
The deviations are due to the difference between the electrical load demand and
generation and cause unwanted effects. Automatic Generation Control (AGC) in
each area regulates the generator set point automatically for the corresponding load
change by calculating and driving Area Control Error (ACE) to zero [1–4].
Therefore, an effort is made for the design of the PIDA controller for AGC which
is tuned by the hSGA-PS technique. The aim of this work is of threefold: (i) to
exhibit the superiority of new powerful computational intelligence technique like
SGA over FA-tuned PID controller for AGC, (ii) to exhibit the advantage of PIDA
structure over PID structure, (iii) to show the effectiveness of hybridization of SGA
and PS, where PS is used to improve the results of SGA-tuned PIDA values [5–9].
Here, three-area reheat thermal system with different capacities with nonlinearities
is considered for investigation.
2 System and Controller
2.1 System Investigated
A three-area test system is considered which contains three thermal units of different
capacities of each area with GDB and GRC as shown in Fig. 1. The most commonly
specified value of GRC in the reheat unit is around 3% per min. Governor dead band
(GDB) is a physical constraint resulted due to mechanical friction and backlash and
due to overlap of valves in the hydraulic relay. The speed of GDB affects the dynamics
of the power system. It causes an increase in undershoot/overshoot and steady-state
apparent speed regulation. The typical width of GDB is 0.06% (0.036 Hz). In the
current research, a GDB of 0.036 Hz and GRC of 3%/min are considered.
2.2 Controller Structure
In Fig. 2, PID controller structure used for AGC is shown with usual notation. In
Fig. 3, the structure of proposed PIDA controller is shown, where KP,KI,KDare
A Hybrid Search Group Algorithm and Pattern … 311
AREA-1
1
F
Reheat 1
Controller 1
Controller 2
Controller 3
Turbine 1 with GRC
Kr3Tr3.s+1
Tr3.s+1
Reheat 3
Kr2Tr2.s+1
Tr2.s+1
Reheat 2
1
s
1
s
1
s
1
TG3.s+1
Governor 3
1
TG2.s+1
Governor 2
1
TG1.s+1
Governor 1
....
2*π*T23
s
..
Controller
.,.
2*π*T13
s
.
Controller
,.,
1
TT3
1
TT2
1
TT1
KPS3
TPS3.s+1
KPS2
TPS2.s+1
KPS1
TPS1.s+1
Power
System-1
Dead Zone 3
Dead Zone 2
Dead Zone 1
a13
a23
a12
1/R3
1/R2
2*π*T12
s
1/R1
B3
Kr1Tr1.s+1
Tr1.s+1
B2
B1
Controller
Power
System-2
1D
P
1
Tie
P
Turbine 2 with GRC
AREA-2
2D
P2
F
2
Tie
P
AREA-3
Power
System-3
3D
P3
F
3
Tie
P
Turbine 3 with GRC
Fig. 1 Three-area test system with reheat, GDB, and GRC
Fig. 2 Structure of PID
controller
Fig. 3 Structure of PIDA
controller
312 S. Nayak et al.
the conventional PID controller gains and KAis acceleration gain. The addition of
new acceleration term to PID controller makes the system response faster with less
overshoot.
2.3 Objective Function
The performance criteria suitable for AGC studies is ITAE as reported in the literature
[6] and hence the same is selected to tune PID/PIDA controller parameters. ITAE is
expressed in Eq. (1).
J=ITAE =
tsim
0
(|Fi|+|PTi e−i|)·t·dt(1)
where Fifrequency variation of area i;PTie−inet transmission line power
variation in area i;tsim is the simulation time range.
2.4 Search Group Algorithm
The key feature of SGA is that it establishes a balance among the exploitation and
exploration stage of the algorithm run. SGA is comprised of five steps; each step is
explained as follows [10,11].
2.4.1 Initial Population
Initial population Pis generated randomly as per Eq. (2)
Pij =Xmin
j+Xmax
j−Xmin
jU[0,1],where j=1,...,n,i=1,...,npop
(2)
Pij is jth design parameter of the ith individual of the population P,nrepresents
the total number of search parameters, npop represents the total no of population,
U[0,1] is a random number from 0 to 1 uniformly distributed, Xmin
jand Xmax
jare
the minimum and maximum limits of the jth search parameter.
A Hybrid Search Group Algorithm and Pattern … 313
2.4.2 Selection of Initial Search Group
Once population is created, each individual population’s objective function is calcu-
lated and a search group Ris constructed by choosing ngpoints from the population
P. The selection is accomplished by a standard tournament selection procedure and
each row of R represents an individual. The members of rare ranked after each iter-
ation. If Rirepresents the ith row of R, then R1represents the best design and Rngis
the worst design in R,ngis the number of members in the search group [12].
2.4.3 Mutation of the Search Group
To enhance exploration capability, nmut individuals are replaced from Rby new
individuals generated according to Eq. (3)
Xmut
j=ERj+tεσRj,for j=1,...,n(3)
where Xmut
jis the jth design parameter of a mutated individual, σand Eare standard
deviation and mean operators, εrepresents a suitable random number, trepresents a
control parameter which decides how far a mutated individual is produced from the
average value of the population. The individual which is to be substituted is decided
by an “inverse tournament” selection procedure in which the design having worst
objective function cost is chosen to be replaced.
2.4.4 Creation of Families of Individual Search Group Member
Perturbation defined by Eq. (4) is used by each member of the search group to
generate a set of individuals known as family
Xnew
j=Rij +αε for j=1,...,n(4)
where perturbation size is controlled by α.Thevalueofαreduces after each iteration
kof the search process. αis updated by Eq. (5):
αk+1=bαk(5)
where bis a parameter that defines the way that the value of αkreduces as the
iterations pass by.
The distance of a newly generated individual from its search group member is
controlled by αkand is accountable for the exploitation and exploration ability of
the SGA. In the first stage of iterations, αkvalue is chosen such a high value such
that, it allows exploring the maximum of the design regions. As the iteration lapses
by, αkreduces and allows exploiting the design regions. The range of αkis given by
314 S. Nayak et al.
α0≤αk≤αmin
where αmin is the minimum value of αk, guaranteeing minimum mobility of the new
points up to the last iterations of the SGA, α0is the value of αthe first iteration.
Let each family be denoted by Fi, where i=1tong. The objective function value
of each population of search group decides the number of individuals to be generated
by that member. Better the quality of objective function, the higher is the number of
individuals it generates. The total number of individuals generated by search group
members at each iteration are kept constant, i.e. equals to npop −ng.
2.4.5 Formulation of New Search Group
In the first itmax iterations, the best member from each family is taken into account
to formulate a new search group and this is called the global phase of algorithm.
When the iteration number is greater than itmax
global, a new search group is created by
taking fittest ngpoints from all the families and this stage is known as local phase
of SGA. Where itmax is maximum iterations, itmax
global is the maximum iteration of the
global phase.
2.5 Pattern Search
Pattern Search (PS) commences with the initial point X0that is the best solution
delivered by the SGA. At the first iteration, the direction vectors or patterns are
created as [0, 1], [1, 0], [−1, 0] and [0, −1] using a scalar =1 known as mesh size
[13–16].
3 Results and Discussions
3.1 Implementation of SGA/HSGA-PS Algorithm
MATLAB/Simulink platform is employed to create the model of the test system.
A simultaneous 10% step load is applied in area −1 and 2 at t=0 s for objective
function calculation. The parameters of SGA play a significant contribution to the
performance of SGA and the range of these parameters is chosen as per literature
[11]. The SGA control parameter values used in the algorithm are presented in
Table 1. The PID and PIDA parameters are tuned separately by using SGA. The
tuned parameters of PID/PIDA are noted down after the completion of a maximum
iteration itmax. The SGA was repeated for 50 times and the best solutions among 50
runs of the optimization process are presented in Table 2. The best solutions provided
A Hybrid Search Group Algorithm and Pattern … 315
Tabl e 1 SGA parameter
values Parameters Value/expression
npop 100
α02
αmin 0.01
itmax 50
itmax
global 0.5×itmax
ng0.2×npop
nmut 0.03 ×npop
t1, 2, 3 respectively for each muted individual
by SGA-tuned PIDA controller parameters are then fine-tuned by PS algorithm. The
best solution for PIDA gains provided by SGA are used as PS algorithm’s initial
points. In PS algorithm, the values of mesh size, mesh contraction factor, and mesh
expansion factor are taken as 1, 0.5, and 2, respectively. The number of generations
and objective function calculations are set as 10 and 100, respectively. The flowchart
of the proposed hSGA-PS approach is illustrated in Fig. 4. The final best value of
hSGA-PS-tuned PIDA controller parameters are recorded in Table 2.
3.2 Analysis of Results
For a simultaneous application of a 10% step rise in area 1 and area 2, load demand
at t=0.0 s, the performance of the system using SGA-tuned PID/PIDA controller
and hSGA-PS-tuned PIDA controller is presented in Table 3. The performances
of the SGA-tuned PID/PIDA controller and hSGA-PS-tuned PIDA controller are
compared with recently published FA-tuned PID controller parameters [9]forthe
same test system. It is observed from Table 3that a reduced ITAE value is achieve
during SGA-tuned PID controller (ITAE =45.0126) as compared to FA-tuned PID
controller (ITAE =90.8448). This shows SGA performs better than FA. In [9],
the superiority of FA was demonstrated compared to some other recently proposed
optimization technique. Again the performance is improved by decreasing ITAE
value by using SGA-tuned PIDA controller (ITAE =36.2459) compared to SGA-
tuned PID controller. Finally, by the application of proposed hSGA-PS algorithm to
PIDA controller outperform the SGA-tuned PIDA controller. Minimum ITAE value
of 35.4556 is obtained by hSGA-PS-tuned PIDA controller as stated in Table 3.
The system dynamic responses for variation in frequency and variation in tie-line
power are shown in Figs. 5,6. It is observed from the Figs. 5,6that best dynamic
performance is achieved by hSGA-PS-tuned PIDA controller, significant enhance-
ment is witnessed with proposed SGA-tuned PIDA controller than SGA optimized
PID controller, also SGA-optimized PID controller gives better response than FA
optimized PID controller (Table 4).
316 S. Nayak et al.
Tabl e 2 Tuned PID/PIDA controller parameters
Technique/controller Proportional gain Integral gain Derivative gain Acceleration gain
KP1KP2KP3KI1KI2KI3KD1KD2KD3KA1KA2KA3
FA PI D [ 9]−1.5293 −1.6612 −0.1738 −1.6021 −0.2000 −0.1257 −1.1051 −1.3578 −0.2705 −−−
SGA PID −0.6413 −0.9158 −0.6726 −0.0998 −0.0955 −0.1211 −0.7065 −0.9549 −1.5670 − − −
SGA PIDA −0.3602 −0.4726 −0.7399 −0.1881 −0.0856 0.0043 −0.5519 −0.3621 −1.0458 −0.0264 −0.0682 −0.1331
hSGA-PS PIDA −0.3602 −0.4726 −0.7399 −0.1881 −0.0856 0.0043 −0.5519 −0.3621 −1.0458 −0.0264 −0.0682 −0.2581
A Hybrid Search Group Algorithm and Pattern … 317
Fig. 4 Flowchart of the hSGA-PS algorithm
Tabl e 3 ITAE value and %
improvement for each
controller for 10% step load
demand in area-1 and area-2
Optimization technique/Controller ITAE
Val u e % Imp.
FA: PID controller [9]90.8448 −
SGA: PID controller 45.0126 50.45
SGA: PIDA controller 36.2459 60.10
hSGA-PS: PIDA controller 35.4556 60.97
Further, only a 10% step rise in area 1 load is considered at t=0 s to analyze the
influence of disturbance location on the performance of system. The system dynamic
response for single disturbance are presented in Figs. 7,8. It is obvious from Figs. 7,
8that the proposed hSGA-PS-optimized PIDA controllers are robust and delivers
satisfactory performance when disturbance location changes. Also improved results
are achieved with the proposed SGA-tuned PID controller in comparison with FA-
optimized PID [9], SGA-tuned PIDA controller has better responses compared to
SGA-optimized PID controller and finally hSGA-PS-optimized PIDA controllers
318 S. Nayak et al.
0 5 10 15 20 25
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
Time (Sec)
F2 (Hz)
FA PID [9]
SGA PID
SGA PIDA
hSGA-PS PIDA
Fig. 5 Area-2 frequency deviance for simultaneous 10% step load demand in area-1 and area-2
0 10 20 30 40 50 60
-0.1
-0.08
-0.06
-0.04
-0.02
0
0.02
Time (Sec)
PTie-2 (P.U)
FA PID [9]
SGA PID
SGA PIDA
hSGA-PS PIDA
Fig. 6 Area-2 tie line power deviance for simultaneous 10% step load demand in area-1 and area-2
have best results compared to SGA-tuned PIDA controller. Hence, hSGA-PS-tuned
PIDA controller parameters obtained from double disturbance are also effective for a
single disturbance. So the proposed design of PIDA controllers is robust and performs
satisfactorily regardless of load disturbance location. Figure 9shows the system
variation with change in load.
A Hybrid Search Group Algorithm and Pattern … 319
Tabl e 4 Sensitivity analysis under varied conditions
Parameter
variation
% Change Settling time Ts(s) ITAE
F1F2F3Ptie−1Ptie−2Ptie−3
Nominal 014.96 14.49 15.59 35.98 34.01 20.74 35.4556
Loading
condition
−50 24.09 27.68 24.47 35.55 32.73 20.49 47.5721
−25 14.54 22.67 15.07 35.8 33.49 20.64 40.5453
+25 15.48 14.87 16.05 36.13 34.38 20.79 32.9753
+50 16.17 15.48 16.75 36.26 34.86 21.13 31.9027
TG−50 14.81 14.32 15.41 36.1 34.05 20.57 34.4011
−25 14.91 14.42 15.52 36.03 34.02 20.66 34.899
+25 14.86 14.5 15.61 35.97 34.01 20.79 36.1316
+50 14.71 14.47 15.61 35.96 34.01 20.83 36.8923
TT−50 15.13 14.59 15.71 36.23 34.07 20.7 35.0917
−25 15.07 14.55 15.65 36.1 34.03 20.72 35.2686
+25 14.76 14.35 15.45 35.88 33.97 20.72 35.7341
+50 14.73 14.25 15.28 35.79 33.94 20.68 36.0302
H−50 30.75 31.73 31.48 35.45 25.2 12.07 62.9446
−25 23.06 26.11 23.63 35.94 31.96 18.44 42.5669
+25 17.47 16.6 18.03 36.77 36.9 22.76 35.9119
+50 20.05 18.79 20.55 37.39 38.44 25.1 37.7925
0 5 10 15 20 25 30 35
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
Time (Sec)
F1(Hz)
FA PID [9]
SGA PID
SGA PIDA
hSGA-PS PIDA
Fig. 7 Area-1 frequency deviance for 10% step load demand in area-1
320 S. Nayak et al.
Fig. 8 Area-2 tie line power deviance for 10% step load demand in area-1
0510 15 20 25 30 35 40
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
Tim e (S ec )
F
1
(Hz)
-50% of Nominal Loading
+50% of Nominal Loading
Nominal Loading
Fig. 9 Area-1 frequency deviance with the variation of loading
4 Conclusion
In this study, a hybrid SGA and PS algorithm has been applied to tune PIDA gains
for AGC of the power system. For the first time, SGA is used in this area to optimize
the PID/PIDA controller gains using ITAE objective functions. Firstly, SGA is used
to tune PID parameters in a three unequal area interconnected thermal power system
with GDB and GRC nonlinearities and the superiority of SGA technique is demon-
strated by comparing the results with a published optimization method which was
A Hybrid Search Group Algorithm and Pattern … 321
proved superior performance compared to GA, DE, BFOA, PSO, hBFOA-PSO, and
Ziegler–Nichols-based controllers. Then in the second stage, PIDA controller gains
are tuned by SGA and the result reveals that SGA-optimized PIDA controller gives
a significant improvement in the response than the SGA-tuned PID controller. In the
third stage, to take advantage of local search capabilities of PS, PS is used to fine-tune
the PIDA controller parameter, in which SGA-tuned values are taken as the initial
starting point. The proposed hSGA-PS-based PIDA controller provides improved
performance compared to SGA tuned PIDA controller. So, it can be concluded that
the proposed hSGA-PS-based PIDA controller is very effective, robust and gives
better system performance compared to FA-tuned PID controller.
Appendix
Nominal parameters of the investigated system are
Three unequal areas thermal system [9, 20]:
f=60 (Hz); D1=D3=0.015, D2=0.016 (p u Hz); B1=0.3483, B2=
0.3827, B3=0.3692 (p u Hz); 2H1=0.1667, 2H2=0.2017, 2 H3=0.1247,
(p u s); TG1=0.08, TG2=0.06, TG3=0.07 (s); R1=3.0, R2=2.73, R3=2.82
(Hz/p u); TT1=0.4, TT2=0.44, TT3=0.3(s);Tr1=Tr2=Tr3=10 (s); Kr1=
Kr2=Kr3=0.5; T12 =0.2, T23 =0.12, T31 =0.25 (p u/Hz), Pr1=2000 MW,
Pr2=4000 MW, Pr3=8000 MW.
References
1. P. Kundur, Power System Stability and Control, 8th edn. (Tata Mc-Graw Hill, New Delhi, 2009)
2. O.I. Elgerd, Electric Energy Systems Theory: An Introduction (Tata Mc-Graw Hill, New Delhi,
2007)
3. H. Bevrani, Robust Power System Frequency Control (Springer, 2009)
4. L.C. Saikia, J. Nanda, S. Mishra, Performance comparison of several classical controllers
in AGC for multi-area interconnected thermal system. Int. J. Electr. Power Energ. Syst. 33,
394–401 (2011)
5. S. Mishra, G. Mallesham, P.C. Sekhar, Biogeography based optimal state feedback controller
for frequency regulation of a smart micro grid. IEEE Trans. Smart Grid 4, 628–637 (2013)
6. H. Shabani, B. Vahidi, M.A. Ebrahimpour, Robust PID controller based on imperialist
competitive algorithm for load-frequency control of power systems. ISA Trans. 52, 88–95
(2012)
7. R.K. Sahu, S. Panda, U.K. Rout, D.K. Sahoo, Teaching learning based optimization algorithm
for automatic generation control of power system using 2-DOF PID controller. Int. J. Electr.
Power Energ. Syst. 77, 287–301 (2016)
8. U.K. Rout, R.K. Sahu, S. Panda, Design and analysis of differential evolution algorithm based
automatic generation control for interconnected power system. Ain Shams Eng. J. 4, 409–421
(2013)
9. S. Padhan, R.K. Sahu, S. Panda, Application of firefly algorithm for load frequency control of
multi-area interconnected power system. Elect. Power Comp. Syst. 42, 1419–1430 (2014)
322 S. Nayak et al.
10. R.K. Sahu, S. Panda, S. Padhan, A hybrid firefly algorithm and pattern search technique for
automatic generation control of multi area power systems. Int. J. Electr. Power Energ. Syst. 64,
9–23 (2015)
11. S.G. Matheus, R.H. Lopez, L.F.F. Miguel, Search group algorithm: a new metaheuristic method
for the optimization of truss structures design. Comput. Struct. 153, 165–184 (2015)
12. Y. Bao, Z. Hu, T. Xiong, A PSO and pattern search based memetic algorithm for SVMs
parameters optimization. Neurocomputing 117, 98–106 (2013)
13. A.K. Al-Othman, N.A. Ahmed, M.E. Al Sharidah, A. Hanan, A hybrid real coded genetic
algorithm-pattern search approach for selective harmonic elimination of PWM AC/AC voltage
controller. Int. J. Electr. Power Energ. Syst. 44, 123–133 (2013)
14. U. Thaiwasin, S. Sujitjorn, Y. Prempraneerach, J. Ngamwiwit, Torsional resonance suppression
via PIDA controller, in 2000 TENCON Proceedings of Intelligent Systems and Technologies
for the New Millennium (2000), pp. 498–503
15. D.Y. Ha, I.Y. Lee, Y.S. Cho, Y.D. Lim, B.K. Choi, The design of PIDA controller with pre-
compensator, in ISIE 2001. 2001 IEEE International Symposium on Industrial Electronics
Proceedings (Cat. No.01TH8570) (2001), pp. 798–804
16. S. Sornmuang, S. Sujitjorn, GA-based PIDA control design optimization with an application
to ac motor speed control. Int. J. Math. Comput. Simul. 4, 67–80 (2010)
Fractional Order PID Controlled PV Fed
Quadratic Boost Converter TZ Source
Inverter Fed Permanent Magnet
Brushless Motor Drive
N. K. Rayaguru and S. Sekar
Abstract Permanent magnet brushless motors are industry friendly motors with a
significant issue with respect to power quality. There is always a necessity to solve
dynamic response related issues in a converter inverter system with less power devices
and minimal complexities. This work proposes combination of QBC and TZSI for the
control of PV fed PMBLDCM. DC from-PV is stepped up utilizing QBC and the yield
of QBC is applied to TZSI. This effort covenants with comparison of closed-loop
proportional resonant and fractional order PID controlled quadratic boost converter
with TZ source inverter fed permanent magnet brushless motor drive systems. The
effort incorporates the control strategies in both open and closed-loop control of an
industry friendly permanent magnet brushless motor drives system incorporating a
conventional FOPID controller and an advanced proportional resonant controller. The
FOPID and PR controllers regulate the speed of revolution of the motor and thereby
reduces the ripples associated in the torque of the motor. Finally the improvement time
response is obtained which is depicted through the results on simulation performed
indicates reduced steady-state error and enhanced dynamic response.
Keywords Permanent magnet brushless motor ·MPPT ·PWM inverter ·
Quadratic Boost Converter (QBC) ·T-Z source inverter (TZSI) ·Motor load ·
PR-proportional resonant ·FOPID-Fractional order PID ·Steady-state error ·Rise
time ·Peak time and settling time
1 Introduction
The solar-based energy source had taken the best opportunity among other ways
of generation of electricity from renewable sources because the energy produced is
N. K. Rayaguru (B)·S. Sekar
Department of EEE, Hindustan Institute of Technology and Science, 1, Rajiv Gandhi Salai OMR,
Padur, Chennai, Tamilnadu 603103, India
e-mail: nkrayaguru29@gmail.com
S. Sekar
e-mail: ssekar.pt@hindustanuniv.ac.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_29
323
324 N. K. Rayaguru and S. Sekar
cordial. Moreover had a longer existence with acceptable payback period, excep-
tionally versatile and compact. A basic method to structure the output voltage of
sun-based energy sources was to associate several solar panels in use and an ordi-
nary VSI. However, the design had some disadvantages like constant low quality and
low efficacy. A simulation model was created that used the essential circuit condi-
tions of the photovoltaic (PV) cells, along with the impacts of temperature change
and irradiation.
The existing various DC/DC converter topologies are buck, boost, and buck-boost;
Sepic, Luo and C’uk [1–3].
Among these power converters, a boost converter that raises the input voltage to
an appropriate level is most utilized because of a few favorable circumstances, for
example, low ripple current waves, high proficiency, quick transient reaction, and an
improved unwavering quality [4–6].
This work likewise investigates the closed-loop execution of the QBC framework
by utilizing different PI, FOPID controller [7–9]. Closed structure which incorporates
two cascaded control circles; inside current control and outside speed control are
utilized.
The above papers do not deal with QBC-TZSI fed PMBLDC Drive system. This
work suggests the QBC for the control of PMBLDC Drive system. The exceeding
writing does not deal with evaluation of FOPID and PR-based TZSI fed PMBLDC
drive systems. This effort mainly presents closed-loop simulation and execution of
QBC-TZSI fed PMBLDC Drive. PRC is proposed for the closed loop control of
QBC-TZSI fed PMBLDC Drive system. The objectives of this work are to minimize
settling time and steady-state error of QBC-TZSI fed PMBLDC Drive system using
FOPID/PR controller.
2 Quadratic Boost Converter
2.1 Circuit-Description
In single QBC has a single switch M1, two inductors L1and L2, two Capacitors
C1,C2, three diodes named D1,D2,D3, and a motor load. A single QBC same as
cascaded boost converter but it uses Single switch. The QBC system consists of two
single QBC in parallel connection (Fig. 1)[10,11].
2.2 Analysis of Single QBC
Mode 1-Operation
This mode begins when MOSFET switch M1is turned on. During this mode D1and
D3Reverse biased connection and D2forward biased connection. Now the inductors
Fractional Order PID Controlled PV Fed … 325
Fig. 1 The proposed circuit diagram of QBC
L1and L2get energized by Vin and C1separately. Subsequently Inductor current,
IL1and IL2increment in size. Capacitor C2supplies energy to the load (Figs. 2,3)
[12].
IL1On =
VsKT
L1
(1)
IL2On =
VSKT
L1
Meanwhile the capacitor currents are expressed by
Fig. 2 Single quadratic boost converter: topology
326 N. K. Rayaguru and S. Sekar
Fig. 3 QBC yopology when SW is turned-on
IC1On =−IL2(3)
IC2On =−I0(4)
where IL1and IL2are the change in inductor current 1 and 2, respectively. K,the
duty cycle of switch, Vin is the input voltage, and Tthe switching period [13,14].
Mode 2 Operation
During this mode switch M1is in off state. Diodes D1and D3are forward biased and
diode D2is in reverse biased condition. The current in the inductor L1and L2now
begins diminishing as the energy put away in the inductor L1is presently moved to
capacitor C1and energy put away in the inductor L2is currently moved to the load
side. Capacitor C1,C2gets energized during this mode. During this time voltage
over the inductor turns around as the pace of progress of current is insignificant, i.e.,
decreasing current. The capacitor current for this state as (Figs. 4,5)[15].
Fig. 4 QBC topology when SW is turned-off
Fractional Order PID Controlled PV Fed … 327
Fig. 5 Capacitor current waveform
IL1Off =
Vs−VC
1(1−K)T
L1
(5)
IL2Off =
VC
1−V0(1−K)T
L2
(6)
IClOff =IL1−IL2(7)
IC2Off =IL2−I0(8)
3 System Configuration
Figure 6delineates the block representation of open-loop QBC TZSI fed
PMBLDCM. The PV Panel is characterized as the connection of PV cells in
series/parallel to create the necessary voltage and current. Produced by PV is DC.
DC is stepped up using QBC. The output of QBC is converted to AC using TZSI.
The TZSI feeds PMBLDCM.
FOPID/PR is outlined in Fig. 7. Motor speed is measured and it is related with
the reference speed to get error in speed. The speed error is implied to the speed
FOPID/PR controller. The output of speed FOPID/PR is utilized to create current
reference. The reference current is compared with the actual current and the current
Fig. 6 Block diagram of open-loop QBC TZSI fed PMBLDCM
328 N. K. Rayaguru and S. Sekar
Fig. 7 Block diagram of closed-loop QBC-T-ZSI fed PMBLDCM with PR and FOPID controllers
error is applied to current FOPID/PR and its output is exploited to update the PWM
of QBC [16–18].
4 Simulation Results
4.1 QBC–TZSI-PMBLDC with Source Disturbance
Simulink results of QBC–TZSI-PMBLDC with source disturbance are outlined here.
Circuit diagram of QBC with T-Z source is appeared in Fig. 8. The pulses for inverter
are generated based on the information from the hall-sensor. Scopes are connected
to measure the QBC voltage, inverter voltage, motor speed, and motor torque.
Voltage across PV of QBC-TZSI-PMBLDC is outlined in Fig. 9and its value
enhances to 95 V at t=1 s. The augment in PV voltage is due to enhance in
insolation. Voltage across QBC is outlined in Fig. 10 and its value enhances to 75 V.
The QBC voltage is enhanced due to enhancement in PV voltage.
Fig. 8 Circuit diagram of QBC with T-Z source
Fractional Order PID Controlled PV Fed … 329
Fig. 9 Voltage across PV of QBC–TZSI-PMBLDC
Fig. 10 Voltage across QBC
Voltage across inverter is outlined in Fig. 11 and its value enhances to 75 V. The
augment in inverter voltage is due to enhancement in PV voltage. Motor speed of
QBC-TZSI-PMBLDC is appeared in Fig. 12 and it’s value enhances to 390 RPM.
The augment in speed is due to enhance in PV voltage.
Motor Torque of QBC-TZSI-PMBLDC is outlined in Fig. 13 and its value
enhances to 1.5 N m. The enhancement in torque is due to boost in voltage.
Fig. 11 Voltage across inverter of QBC-TZSI-PMBLDC
330 N. K. Rayaguru and S. Sekar
Fig. 12 Motor speed of QBC-TZSI-PMBLDC
Fig. 13 Motor Torque of QBC-TZSI-PMBLDC
4.2 Closed-Loop PR Controlled QBC-TZSI-PMBLDC
Simulink diagram of PR controlled closed-loop QBC-TZSI-PMBLDC is outlined
in Fig. 14. First PR controller is used for both voltage control mode as well as
current control mode. Then PR controllers are replaced with FOPID controllers and
the speed, torque, and power are measured. Here the dynamic performance of the
controllers is compared. The error signal is coursed by FOPID as well as PR to
sustain the speed of QBC-TZSI-PMBLDC constant and diminish the steady-state
errors [19,20].
Voltage across QBC-TZSI-PMBLDC is presented in Fig. 15 and it’s value is 70
V. Voltage across motor load of PR controlled closed-loop QBC-TZSI-PMBLDC is
presented in Fig. 16 and it’s value is 60 V.
Motor speed of PR controlled closed-loop QBC-TZSI-PMBLDC is presented
in Fig. 17 and it’s value is 300 RPM. Motor torque of PR controlled closed-loop
QBC-TZSI-PMBLDC is outlined in Fig. 18 and it’s value is 1.7 N m.
Fractional Order PID Controlled PV Fed … 331
Fig. 14 Circuit diagram of PRC closed-loop QBC-TZSI-PMBLDC
Fig. 15 Voltage across QBC-TZSI-PMBLDC
Fig. 16 Voltage across motor load of PRC closed-loop QBC-TZSI-PMBLDC
332 N. K. Rayaguru and S. Sekar
Fig. 17 Motor speed of PRC closed-loop QBC-TZSI-PMBLDC
Fig. 18 Motor Torque of PRC closed-loop QBC-TZSI-PMBLDC
4.3 Closed-Loop FOPID controlled QBC-TZSI-PMBLDC
Voltage across QBC-TZSI-PMBLDC is depicted in Fig. 19 and it’s value is 100 V.
Voltage across motor load of FOPID controlled closed-loop QBC-TZSI-PMBLDC
with TZSI is presented in Fig. 20 and it’s value is 60 V.
Motor speed of FOPID controlled closed-loop QBC-TZSI-PMBLDC is presented
in Fig. 21 and it’s value is 300 RPM. Motor torque of FOPID controlled closed-loop
QBC-TZSI-PMBLDC is outlined in Fig. 22 and it’s value is 1.2 N m.
Fig. 19 Voltage across QBC
Fractional Order PID Controlled PV Fed … 333
Fig. 20 Voltage across motor load of FOPID controlled closed-loop QBC-TZSI-PMBLDC
Fig. 21 Motor speed of FOPID controlled closed-loop QBC-TZSI-PMBLDC
Fig. 22 Motor Torque of FOPID controlled closed-loop QBC-TZSI-PMBLDC
Comparsion of time domain parameters for motor speed using PR-PR and FOPID-
FOPID is presented in Table 1. By using FOPID-FOPID controller, ‘RT (rise time)’
is diminished from 1.15 to 1.14 s; ‘PT (peak time)’ is diminished from 2.45 to 1.34 s;
Tabl e 1 Comparsion of time domain parameters for motor speed using PR and FOPID
Controller Rise time (s) Peak time (s) Setting time (s) Steady-state error (RPM)
Dual-PR 1.15 2.45 3.02 3.6
Dual-FOPID 1.14 1.34 2.23 2.8
334 N. K. Rayaguru and S. Sekar
Tabl e 2 Comparison of time domain parameters for motor torque using PR-PR and FOPID-FOPID
Controller Rise time (s) Peak time (s) Setting time (s) Steady-state error (N m)
Dual-PR 1.13 2.82 2.93 0.8
Dual-FOPID 1.11 1.29 1.82 0.7
‘ST (settling time)’ is diminished from 3.02 to 2.23 s; ‘SSE (steady-state-error)’ is
diminished from 3.6 to 2.8 RPM for motor speed (Table 2).
Comparsion of Time domain parameters for motor torque using PR-PR and
FOPID-FOPID is presented in Table 1. By using FOPID controller, ‘RT (rise-time)’
is diminished from 1.13 to 1.11 s; ‘PT (peak time)’ is diminished from 2.82 to 1.29 s;
‘ST (settling time)’ is diminished from 2.93 to 1.82 s; ‘SSE (steady-state error)’ is
diminished from 0.8 to 0.7 N m for motor torque.
5 Conclusion
Closed-loop proportional resonant and fractional order PID controlled QBC-TZSI-
PMBLDC fed permanent magnet brushless motor drive systems are designed,
analyzed, and simulated utilizing MATLAB-simulink. By using FOPID controller,
‘ST (settling time)’ is diminished from 3.6 to 2.8 s and ‘SSE (steady-state error)’
is diminished from 0.8 to 0.7 N m. Hence, the outcome represents that the dual-
loop FOPID controlled closed-loop QBC-TZSI-PMBLDC is superior to dual-loop
PR controlled closed-loop QBC-TZSI-PMBLDC. The benefits of proposed QBC-
TZSI-PMBLDC system are diminished number of switches and enhanced-response.
The drawback of QBC-TZSI-PMBLDC is that the number of passive components is
increased.
The present effort covenants with the comparison of closed-loop proportional
resonant and fractional order PID controlled QBC-TZSI-PMBLDC. The comparison
of closed-loop Fractional PID and Fuzzy logic controlled QBC-TZSI-PMBLDC can
be done in future.
References
1. A. Mudhol, A.J.P. Pius, Design and implementation of boost converter for photovoltaicsystems.
IJIREEICE 4(2) (2016)
2. P. Wang, L. Zhou, Y. Zhang, J. Li, M. Sumner, Input-parallel output series DC-DC boost
converter with a wide input voltage range, for fuel cell vehicle. IEEE Trans. Veh. Technol.
66(9), 7771–7781 (2017)
3. C.C. Chan, A. Bouscayrol, K. Chen, Electric, hybrid, and fuel-cell vehicles: architectures and
modeling. IEEE Trans. Veh. Technol. 59(2), 589–598 (2010)
4. T.R. Choudhury, B. Nayak, Comparison and analysis of cascaded and quadratic boost converter.
IEEE. 978-1-4799-7455-9/15/$31.00 (2015)
Fractional Order PID Controlled PV Fed … 335
5. M.A. Al-Saffar, E.H. Ismail, A.J. Sabzali, High efficiency quadratic boost converter. IEEE
(2012). 978-1-4577-1216-6/12/$26.00 (2012)
6. Y. Wang, Y. Qiu, Q. Bian, Y. Guan, D. Xu, A single switch quadratic boost high step up DC-DC
converter. IEEE Trans. Ind. Electron. https://doi.org/10.1109/tie.2018.2860550 (2018)
7. L.S. Vilefort, W.G. Silva, D.A. de Andrade, F.V.R. da Silva, L.C.G. de Freitas, E.A.A. Coelho,
J.B. Vieira Jr, Quadratic boost converter using a soft single switch. IEEE. 978-1-4799-0272-
9/13/$31.00©2013
8. P. Priya, G.S. Basha, S.V.S. Niranjan, R. Seyezhai, Analysis and experimentation of quadratic
boost converter for photovoltaic applications. Indian J. Sci. Technol. 10(37) (2017). https://doi.
org/10.17485/ijst/2017/v10i37/117553
9. M. Forouzesh, Y.P. Siwakoti, S.A. Gorji, B. Lehman, Step-up DC–DC converters: a compre-
hensive review of voltage-boosting techniques topologies, and applications. IEEE Trans. Power
Electron. 32(12) (2017)
10. A.M.S.S. Andrade, L. Schuch, M.L. de S. Martins, High-stepup PV-module-integrated
converter for PV-energy harvest in-FREEDM-systems. IEEE Trans. Ind. Appl. 53(2), 1138–
1148 (2017)
11. N.R. lingam, S. Priyanga, Implementation of PV-based boost converter using PI controller with
PSO algorithm (2017)
12. K.B. Tawfiq, A.S. Mansour, H.S. Ramadan, M. Becherif, Wind energy conversion system
topologies and converters: comparative review. 162, 38–47 (2019)
13. C.-C. Hua, C.-H. Cheng, Design and implementation of power converters for wind energy
conversion system. IEEE (2010)
14. J. Bauman, M. Kazerani, Acomparative study of fuel-cell battery, fuel cell-ultra capacitor, and
fuel-cell-battery-ultra capacitor vehicles. IEEE Trans. Veh. Technol. 57(2), 760–769 (2008)
15. C.-M. Lai, Y.-H. Cheng, M.-H. Hsieh, Y.-C. Lin, Development of a bidirectional DC/DC
converter with dual-battery energy storage for hybrid electric vehicle system. IEEE Trans.
Veh. Technol. 67(2), 1036–1052 (2018)
16. Y. Zhang, L. Zhou, M. Sumner, P. Wang, Single-switch, wide voltage gain range, boost DC–DC
converter for fuel cell vehicles. IEEE Trans. Veh. Technol. 67(1), 134–145 (2018)
17. F. Zhang, X. Zhang, M. Zhang, A.S.E. Edmonds, Literature review of electric vehicle tech-
nology and its applications, in 2016 5th International Conference on Computer Science and
Network Technology (ICCSNT), Changchun, China, Dec. 2016, pp. 10–11
18. Y.B. Koca, A. Yonetken, Y. O˘guz, Boost converter design and analysis for photovoltaic systems
(2019)
19. A. Khaligh, Z. Li, Battery, ultra capacitor, fuel cell and hybrid energy storage systems for
electric, hybrid electric, fuel cell, and plug-in hybrid electric vehicles: state-of-the-art. IEEE
Trans. Veh. Technol. 59(6), 2806–2814 (2010)
20. Z. Chen, P. Yang, G. Zhou, J. Xu, Z. Chen, Variable duty cycle control for quadratic boost PFC
converter. IEEE Trans. Ind. Electron. 63(7) (2016)
Peformance Analysis of Joysticks Used
in Infotainment Control System
in Automobiles
Vivek Ram Vasan, S. Prasanna Bharathi, Arjun Sharma,
and G. Chamundeeswari
Abstract Distraction while driving can be attributed as an important cause for acci-
dents and the use of smartphones while driving is identified as the major reasons for
distracting the driver. In order to reduce distractions due to the usage of mobile phones
in automobiles, latest technologies are available to operate the mobile phones through
the dashboard HUD’s and steering wheel housed controls. Android Auto is one such
technology developed at Google that allows the android phones with android version
5 and above to be interfaced and controlled by the Car dashboard controls. Mirror
link is another standard available to interface smart phones to a car’s infotainment
system. The driver and passengers can interact and use them for assistance while
driving by using the steering wheel controls and the dashboard buttons which is part
of the automotive infotainment system. However, these technologies are available
only in some of the latest high-end premium cars released after 2015. There are
many Pluggable Hands-free devices available, that enable the users to control the
mobile phones in the car. These devices are mountable on cars’ components such
as steering wheel and are connected to the phone using Bluetooth protocol. They
provide key-based input interfaces using which the user can control specific features
of the mobile phone.
Keywords Android Client-Server application ·Atmega microcontroller ·HC-05
module ·HUD
V. R. Va s an ( B)
ESP-AS2, Robert Bosch Engineering and Business Solutions, Adugodi, Bangalore, India
e-mail: vivek435@gmail.com
S. Prasanna Bharathi
Department of ECE, SRM Institute of Science and Technology, Vadapalani, Chennai, India
e-mail: prasanns@srmist.edu.in
A. Sharma
ESP-AS13, Robert Bosch Engineering and Business Solutions, Adugodi, Bangalore, India
e-mail: arj.sharma1997@gmail.com
G. Chamundeeswari
Saveetha School of Engineering, SIMATS, Chennai, India
e-mail: easwarig@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981-15-5566- 4_30
337
338 V. R. Vasan et al.
1 Introduction
Distraction of Driver is considered as one of the primary causes of accidents and the
use of smartphone is identified as a common practice that can potentially distract
the driver during driving. Distraction caused by mobile phones includes talking
on the mobile phone, chatting, voice recognition, use of maps for navigation and
tuning the sound system. It would be very tedious and not realistic to estimate how
many distracted drivers drive on a particular section of the road because surveys and
cameras do not aid to come to a conclusion of such distraction. However, it is clear
that distraction of driver is a major cause of road traffic accidents because a distracted
driver tends to fail spotting a stop sign, or see a red light at a traffic junction, and
also could violate the permitted speed limit that might put their own safety or that of
others at risk.
Research literature has highlighted that during driving speaking on the phone is
a very risky in terms of road safety and it contributes to increase in road accidents
[1], thus, presenting as a significant threat to the public. Smartphone usage while
driving is predominant in adolescent and amateur drivers. Recent studies shows [2]
that approximately 50% of the young drivers of the age in between 18 and 25 years,
indulge in the practice of the use smartphone while driving, in which about 25% of
them read emails and browse the World Wide Web, and around 65% of them send
text messages.
It was observed from a study report that among 2500 distracted driver-related
incidents is the considered the highest of smartphone caused harmful accidents corre-
sponded to young drivers [3], Horberry et al. [4]. Another report shows that more than
65% of drivers indulging in smartphone usage while driving constituted of drivers
less than 40 years of age. A conversation over the mobile phone distracts the driver
by immersing the individual into the contents of the phone rather than concentrating
on the task which is driving. Thus, particular research was undertaken to study the
reaction times of drivers as a measure of the crash probability due to distraction
from mobile phone usage. The research considered various study conditions which
include a test drive in a simulator, laboratory and field trials.
An experiment conducted [5] to observe the braking maneuvers of drivers who
use a mobile phone at the blink of a red light in a Test bench setup deduced that
hand-held as well as hands-free smart phone tasks results in slower response times
to perform the task of braking. A desktop simulator experiment conducted where the
distracted driver has to commence driving after a green light while conversing over
a smartphone deduced that the distracted drivers took almost 1/3 of a second more
to begin driving [6].
Many studies have targeted to study the negative effects of mobile phone usage on
Car-tailing maneuvers. Car-following is the behavior of a driver to follow a vehicle
in the front. It is a predominant driving situation as well as a prerequisite for safe
driving [7]. In car simulation, [8] the adverse effect of following a car while the use of
smartphones and in-car infotainment system was studied. A study of 40 active drivers
performed a driving maneuver for 80 km where a total of 20 car-tailing scenarios
Peformance Analysis of Joysticks Used … 339
occurred arbitrarily. The drivers were tested in simulated driving conditions where
they are exposed phone conversation task in 8 of these car-following situations. It
showed that there was an increase response time for mobile phone conversations
while driving. Furthermore, it was found that the drivers did not compensate for the
increase in response time by increasing the spacing between cars while conversing
over the mobile. However, recent reports show a decrease in driving speed when a
phone call is received by the driver, an act known as driving risk compensation and
avoidance [9]. For example, significant decrease in speed when driving on call with
the phone on the driver’s hand was observed [10], compared to other types of mobile
conversations.
A slow braking response time was observed for the driver conversing over a
mobile phone when compared to driving without a phone [11]. The distracted drivers
additionally took more time to recuperate their lost speed after braking. It was also
observed that the participants indulging in conversation over the mobile phones
appeared to have a more cautious driving profile and maintained lower speed and
higher spacing with respect to following distance compared to non-distracted drivers
[12]. However, the crash rates were still higher for distracted driving compared to
non-distracted driving [13]. Particularly within the distracted driving conditions,
there was no significant difference observed between driving with hands-free and
hand-held mobile phone conversations.
Driver response time to an unforeseen event, the speed and the spacing are the
key factors to ensure stability and flow of road traffic. Talking over the phone while
driving can greatly affect such factors, thus, causes the driver to drive poorly while
following the leading vehicle [14]. Studies have tried to summarize the danger of
mobile phone usage in a car in the above-mentioned scenarios, however, the data is
limited and the knowledge we have on this critical issue remains elusive.
2 Feasible Mechanical Design for the Novel Car Accessory
2.1 Button/Key Pad-Based Input Interface
The Bluetooth kits available in market have buttons as the input interface through
which the driver can interact with the device to control the phone’s apps. These
devices would require the driver to use his/her index finger or the thumb to press the
buttons and use the device in order to make or receive a call.
The button size is expected to be big enough to enable the driver to distinguish
between the various button options available. Also, key-based input interfaces would
require the driver’s attention for at least a fraction of second which is sufficient to
cause another distraction which might prove fatal. The devices are mounted on the
steering wheel. Hence the size of the device is the function of available number of
input buttons. This would limit the number of controls the device can provide. For
example, a device with four buttons can provide four functionalities like
340 V. R. Vasan et al.
Fig. 1 2012 steering wheel
button system [R] to the
2018 steering wheel [L]
•Receive Incoming Call
•End a Call
•Increase Volume
•Decrease Volume.
In order to provide extra features, extra buttons have to be added which would in
turn increase the size of the device. As the size of the device increase, it becomes
more of a obstruction to the driver handling the steering wheel. Therefore, it could
be clearly concluded that currently available devices with button input interface have
a series of limitations and disadvantages. Figure 1illustrates the increase in the size
of the device as the number of buttons increases.
2.2 Joy Stick-Based Input Interface
A joystick is a lever that has the ability to move in all directions. It is flexible in nature
and provides many degrees of motion. The idea to use a joystick is to translate the
movement of a flexible vertical stick which is made of plastic into electronic signal
that a controller can process. To put into perspective, translating physical movement
into electric signal. The current work proposes a joystick-based design for the car
accessory using which mobile phone can be controlled. The design is illustrated in
Fig. 2.
Fig. 2 Joystick device
mounted on steering wheel
and a conventional joystick
Peformance Analysis of Joysticks Used … 341
3 Advantages of Selecting Joystick-Based Design
Joystick has a set of advantages over the keypad/button-based input interface. The
advantages are listed below:
1. A joystick occupies relatively less space compared to the latter. The base of the
joystick is fixed and hence the design can be made as optimal as possible to make
the entire device occupy less space on the steering wheel.
2. It is easy to train the users on the usage. The joystick movement is either towards
the x-axis or y-axis. The human cognitive ability enables the user to pick up such
actions easily. The drivers need not have to spend even a fraction of second in
identifying the key to be operated in order to provide a specific Input.
3. The ease of handling a joystick, the design can be provided as slender and flexible
as possible. This would enable the driver to control the joystick with the help
of the region between the thumb and index finger without having to move the
finger from the steering wheel. This would ensure perfect grip while handling
the device.
4. The joysticks can be designed in such a way that they are able to move in z
dimension as well. Hence the number of Input controls that can be provided by
a joystick is higher. A minimum of 4 to a maximum of 9 input signals can be
generated using a single joystick. This is a space saving factor and definitely
overcomes the disadvantage of keypad-based input as the size of the device
increases with increase in keys.
5. The joystick can be used as a base for keypad as well. The column and the head
of the joystick can be embedded with keypads if required in order to increase the
number of input signals that can be generated using a single device.
6. The joystick inputs can be designed for a particular pattern, for example, moving
the joystick towards x-axis twice within a small-time interval can be decoded
as another input signal thus increasing the number of input signals that can be
generated.
7. The joystick occupies less space and hence the manufacturing material cost would
be relatively lesser when compared to a device with keypad providing the same
set of features.
3.1 CAD Model Using Solid Works
A CAD model has been designed for the prototype device using solid works. The
mechanical design proposes a device with a flexible head and rectangular base. The
rectangular base can be used to house the electronic components. The dimension
of the rectangular base is 5 cm ×3cm×3 cm. It is 5 cm wide, 3 cm long and
has a thickness of 3 cm. This would enable the device to occupy very less space on
the steering wheel. The rectangular base is provided with facility to attach a flexible
rubber strap by means of which the device can be wrapped around the steering wheel
342 V. R. Vasan et al.
or one of the rims of the wheel or the connector between the hub and the rim as per
the user’s preference. The length of the Joystick head is about 2.5 cm (Figs. 3,4,5,
6).
Fig. 3 A CAD model for the
prototype
Fig. 4 The mounting of the
device on the steering wheel
is illustrated
Fig. 5 Front view of the
mounted device
Peformance Analysis of Joysticks Used … 343
Fig. 6 Thesideviewofthe
mounted device
4 Electronic System Design of the Novel Car Accessory
This chapter is dedicated to designing the electronic system of the car accessory
proposed as part of the work. The device can be built as a microcontroller-based
device. The joystick is to be interfaced to the input pins of the controller. The inputs
should be decoded based on the signal strength obtained at the input lines. A Bluetooth
module is used to transmit Bluetooth messages that are intercepted by the android
phone. These messages are used to control the features in the phone using an android
service (Fig. 7).
4.1 Interfacing BT and Joystick to Arduino
The joystick was interfaced to Analog Pin A0(x-axis) and A1(y-axis) on Arduino
board. The power of 5 V required by the joystick was drawn from the power rails
available on the Arduino board. The GND was connected to GND point on the board.
Fig. 7 The system is represented using the block diagram
344 V. R. Vasan et al.
Fig. 8 BT and Joystick
interfaced to Arduino
The HC-05 Bluetooth module was powered using the onboard power supply from
the Arduino board, while Arduino board drew its power from USB port of the Laptop
or can be given through a battery pack. The TXD pin of HC05 was interfaced to
PD0(RXD),andtheRXDpintoPD
1(TXD) of the Arduino board. The GND point
of HC-05 was grounded.
The Arduino board was programmed to send BT data. Tera Term Software was
used to verify if the data was sent by the BT module accurately. The data transmitted
by the BT module based on the inputs from joystick is observed using Tera Term
emulator. 1 is sent for Positive X-axis, Data 2 is sent for Negative X-axis, Data 3
is sent for Positive Y-axis, and Data 4 is sent for Negative Y-axis. The Tera term
emulator is shown in Fig. 9(Fig. 8).
Fig. 9 Bluetooth message received on Tera term emulator
Peformance Analysis of Joysticks Used … 345
4.2 The Bluetooth Android Service
The android platform supports Bluetooth network stack, using which a device can
exchange data wirelessly with another Bluetooth device. Android Bluetooth APIs
which are part of the application framework are used to build Bluetooth-based
services and applications in android. The use of Bluetooth function call’s (API’s),
an application can perform the following:
•Search for accessible Bluetooth-enabled devices
•RFCOMM channels Establishment
•Connect to other Bluetooth devices
•Transfer data and control apps.
Bluetooth-enabled devices should first form a communication channel to transmit
data between each other. This is called PairingProcess. One device being discoverable
makes it available for incoming requests for connection. The application then finds
the detected device using a SDP process. Once the discoverable device accepts the
request for pairing, they exchange the security keys.
5 Experiments and Result Analysis
In the proposed work, a joystick has been used as an input device instead of a
traditional key-based input. The space advantage is obvious in the mechanical design
of the device. However, in order to test the effectiveness of the usage of the device,
an experiment was designed. The time taken by a user to respond to a request, for
example, the user has a hands-free device plugged in his car and a call arrives, the
user gets distracted and makes an effect to receive the call by pressing the key on
the hands-free device that is mounted on the steering wheel. If the number of keys
available on the device is more, the user needs to spend some effort to identify the
right key to be pressed. This adds on to the response time. The response time is
directly proportional to the amount of distraction a user experience.
5.1 Experiment Design
To verify the difference in response time while using a key-based device and Joystick-
based device, two MATLAB GUI applications were created. The first GUI simulated
the behavior of a key-based device (Fig. 10).
The key GUI consisted of 6 keys labeled randomly as A key, B key, M key, T key,
V key, and Y key. A Start key and End key were provided to start the experiment. On
346 V. R. Vasan et al.
Fig. 10 Key GUI experiment
pressing the Start key, a message was displayed to press Key A and a sequence of
keys randomly. The message was displayed at random time intervals and sequences
(Fig. 11).
Ten people of two different age groups 18–35, 35–60 response time was measured
while using the GUI. Each User was given three trials and their response time, the
time from when the message was displayed to the time when the key was pressed for
each of the keys was measured and average response time was recorded (Fig. 12).
Another GUI to simulate joystick behavior was designed. It consisted of only
four input points one at north, south, east, and west directions, the input points did
not have any labels. The same group of people who used Key Experiment GUI was
Fig. 11 Key experiment sequence
Peformance Analysis of Joysticks Used … 347
Fig. 12 Joystick GUI experiment
made to use the joystick experiment GUI. Commands, like up, down, left, right, were
displayed in a random sequence as shown in Fig. 13 and the time taken by the user
to respond to those messages were recorded. Each user was again given three trials
and the average response time was recorded (Tables 1,2,3, and Fig. 14).
Fig. 13 Joystick experiment sequence
348 V. R. Vasan et al.
Tabl e 1 Key GUI experiment results
Person Age
group
Akey
response
time (s)
Bkey
response
time (s)
Mkey
response
time (s)
Tkey
response
time (s)
Vkey
response
time (s)
Ykey
response
time (s)
Average
response
time (s)
118–35 1.701834 1.519811 1.474536 0.483003 1.248641 1.994034 1.4036431
218–35 1.657443 1.255486 1.500717 0.614673 1.392545 1.403633 1.3040828
318–35 2.357872 2.295666 2.570059 1.231183 2.023168 2.071644 2.0915986
418–35 1.463217 1.607738 1.686016 1.063738 1.549565 1.643458 1.5022887
518–35 1.896672 2.159435 1.584615 0.930515 0.31873 1.10495 1.3324862
636–60 3.371246 5.602098 2.356469 3.190361 2.276751 3.436133 3.3721764
736–60 1.576033 3.575897 3.281622 1.558561 3.072469 2.535655 2.600039
836–60 2.129173 3.695737 2.423518 1.626736 2.506214 3.021394 2.5671287
936–60 2.202831 3.839781 1.932216 1.458182 2.153239 7.042369 3.1047698
10 36–60 2.301473 1.663203 1.112434 1.621941 1.259696 1.453597 1.5687241
Tabl e 2 Joystick GUI experiment results
Person Age
group
Up
response
time (s)
Down
response
time (s)
Left
response
time (s)
Right
response
time (s)
Up twice
response
time (s)
Up and
right
response
time (s)
Average
response
time (s)
118–35 1.694997 1.245673 1.131581 1.214475 1.196757 1.016823 1.250051
218–35 1.82407 1.172241 0.703676 0.759536 1.114881 1.031267 1.100945
318–35 1.719633 1.054821 0.930883 1.503861 1.552299 1.832409 1.432318
418–35 1.839778 1.03758 1.876124 1.039421 1.698226 1.100322 1.431909
518–35 1.001523 1.113212 0.891723 0.711242 1.27988 1.823234 1.136802
636–60 2.83941 1.837578 1.769216 2.823801 2.592501 2.01351 2.312669
736–60 1.839778 1.03758 1.876124 2.039421 1.698226 1.701386 1.698753
836–60 2.316238 1.29518 3.198072 1.58362 2.323278 2.000185 2.119429
936–60 1.999988 1.910189 2.563862 2.583041 1.91427 2.135521 2.184479
10 36–60 2.687908 0.648679 0.169645 1.615324 1.280389 1.117162 1.253184
5.2 Result Analysis
X-axis represents the individuals and they are identified by identification number
1–10. The Y-axis represents the response time of the device in seconds.
Red-colored line indicates each individual’s response time in Key GUI Experiment
and Blue colored line indicates each individual’s response time in Joystick GUI
Experiment.
From the data statistics it is clearly evident that the response time in case of key-
based inputs is relatively higher when compared to joystick-based Input interface.
Peformance Analysis of Joysticks Used … 349
Tabl e 3 Data statistics for
each of the experiment Joystick GUI response Time (s)
Min 1.1009
Max 2.3127
Mean 1.5921
Median 1.4321
Mode 1.1009
Std 0.4584
Range 1.2117
Key GUI response Time (s)
Min 1.3041
Max 3.3722
Mean 2.0847
Median 1.8302
Mode 1.3041
Std 0.7775
Range 2.0681
1 2 3 4 5 6 7 8 910
1
1.5
2
2.5
3
3.5
Individual Identification Number
Respons Time in Seconds
Key Expt
JoyStick Expt
Fig. 14 A comparison of the response time
350 V. R. Vasan et al.
The maximum response time for joystick was 2.3127 s and that for key input was
3.3722 s almost 50% more than the response time of joystick experiment.
6 Summary
Many Hands-free devices are available in the market that could be mounted in the
car and they enable us to control some of the features of the Android phone like
make/receive call, play music, etc. Keys are provided on the joystick by means of
which a user can perform a particular action. Some of these devices have built-
in speakers/displays that play/display the phone data. However, the smartphones
these days are big and powerful enough that eliminates the need for separate
speaker/display units in hands-free device. The current work proposes a hands-free
device that would use a unique design—A joystick-based Input Interface.
A CAD model has been designed for the prototype device using solid works. The
mechanical design proposes a device with a flexible head and rectangular base. The
rectangular base can be used to house the electronic components. The dimension of
the rectangular base is 5 cm ×3cm×3 cm. It is 5 cm wide, 3 cm long and has a
thickness of 3 cm. This would enable the device to occupy very less space on the
steering wheel. Due to the Axial movement provided by the joystick, it is possible
to generate many input actions using a single joystick. This would offer the user to
control many functionalities on the phone using different input actions. Also, the
distraction level is directly proportional to time taken by the driver in performing
a specific action using the hands-free, say connect a call by pressing a key. An
experiment was conducted to measure the response time of the users with respect
to handling key-based device and joystick-based devices. It was observed that the
response time reduces by almost 50% in case of the joystick device as the human
brain finds it much easier to respond to repetitive common actions rather than the
search action.
The design and production cost of the proposed device is relatively cheaper when
compared to other devices available in the market as it uses minimal manufacturing
material due to its size. The electronic components chosen: ATMEGA 328 Microcon-
troller, HC05 Bluetooth are relatively cheaper and hence the device can be introduced
to the market at a price almost equal to Rupees 500. The prototype for the device
was developed on an Arduino board. A two-axis joystick was used as the input
device. A HC-05 Bluetooth module was used to communicate to the android device.
An Android Bluetooth service was created which received the commands from the
Bluetooth module and controlled certain features of the phone which in turn reduces
the driver’s distraction level and helps him/her drive safely on the road.
Peformance Analysis of Joysticks Used … 351
References
1. D.A. Redelmeier M.D., R.J. Tibshirani, Association between Cellular-Telephone Calls and
Motor Vehicle Collisions.https://doi.org/10.1056/NEJM199702133360701
2. AAMI, AAMI young driver’s annual road safety index, AAMI, Melbourne (2010). Retrieved
27 September 2012, from http://aami.com.au/sites/default/files/fm/news/2010-AAMI-Young-
Drivers-Index.pdf
3. L.T. Lam, Distractions and the risk of car crash injury: The effect of drivers’ age. J. Saf. Res.
https://doi.org/10.1016/S0022-4375(02)00034-8
4. T. Horberry, Driver distraction: The effects of concurrent in-vehicle tasks, road environment
complexity and age on driving performance. Accid. Anal. Prev. 38, 185–191 (2006)
5. W. Consiglio, Effect of cellular telephone conversations and other potential interface on
reaction time in a braking response.https://doi.org/10.1016/S0001-4575(02)00027-1
6. K.E. Beede, Engrossed in conversation: The impact of cell phones on simulated driving
performance. https://doi.org/10.1016/j.aap.2005.10.015
7. M. Saifuzzaman, Z. Zheng, Incorporating human-factors in car-following models: a review of
recent developments and research needs. Transp. Res. Part C Emerg. Technol. (2014). https://
doi.org/10.1016/j.trc.2014.09.008
8. H. Alm, L. Nilsson, The effects of a mobile telephone task on driver behaviourin a car following
situation. Accid. Anal. Prev. 27, 707–715. http://dx.doi.org/10.1016/0001-4575(95)00026-V
9. J. Törnros, Mobile phone use—effects of conversation on mental workload and driving speed
in rural and urban environments. Transp. Res. Part F Traffic Psychol. Behav. (2006). https://
doi.org/10.1016/j.trf.2006.01.008
10. T. Ranney, G. Watson, E.N. Mazzae, Y.E. Papelis, O. Ahmad, J.R. Wightman, Examination
of the Distraction Effects of Wireless Phone Interfaces Using the National Advanced Driving
Simulator-Preliminary Report on Freeway Pilot Study (2004). Paper No. DOT 809 737
11. D. Strayer, J. Watson, F. Drews, Cognitive distraction while multitasking in the automobile.
Psychol. Learn. Motiv. 54, 29–58 (2011). https://doi.org/10.1016/B978-0-12-385527-5.000
02-4
12. R. Matthews, S. Legg, S. Charlton, The effect of cell phone type on drivers subjective workload
during concurrent driving and conversing. Accid. Anal. Prev. 35(4), 451–457 (2003)
13. L. Nilsson, Behavioural research in an advanced driving simulator—experiences of the VTI
system, in Paper Presented at the 37th Annual Meeting of the Human Factors and Ergonomic
Society, Seattle, USA, 11–15 October 1993
14. C. Patten, A. Kircher, J. Östlund, L. Nilsson, Using mobile telephones: cognitive workload and
attention resource allocation. Accid. Anal. Prev. 36, 341–350 (2004)
Probabilistic Principal Component
Analysis (PPCA) Based Dimensionality
Reduction and Deep Learning for Cancer
Classification
D. Menaga and S. Revathi
Abstract Cancer is a threatening disease that causes the death of major humans. The
DNA microarray-based gene expression profiling is an effective technique employed
for cancer classification, diagnosis, and treatment. The early detection creates a need
for a better and accurate method that offers the information of cancer in the patient
that enables better decision-making by the clinicians and treating them. Hence, in
this work, a cancer classification technique is devised by utilizing the gene expres-
sion data collected from the patients. Since the input gene expression data is huge,
the dimensionality of the data is reduced, undergoing transformation using Proba-
bilistic Principal Component Analysis (PPCA), for the efficient computation of the
maximum likelihood estimates. Once the dimensionality of the data is reduced, it
is classified using Deep Belief Network (DBN) to perform the cancer classification.
DBN classifies the dimensionality reduced gene data, providing two classes, namely,
normal and abnormal. The comparative performance analysis of the proposed tech-
nique is evaluated based on the three metrics, such as sensitivity, specificity, and
accuracy. The effectiveness of the devised methodology was deliberated through the
achieved accuracy value of 0.8085, specificity value of 0.8333 and sensitivity of
0.7981.
Keywords Cancer classification ·Gene expression data ·Dimensionality
reduction ·PPCA ·DBN
1 Introduction
The cancerous tumors can be treated appropriately only when the exact type of cancer
is identified earlier. The conventional cancer diagnosis was done on the basis of the
D. Menaga (B)·S. Revathi
Department of CSE, B.S. Abdur Rahman Crescent Institute of Science and Technology, Chennai,
India
e-mail: dev.menaga@gmail.com
S. Revathi
e-mail: srevathi@crescent.education
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_31
353
354 D. Menaga and S. Revathi
clinical and morphological appearance [1]. Since cancer is caused by various factors,
these traditional methods were not capable of identifying the cancer effectively [2,
3]. The advancement in cancer cell classification has gained a lot of research interest
due to the emerging need for cancer treatment. The cancer classification is also
based on the tumor’s morphological appearance, but it is affected by the limitation of
incorrect prediction of the cancer cell. The incorrect prediction is mainly due to the
nature of cancer cell, which portrays distinct symptoms at distinct conditions. With
the aim of resolving this limitation, the DNA microarray technology was adopted
for performing the gene expression data-based classification as it provided highly
accurate classification [4,5]. It is an essential platform widely utilized for analyzing
the gene expression data of millions of genes in various experimental studies [6]. The
microarray technology was evolved estimating the variation in all the genes available
in the provided biological sample by the utilization of the hybridization mechanism.
The gene expression patterns in the microarray data provide the useful information
related to the significant advancement in the fields of biology and clinical medicine for
the medical prognosis and diagnosis [7]. The exploration of the molecular variation
existing in the cancer cells for the categorization of the tissue into cancerous or
normal tissues is one of the prevalent applications of the microarray technology [8].
The DNA microarray methodology performs the identification, classification, and
prediction of the distinct cancer types [9,10]. The proliferation of gene expres-
sion is done for estimating the enormous genes available within a single RNA by
practicing hybridization process. The DNA microarray is created by integrating the
healthy DNA reference and the testing DNA [11,12]. It is utilized for expressing
and comparing the additional insight about the genes at specified time and status for
carrying out the identification of cancer or any other disease [13,15]. The color spot
matrix comprising of the gene expression level’s quantitative values is generated by
the utilization of the laser and fluorophores [14]. The obtained gene expression level
is defined as a signature beneficial for the diagnosis of the distinct diseases, which
occur mainly due to the alteration in the gene expression level [16]. The DNA microar-
rays when concatenated with the other computational approaches become capable of
providing an efficient diagnosis. The DNA microarray classification comprises three
steps, they are prediction of the gene, discovery of class, and prediction of class [17,
19]. The basic task of microarray data classification is the determination of the class
for the concerned unknown samples. The millions of genes together constitute the
DNA microarray samples and the selection of best gene set is a tough task, so the
modeling of a trustworthy classification methodology is an immediate need. The
presence of unwanted genes increases both the noise and the dimensionality of the
obtained gene expression matrix; this further leads to high computational complexity
[14,18].
Various classification methodologies based on the gene expression data were
developed for performing the cancer identification. Anyhow, the gene expression
data-based cancer classification suffers from the issues, like high dimensionality and
multi-class imbalance [8]. Hence, in this paper an effective cancer cell classification
methodology and feature selection scheme are to be devised for the classification
of the given microarray data with the intention of identifying cancer. The distinct
Probabilistic Principal Component Analysis … 355
research works related to the microarray dataset-based cancer cell classification were
not able to correctly identify the reduced number of informative genes with maximum
classification accuracy. Initially, the dimensionality of the gene expression data is
reduced by employing the PPCA [21] through the effective estimation of maximum
likelihood. This reduces the dimensionality related complexity issues during clas-
sification. Subsequently, the gene expression data with the reduced dimension is
classified utilizing the DBN classifier. From the obtained classification results, the
cells under examination can be either declared as a normal cell or abnormal cell.
The rest of this paper is organized as follows: The brief introduction on the gene
expression-based cancer cell classification is presented in Sect. 1and the survey
on the relevant works is discussed in Sect. 2. Section 3elaborates the proposed
method of cancer cell classification by adopting the PPCA and DBN classifier and
the simulation results of the proposed classification methodology are demonstrated
in Sect. 4, and finally, in Sect. 5, the conclusion of the research is presented.
2 Motivation
2.1 Literature Survey
Here, four literature works related to the gene expression-based cancer cell
classification are presented.
Liu et al. [2] developed a simple Principal Component Accumulation (PCAcc)
scheme for resolving the issues in the gene expression data-based cancer classifi-
cation. The main intention was to reduce the dimensionality through the utilization
of multiple Principal Component (PC) subspaces rather than the first few PCs. In
spite of this advantage; this method could not provide the biological interpretations.
Patricia Melin Oscar Castillo [3] presented type-2 fuzzy logic for resolving the issues
faced by clustering, pattern recognition, and classification. The type-2 fuzzy rule’s
structure is the same as the structure of the type-1 fuzzy rule case; because the only
difference between the type-1 and type-2 is the membership functions nature. The
devised scheme was applicable in a wide range of application with high complexity
and the maximum degree of uncertainty and is computationally inexpensive.
Alshamlan et al. [15] presented a Bio-Inspired evolutionary gene selection
methodology for performing the microarray-based cancer classification. The devel-
oped algorithm was an effective wrapper gene selection scheme; it has the ability
to search the near-optimal or optimal solutions in the large dimensional space. This
scheme achieved maximum classification accuracy with a limited number of selected
genes. Van De Vijver et al. [14] developed a complementary DNA (cDNA) microarray
analysis for the treatment of breast cancer. The prognosis profile is a strong devel-
opment predictor for performing the patient’s distant metastases with the lymph-
node positive disease. This method suffered from the risks of under treatment and
overtreatment because of the improper patient selection.
356 D. Menaga and S. Revathi
2.2 Challenges
The distinct challenges faced by the gene expression-based cancer cell classification
are elaborated as follows,
•The microarray dataset is affected by number of irrelevant genes and the level of
noisy genes; these made the classification task highly challenging [15].
•Cancer classification is a tedious process since it is historically dependent on the
biological knowledge rather than the unbiased and systematic schemes for the
effective recognition of the tumor subtypes [4].
•The availability of unwanted genes increases both the noise as well as dimen-
sionality to the gene expression data; these leads to hike in the complexity of
computation during clustering and classification [19].
•The numbers of features present are comparatively larger than the available
number of cancer samples, this leads to the risk of overfitting during the classifier’s
training phase [20].
3 Gene Expression-Based Cancer Cell Classification
by Adopting the Proposed PPCA-Based DBN Classifier
The proposed PPCA-based DBN classifier for classifying the cancer cell based on
the gene expression data is elaborated in this section. Figure 1portrays the block
diagram of the cancer cell classification employing the proposed PPCA-based DBN
classifier. As a first step; the features are extracted from the provided input gene data.
In general, the extracted features will be fed to the classifier for training; but here the
dimensionality reduction phase is included for reducing the dimensionality of the
Fig. 1 Block diagram of cancer cell classification using the proposed PPCA-based DBN classifier
Probabilistic Principal Component Analysis … 357
feature space. The methodology adopted for the dimensionality minimization is the
PPCA. Now, the feature vector with the reduced dimensionality will be trained and
classified using the DBN classifier. Eventually, the provided input gene expression
data will be classified as either normal cell or abnormal cell conveying the absence
or presence of cancer accordingly.
3.1 Input Gene Expression Data
Let the input gene expression data be indicated as and the gene data’s dimension
carrying the gene information be. The considered gene expression data comprises
gene information from number of individuals. The millions of genes in a person
together constitute the gene expression data; the mechanism of learning the gene
characteristics supports in better understanding of human life. The classification
of gene expression data must be done for the earlier diagnosis and prevention of
diseases.
3.2 Dimensionality Reduction Using PPCA
The provided input data is subjected to feature extraction for the extraction of relevant
features. As the input database is of large size, probably there will be a traffic hike in
the feature space dimensionality. This further leads to the upsurge in the complexity
during the classifier’s training phase. Hence, PPCA [21] is employed in this work,
for reducing the dimensionality of the feature space. The PPCA utilizes the minor
algebraic manipulation for the dimensionality reduction; it is denoted as B.The
dimensionality reduction is done by practicing the conditional probability on the
time-space; it is expressed as shown in the below equation,
ru|su=P−1BT(su−μ) (1)
Here, the mean value is denoted as μ, the conditional probability over the time su
is denoted as ru, and the projection measure having the value of BTB−1is indicated
as P−1. Eventually, the feature space dimensionality is reduced and is provided to the
classifier. Let Sbe the selected genes obtained as a result of dimensionality reduction
and is of size M×h, such that h<N, and the database is represented as
S=Sxy;1<x≤M;1<y≤h(2)
358 D. Menaga and S. Revathi
3.3 DBN-Based Classification of Cancer
The next step after the dimensionality reduction is the classification of the provided
feature vector, for the identification of cancer. The classification scheme adopted for
the cancer cell classification is the DBN.
(i) DBN Architecture
The DBN [22] is a generative neural network comprising of many layers. Every layer
is composed of the input layer represented by the visible neurons and the output
layer represented by the hidden neurons. One of the extraordinary characteristics of
the DBN is that there are no links among the visible neurons and no links among
the hidden neurons. The connections are symmetric and the visible neurons are
completely interlinked with the hidden ones. The employed DBN comprises Multi-
Layer Perceptron (MLP) layer and multiple layers of Restricted Boltzmann Machines
(RBM). The MLP layer comprises the input layers, hidden layers, and output layers;
and the RBM’s hidden layer acts as the output layer of the successive layers. Here,
the DBN architecture is designed by considering two RBMs and an MLP layer, and
the architecture is sketched in Fig. 2. The DBN is provided with the feature vector
with reduced dimension from the PPCA for the effective cancer cell identification.
Because of the dimensionality reduction through PPCA, the developed classification
scheme achieves maximum classification accuracy.
The input layer and hidden layer of the RBM1 is given as
Fig. 2 DBN architecture
Probabilistic Principal Component Analysis … 359
LR1
1=LR1
1,LR1
2,...,LR1
j,...,LR1
h=Sxy;1<j≤h(3)
GR1
1=GR1
1,GR1
2,...,GR1
φ,...,GR1
n;1<φ≤n(4)
where the number of input neurons in RBM1 of DBN is denoted as hand the number
of hidden neurons in RBM1 of DBM is denoted as n. The RBM1’s weight is denoted
as WR1
1=WR1
jφand the RBM1’s weight dimension is indicated as h×n.The
output of RBM1 layer is given by the below equation,
GR1
φ=ε⎡
⎣ψR1
φ+
j
LR1
j×WR1
jφ
⎤
⎦(5)
Here, εis the activation function, the bias of the φth hidden layer of RBM1 is
denoted as ψR1
φ, and the jth neuron in the RBM1 layer is indicated as LR1
j.
Likewise, RBM’s input layer and hidden layer are represented as
LR2
1=LR2
1,LR2
2,...,LR2
n=GR1
φ;1<φ≤n(6)
GR2
1=GR2
1,GR2
2,...,GR2
φ,...,GR2
n(7)
The total number of hidden neurons in the RBM1 layer is same as the total
number of input neurons in the RBM2 layer. The RBM2 layer’s weight is indicated
as WR2
2=WR2
φφ . The existing neuron weight between the RBM1’s φth visible
neuron and the RBM2’s φth hidden neuron is denoted as WR2
φφ . The RBM2 layer’s
output is expressed by the below equation
GR2
φ=ε⎡
⎣ψR2
φ+
j
LR2
j×WR2
φφ
⎤
⎦∀LR2
j≈GR1
φ(8)
Now, the output from the RBM2’s hidden layer is fed as the input to the MLP
layer of the DBN classifier. The MLP’s input layer and hidden layer is represented
as shown in Eqs. (9) and (10), respectively.
Q=Q1,Q2,...,Qφ,..., Qn=GR2
φ;1≤φ≤n(9)
GQ=GQ
1,GQ
2,....,GQ
u,...,GQ
p;1≤u≤p(10)
The MLP layer’s output is represented as
o=o1,o2,...,of,...,oc(11)
360 D. Menaga and S. Revathi
Here, the total count of output is denoted as cand the fth output neuron’s output
is denoted as of. The MLP layer’s output is expressed as shown below
of=
p
u=1
Wuf ∗GQ
u;1≤u≤p;1<f≤c(12)
Here, the weight between the fth output neuron and uth hidden neuron is denoted
as Wuf and the dimension of which is given by (p×c). The output at the hidden
layer is expressed as shown in Eq. (13)
GQ
u=⎡
⎣
b
φ=1
Wφu∗Qφ⎤
⎦Ku∀Qφ=GR2
φ(13)
Here, the hidden neuron’s bias is denoted as Kuand the weight between the φth
input neuron and the uth hidden neuron is denoted as Wφuand its dimension is given
by (b×p).
(ii) Training Phase
The training phase in the MLP layer of DBN is carried out by practicing the supervised
steepest descent algorithm. The RBMs are trained by utilizing the unsupervised
learning of steepest descent algorithm. The series of steps followed during the training
phase of DBN are explained below:
Step 1 Training of RBM1 layer and RBM2 layer
The first step is the training of the RBM layer. The DBN classifier employed in this
research comprises two RBMlayers. The training sample with the selected genes of
reduced dimension is provided as the input to the RBM1 layer for estimating the data’s
probability distribution. Then, the computed probability distribution is encoded with
the weights and the determined output is provided as the input to the RBM2 layer.
The similar procedure is carried out by the RBM2 layer and the obtained output is
fed to the MLP layer.
Step 2: Training Phase of MLP layer
The steps followed in the DBN’s MLP layer training are given below
(b) Read the input sample: The output from the RBM2 layer indicated as GR2
φis
provided as the input to MLP layer.
(c) Estimation of MLP layer’s output: The MLP layer’s outputs are given by Gu
and ofand the outputs are dependent on the steepest descent algorithm’s weight
update equation.
(a) Initialization: The first step is the random initialization of neuron weights of
the MLP layer. Let the visible layer weight be indicated as Wuf and the hidden
layer weight be indicated as Wφu.
Probabilistic Principal Component Analysis … 361
Step 3: Error computation
The network’s error calculated on the basis of the average Mean Square Error (MSE)
is expressed using the below equation
EAverage =1
h×
h
j=i
oj
f−oj
grd;1≤f≤c(14)
Here, the total number of data samples is denoted as h, the output of the DBN
classifier is denoted as oj
f, and the desired output of the DBN classifier is indicated
as oj
grd .
Step 4: Updating the weights in the MLP layer By concerning the partial deriva-
tive of the average error EAverage, the weights of the visible neurons and the hidden
neurons are updated. The incremental weights are determined by utilizing the
derivatives of Eq. (14) are given in the below equations
WSD
φu=−ξ×αEAverage
αWφu
(15)
WSD
φj=−ξ×αEAverage
αWuj
(16)
Here, the learning rate is denoted as ξ. The steepest descent algorithm’s weight
update equation is expressed as
WSD
φu(T+1)=WSD
φu(T)+WSD
φu(17)
Step 5: Re-computation of the average error of MLP outputs The average error
of the MLP output is determined by utilizing Eq. (14). If the obtained error value is
high, then the above steps are repeated. If the obtained error value is low, then the
neuron weight is updated using Eq. (17), which is the weight update equation of the
steepest descent algorithm.
Step 6: Termination The above steps are iteratively repeated for a maximum
number of iterations, denoted as T. Eventually, the optimal neuron weights are
found and on the basis of the obtained best solution, either the absence or pres-
ence of cancer is identified. The decisions are made by the genes, which vary from
the normal behavior of genes.
4 Results and Discussion
The experimentation results of the proposed PPCA-based DBN classifier for the
cancer cell classification are demonstrated in this section.
362 D. Menaga and S. Revathi
4.1 Experimental Setup
The experimentation is done in a PC configured with Intel core-i3 processor,
Windows 10 OS, and 4 GB RAM. The implementation tool adopted in this research
for carrying out the experiment is JAVA.
4.2 Dataset Description
The Standard databases, like Colon dataset [23] and Leukemia dataset [24], are
utilized in this research for performing the experimentation. The Leukemia dataset
contains gene expression measurements on 72 leukemia patients. The Leukemia
dataset comprises 7128 genes; it is stored as a matrix of dimension 7128 ×72 (10 MB)
with the column names representing the class labels. There is also a smaller subset of
these data, consisting of 3571 genes, which is stored as the 3571 ×72 matrix (5 MB),
with the column names representing the class labels. The colon dataset designed for
studying colon tumors has 6500 genes comprising of 22 samples of normal colon
tissue and 40 samples of colon tumor.
4.3 Evaluation Metrics
The evaluation metrics concerned for the analysis of the devised PPCA-based DBN
are sensitivity, specificity, and accuracy; these metrics are defined in this part as
follows:
Sensitivity: The sensitivity metric estimates the true positives attained by the
developed classifier. It is formulated as
Sensitivity =TP
TP +FN (18)
where the true positive is denoted as TP, the true negative is indicated as TN, the
false positive is denoted as FP, and the false negative is indicated as FN.
Specificity: The specificity metric computes the true negatives attained by the
devised PPCA-based DBN classifier, and is expressed as
Specificity =TN
TN +FP (19)
Accuracy: The accuracy is defined as the percentage of accurate classification for
the identification of the presence/absence of cancer. It is computed as
Probabilistic Principal Component Analysis … 363
Accuracy =TN +TP
TN +TP +FN +FP (20)
4.4 Methods Taken for the Comparison
The different classifiers chosen here for the comparative performance analysis are
NN [9], DBN [22], PPCA +NN [21, 9]. The performance of these compara-
tive techniques will be compared with that of the proposed method to evaluate its
effectiveness.
4.5 Comparative Analysis
The results of comparative analysis of the devised PPCA-based DBN classifier with
the comparative models are demonstrated in this section by varying the percentage
of training data. When the 50% of data is used for training, remaining 50% of data
is used for testing. Similarly, for 60, 70, 80, and 90% of the training data, the testing
data is 40, 30, 20, and 10%.
(i) Using Colon dataset
The comparative analysis results of comparative classifiers on the colon dataset are
described here.
The comparative analysis based on varying percentage of training data (50–90%)
on the colon dataset is depicted in Fig. 3. Figure 3.a displays the comparative anal-
ysis of the distinct classifiers based on sensitivity. At 90% training data, sensitivity
measure attained by the NN classifier is 0.7384, the DBN classifier is 0.74, the PPCA
+NN classifier is 0.7977, and the PPCA +DBN classifier is 0.7984. Figure 3.b
sketches the comparative analysis of the distinct classifiers based on specificity. For
90% training data, the specificity values of the NN classifier, DBN classifier, PPCA
+NN classifier, and PPCA +DBN classifier are 0.5416, 0.6521, 0.6994, and 0.7672,
respectively. The comparative analysis of the distinct classifiers based on accuracy
is pictorized in Fig. 3.c. The accuracy measure attained at 90% training data by the
NN classifier is 0.5416, the DBN classifier is 0.6521, the PPCA +NN classifier is
0.6994, and the PPCA +DBN classifier is 0.7672. The obtained maximum accuracy,
high specificity, and sensitivity deliberate the effectiveness of the proposed PPCA +
DBN classifier.
Figure 4depicts the comparative analysis based on varying number of hidden
neurons on the colon dataset. The comparative analysis of the different classifiers
based on sensitivity is displayed in Fig. 4.a. The sensitivity values for the number
of hidden neurons =500 are 0.5803 for NN classifier, 0.6782 for DBN classifier,
0.7973 for PPCA +NN classifier, and 0.7981 for PPCA +DBN classifier. Figure 4.b
364 D. Menaga and S. Revathi
(a) (b)
(c)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
50 60 70 80 90
Sensitivity
Percentage of training data
NN
DBN
PPCA+NN
PPCA+DBN
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
50 60 70 80 90
Specificity
Percentage of training data
NN
DBN
PPCA+NN
PPCA+DBN
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
50 60 70 80 90
Accuracy
Percentage of training data
NN
DBN
PPCA+NN
PPCA+DBN
Fig. 3 Comparative analysis using colon dataset based on training data percentage asensitivity,
bspecificity, and caccuracy
portrays the comparative analysis of the different classifiers based on specificity. As
the number of hidden neurons =500, the achieved specificity values are 0.7058
for NN classifier, 0.7328 for DBN classifier, 0.7487 for PPCA +NN classifier,
and 0.8333 for PPCA +DBN classifier. The comparative analysis of the distinct
classifiers based on accuracy is sketched in Fig. 4.c. The accuracy measures attained
when the number of hidden neurons =500 are 0.6714 for NN classifier, 0.7285 for
DBN classifier, 0.7686 for PPCA +NN classifier, and 0.8085 for PPCA +DBN
classifier. The superiority of the proposed PPCA +DBN classifier over the other
classifiers is conveyed through the obtained values.
4.5.1 Using Leukemia Dataset
The comparative analysis results of the distinct classifiers on the Leukemia dataset
are elaborated in this part.
The comparative analysis based on varying percentages of training data on the
Leukemia dataset is depicted in Fig. 5. Figure 5a displays the comparative analysis of
the distinct classifiers based on sensitivity. For 90% training data, sensitivity measure
attained by the NN classifier is 0.3826, the DBN classifier is 0.6, the PPCA +NN
Probabilistic Principal Component Analysis … 365
(a) (b)
(c)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
100 200 300 400 500
Sensitivity
Number of hidden neurons
NN
DBN
PPCA+NN
PPCA+DBN
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
100 200 300 400 500
Specificity
Number of hidden neurons
NN
DBN
PPCA+NN
PPCA+DBN
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
100 200 300 400 500
Accuracy
Number of hidden neurons
NN
DBN
PPCA+NN
PPCA+DBN
Fig. 4 Comparative analysis using colon dataset based on the number of hidden neurons
asensitivity, bspecificity, and caccuracy
classifier is 0.7946, and the PPCA +DBN classifier is 0.7951. Figure 5b sketches the
comparative analysis of the distinct classifier based on specificity. For 90% training
data, the specificity value of the NN classifier, DBN classifier, PPCA +NN classifier,
and PPCA +DBN classifier is 0.6395, 0.7246, 0.7374, and 0.75, respectively. The
comparative analysis of the distinct classifier based on accuracy is pictorized in
Fig. 5c. The accuracy measures attained at 90% training data by the NN classifier
is 0.4438, the DBN classifier is 0.6286, the PPCA +NN classifier is 0.6534, and
the PPCA +DBN classifier is 0.7057. The determined accuracy, specificity, and
sensitivity elucidate the effectiveness of the proposed PPCA +DBN classifier over
the other classifiers.
Figure 6depicts the comparative analysis based on varying number of hidden
neurons on the Leukemia dataset. The comparative analysis of the different classifiers
based on sensitivity is displayed in Fig. 6a. The sensitivity values at number of hidden
neurons =500 are 0.3826 for NN classifier, 0.6 for DBN classifier, 0.7926 for PPCA
+NN classifier, and 0.7938 for PPCA +DBN classifier. Figure 6b portrays the
comparative analysis of the different classifiers based on specificity. At the number
of hidden neurons =500, the achieved specificity values are 0.6309 for NN classifier,
0.7355 for DBN classifier, 0.75 for PPCA +NN classifier, and 0.8550 for PPCA +
366 D. Menaga and S. Revathi
(a) (b)
(c)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
50 60 70 80 90
Sensitivity
Percentage of training data
NN
DBN
PPCA+NN
PPCA+DBN
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
50 60 70 80 90
Specificity
Percentage of training data
NN
DBN
PPCA+NN
PPCA+DBN
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
50 60 70 80 90
Accuracy
Percentage of training data
NN
DBN
PPCA+NN
PPCA+DBN
Fig. 5 Comparative analysis using leukemia dataset using based on training data percentage
asensitivity, bspecificity, and caccuracy
DBN classifier. The comparative analysis of the distinct classifiers based on accuracy
is sketched in Fig. 6c. The accuracy measures attained when the number of hidden
neurons =500 for NN classifier, DBN classifier, PPCA +NN classifier, and PPCA
+DBN classifier are 0.4438, 0.6283, 0.6978, and 0.7643, respectively. The excellent
performance of the proposed PPCA +DBN classifier over the other classifiers is
conveyed through the determined values.
5 Conclusion
In this paper, an effective cancer cell classification methodology, entitled PPCA-
based DBN, is developed for the classification of cells. Initially, for the gene expres-
sion data provided as the input, the dimensionality of the feature space is reduced by
adopting the PPCA. Then, the dimensionality reduced features are fed to the DBN
classifier for the classification of the data into normal and abnormal. Eventually, the
presence or absence of the cancer is discovered by the proposed PPCA-based DBN
classifier. The experimentation of the proposed PPCA-based DBN is evaluated by
utilizing the metrics like sensitivity, specificity, and accuracy. The simulation results
Probabilistic Principal Component Analysis … 367
(a) (b)
(c)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
100 200 300 400 500
Sensitivity
Number of hidden neurons
NN
DBN
PPCA+NN
PPCA+DBN
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
100 200 300 400 500
Specificity
Number of hidden neurons
NN
DBN
PPCA+NN
PPCA+DBN
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
100 200 300 400 500
Accuracy
Number of hidden neurons
NN
DBN
PPCA+NN
PPCA+DBN
Fig. 6 Comparative analysis using leukemia dataset based on number of hidden neurons
asensitivity, bspecificity, and caccuracy
based on the number of hidden neurons deliberated that the proposed PPCA-based
DBN outperformed the other classifiers by achieving the sensitivity value of 0.7981,
specificity value of 0.8333, and accuracy measure of 0.8085.
References
1. Z. Liu, D. Tang, Y. Cai, R. Wang, F. Chen, A hybrid method based on ensemble WELM
for handling multi-class imbalance in cancer microarray data. Neurocomputing 266, 641–650
(2017)
2. J.J. Liu, W.S. Cai, X.G. Shao, Cancer classification based on microarray gene expression data
using a principal component accumulation method. Sci. China Chem. 54(5), 802–811 (2011)
3. Patricia Melin, Oscar Castillo, A review on type-2 fuzzy logic applications in clustering,
classification and pattern recognition. Appl. Soft Comput. 21, 568–577 (2014)
4. T.R. Golub, D.K. Slonim, P. Tamayo, C. Huard, M. Gaasenbeek, J.P. Mesirov, H. Coller et al.,
Molecular classification of cancer: class discovery and class prediction by gene expression
monitoring. Science 286(5439), 531–537 (1999)
5. Z. Zainuddin, P. Ong, Reliable multiclass cancer classification of microarray gene expression
profiles using an improved wavelet neural network. Expert Syst. Appl. 38(11), 13711–13722
(2011)
368 D. Menaga and S. Revathi
6. V. García, J.S. Sánchez, Mapping microarray gene expression data into dissimilarity spaces for
tumor classification. Inf. Sci. 294, 362–375 (2015)
7. R.M. Luque-Baena, D. Urda, J.L. Subirats, L. Franco, J.M. Jerez, Analysis of cancer microarray
data using constructive neural networks and genetic algorithms, in Proceedings of the Inter-
national Work-Conference on Bioinformatics and Biomedical Engineering (IWBBIO) (2013),
pp. 55–63
8. A. Antoniadis, S. Lambert-Lacroix, F. Leblanc, Effective dimension reduction methods for
tumor classification using gene expression data. Bioinformatics 19(5), 563–570 (2003)
9. B.A. Garro, K. Rodríguez, R.A. Vázquez, Classification of DNA microarrays using artificial
neural networks and ABC algorithm. Appl. Soft Comput. 38, 548–560 (2016)
10. C. Sotiriou, M.J. Piccart, Taking gene-expression profiling to the clinic: when will molecular
signatures become relevant to patient care? Nat. Rev. Cancer 7(7), 545 (2007)
11. A. Abderrahim, E.-G. Talbi, M. Khaled, Hybridization of genetic and quantum algorithm for
gene selection and classification of microarray data. Int. J. Found. Comput. Sci. 23(02), 431–444
(2012)
12. E. Alba, J. Garcia-Nieto, L. Jourdan, E.-G. Talbi, Gene selection in cancer classification using
PSO/SVM and GA/SVM hybrid algorithms, in The Proceedings of the IEEE Congress on
Evolutionary Computation (CEC 2007) (2007), pp. 284–290
13. C.J. Alonso-González, Q.I. Moro-Sancho, A. Simon-Hurtado, R. Varela-Arrabal, Microarray
gene expression classification with few genes: criteria to combine attribute selection and
classification methods. Expert. Syst. Appl. 39(8), 7270–7280 (2012)
14. M.J.VanDeVijver,Y.D.He,L.J.Van’tVeer,H.Dai,A.A.M.Hart,D.W.Voskuil,G.J.Schreiber
et al., A gene-expression signature as a predictor of survival in breast cancer. N. Engl. J.
Medicine. Eng. J. Med. 347 (25), 1999–2009 (2002)
15. H.M. Alshamlan, G.H. Badr, Y.A. Alohali, The performance of bio-inspired evolutionary gene
selection methods for cancer classification using microarray dataset. Int. J. Biosci., Biochem.
Bioinform. 4(3), 166 (2014)
16. M. Dashtban, M. Balafar, Gene selection for microarray cancer classification using a new
evolutionary method employing artificial intelligence concepts. Genomics. 109(2), 91–107
(2017)
17. T.K. Karakach, R.M. Flight, S.E. Douglas, P.D. Wentzell, An introduction to DNA microarrays
for gene expression analysis. Chemometr. Intell. Lab. Syst. 104(2), 28–52 (2010)
18. S. Ghorai, A. Mukherjee, S. Sengupta, P.K. Dutta, Multicategory cancer classification from gene
expression data by multiclass NPPC ensemble, in The Proceedings of the IEEE International
Conference on Systems in Medicine and Biology (ICSMB) (2010), pp. 41–48
19. H.M. Alshamlan, G. H. Badr, Y. Alohali, A study of cancer microarray gene expression profile:
objectives and approaches, in Proceedings of the World Congress on Engineering, vol 2 (2013),
pp. 1–6
20. H. Hijazi, C. Chan, A classification framework applied to cancer gene expression profiles. J.
Healthc. Eng. 4(2), 255–283 (2013)
21. M.E. Tipping, C.M. Bishop, Probabilistic principal component analysis. J. Roy. Stat. Soc.
Series B (Statistical Methodology) 61(3), 611–622 (1999)
22. M.A. Salama, A.E. Hassanien, A.A. Fahmy, Deep belief network for clustering and classifi-
cation of a continuous data, in Proceedings of the The 10th IEEE International Symposium on
Signal Processing and Information Technology (2010), pp. 473–477
23. U. Alon, N. Barkai, D.A. Notterman, K. Gish, S. Ybarra, D. Mack, A.J. Levine, Broad
patterns of gene expression revealed by clustering of tumor and normal colon tissues probed
by oligonucleotide arrays. PNAS. 96(12), 6745–6750 (June 8, 1999)
24. Leukemia data. https://web.stanford.edu/~hastie/CASI_files/DATA/leukemia.html. Accessed
on 07 Nov 2017
Blockchain Technology for Data Sharing
in Decentralized Storage System
D. Praveena Anjelin and S. Ganesh Kumar
Abstract In Cloud architecture, the storage systems can be centralized and decen-
tralized environments. Centralized storage systems stored data in private cloud, main-
tained in a single location, can be accessed by one or more user, whereas in decen-
tralized storage system, the data is stored on more than one or multiple servers. Users
or companies are being a part of the decentralized cloud storage cloud to host the
servers. The data may be stored on any server, i.e., no dedicated server for data storage
and which can be accessed easily. In decentralized storage system, the files are stored
and protected with the help of blockchain technology. Blockchain is nothing but a
chain of blocks (computers or servers), which are connected by using cryptography.
Each block contains a set of transactions which has been encrypted and shared in a
secured way among multiple servers. Blockchain technology ensures confidentiality
and integrity, which can be implemented by using peer-to-peer network. In this paper,
we discussed how the technology improves security used by different applications
in distributed environment. Further, we proposed adaptive encryption algorithm to
improve security and access control.
Keywords Adaptive encryption ·Cryptography ·Decentralized storage ·
Blockchain
1 Introduction
In the past few years, most of the organizations have shown concern on outsourcing,
it outsources the data and as well as gives functional services to clouds. Actually, the
classical cloud storage is a centralized storage system acted as a trusted third-party
storage provider to store and transfer the data. This system consumes high operational
D. Praveena Anjelin (B)·S. Ganesh Kumar
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
SRM University, Kattankulathur, Chennai 603203, India
e-mail: pd6231@srmist.edu.in
S. Ganesh Kumar
e-mail: ganeshk1@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_32
369
370 D. Praveena Anjelin and S. Ganesh Kumar
cost, less security, poor performance, and lack of availability to transfer. To overcome
the drawbacks of centralized cloud storage, a decentralized storage system (DSS)
has been introduced with many features. Blockchain-based cloud storage solutions
allow users to store data safely and give access to all participants in digital activities
with most effect. Filecoin, storj, ppio, ochain, opacity, bittorrent, neo, and Dfinity
are the various platforms of decentralized storage solutions.
Instead of central server, [1] InterPlanetary File System (IPFS) is a distributed file
system which stores data on large number of computers that are connected by peer-to-
peer networks. IPFS can communicate through Transmission Control Protocol (TCP)
which has certain rules for data storage and transmission among connected nodes in
the network. Node addressing system identifies nodes and content addressing system
constructs the index of the content. Both node address and content address are stored
in InterPlanetary Naming System. Files stored on IPFS can be accessed and shared
by the users connected to an ipfs node. Files are split into small blocks, those blocks
are hashed and distributed to storage nodes in the network. In order to consider
privacy and data availability, DSS is better than centralized cloud storage systems, in
which operational cost is much lesser than centralized storage. Existing decentralized
storage systems supports easy implementation of encryption and storage of data,
but the difficulty of data sharing is done secretly. Hence, we proposed a distributed
framework to control access and search data using Ethereum Block Chain blockchain
technology. Further to ensure security, we used ELGAMAL for encryption which
protects the data in cloud storage.
The rest of the paper is sectioned as follows: Sect. 2consists of literature review
of existing technologies that have been used for data sharing in DSS. Section 3
presents reviews storage solutions of blockchain technology. Section 4, elaborate the
user access control policy in Blockchain technology. Section 5presents comparative
analysis of various security algorithms in blockchain with DSS. Section 6presents
review of application areas of blockchain. Section 7concludes the chapter.
2 Literature Review
Decentralized Cloud Storage In Decentralized Cloud storage system, the data is
stored on multiple computers or servers connected by P2P network. These servers
are hosted by a general user or an organization which can contribute to this decen-
tralized cloud. The files are encrypted by cryptography and protected with the help
of blockchain technology. The participants in this system can also earn money in
cryptocurrency by sharing their unused space. Each computer in this network stores
encrypted form of user data, the user is the only authenticated person to access
and manage the files through their own public or private key. There are so many
advantages in this technology,
•No dedicated servers for data storage.
•Peer-to-peer system network.
Blockchain Technology for Data … 371
•Always availability of data.
•User can earn extra money by allowing free space to share.
•Since files are encrypted, it achieves higher rate of data security.
•Operational cost is lesser.
•Load balancing is very flexible.
•It requires less computing power and bandwidth.
Existing Technologies in DSS Distributed storage systems have huge volume of
structured data spread over many nodes in the network, which provide highly obtain-
able service with no distinct end of failure. Some organizations provide distributed
storage systems such as Sia, Storj, Maidsafe, and Ethereum, which are based on
blockchain technology and a peer-to-peer architecture.
SIA: Sia is a decentralized peer-to-peer cloud storage scheme with blockchain
technology. When the user chooses to store data on Sia blockchain. Sia first splits
whole data into blocks, encrypts and distributes the data into various nodes connected
in the network. The user can retrieve data by making request with a private key.
User can get paid in Siacoin by renting out the extra space on their PC to the Sia
decentralized network. Similarly, users who wish to use storage space need to pay
Siacoins to a host. Sia does this by creating a marketplace for hosts and users via
the token economics of Siacoin. Siacoin currency is used to execute a file storage
contracts on the Sia blockchain and aim to reduce the cost of cloud storage. A host in
this network puts a file contract agreement for keeping files, space required, and to
invest Siacoins as collateral. A host can be punished if he does not serve to the user
who made the data request. Once the agreement is over, the sum will be credited to
the host in Siacoin account by the file contractor.
Storj: This storage platform is built on the Ethereum network best suite for cryp-
tocurrency, store data in decentralized in asecure manner. [2] The user can upload
the file with their own private key in order to ensure authenticate person and file is
encrypted before sharing in the network. It separated the files from the user and stored
in decentralized storage model. When user wants to access the file, he has to make
a request, then storj uses distributed hash tables to locate all the shards and piece
them together. Storj rents its own network to many users and puts charges for those
who had used the network. This technology creates revolution in file sharing due to
availability of data among connected nodes. Storj has partnerships with Microsoft
Azure and Heroku to deploy its own development tools which creates great initiative
for the open-source developer ecosystem. User authentication with direct payment
is a major success in storj.
Swarm: Swarm is another decentralized storage system built on Ethereum
network. It is one of the best content distributor and provides redundant store for
web application. It provides various base layer services for web3 such as media
streaming, none-to-node messaging, database services, etc. Actually the content is
hosted on peer-to-peer network instead of individual computers or servers. This
network is responsible for user authentication and allows them to access resources
with payment. It directly integrated with Ethereum blockchain for service payments
and data availability.
372 D. Praveena Anjelin and S. Ganesh Kumar
Cassandra: Cassandra is a distributed storage system used for storing and running
a huge volume of structured data across various commodity servers. Cassandra
supports dynamic control over data layout for clients with a simple data model
[3]. The major distributed system techniques used in Cassandra are partitioning,
scaling, replication, creating membership, and failure handling. To handle, either
read or write requests, all these modules work under synchrony mode. Partitions the
data dynamically, encodes it by hashing method, and assigns it to all nodes in the
network. This distributed system uses replication algorithm to improve data avail-
ability and durability. Among multiple nodes, one can act as coordinator node which
is responsible for keeping redundant data files whenever failure takes place in the
network. It provides various replication polices such as Rack Aware, Rack Unaware,
and Datacenter Aware to their clients stating that how and when data needs to be
replicated. In Bootstrapping module, whenever a node wants to join in the group,
first that node has to read its configuration file and membership identity card will be
issued. The configuration file contains a list of linking points in it, with the help of
this a node can join in the group and share files. It achieves high throughput, high
performance, and providing better scalability.
Data Storage systems
Meta Product manufacturers always use their own private cloud to store data due
to lack of security and privacy issues. Manufacturers can’t trust another network
and avoid data sharing [4]. Shrestha, A. K. and Vassileva, J proposed decentralized
storage platform for Meta products, such as smartphones, wearable sensor devices,
and smart cars. Meta Products must store group user data, later these data can be
reused. It is a trusted system for sharing user information across various domains
deployed by different organizations. It maintains distributed file account which holds
the following information, who can contribute with sharing systems, what file may
access, when that file retrieved, etc. This storage system has inbuilt trust mechanism,
so it easily identifies the threat whenever a user tries to access an unwanted file. It
supports personalized and content-oriented services to their clients.
Replication is a key technology of DSS [5] Yijie wang and Sijun Li proposed
indirect replication algorithm. In the proposed model, the data is split into different
data blocks, these blocks are encoded and distributed among several storage systems.
Since data is redundant among data blocks, bipartite graph is used to encode the data
blocks. The major advantage of using this algorithm is it provides security, durability,
and availability of data. This model can be implemented with less operational cost
and storage cost.
3 Blockchain Technology
Cloud storage is a centralized database system, in the extension of a centralized
one, a decentralized storage framework is designed with blockchain technology. As
we know that blockchain is a distributed database, a user can store any kind of
Blockchain Technology for Data … 373
information across different blocks (personal computer or server) connected in the
network. Blocks are connected in chronological chain model, hence it is named as
blockchain. It can be divided into three categories.
•Cryptocurrency bitcoin
•Smart contracts
•Application.
Cryptocurrency bitcoin: It is nothing but digital cash and as well as currency paid
through online modes. It was established in the year 2008 with a group of people under
the pseudonym Satoshi nakamoto, effectively solved the Byzantine (cryptographic
research) problem. It is the most useful tool for value transfer underlying blockchain
technology. The blockchain technology replicated all transactions done in the Bitcoin
network. Bitcoin uses proof-of-work mechanism to prevent double-spending in the
network; i.e., one who may spend same funds twice. The proof-of-work solved by
miners in the bitcoin. Miners are the bitcoin nodes which checks its blockchain
history and verifies all transaction made in the network. If anyone in the network
tries to change the history, it takes more computational power in the network in
order to verify. Actually proof-of-work mechanism is very expensive but, it is the
only prevention method against Sybil attack. Sybil attack means, a node in network
claims multiple fake identities and gain resources without spending currency. This
attack highly influenced on following applications, voting systems, location-aware
routing, data aggregation, and reputation evaluation.
Smart contracts: Smart contract is nothing but crypto contract, maintained by
peer-to-peer network. In 1994, smart contracts idea was first proposed by Nick Szabo
in 1994. He was interested in starting digital currency. Then later, in 2008, the
cryptocurrency bitcoin was developed via a blockchain platform, this technology
enabled smart contract. This tool can provide some coordination and assigned agree-
ments among nodes who participate in the network. This tool can directly control
digital payment, currency transfer, or sharing assets between nodes in the network. It
enforces the contractor to follow all rules and obligations related to the agreement as
the same way what traditional contract did by generating tokens. Blockchain managed
the self-enforcing agreement which is embedded in some other computer code. The
code contains smart contract agreement policy which parties need to interact; what
file needs to be shared and cost details. The agreement is enforced automatically
when the rules are met, i.e., it unlocks the access and manages tokenized assets.
In technical aspects, the merit of the smart contract is self-verifying, self-executing,
and tamper resistant. In economic aspects, decentralized ledger initiated transactions.
So, low expenses and maintains higher transparency. In legal aspects, security is the
major issue that needs to be resolved.
Blockchains: In general, blockchain requires a huge computational resource to
run, since it has been available in public and accessed by many nodes in the network.
So, the system is almost utilized its maximum potential but the user should have
limitations to access. Sometimes an unauthorized user may access the network, he
can download the file, i.e., confidentiality of data, secure sharing, and redundancy are
major issues in blockchain. In distributed systems, private data must be shared by all
374 D. Praveena Anjelin and S. Ganesh Kumar
nodes connected on the network. So, data may be completely exposed. In centralized
storage systems, data can be stored securely with less transparency. This is one of the
primary contradiction between centralized storage system and decentralized public
blockchain. Based on access mechanism, blockchain can be classified as,
Public blockchain: A blockchain grants permissions to issue transactions and to
read access for all data of the chain to all users.
Private blockchain: A blockchain is available for a predefined list of entities
to access stored data and create a transaction on the chain. There are two types of
blockchains.
Permissionless blockchain: This type of blockchain does not restrict the identities
of transactions.
Permissioned blockchain: This type of blockchain accepts only the predefined
entities performed transactions with known identifications [6,7].
Both private and permissioned blockchains are not the same based on storage
mechanism. For example, a permissioned blockchain can be available to the public
to read the data from blockchain, but the general users might not have permission
to create transactions. It has distributed ledgers to maintain storage details, access
details and protocols can be used to provide security and increase the availability of
data. But, in private blockchains may have full permission to access the blockchain
by preconfigured entities. Moreover blockchain may be any type, must have multi-
level permissions such as connection establishment, create storage block, send and
receive transactions, read data from decentralized storage systems. All these tech-
nologies support security, data consistency, and easy access. The following table
shows comparison among various blockchains (Table 1).
Blockchain-based storage solutions: Blockchain technology provides decen-
tralized storage structure with a secure manner when compared to traditional cloud
storage. There has been significant growth in technology with respect to storage,
security, authentication, and access control. Whenever a node in network wants to
upload a file, first that must be encrypted before sharing and also ensure availability
Tabl e 1 Comparisons among public blockchain, consortium blockchain and private blockchain
[6]
S. No Property Public blockchain Consortium
blockchain
Private blockchain
1Consensus
determination
All miners Selected set of nodes One organization
2Read
permission
Public Could be public or
restricted
Could be public or
restricted
3Immutability Nearly impossible to
tamper
Could be tampered Could be tampered
4 Efficiency Low High High
5Centralized No Partial Ye s
6Consensus
process
Permissionless Permissioned Permissioned
Blockchain Technology for Data … 375
of data [8]. Metadisk is an open-source software to provide decentralized storage
system which is more secure and efficient. Its main objective to establish a constant
platform for P2P cloud storage network (Storj). This application provides an inter-
face for nontechnical users also. Using this API, users can upload and download
files in a secure way. Client-side encryption has taken place on it. Users can generate
private key and assign to files while uploading files on the network. Multiple nodes
in a network may have the same file according to data availability.
Once the file has been encrypted, the SHA-256 hash serves a unique identity
and way to detect file tampering. [9] In blockchain distributed storage, first data
storage provides decentralized platform where, ana organization can register in order
to access distributed data and further it can be shared. Anonymous access control
component ensures the client access is provided by data owner itself. The data owner
may not aware of who will be the data user and how many times the data consumer
has to show the credential proof. Private keyword search mechanism helps a data
user to recognize the encrypted information. The consumer can directly interact
with blockchain node by giving a single keyword to get the fingerprint of specific
data content. At last, data consumer retrieves encrypted documents from decentral-
ized storage system. Keyword search cryptosystem consists of KeyGen, Trapdoor,
Encrypt, and test algorithms. KeyGen generates private search key, encrypt algorithm
produces ciphertext for keyword which are performed by the data owner. KeyDerieve
algorithm is used by the owner to produce secret search key for data consumer. Trap-
door is run by data user for keyword. Test algorithm is executed by smart contracts
on blockchain.
4 Access Control Mechanism in BlockChain Technology
Samarati and de Vimercati [10] defined Access Control as “The process of mediating
every request to resources and data maintained by a system and determining whether
the request should be granted or denied.” The primary goal of access control mech-
anism is to verify and limit the actions that authorized users can perform within a
computer system. Access control has set certain limitations to users, as what a user
could perform directly with the system, what programs can be executed on behalf
of the users who are permitted to execute those programs with approval. An access
control community should have an object, subject, operations, permissions, control
list, and matrix.
Access control Models: The access control system has been developed based on
security policy and security model. Security mechanism is a model which comprises
of low-level implementation of both hardware and software. Access control models
can be categorized as,
1. Mandatory Access Control
2. Discretionary Access Control
3. Role-based Access Control
376 D. Praveena Anjelin and S. Ganesh Kumar
4. Organization-based Access Control
5. Attribute-based Access Control
6. Centralized and Decentralized Access Control.
Laurent et al. [11] They proposed a new access control method. In that, a smart
auditable contracts deployed in blockchain infrastructures which is a transparent one
and provides controlled access to outsourced data, i.e., Malicious entities should
not access data without the knowledge of the data owner. They would implement
this proposed new solution on Ethereum blockchain network. The proposed access
control mechanism has four entities which are defined as Data Storage Provide (DSP):
It is not an active client having read access alone and governs distributed remote
servers and hosts application services. Data Owner (DO): Ensures authorized entities
and makes uses of resources provided by data storage provider in order to store and
share data in network. Data retriever (DR): DR is a client of blockchain, he can access
the content stored in remote servers. Blockchain Infrastructure (BC): It is a mediator,
permits to authenticate DRs. It ensures authenticated access control with respect
to efficient whitelist definitions and preserves privacy among entities. Unlinking
properties can support one-time access control which ensures unlinkability between
different access sessions by users.
Ronghua and Chen[12] A blockchain-enabled decentralized access control
(BlendCAC), is suitable for decentralized storage in a secured manner. Its major
aim is to provide effective access control mechanism to devices, services, and infor-
mation especially in large scale IoTs (Internet of Things). It efficiently does process
related to IoT and also provides granularity, scalability, and dynamicity of AC strate-
gies for IoTs. The proposed system, IoT devices acts as a master, controls all the
resources used in the network instead of centralized authority. The operations are
carried out in Ac as listed below. Registration: Entities that participate in blockchain
network must create an account with a pair of keys for authentication purpose. Smart
Contract Deployment: Domain owner developed and deployed capability tokens on
blockchain network, those tokens are managed by smart contract. Capability Prop-
agation: Initially, an entity sends access rights request to the domain owner to get
the capability token. Then the system checks for authorization, if it is found to be
right entity, then capability token is issued by data owner. Capability token encodes
the access rights, then initiates transactions in order to update token data in the
smart contract. Authorization Validation: Service providers performed authorization
validation on receiving service request from an entity.
Steichen et al. [13] IPFS uses Ethereum smart contract to provide access-
controlled file sharing. The proposed mechanism allows users to do all financial trans-
actions in its own cryptocurrency area and custom currencies called tokens. Access
control system has some privileges regarding addition, modification, and removal of
permissions that are recorded on blockchain. The access control contract contains
four functions such as, AMIOwner, MIOwnerMultiple, CheckAccess, and Check-
AccessMultiple. AdBlock function used to register an IPFS file chunk in control list.
Cryptographic hashes identify IPFS file chunks. This hash value is used as a key
stored along with file chunks. GrantAccess function accepts request given by data
Blockchain Technology for Data … 377
owner, check for hash value. If value exists, then it sends the key to owner. Remove
access function used to remove the user whenever hash value mismatched or different
being posted by data owner. AMIOwner: It accepts the request and checks the hash
value of file chunks. If it has the same value what a data owner already registered, then
it gives access rights to data owner to get access of file chunks of storage blockchain.
Its values are mismatched, then it returns false message. DeleteBlock function reverts
if the hash value is empty, soon it removes files from data storage since the file has
not been requested so far from the owner’s side. fhash value resides, it knows that
file may be accessed and retrieved soon.
Thwin and Vasupongayya [14] Personal health record system (PHR system) stores
individual health-related information. This system allows data owners to share and
manage data among selected individuals. Information stored in database may be
incorrect because of immutability, irreversibility properties, and originality of content
may be inequitable. Fortunately, blockchain technology gives potential solutions to
solve this issue. The proposed system supports tampering resistance feature which
includes fine-grained access control, auditability, tamper resistance, and revocability
of consent. To preserve privacy, proxy reencryption and cryptographic techniques are
used. Proxy reencryption scheme is an asymmetric cryptosystem; enables its users
to share their decryprion capabilities with others. The ciphertext is encrypted with
user public key, that file can be decrypted by another user by using their own private
key but the content cannot be retrieved fully during transactions. Since the encrypted
file can be constructed in such a way, ensure secure data sharing scheme. The data
owner must send reencryption key to proxy in order to share data. The reencryption
key is generated with the combination of owners secret key and users public key.
Moreover the proxy cannot read any information from original data with the help of
reencryption key.
5 Privacy/Security in Blockchain
(i) Existing Algorithm:
Decentralized storage solutions create massive attraction among storage providers.
Data can be stored in different nodes in the network and being shared with entire
network. However, uploading, downloading, and sharing of files may not be secure
in a distributed platform. Since, authentication and security can be a big challenge
in distributed architecture. Existing decentralized architectures are constructed to
support huge volume of data storage and shard, but, still there are no effective
solutions for security issues. P2P network has been widely implemented over the
distributed systems. We compared different blockchain technologies so far that had
been implemented, security algorithms used, and their performance with evalua-
tion metrics. The comparison of different blockchain platforms along with network
performance has been shown in Table 2.
378 D. Praveena Anjelin and S. Ganesh Kumar
Tabl e 2 Comparisons of blockchain technologies, security algorithm used
S. No Platforms/applications Business logic Configuration
mode
Encryption Wor k l oad Transaction data evaluation metrics
1Ethereum Smart contract Private network Single public
key
Random Deployment time,
completion time
Execution time,
throughput
2Hyperledger Fabric Chaincode Private network Different
private key
Random Completion time Latency and
throughput
3Smart home systems Bitcoin Private
blockchain
Signcryption Asynchronous Completion time Execution time
4 Peer-to-peer Medibchain
protocol (smart
contract)
Private
blockchain
Elliptic curve
cryptography
5P2P network Bitcoin Private Digital
signature
algorithm
Random Transaction time Latency
Blockchain Technology for Data … 379
(ii) Adaptive Encryption Algorithm:
Adaptive encryption techniques can be applied on decentralized storage systems
because of data confidentiality. This algorithm permits the cloud server to do a large
set of SQL operations over encrypted data. The below listed encryption schemes are
used in it.
•Deterministic
•Order Preserving encryption
•Random
•Search
•Plain
•Sum.
Each client can participate in DSS to get services by direct execution of SQL oper-
ations. This algorithm guarantees the same level of scalability and data availability.
Initially, data is encrypted using key, then it is stored in cloud server. Whenever the
client wants to get data, he has to decrypt with key value after downloading a file
from the server. This algorithm is well-suited for encryption since it does not allow
even a trusted party to manage encryption details. It also simplifies database server
configuration because of operation automation.
6 Application Areas
Blockchain technology is very proficient and gives copious reimbursement such as
decentralization, persistency, auditability, and anonymity. Cryptocurrency, financial
services, Internet of Things, risk management are various sectors where blockchain
technology can be applied.
Finance:
(i) Financial Services: Bitcoin is one of mostly widely used cryptocurrency
method. Blockchain technology can be used for financial transactions, clearing
and settlement of financial assets, and reducing cost and risk. Microsoft Azure,
IBM are tremendous software companies offering blockchain services to the
general public.
(ii) P2P financial market: Blockchain builds a P2P financial network to establish
financial services among network in a secure and reliable way. Blockchain
creates shared computation protocols to create a P2P financial shared compu-
tational market. This MPC market allows a P2P network to do computational
task in offload mode.
(iii) Enterprise transformation: The traditional organization has completed all its
transactions smoothly with help of blockchain technology. For example, the
traditional postal operators (Pos) acts as an intermediator between customers
380 D. Praveena Anjelin and S. Ganesh Kumar
and merchants. Pos make use of cryptocurrency techniques to extend both
financial and nonfinancial services.
(iv) Risk management: Blockchain technology provides risk-management frame-
work used to decide the investments plan, collaterals, and analyze investment
risk also. Smart contract enables decentralized autonomous organizations to
manage with all its business activity collaborations.
Internet of things (IoT): IoT is nothing but integration of objects (electronic
devices and smart objects) with internet to provide different services to users. Some
of the applications of IoTs as logistic management with Radio-Frequency Identifica-
tion (RFID), smart homes, smart grids, e-health,maritime industry, etc. Blockchain
with IoT potentially improves e-business model. This model has DAC, the people
who trade with DAC to get transactions quickly, get coins, and the people can also
exchange data with intermediary. Blockchain can improve privacy in IoT applica-
tions. Hardjono and Smith (2016) proposed a privacy-preserving method for IoT
device into cloud ecosystem. This system ensures security by not allowing unau-
thorized access of devices. IBM ADEPT system developed to build a distributed
platform, where devices are connected.
Public and Social Services: Blockchain can be widely used for public services
such as patent management, income tax generation and monitoring system, digital
signatures, marriage registrations, etc. Decentralized storage access provided to all
users connected in network which reduces hard copy system and easy access to user.
(i) Education: Blockchain learning is proposed by Devine (2015) which achieves
online educational marketing strategy. In this learning model, the teachers can
segregate the learning materials into blocks; those blocks are packed and placed
into the blockchain. Any user can access blocks on blockchain by paying coins
through cryptocurrecy.
(ii) Energy saving: Gogerty and Zitoli proposed solarcoin. Solarcoin encour-
ages the usage of renewable energies by rewarding digital currency to energy
producers.
(iii) Land Registration: Blockchain technology highly contributes in public
services and as a proof land registration is major application of it. The land
information such as physical status, registered details, authorized person, land-
mark, cost details are publicized on blockchains. Whenever there is any changes
in details, that are updated lively on blockchains to improve effectiveness of
communal services consequently.
Reputation System: Reputation is nothing but trustworthy. Simply saying, an
Individual’s reputation can be evaluated based on the previous transaction done by
him. Individuals overall interactions can be taken place within community. In e-
commerce applications, to achieve high reputation, many service providers enrolled
large number of fake customers. This creates problem for vendors also. This issue
can be solved with the support of blockchain. Reputation is an important thing for
an academician [15]. Blockchain-based distributed system is proposed to manage
educational record and reputation. In this model, initially each institution and their
Blockchain Technology for Data … 381
professionally qualified workers will be awarded (educational reputation currency).
An institution also awards the intellectual worker by transferring record status to
them. All transactions are stored dynamically. So, changes in reputations can be
detected easily. In web community, members can be reputed based on their qualities.
Carboni proposed a reputation model based on blockchain, the service providers in
web community may be reputed by getting feedback from customers. Reputation
values are stored on blocks across a distributed network.
7 Conclusion
Decentralized storage system is created a revolution in storage area where the public
can interact with internet. It provides distributed data storage, though there may issues
on security, no safe on stored data, trustless environment. In this cyber world, no secu-
rity, difficult to predict third party access, no privacy of sharing data. Blockchain is
emerging technology which gives promising potential solutions for these issues. Data
stored on this blockchain is more secure and ensures confidentiality by using cryp-
tographic techniques. It provides high-quality services to all connected in network
and set check constraints for third party access. The aim of this paper is to show a
comprehensive view of blockchain technology and its merits. This paper includes
study of existing technologies in DSS, storage solutions of DSS with blockchain
technology, various access control mechanisms implemented for blockchain. In this
paper, we have also discussed about security algorithms used in blockchain, and
comparisons among various public blockchain used, consortium blockchain used and
private blockchain used in network. Further we discussed about various application
areas, where blockchain technology could be implemented.
References
1. J. Benet, IPFS—Content addressed, versioned, p2p file system (draft 3), (2014). https://ipfs.
io/ipfs/QmR7GSQM93Cx5eAg6a6yRzNde1FQv7uL6X1o4k7zrJa3LX/ipfs.draft3.pdf
2. Storj—A Peer-to-Peer Cloud Storage Network. https://storj.io/storj.pdf
3. A. Lakshman, P. Malik, Cassandra—A Decentralized Structured Storage system, from
Facebook Org
4. A.K. Shrestha, J. Vassileva, Towards decentralized data storage in general cloud platform for
meta-products, in Proceedings of the International Conference on Big Data and Advanced
Wireless Technologies (BDAW ‘16) (ACM, New York, NY, 2016)
5. Y. Wang, S. Li, Research and performance evaluation of data replication technology in
distributed storage systems. Comput. Math. Appl.
6. Z. Beaven, D. Neilson, R. Osborne, P. Pacifico, Blockchain for creative industries music on
the blockchain (2016)
7. Public versus Private Blockchains Part 1: Permissioned Blockchains (2015)
8. S. Wilkinson, J. Lowry, Metadisk: Blockchain-Based Decentralized File Storage Application.
http://storj.io. August 20, 2014
382 D. Praveena Anjelin and S. Ganesh Kumar
9. H. Giang, W. Keong, Blockchain-based system for secure data storage with private keyword
search, in on 13th World Congress on Services (IEEE, 2017)
10. P. Samarati, S. De Capitani, D. Vimercati, Access control: policies, models, and mechanisms
(2000)
11. M. Laurent, N. Kaaniche, C. Le, M. Vander Plaetse, An access control scheme based on
blockchain technology
12. E.B. Ronghua, G. Chen, BlendCAC: a Blockchain–Enabled Decentralized Capability-based
Access Control for IoTs (April 2018)
13. M. Steichen et al., Blockchain-based, Decntralized Access Control for IPFS. IEEE conf.
Internet of Things (2018)
14. T.T. Thwin, S.Vasupongayya, Blockchain-based access control model to preserve privacy for
personal health record systems. Secur. Commun. Netw. Article ID 8315614, (2019)
15. Z. Zheng, H.-N Dai, Bloakchain challenges and opportunities: a survey. Article ID 101504,
(Oct2018)
Understanding Concepts of Blockchain
Technology for Building the DApps
P. Shamili, B. Muruganantham, and B. Sriman
Abstract Blockchains are the digital ledgers with tamper resistance and are imple-
mented in the distributed networks with the consensus protocols which were intro-
duced by Satoshi Nakamoto. A high-level secured technology for users both in the
public and private networks. The peers in the network who share their transactions
with the other peers are recorded and once recorded transactions cannot be modified
by any of the peers in the blockchain technology. Here the digital signatures are used
to provide the trusted third party to prevent the double-spending attacks. Due to the
longest chain generated by the user by using the high computational power, there
may be possibility of an attacker to access the whole network and this is prevented
by the proof of work consensus protocol as forking concept. This paper provides a
review of the blockchain network and helps the readers to understand the working
of blockchain technology.
Keywords Blockchain ·Consensus protocols ·Smart contracts ·Distributed
network ·Ethereum ·Hyperledger composer
1 Introduction
The signed transactions that are arranged into the blocks and distributed and stored
in the digital ledgers are named as the blockchains. After the consensus decision
and the validation of the blocks linked to one another, they are linked through the
P. Shamili (B)·B. Muruganantham ·B. Sriman
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
Kattankulathur, Chennai, Tamil Nadu 603203, India
e-mail: shamilip@srmist.edu.in
B. Muruganantham
e-mail: muruganb@srmist.edu.in
B. Sriman
e-mail: srimanb@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_33
383
384 P. Shamili et al.
cryptographic value of the previous block [1]. The blocks created before will become
more difficult to tamper, if the new blocks are added to it. The newly added blocks
data will be copied to all the nodes in the blockchain network according to the
predefined rules.
In the centralized system, the data can be altered or modified by the central
authority. But in the blockchain network, none of the blocks could be modified
(tamper resistant) once it is created, because of the distributed fashion [2]. In 2008,
the combination of few of the technologies and the computation concepts came out
to create the cryptocurrencies as Electronic cash and are protected through the puzzle
solving mechanism.
The first modern cryptocurrency in the blockchain network was widely spread in
the year 2009 namely the Bitcoin. In the distributed system, the transfer of informa-
tion in a digital form represented as Electronic cash takes place in the bitcoin systems.
The nodes in the bitcoin network can transfer and digitally sign their rights in that
data to all the nodes and the blockchain record the transaction publicly [3]. So that all
the nodes in the network are allowed to verify the validity of the transactions. These
details were maintained independently in the distributed group of nodes. Along with
this mechanism, the cryptographic values are made tamper proof.
The blockchain technology enabled many of the cryptocurrencies like and the
Ethereum and Bitcoin. Due to this, the blockchain is seen as bound to the cryptocur-
rency solution. It is being investigated for many of the sectors and also available
for many of the applications. It is a challenge for the investigators to understand the
distributed systems and the cryptographic primitives.
2 Overview of Blockchain
The tamper-proof digital ledgers were implemented in the distributed fashion without
the centralized authority which are the blockchains. Initially, it is to enable the
nodes to record the transactions in the network. In this way no such transactions
in the network could be modified. This idea was implemented in the concept of
cryptocurrencies of electronic cash such as bitcoin.
Bitcoin was the first blockchain-based cryptocurrency. The information stored in
the bitcoin blockchain represents the electronic cash and are attached to the digital
address [4]. The nodes in the bitcoin blockchain use the digital signatures and transfer
the data to the other nodes in the blockchain network. The data recorded are available
transparently for all the nodes in the network to validate the transactions. Along with
the cryptographic mechanism, the blockchain is made tamper proof for the nodes
which try to modify or forge the data.
The blockchain creates the trust in the environment where all the nodes could not
easily be identified. This trust is typically delivered to all the nodes in the blockchain
network. The characteristics that are needed for the trust for blockchain technology
is described in Fig. 1as follows,
Understanding Concepts of Blockchain Technology … 385
Fig. 1 Block chain architecture
Ledger—like the traditional ledgers for storing the transactions and the data, the
blockchain uses the digital ledgers to store the transactions in the network that
could not be overwritten.
Distributed—the blockchain is distributed. This allows the nodes in the
blockchain to join the blockchain network and prevents from the attackers [5].
The possibility of involving the attacker may happen if the number of nodes in
the blockchain increases.
Secure—the blockchain is secured by a cryptographical puzzle, so that the
transactions are tamper proof.
Shared—the digital ledgers are shared among all the nodes in the blockchain
network. It provides the nodes to transparently available for all the nodes.
3 Hash Functions
Cryptographic hash is the most important component in the blockchain technology
[6]. The method for applying the cryptographic hash function which gives the relative
outputs for the inputs given. The nodes are allowed to check the data, hash of that
data, and finding the results of that data for proving that the data is immutable. If
there is any change in the input, there will be different output digest. This is shown
in Table 1.
Tabl e 1 Hash function
SHA-1 SHA-224 SHA-256 SHA-384 SHA-512
Message digest size 160 224 256 384 512
Message size <264 <264 <264 <2128 <2128
Block size 512 512 512 1024 1024
Wor d s i z e 32 32 32 64 64
Number of step/round 80 64 64 80 80
386 P. Shamili et al.
Fig. 2 Block header
The more specifically used algorithm for securing the data in the blockchain
technology is the Secured Hash Function (SHA) which is of 256-bits [7]. Most of
the computers nowadays use this algorithm in hardware for more high computation.
Few of the important security properties of the cryptographic hash functions are:
•Preimage resistant: it is also called the one-way function, where the input hashes
cannot be found for the inputs.
•Collision resistant: it means that one could not find the two of the input hash to
the same outputs.
A. NONCE
Nonce is cryptographically generated random number by using the algorithms. The
different hash values were hash digest [8] which were produced by combining the
data and the nonce as shown in Fig. 2.
DIGEST =HASH(DATA +NONCE)
B. Ledger
The data about the transactions that are stored in a place is named as ledger. In
the traditional ledger, the papers and pens were used for keeping the records of all
types of transactions. By using the computers, the ledgers were updated to digital
ledgers that are operated by centralized organizations. These were under centralized
control. It is now updated to decentralized system, especially in blockchain.
C. Block
Each and every transaction in the blockchain network is stored in the digital ledger
called the blocks in the blockchain network. The block consists of data about all the
Understanding Concepts of Blockchain Technology … 387
transactions that occurred in the distributed network. The copy of a single block data
is shared with every other node in the network by the consensus protocols. The block
contains the nonce value (hash value), the timestamp of every transaction.
D. Block header
The block header in the blockchain holds the following data,
a. Consists of the previous block hash value.
b. The hash value representation shows the root hash for that block.
c. The Timestamp.
d. Block size.
e. Value of nonce (generated by solving the mathematical crypto puzzle).
E. Data in block
Like block header, the block data consist of all the data or the transactions that
occurred in a block in the blockchain network.
F. Merkle tree
The hash tree is formerly known as Merkle tree or Merkle root. For the data synchro-
nization and the verification of the data structure is used. The tree data structure is
holding the values of hash of non-leaf nodes and that represents the values of its child
nodes. The leaf nodes in the Merkle tree maintains the same depth as possible. The
hash functions are maintained for its integrity.
G. Digital signatures
The digital signatures are the keys for the security of data in the blockchain tech-
nology, where it works under the algorithm of SHA-256. In the blockchain network,
the signed and verified transaction is shared among all the nodes in the network [9].
It has two phases, they are Verification Phase and Signing Phase. For example, two
nodes in the blockchain network agree under the consensus protocol and send the
data to each other. That data is signed and encrypted by the user who sends. This is
the Signing phase. Then that data is being Verified by the user who receives it. This
is the Verification phase. So that the data is highly secured and tamper proof.
H. Forking
In the blockchain technology, the changes and updating of data will become difficult
for some time. In the permission-less blockchain networks, due to heavy competi-
tion and increase in the nodes agreed by the consensus protocols, it becomes much
difficult. Few of the changes to the blockchain network and to the data are called
Forking as shown in Fig. 3. It is divided into two types, they are soft forking and hard
forking. It is compatible for the nodes to change backwards, the need not be updated
in the soft forks. But in hard forking, it is not compatible for the nodes to go back and
the blocks will not be updated and also be rejected. In the permissioned blockchain
network, as all the nodes come under the consensus protocol and all nodes were
known to each other, there will not be any chances for forking.
388 P. Shamili et al.
Fig. 3 Forking
4 Proof of Work Consensus Model
The Proof of Work consensus protocol is to solve the crypto puzzle by using all the
computational powers by the nodes. This is a proof that the node has performed the
work by solving the puzzle. It is designed in such a way that the nodes find it difficult
to verify the block. It proposes the nodes to proceed to create a new block and the
proposed block will be rejected if it is not satisfied.
A. Crypto Puzzle
The crypto puzzle is a method of solving a mathematical problem for creating a new
block by verifying the previously created block’s transaction. The puzzle is a hash
value that is generated as the nonce (the block header) by the block. For each and
every attempt tried for creating a block by the node to compute the hash value to
that block header, it takes a minimum of 10 min for creating a block by solving the
crypto puzzle. For increasing the difficulty, the target value may be modified.
B. Computation Power
In the proof of work consensus protocol, the node uses its computational power for
solving the mathematical crypto puzzle to create a new block. As mentioned earlier,
it takes a minimum of 10 min to solve the crypto puzzle by using the computational
power. By using this method, the nodes in the network can influence others by having
high computational powers. This type of attack is called the Sybil attack.
C. Double-Spending Attack
The digital assets are transacted more than once at the same set are known to be
double spending. The group of nodes in the blockchain network do double spending
for verifying a transaction to create a new block is called 51% attack. This attack can
be prevented by using the timestamp for all the nodes in the bitcoin network.
Understanding Concepts of Blockchain Technology … 389
5 Distributed Applications in Blockchain
There are three types of blockchain technology [10]. They are,
1. Public
2. Private
3. Consortium.
A. Public Blockchain
The platform where every node in the blockchain network are able to access the data
and have to submit the proof of work for accessing the same data. Bitcoin is the best
example for the public blockchain, where it is a decentralized network as the number
of nodes uses this technology.
Few of the other example for the public blockchain are:
•Ethereum: helps running the smart contracts and also allows the users to build
the dapps.
•Factom: the record used to store the details for the use of governance and business.
•Blockstream: provides the focus on extending the bitcoins for the sidechain
technology.
B. Private Blockchain
In the private blockchain network, only the authorized nodes are allowed to access the
data. Similar to the centralized servers, only the owner of the block has permission
to grant access to the other nodes in the private blockchain network. Many of the use
cases were built using this network. A shared software is provided by Eris industries
using this blockchain technology. The operations based on the financial institution
including the settlement and the clearance were provided by this private blockchain.
For the financial transactions, the distributed opensource database is provided by
Multichain.
C. Consortium Blockchain
The consortium blockchain is a combination of both public and private blockchain.
But it usually follows the private blockchain consensus protocols. By collaborating
with other networks, the users create different decentralized applications for many
of the use cases. Even though they restrict access to users, it is maintained that the
solutions and the property rights within the consortium is maintained.
D. Smart Contract
The collection of data and the codes that are deployed by using the algorithms (SHA-
256) signed for the transactions in the blockchain networks [11]. The Ethereum uses
the programming language named Solidity. It is more or less similar to JavaScript
and C++. Few other smart contracts are Vyper and Bamboo. The languages like
390 P. Shamili et al.
Serpent and Mutan were used before Solidity language. For example, Ethereum
smart contract and the Hyperledger fabric chain codes. The blockchain technology
is maintained and leveraged to an extent that is done by the Smart contracts [12].
Within the blockchain network [13], the smart contracts are executed and the nodes
that execute the smart contracts should proceed for the same results and results are
recorded in the blockchain.
6 Ethereum
Ethereum is an open-source software developed by Vitalik Buterin. It is executed
with the help of smart contracts for computing the decentralization based blockchain
technology. Vitalik has extended the version of Ethereum in the blockchain tech-
nology and the bitcoin protocols to enable the Ethereum applications most of the
cryptocurrencies [14,15]. Ethereum has the inbuilt cryptocurrency as Ether and the
language used here is Solidity. It runs on the network named as Ethereum Virtual
Network (EVM) and it can deploy and write the Ethereum smart contracts in a short
period of time [16]. The solidity cannot be used directly to connect with Ethereum
Virtual Machine, so the opcodes were used to connect with Ethereum.
A. Opcodes
As mentioned above, Ethereum Virtual Machine uses the set of rules to execute the
specified tasks that are named as Opcodes. While writing the opcodes, 140 unique
opcodes were used. It means that the Ethereum virtual machine can compute any type
of task with enough resources. It is because of 1 byte opcodes. There is a maximum
of 256 (162) opcodes. Few of the opcodes are as follows;
•Opcodes for stack manipulation (push, pop, dup, swap)
•Opcodes for comparison/arithmetic/bitwise (add, sub, gt, lt, and or)
•Opcodes for environment (caller, call value, number)
•Opcodes for memory manipulation (mload, mstore, mstore8, msize)
•Opcodes for storage manipulation (sload, sstore)
•Opcodes for program counters (jump, jumpi, pc, jumpdset)
•Opcodes for halting (stop, return revert, invalid, selfdestruct).
B. Deploying the Smart Contract
A transaction is created while deploying a smart contract without the to address. Few
of the bytecodes were into the input data named as constructor, which are needed
to write the starting variables to store before copying the runtime bytecode for the
contract’s code. During the runtime it will run on every call for contract but runs only
once when the bytecode creation is deployed.
Understanding Concepts of Blockchain Technology … 391
Fig. 4 Truffle migration
C. Migrating and Compiling the Smart Contract
Guiding the way for the developers to automate deployment data and their supporting
structures are named as migrations. For managing the new version of software, it is
very much useful and are not exclusive for blockchain development.
a. Truffle Migration
Migrating the truffle enables the smart contracts to push it into Ethereum blockchain
and to set the prerequisites for linking the contracts to the other nodes with the
initial data. Where the migrations are the management for the contract address on
the blockchain network. It all happens in the truffle for entire migration as shown
in Fig. 4. For this truffle migrations, the prerequisites should be verified that truffle
framework and the ganache cli has been installed.
With the migrations, the smart contracts were deployed, for this the artifacts
needed to be accessed. The files to describe the migrations are Contact addresses,
the networks (where the contracts to be deployed), and the functions.
The description for the smart contracts in the .json files
•Name of the contract
•Application Binary Interface of the contract
•Bytecode of the contract
•Deployed bytecode of the contract
•The last compiled version of the contract data
•Each of the contract addresses and the networks on which the contracts deployed.
With these files, the communication with the smart contract is enabled by the
JavaScript to the Truffle. For an illustration, when a contract. Address in the JavaScript
code called the .json file is read by the Truffle framework and the networks and the
contracts versions were enabled.
392 P. Shamili et al.
7 Hyperledger
The Linux foundations have developed an open-source platform named Hyperledger.
The Hyperledger comes under the permissioned blockchain technology [17]. It is also
a customizable application for few of the organizations. According to the needs and
the necessity of the organizations, it is customizable as plug and play. Scalability,
resilience, and confidentiality were delivered by Hyperledger. It uses the container
technology in the blockchain to host the smart contract. In comparison with the
Ethereum, it is more secure by authorized nodes [18].
A. Components in Hyperledger Fabric
1. Ledger: the fabric ledger consists of the state about the blockchain and the world
state. In blockchain, the history of all the state changes were stored. Where the
world state holds the status of all assets which are tracked. All the nodes in the
channel or network has the ledgers and are stored on the blockchain network.
2. Peers: peers are participants in the organizations who make the physical structure
of the blockchain networks. It is maintained by the particular organization that
are known well to the blockchain network. The peers in the network can endorse
the ledger and maintains the multiple ledgers in the network.
3. Channels: the peers in the blockchain network want to get connected or to
communicate with the other peers in the network privately. Where the ledgers
are channel specific, the peers in the network willonly be able to access and see
the transactions of their ledgers.
B. Hyperledger Composer
Similar to Hyperledger, the Hyperledger composer is an open-source software devel-
opment tool and the framework for building the blockchain technology. The composer
provides a comfortable layer and the business abstraction for implementing the smart
contracts by using the fabric instead of using the smart contracts. It is easier for the
users to connect with the business networks through mobile applications or web
services. This network consists of all the assets, transactions, and the participants. It
provides the access granting control to all the peers in the network.
C. Hyperledger Composer Deployment
The development of a business network is provided with multiple tools by the Hyper-
ledger composer. It permits to develop the tools both local and online. The deploy-
ment, configuration, and the manual testing of the network were provided for the user
interface by the composer playground. It is not intended for the development of the
automated testing for the use of source code control of the versions. The following
diagram explains the business network in Hyperledger composer.
Hyperledger composer is the fastest way for creating new business networks [19].
After creating all the above files like model file, script file, access control file, and
query file the structure will be formed and generated to compose the blockchain
network as shown in Fig. 5.
Understanding Concepts of Blockchain Technology … 393
Fig. 5 Hyperledger composer
8 Hyperledger and Ethereum Comparison
Characteristics Ethereum Hyperledger
Platform description Platform with generic blockchain Platform with modular
blockchain
Operations types Ethereum for developers By the Linux foundation
Consensus Ledger levels by the proof of work
(PoW)
Approaches with multiple level
transactions
Smart Contracts Smart contract (solidity) Chain code (Go, java)
Currency Ether No currency needed
The comparison of the Ethereum and Hyperledger [20] are as follows.
9 Conclusion
In this paper, the high-level overview of the blockchain technology was provided. The
many ways of implementing the blockchain have been discussed and the components
and few diagrams were provided along withthe mechanism how consensus protocols
were working. The Smart contracts for deploying the network in the blockchain
and the forking techniques were also explained. Some of the comparisons between
the Ethereum and the Hyperledger was also shown. This paper also presents some
of the areas where the blockchain technology can be implemented. To conclude,
blockchain is a high-level secured technology for sharing the transaction both in
public and private networks.
394 P. Shamili et al.
References
1. S. Nakamoto, Bitcoin: a peer-to-peer electronic cash system (2008)
2. N. Harish, S. Krause, H. Gradstein, Distributed ledger technology (DLT) and blockchain (2017)
3. Hash per Second. Bitcoin Wiki, http://en.bitcoin.it/wiki/Hash_per_second
4. W. Akins, J.L. Chapman, J.M. Gordon, A whole new world: Income tax considerations of the
bitcoin economy (2013). Online Available: https://ssrn.com/abstract=2394738
5. M. Sharplesand, J. Domingue, The blockchain and kudos: a distributed system for educational
record, reputation and reward, in Proceedings of 11th European Conference on Technology
Enhanced Learning (EC-TEL 2015) (Lyon, France, 2015), pp. 490–496
6. National Institute of Standards and Technology, Secure Hash Standard (SHS), (Federal Infor-
mation Processing Standards (FIPS) Publication 180–4, August 2015). https://doi.org/10.6028/
NIST.FIPS.180-4
7. National Institute of Standards and Technology (NIST), Secure Hashing website. https://csrc.
nist.gov/projects/hash-functions
8. National Institute of Standards and Technology, SHA-3 Standard:Permutation- Based Hash and
Extendable-Output Functions (Federal Information Processing Standards (FIPS) Publication
202, 2015). https://doi.org/10.6028/NIST.FIPS.202
9. National Institute of Standards and Technology (NIST), Digital Signature Standard (Federal
Information Processing Standards (FIPS) Publication 186–4, July 2013. https://doi.org/10.
6028/NIST.FIPS.186-4
10. Hertig, What is a decentralized application? Coindesk (2017). Online available: https://www.coi
ndesk.com/information/what-is-a-decentralized-application-dapp/. Accessed on 14 May 2018
11. G. Foroglou, A.-L. Tsilidou, Further applications of the blockchain (2015)
12. A. Kosba, A. Miller, E. Shi, Z. Wen, C. Papamanthou, Hawk: the blockchain model of cryptog-
raphy and privacy-preserving smart contracts, in Proceedings of IEEE Symposium on Security
and Privacy (SP) (San Jose, CA, USA, 2016), pp. 839–858
13. N. Singh, Ethereum or hyperledger fabric? Medium (2018). Online available: https://medium.
com/quillhash/ethereum-or-hyperledger-fabric-259f3c9b8da6. Accessed on 02 May 2018
14. Exchange Ethereum Stack https://ethereum.stackexchange.com/questions/760/how-is-the-add
ress-of-an-ethereum-contract-computed
15. Y. Zhang, J. Wen, An iot electric business model based on the protocol of bitcoin, in Proceedings
of 18th International Conference on Intelligence in Next Generation Networks (ICIN) (Paris,
France, 2015), pp. 184–191
16. H.-H. Buerger, Ethereum white paper, Github.Com (2016). Online available: https://github.
com/ethereum/wiki/wiki/White-Paper. Accessed on 18 Feb 2018
17. P. Sajana, M. Sindhu, M. Sethumadhavan, On blockchain applications: hyperledger fabric and
ethereum. Int. J. Pure Appl. Math. 118, 2965–2970 (2018)
18. LDAP.com. https://www.ldap.com
19. S. Pongnumkul, C. Siripanpornchana, S. Thajchayapong, Performance analysis of private
blockchain platforms in varying workloads, in 2017 26th International Conference on
Computer Communication and Networks (ICCCN) (2017), pp. 1–6
20. M. Valenta, P. Sandner, Comparison of ethereum, hyperledger fabric and corda (2017)
Blockchain Technology: Consensus
Protocol Proof of Work and Proof
of Stake
B. Sriman, S. Ganesh Kumar, and P. Shamili
Abstract The core technology used for many of the cryptocurrencies is blockchain.
The extensive research attention has been received by the blockchain, a distributed
ledger technology. The peer to peer network and the cryptography are playing a
major role in blockchain technology. In addition to that, the consensus protocols
are also playing an important role in the fundamentals of the blockchain tech-
nology. In blockchain systems, the security and the fault tolerance were guaran-
teed by the consensus protocols. There are two broad categories of the blockchain
consensus protocols. They are Probabilistic finality consensus and the absolute
finality consensus protocols. In this paper, the consensus protocols of the above
two categories and the strengths and the weakness of blockchain types that could be
applicable are presented.
Keywords Blockchain ·Consensus protocols [POW, POS] ·Cryptocurrency ·
Cryptography
1 Introduction
A new decentralized cryptocurrency was described by Nakamoto in the white paper
for the first time in blockchain. The digital cryptocurrency [1] like bitcoin took the
blockchain technology attention among all the people. After that, most of the ideas
came out with the blockchain technology. With the combination of all the techniques
like cryptography, peer to peer network [1], and distributed system technology, the
B. Sriman (B)·S. Ganesh Kumar ·P. S h a m il i
Computer Science and Engineering, SRM Institute of Science and Technology, Chennai 603203,
India
e-mail: srimanb@srmist.edu.in
S. Ganesh Kumar
e-mail: ganeshk1@srmist.edu.in
P. Shamili
e-mail: shamilip@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_34
395
396 B. Sriman et al.
blockchain uses it. Blockchain provides a secure framework to the cryptocurrencies,
where no one could tamper the data of the transactions. Due to this, the wide use of
blockchain technology [5] is used in the fields of Supply chain, Internet of Things
(IoT), medical, and the financial fields.
There is a big issue based on applying the blockchain technology and many of
the challenges; how to design a suitable consensus protocol. Here the consensus
is the distributed ledger, [6] that is maintained by all the nodes in the network.
In the existing system (Traditional software), the main problem is maintaining the
consensus because of the centralized servers [7]. But, in the blockchain distributed
system, each and every node itself is a host and a server. Moreover, it needs to share
the information to all the nodes in order to complete the consensus protocols.
This is an excellent consensus protocol that could tolerate the phenomena and
minimize the dangers which are not affected by the consensus results. The consensus
protocols should be adopted by the systems that are suitable for the system based
on blockchain. The three types of blockchain are public, private, and consortium
blockchain [19]. Each one has its own application scenarios. In this paper, the
different consensus protocols of the blockchain are analyzed and their applications
and the performances based on the scenarios.
2 Consensus Protocols Proof of Work [Pow]
Satoshi Nakamoto [1] applied the bitcoin white paper with the concept of Proof of
Work (PoW) in 2008, which was first published by Bonneau et al. [7] Cynthia Dwork
and Moni Naor in 1993. The most popular consensus mechanism for cryptocurrency
is the Proof of Work. In 1999, the term “Proof of Work” was initially used by Markus
Jakobsson and Ari Juels [3]. The principle followed here is “Easy to verify butdifficult
to find”.
Principle: A solution that is easy to verify but difficult to find
A. Miners
The process of solving the crypto puzzle is called mining [12]. The mining activity
performed by the node in the network is named as minor. Then the new block is added
to the blockchain as shown in Fig. 1. Here comes the proof of work as an answer to
the mathematical problem (crypto puzzle) which is considered as the solution that is
difficult to find but it is easily verified.
B. Mining Pool
For finding the new blocks, the amount of work done by the miners over a network
with the shared processing power equally in the pooling of resources are the mining
pool [14]. The members in the mining pool are awarded with the amount of “Share,”
who provides the partial verified Proof of Work. The difficulty level for the minors
is increased for mining in the pools where it takes more time for the slower miners
Blockchain Technology: Consensus … 397
Fig. 1 Flow of POW
to create a block as shown in Fig. 1. Rather than randomly finding once few years, it
is better to get rewards for every newly generated block. This is the solution for the
problem to the miners to pool in their resources.
C. Solving the Crypto Puzzle
The puzzle is solved by guessing at random hash value [1]. Predicting the output
is made impossible by using the hash function. For that difficulty, the minors choose
the value randomly and apply it to the hash function and the data in a particular block.
The number of zeros that are preestablished will be the starting for the resulting hash.
There may be many results with the combination of two consecutive integers and it
leads to numbers that are impossible to find. Now, what to do if there is more than
one nonce produced? (all minors trying for the same block).
If a particular miner as shown in Table 2, in the peer to peer network solves the
crypto puzzle [10] or finds the hash within the time period and the announcement
takes place. Then the other miners will stop their work and go for the next nodes
to solve the crypto puzzle. As a result, the miners will be rewarded as they get new
bitcoins for solving the puzzle.
3 Cryptography SHA-256
A. Cryptography Major Role in Proof of Work
SHA 256 is the most popularly used Proof of Work consensus and was announced
as the portion of the Bitcoin. [1] The others were SCRYPT, SHA-3, SCRYPT JANE,
SCRYPT-N, etc.
398 B. Sriman et al.
Tabl e 1 SHA HASH calculation
SHA-1 SHA-224 SHA-256 SHA-384 SHA-512
Message digest size 160 224 256 384 512
Message size <264 <264 <264 <2128 <2128
Block size 512 512 512 1024 1024
Wor d s i ze 32 32 32 64 64
Number of step/round 80 64 64 80 80
With the help of the hash value, the miners try to find the random nonce [14]
(random data of small size) and find the block that holds the hash (of binary values)
as shown in Table 1, with particular 0’s. A hash with number of zeros that has the
rare hashes to find. If it’s a good hash, the data is not found and have to try many
times to find a perfect nonce [16].
The number of zeros were based on the difficulty [17] faced by the miners to
search for a perfect block. For every 10 min, a new block is created by an average of
how many blocks that are previously added to it.
B. Nonce
The cryptocurrencies and the blockchain are working under the mining algorithm
like Proof of Work with the concept of the central part as nonce [9]. A nonce, which
produces the hash which is lower than or equal to the hash value is fixed by the
network as difficult as shown in Fig. 2. For finding these values the miners do mining
and compete with each other. In that network, the miner who finds a nonce is named
as the golden nonce. Then the miner gets the reward and adds that block to the
Fig. 2 Hashing techniques
Blockchain Technology: Consensus … 399
Fig. 3 Block summary
blockchain network. It all happens during the Proof of Work mining. Now let us see
how the process of nonce is done in mining.
C. Building Blocks with Nonce
Like timestamp and the difficulty target, the block header stays with its key data
value. A key with 32-bit in block is the nonce as shown in Fig. 3. For creating a
new block, the miners choose a random nonce and add it to the block header while
building a block. As stated before in nonce, if the miners are not having the number
of zeros, then they avoid the hash and go for the new nonce. It will be repeated until
a nonce is discovered by the miners.
D. Block Reward
Every block that is mined successfully by the miners in the blockchain network are
rewarded with the Bitcoin block.
•The number of bitcoins a miner gets, is the number of rewards that miner mines
a block.
•The reward becomes half for every four years or every 210,000 blocks.
•The expected reward that hit be zero to around 2140
E. Concept of Block Reward
The size of every bitcoin block is 1 MB and holds the data of transaction informa-
tion. Illustration, when a transaction occurs between two nodes in the network and
information about transaction are stored in a block [16].
At the initial stage (in 2009), for each and every bitcoin block was rewarded as a
worth of 50 BTC (bitcoin). In February 2019, its reward was 12.5 BTC for one block
and bitcoin price is $3500, which is of 12.5 ×3500 =$42,000. The block rewards
400 B. Sriman et al.
were given to the miners those who do mining by using their computation powers
to find a new block as shown in Fig. 3Similarly, the other cryptocurrencies in the
blockchain have the same mechanism for rewarding the miners. The miner who wins
the block reward adds the first transaction on the block.
4 Main Issues with Proof of Work Consensus
The following are some of the issues for the Proof of Work consensus mechanism.
A. The 51% Attack
In the blockchain network, if any node gains 51% or more than 51%, [8] the nodes
could influence the blockchain by gathering most of the network as shown in Table 2.
Time consuming: for solving the crypto puzzle, the miners have to check for the
nonce which must be solved to mine the block. This helps in time consumption.
•Resource consumption: in order to solve the mathematical crypto puzzle, the
miners consume high computational power. This results in wastage of resources
like hardware, space, money, and energy. It is estimated that the world’s electricity
spent for verifying transactions in 2018 was 0.3%.
•It takes a minimum of 10–60 min for the confirmation of any of the transactions in
the blockchain network. This is due to the time taken for mining the transactions
and adding it to the blockchain.
B. Computation Power
•Carbon footprint: 34.73 Mt CO2In comparison with the carbon footprint
of Denmark over 723,140 VISA transactions with 48,872 h of time watching
YouTube.
•Electrical Energy: 73.12 TWh In comparison with the power consumption of
Austria, the equivalent power consumption has an average of over 20.61 days in
US as shown in Fig. 4.
Tabl e 2 Analysis and comparison
System function Proof of work [POW] Proof of stake [POS]
Mining power The work (solving the crypto
puzzle) done by the miner
Depends on the stake of the
miners
51% attack Level incentive to avoid 51%
attack
Highly expensive in 51% attack
Energy consumption Higher Lesser
Decentralized versus
centralized
Becomes powerful when nodes
which tend to be centralized
over a period of time
Mining communities will
become decentralized based
upon their stake
Target time for a block For every 10 min For every 15 s
Blockchain Technology: Consensus … 401
Fig. 4 Energy consumption chart
•Electronic Waste: 11.49 kt In comparison with the e-waste generation of Luxem-
bourg, as shown in Fig. 4, the equivalent weight of 1.48 “C”-size batteries or golf
balls around 2.09
C. Definition of 51% Attack
This attack on blockchain is the 51% attack. [15] Usually the bitcoins, the attacks
which are mainly by a group of miners who control more than 50% of the mining
hash in the network or by the computational power. The ability as shown in Table 2,
of the attackers is to prevent the transactions that gain confirmations and can halt the
payments between the users [19]. They can also reverse the completed transactions
of any user in the network and this leads to Double Spend coins. The attackers could
not alter old blocks or create new coins in the blockchain-based cryptocurrency.
D. Double–Spending
The risk at which the spending of digital currency is twice is called Double spending.
The savvy individuals, who understand about the blockchain network and the compu-
tational power needed were reproduced to obtain the digital currencies that are unique
potential problem [15]. This type of issue will not happen during the cash transac-
tions, the parties involving in transactions can easily verify the authenticity and the
ownership of the physical currency. But in digital currency, the digital token can be
copied by the holder and it might be sent to any of the parties or the merchants with
the original one.
E. Proof of Work System Features
The two features that could contribute to the wide spread of the consensus protocol
are
402 B. Sriman et al.
•It is difficult to do the mathematical crypto puzzle problem.
•The correctness of the solution could be easily verified.
5 Consensus Protocols Proof of Stake [Pos]
The complete virtual consensus mechanism was made by the Proof of Stake. The
way of achieving the goal differs and the process remains the same as proof of work.
While in PoW, the miners solve the mathematical crypto puzzle with the help of the
high computing resources.
A. Validators
The validators will be there in PoS instead of miners in PoW [13]. In the ecosystem,
the validators lock some of their Ether [2] as a stake. Like betting, the validators as
shown in Fig. 5, will bet on those blocks that are likely to be added in the chain. In a
proportion to the stake, the validators will be rewarded when a new block is added.
1. One who becomes a validator may hold the cryptocurrency and also sometimes
locked up deposit were required.
2. It is also done based on how much stake or cryptocurrency the validators having
are the chances to mine a new block.
3. In the PoS protocol, it will randomly assign the validator and give the right to
create a new block in between as shown in Fig. 5, the validators and that is based
on the stake value.
4. The reward as shown in Fig. 5, will be provided for the chosen validator.
The Proof of Stake is mainly suitable for resolving the BFT (Byzantine Fault
Tolerance) [4] as the validators were tracked in the network and the known identities.
Fig. 5 Flow of POS
Blockchain Technology: Consensus … 403
For example, the list wallet address. In Byzantine Fault Tolerance, it requires 2/3 of
the validators to be honest and keeping all these individuals helps to maintain the
status.
B. Advantage
•Efficiency in energy
The PoS algorithms were efficient in energy in comparison with PoW [13]. The
mining process in PoS makes it a greener option in cutting out the energy intensively.
•Security
For security in PoS, the attackers should proceed with their stakes and the assets
in order to attempt 51% attacks. In comparison with PoW, the attackers will not lose
their hardware when attempting the 51% attacks.
•Decentralization
Leading to the real threat of centralization, the group of miners mining their resources
(mining pool), could control over 51% [15] of the networks that are running in PoW
systems. As a result, there is an exponential increase in the reward for a single
investment in PoW system because of opposition in the linear increase in the PoS
system. For an example, if the validator in PoS network invests twice as much as any
other validator, they will be granted the control.
C. Conclusion
The PoS made the validators not to have their own computing power as the factors
that may influence the validators to win number of their own coins and the complexity
of the network.
The benefits that switched from PoW to PoS are as follows.
•Savings in energy
•If it becomes highly expensive, there will be safer network.
D. Casper Blockchain
The consensus mechanism in which both the POW/POS works together is named
as Casper Blockchain. Where the Proof of Stake is overlaying on the ethash of the
Proof of Work protocols. In this way, the Casper is a security deposit protocol that
are based on the economic consensus system. The new consensus will be thanked
for making the validators to deposit the security amount to be a part of consensus
protocol. The Casper protocol is a control over security deposit that determines the
specified amount of rewards the validators receive. The validators security deposit
will be deleted and the privilege to be a part of the consensus network will be removed,
if the validator creates the “invalid” block. The Casper security can also be defined
as doing something like Bets. In the PoS protocol, the validators who won the bet
will be rewarded with the money prize based on the transactions.
404 B. Sriman et al.
6 Analysis and Comparison
A. Fault Tolerance
In PoW and PoS protocols [10], the attacker needs a large amount of computing
power or the stake [18] for the creation of a long chain instead of the valid chain. In
bitcoin, the double-spending attack is completed by creating the longer private chain
with a fraction of 50% by using the high computing power. The blockchain network
becomes undetermined, if the attackers fraction becomes 50% or more than that. The
stakeholder who has less than 50% stake only be allowed in the PoW and PoS.
B. Limitation
The PoW consumes most of the computational power in comparison with the other
consensus protocol and the output of PoW per second in the bitcoin, which limits
the applicable transactions of PoW in the payments [20]. The PoW and PoS have the
same features, even though the consumption of the computation power is reduced
and the stakeholders get the reward for the block. Each node communicating in the
PoS requires the message transfer in every round of the consensus, thus it has a
high performance for the network. There will be no anonymity due to the nodes
participating in the consensus. In the PoS, the process in the consensus is completed
in a few seconds, that is suitable for the scenario in payments.
C. Scalability
The PoW and PoS are having good scalability. But there are some ways to improve
the scalability. For example, in Bitcoin it adopts the network of lightning to give
the Off-Chain system to improve the scalability. The plasma layers of layer 1 and
layer 2 are the scaling solution and the sharding technology were proposed by the
Ethereum [8], respectively. With the small number of nodes, the PoS is limited to
the high-performance networks.
D. Scenarios
The current blockchain system is of three parts. They are public, private, and consor-
tium blockchain as mentioned above. Everyone in the public blockchain will partic-
ipate in the distributed ledger and the consensus process. The PoW and PoS [11]
are applicable to the public blockchain. In the private and consortium blockchain,
the nodes in the network participates in the consensus process that belongs to the
permissioned blockchain. Even though, the public and the consortium are not like
a public blockchain, due to the high efficiency and their strong consistency of the
consensus, it will be applicable in the medical and the commercial scenarios.
Blockchain Technology: Consensus … 405
7 Conclusion
In the blockchain system, the consensus protocols are guaranteed for the perfect
operations in the blockchain systems. The nodes in the blockchain network agree
[12] to the transaction or the values through the consensus protocols. In this paper,
the blockchain consensus protocols and the strengths and the weaknesses were
discussed through the application scenarios through the analysis and the compar-
isons. It is concluded that the consensus protocols were of best use and applicable
in the appropriate application scenarios.
References
1. S. Nakamoto et al., Bitcoin: a peer-to-peer electronic cash system
2. V. Buterin, On public and private blockchains (2015). https://blog.ethereum.org/2015/08/07/
on-public-and-private-blockchains/
3. M. Swan, Blockchain: Blueprint for a New Economy (O’Reilly Media, Newton, MA, USA,
2015); A.M. Antonopoulos, Mastering Bitcoin: Unlocking Digital Cryptocurrencies (O’Reilly
Media, Newton, MA, USA, 2014)
4. J. Fan, L.T. Yi, J.W. Shu, Research on the technologies of Byzantine system. J. Softw. 24(6),
1346–1360 (2013). Online available: http://www.jos.org.cn/1000-9825/4395.htm
5. T.T.A. Dinh, R. Liu, M. Zhang, G. Chen, B.C. Ooi, J. Wang, ‘Untangling blockchain: a data
processing view of blockchain systems’. IEEE Trans. Knowl. Data Eng. 30(7), 1366–1385
(2018)
6. F. Tschorsch, B. Scheuermann, Bitcoin and beyond: A technical survey on decentralized digital
currencies, IEEE Commun. Surv. Tuts. 18(3), 2084–2123 (2016). 3rd Quart
7. J. Bonneau, A. Miller, J. Clark, A. Narayanan, J.A. Kroll, E.W. Felten, SoK: research perspec-
tives and challenges for bitcoin and cryptocurrencies, in Proceedings of IEEE Symposium on
Security and Privacy (San Jose, CA, USA, 2015), pp. 104–121
8. G. Wood, Ethereum: a secure decentralised generalised transaction ledger, Yellow paper (2015)
9. C. Natoli, V. Gramoli, The blockchain anomaly, in Proceedings of the 15th IEEE International
Symposium on Network Computing and Applications, NCA’16 (2016), pp. 310–317
10. C. Natoli, V. Gramoli, The balance attack against proof-of-work blockchains: the R3 testbed
as an example, Tech. Rep. 1765133 (2016)
11. A. Black, Hashcash-Adenial of service counter-measure, Tech. Rep. Cypherspace, (2002).
URL http://www.hashcash.org/papers/hashcash.pdf
12. Alyssa Hertig, What Is Ether? Coin Desk, 21 Apr (2017). https://www.coindesk.com/inform
ation/what-is-ether-ethereum-cryptocurrency
13. F.B. Crane, Proof of work, proof of stake and the consensus debate. Cointelegraph, 20 Dec.
2014. https://www.cointelegraph.com/news/proof-of-work-proof-of-stake-and-the-consensus-
debate
14. M.A Alahmad, I. Al-shaikhli et al., Joux multicollisions attack in sponge construction, in
The 6th International Conference on Security of Information and Networks (SIN), 2013 6th
International Conference on, ACM (2013)
15. M.A Al-Ahmad, I.F. Alshaikhli, Broadview of cryptographic hash functions. Int. J. Comput.
Sci. Issues 10(4), 239–246(2013)
16. S. Wu, D. Feng, W. Wu, J. Guo, L. Dong, J. Zou, (Pseudo) preimage attack on round-reduced
Grøstl hash function and others, in Canteaut, vol. 9, pp. 127–145
17. X. Wang, H. Yu, Y.L. Yin, Efficient collision search attacks on SHA-0, in Advances in
Cryptology–CRYPTO 2005 (Springer, Berlin, 2005)
406 B. Sriman et al.
18. Merkle tree URL https://en.wikipedia.org/wiki/merkletree
19. J.R. Douceur, The Sybil Attack, in: International Workshop on Peer-to-Peer Systems (Springer,
Berlin, 2002), pp. 251–260
20. Total cryptocurrency market cap hits new all-time high (2018). URL https://www.coindesk.
com/150-billion-total-cryptocurrency-market-cap-hits-new-time-high.Accessedon14May
2018
Non-Invasive Techniques of Nutrient
Detection in Plants
Amit Singh and Suneeta V. Budihal
Abstract Plants are really important for the earth and for all living beings, as they
absorb carbon dioxide released by the human being and release oxygen for their
survival. Plants require major nutrients such as Nitrogen, Phosphorus and Potas-
sium (NPK) in a proper ratio to keep them healthy and more resistant to pests. The
deficiency of nutrients in plants makes it impossible for the plants to complete the
vegetative and reproductive stage of their life cycle. It is needed to monitor the amount
of nutrients present in plants, so that if there is a deficiency of nutrients, manure and
fertilizers can be added to the desired extent. There can be many methods to monitor
the amount of nutrients in plants such as invasive and non-invasive methods which
include testing the sample in the laboratory, spectroradiometers, reflectometers and
image processing techniques. This paper proposes a survey work carried out and
guidelines for researchers over the various non-invasive methods for the detection
and estimation of macronutrients.
Keywords NPK concentration ·Non-invasive ·Machine learning ·Neural
networks
1 Introduction
The crop yield relies on the continuous interaction amongst the soil and plant features.
A healthier plant provides good crop yield. Hence fertile soil is the main source of
nutrients’ supplier from the plants. There is an increase in demand for food produc-
tion. In order to have plants and crops healthy, proper proportions of nutrients in the
plants are needed. In reality, majority of macronutrients are acquired from the soil
A. Singh (B)·S. V. Budihal
School of ECE, KLE Technological University, Hubli 580031, India
e-mail: amitsingh11097@gmail.com
S. V. Budihal
e-mail: suneeta_vb@kletech.ac.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_35
407
408 A. Singh and S. V. Budihal
such as NPK. Any kind of deficiency in micronutrients can be managed but a defi-
ciency of macronutrients results in a major damage to the growth of the plants and
yield of the crops. Any improper use of fertilizers will affect the plants’ growth and
fertility leading to reduced crop yield, reduced fruit and vegetable size, changed taste
of fruits. Many times, improper utilization of the fertilizers will affect the quality of
the fertile soil also. Excessive use of external fertilizers will contaminate the surface
and groundwater resources. Hence the discussions by the agricultural researchers
and engineers realized that the following topics are to be investigated further.
1. Soil-based agriculture.
2. Timely testing of the soil.
3. Need for pesticides and fertilizers.
4. Early detection of plant diseases and its remedies.
5. Estimation of the macro- and micronutrients.
Nitrogen is abundantly available in the soil but all other nutrients are affected by
the pH value, temperature, organic materials and moisture of the soil. We can monitor
the amount of nutrients in many ways like visual deficiency, soil test, plant tissue
test and the behavior of the plant’s response to the fertilizers or organic manure.
The symptoms of nutrients’ deficiency can be visible in the leaves. Some symptoms
include interveinal chlorosis, marginal chlorosis, edge distortion, changes in the
shape of the leaf, burning and holes in the leaf and change in the colour of the leaf.
Therefore, there is a requirement of nutrient analysis on a timely basis which should
be non-harmful to plants.
2 State-of-the-Art Techniques
A study over the existing literature demonstrated that the NPK nutrients are identified
not only through the health of the leaves, but also through the quality of the soil.
This is carried out by processing the soil and leaves chemically to detect the pH
values. Based on the values, the nutrient concentrations are detected and further
analyzed for actions. Here the leaves of the plants need to be crushed and dried
and processed chemically for further initiatives. Such traditional modes of nutrient
deficiency detection and estimation are termed as an invasive mode in agricultural
language. As these are completely dependent on the amount of the chemical to be
used, type of the chemical to be used, the temperature and lighting conditions change
in the colour of the ingredients after the chemical reaction, the expertise in handling
the laboratory experimentation, human observation errors, etc.
It led to the development of non-invasive mode of nutrient concentration detection
without harming the plants and to provide results that are close to actual values. It
is needed to detect NPK concentrations using sensors, micro-controllers, IoT-based
techniques, Image processing techniques, spectrometers, optical sensors, Machine
Learning techniques, etc.
Non-Invasive Techniques of Nutrient Detection in Plants 409
Currently, large groups of researchers are evaluating the nitrogen content through
non-destructive techniques. Here the leaf chlorophyll content is mapped to nitrogen
concentration as in [1], such as the Soil Plant Analysis Development (SPAD), which
was used to investigate N concentration [2]. In order to monitor the plant nutrition
concentration [3] in large scale, spectral remote sensing analysis technology was
developed. It is very difficult to capture and quantify micronutrients with hyper-
spectrometer [4]. Added to it is the difficulty with the data processing and computa-
tion processes are challenging [5], which restrict the application. However, the high
cost of [6] fluorescence detection sensors for N nutrition detection limits its usage.
Another technique called as phenotyping based on imaging processing was employed
to evaluate N nutrition. Most of the research was focused on colour traits [7] and
[8]. It is indeed needed to help in identification [9], classification [10] and prediction
[11] in plants. ML approaches are applied when a large data set is available, relating
inputs to output quantities. The ML algorithms Support Vector Regression (SVR),
Random Forest (RF) and Neural Network (NN) that supports a non-destructive, easy,
low-cost realization.
3 Non-Invasive Techniques
There are many studies done by the researchers for the nutrient’s analysis. Electro-
chemical sensor has been developed to determine NPK ratio. Multispectral imaging
which is for a few discrete spectral channel or wavebands, Hyperspectral imaging
which is used with a sequence of continuous waveband covering a specific region,
digital image processing, colour histogram were reviewed by many of the researchers
and they all are used as a non-invasive technique for the analysis of nutrients. Many
methods have been proposed by the researchers for diagnosing the nutrient deficien-
cies in plants and crops. In these papers, the various methods and algorithms for
monitoring the nutrient in plants are discussed. Some of the techniques have been
discussed below.
3.1 Detection Using Sensor
In [12], the author proposed the system for monitoring nutrient and moisture level of
the soil using antimony electrode for pH measurement, and the temperature of soil is
measured using DS18B20 sensor; the temperature sensor works on a Dallas one wire
protocol architecture. In [13], the author has made the project which includes various
devices like GPS-based remote controlled device, moisture and temperature sensing,
leaf wetness and security facilities. A wireless sensor is used for reading the soil prop-
erties. In the control system, data received is compared with data set. According to
that, the values are generated in the android application and farmers get the informa-
tion. In [14], the authors take important features like temperature, sunlight, humidity,
410 A. Singh and S. V. Budihal
Fig. 1 A block diagram of NPK detection using sensors and microcontrollers
soil moisture, pH and design a plant monitoring and smart gardening system with
the help of IoT. All the features are measured using sensors and interface with the
raspberry pi microcontroller. All the information of the garden can be directly moni-
tored by the owner of the farm/garden in his/her mobile phone using IoT. In [15],
the author has proposed a system to determine NPK using electrochemical sensor
dipped under soil. Electrochemical sensor works under the principle of absorption
of ions from the soil. Sensors are connected with Arduino for monitoring nutrients
deficiency.
In [16], the author proposed a system which has a hand held device which will give
the pH value and estimate NPK from the pH of that soil. A classification algorithm
to predict the crops based on the values got from the device is used. In [17], a plant
watering system is proposed where the sensor is used to check soil moisture content,
and then Arduino switches ON water pump if the level is less and the information is
sent to the users using IoT modules. In [18], the author proposed a system by using
features like pH level and temperature (Fig. 1).
3.2 Detection Using Image Processing
To analyze the nutrients using image processing technique required the following:
1. The area of the leaf to be processed.
2. Integration of edges and veins of the leaf.
3. Processing of the area of the leaf taken.
4. Extraction of colour feature in the leaf.
5. Identification of deficient nutrients.
Non-Invasive Techniques of Nutrient Detection in Plants 411
A rapid, non-destructive, repetitive and relatively inexpensive imaging technology
is better compared to optical techniques to provide an approach to non-invasive
sensing of N in plants. In [19], a system to measure the pH value, temperature,
various spots and patterns is developed using image processing technique and using
sensors like soil sensor and temperature sensor. The input pixel of the colour images
in three different planes RGB and stored into three variables. After nutrient level is
detected all the data is sent to the cloud server.
In [20], the author discussed the detection of the leaf diseases using the leaf images
and process the image. Nutrient detection is using a web application which is using
image processing and CNNs for finding diseases in leaves. The author’s objective
was to introduce algorithms which can be operated on leaf images include RGB
to HIS. After feature extraction, clustering methods and CNN are used to detect the
diseases in plants. Colour, Hu’s moment and flood fill algorithm are the three features
vector used for the feature extraction purpose. In [21], the author proposes a method
using colour image analysis for monitoring the nitrogen contained in the grapes. The
first step is image acquisition, then the sample is processed in the lab, the method
used in the lab is Kjeldahl digestion method.
In [22], the authors propose the accurate detection of diseases by using plant leaf
images. The features like spots in the leaf, various leaves pattern, edge detection,
colour detection are some of the features that were taken by the author for detecting
the diseases present in the leaves. The technique is inexpensive compared to other
techniques. In [23] the rice leaf features were extracted and stepwise discriminate
analysis for the cross-validation is used to identify NPK. In [24], the author proposed
a method using colour image analysis for the estimation of nitrogen for the cotton
plants. The image processing technique is used for the accurate and faster result and
the technique is inexpensive also. Instead, DIP technique is better than the hyperspec-
tral imaging because it is inexpensive and it is easy to operate. DIP can extract spectral
image, can give morphological, colour and texture information. Image processing is
a non-destructive technique in agriculture field that can be used for greater accuracy
and for faster results. The first step is image acquisition where images are captured.
The second step is pre-processing, where removal of noise takes place. For this
purpose, various pre-processing techniques are used like image clipping, etc. The
third step is image segmentation where the image is separated into various parts
having same features. K-means clustering is used or segmentation, converting RGB
to HSI model. The fourth step is feature extraction, used for identification purpose.
The last step is the classification stage; for classification stage SVM is used (Fig. 2).
3.3 Detection Using Spectral Imaging
Concentration of elements in plants is determined using advanced spectroscopy.
The commonly employed methods are Ultra Violet, Visible and Infrared reflectance
spectroscopy. In [25], the author reviewed existing methods and techniques used
for plants’ nitrogen detection. Estimation of nitrogen in plants is carried out through
412 A. Singh and S. V. Budihal
Fig. 2 A block diagram of
NPK detection using image
processing technique
many techniques such as spectroradiometer, reflectometer, digital camera to measure
optical properties such as chlorophyll and polyphenol fluorescence, crop canopy
reflectance, leaf transmittance, etc. The information of the nutrients is easily trans-
ferred to the farmer in android application using IoT module. In [26], the normal plant
data is collected and compared with the data collected by the sensor. When illumi-
nated with a source of appropriate wavelength, based on diffused reflectance, soil
sample is obtained. In the paper [27], the author discussed estimation using visible
near-infrared source and reflectance value was observed for NPK. In [28], the author
proposed a system for the estimation of NPK nutrients depending on various factors
like soil type, soil microbes and soil pH value. The samples were analyzed using Kjel-
dahl method, UV Spectrophotometer and Atomic Absorption Spectroscopy (AAS)
for the estimation of NPK, respectively. The spectral data is collected using the
assembled handheld spectroradiometer, which consists of a transmitter/receiver setup
(Fig. 3).
Fig. 3 A block diagram for
spectral imagining for
nutrient detection
Non-Invasive Techniques of Nutrient Detection in Plants 413
3.4 Detection Using Sensing Techniques
The nutrient concentration detection and estimation using various sensing methods
are discussed in [29]. Most of the soil nutrient sensing techniques described in liter-
ature involve either an optical sensing using reflectance spectroscopy or an elec-
trochemical sensing using ion-selective electrodes. However, this review focuses
mainly on optical sensing methods and principles for measuring the macronutrients
and other properties in soil. The advantages associated with the optical sensing over
electrochemical technology are non-destructive measurement and no need to take the
sample [30] which has attracted the researchers. In order to evaluate the properties of
the material, the reflectance, absorption/transmittance are used [31]. Optical sensors
are frequently affected by the different soil properties and exhibit different responses
in different regions of the spectral field.
In [32], author used Near Infrared Reflectance (NIR) spectroscopy to detect the
nitrogen concentration through Nitrates and soil moisture content, respectively. The
authors in paper [33] demonstrated usage of admittance spectroscopy to detect the
variation in the nutrient content in plants. The soil organic carbon, nitrate and mois-
ture content is detected using VIS-NIR MIR spectroscopy by Bogrekci and Lee
[34]. The phosphorous is detected through NIR absorbance spectroscopy by authors
in paper [35]. The VIS and NIS methods are used to detect phosphorous and pH
values in [36]. The Raman spectroscopy is used to detect phosphorous in [37]. Lee
et al. [38] discussed and analyzed the Raman scattering, Reflectance spectroscopy
to detect nitrogen and phosphorous. The paper [39] is considered to discuss the
optical sensor for nutrient detection. A MEMS-based technique is explored in paper
[40] for nitrogen detection. Diffuse reflectance spectroscopy is elaborated in [41]for
the detection of nitrogen and organic matter. The authors in [42] explained about
the concentrations of potassium and phosphorous using diffuse reflectance spec-
troscopy. It includes optical, acoustic, electrical and electromagnetic, electrochem-
ical and mechanical [43] techniques. Its sensitivity and quick response [44], the
optical detection method has higher potential towards real-time scenarios. Optical
method was used in several reports on NPK soil detection [45] and [48]. Most of the
devices used additional optical components such as fibre optic to drive light to soil
[46] and [47].
In Fig. 4, NPK detection of soil with optical transducer is overviewed. Light
transmission, detection integrates to form optical transducer. An Arduino was used
to operate the light source in a transmission system as data acquisition from the light
detection system and provides LCD display.
414 A. Singh and S. V. Budihal
Fig. 4 A block diagram for integrated microcontroller with optical transducer
3.5 Future Research Directions
A survey over the state-of-the-art techniques provides great insights for the NPK
detection using non-invasive method. The following research problems are identified
as outcomes of the work.
1. The deep learning techniques can be applied when the data set is available for
the plant leaves.
2. A prediction framework may be proposed to detect the value of one nutrient
having known with the other two.
3. Advanced controllers may be used to process data and speedy operations.
4. A common framework for nutrient detection and disease detection may be
proposed using plant leaf as reference data.
5. A strong data set may be generated for nutrient detection in future use.
6. Using MEMS, a sensor may be developed to detect the NPK nutrients in one
stretch.
4 Conclusion
Plants are an important part of every living being on the earth. It is the responsibility
of humans to take care of plants and get good yield by proper usage of pesticides
and fertilizers. The plants need proper proportions of macro- and micronutrients for
their proper growth. The study in this paper concentrates on the detection of NPK
the macronutrients at the early stage for proper handling. The paper provides state-
of-the-art techniques for the NPK concentration detection in plants. Deficiency of
nutrients in plants makes it impossible for the plants to complete the vegetative and
reproductive stage of their life cycle. It is needed to monitor the amount of nutrients
present in plants, so that if there is a deficiency, manure and fertilizers can be added to
the desired extent. There can be many methods to monitor the amount of nutrients in
plants such as invasive and non-invasive methods. This paper proposed a survey work
Non-Invasive Techniques of Nutrient Detection in Plants 415
carried out and guidelines for researchers over the various non-invasive methods for
the detection of macronutrients. The paper concludes by suggesting a few research
directions in non-invasive nutrient detection techniques.
References
1. R.H. Fox, C.L. Walthall, Crop monitoring technologies to assess nitrogen status. Nitrogen
Agriculture Systems Nitrogeninagric (2008) pp. 647–674
2. S.T. Ata-Ul-karim et al., Non-Destructive Assessment of Plant Nitrogen Parameters Using Leaf
Chlorophyll Measurements in Rice (Frontiers in Plant Science, 2016), pp. 1829–1836
3. T.J. Nigon et al., Hyperspectral aerial imagery for detecting nitrogen stress in two potato
cultivars. Comput. Electron. Agri. 112, 36–46 (2015)
4. L. Chen, L. Lin, G. Cai, Y. Sun, T. Huang, K. Wang, J. Deng, Identification of nitrogen, phos-
phorus, and potassium deficiencies in rice based on static scanning technology and hierarchical
identification method. PLoS ONE 9(11), 113–200 (2014)
5. F.N. Gehan et al., Lights, camera, action: High-throughput plant phenotyping is ready for a
close-up. Curr. Opin. Plant Biol. 24, 93–99 (2015)
6. G. Agati, L. Tuccio, B. Kusznierewicz, T. Chmiel, A. Bartoszek, A. Kowalski, A. Grzegor-
zewska, R. Kosson,S. Kaniszewski, Nondestructive optical sensing of flavonols and chlorophyll
in white head cabbage grown under different nitrogen regimens. J. Agriculture. Food Chem.
64, 85–94 (2015)
7. S.B. Sulistyo et al., Building a globally optimized computational intelligent image processing
algorithm for on-site nitrogen status analysis in plants. IEEE Int. Syst. 33(3), 15–26 (2018)
8. Sulistyo et al., Regularized neural networks fusion and genetic algorithm based on-field nitrogen
status estimation of wheat plants. IEEE Trans. Ind. Inform. 13(1), 103–114 (2017)
9. F. Alvaro, Y. Sook, K. Sang, P. Dong, A robust deep-learning-based detector for real-time
tomato plant diseases and pests recognition. Sensors 17(9), 20–22 (2017)
10. A. Picon et al., Deep Convolutional neural networks for mobile capture device- based crop
disease classification in the wild. Comput. Electron. Agric. 161, 280–290 (2019)
11. X. Yu et al., Prediction of regulatory interactions in Arabidopsis using gene- expression data
and support vector machines. Plant Phys. Biochem. 49, 280–283 (2011)
12. N.A. Abdullah et al., An IoT based system for remote monitoring of soil characteristics, in
International Conference on Information Technology (InCITe) the Next Generation IT Summit
on the Theme-Internet of Things: Connect your Worlds (2016), pp. 316–320
13. N. Suma et al., IOT based smart agriculture monitoring system. Int. J. Recent. Innov. Trends
Comput. Commun. 5(2), 177–181 (2017)
14. R. Elangovan, R. Rozario, A Plant monitoring and smart gardening system using IoT (2018)
15. R. Sindhuja, B. Krithiga, Soil nutrient identification using Arduino. Asian J. Appl. Sci. Technol.
(AJAST) 1(4), 40–42 (2017)
16. K. Abhang, S. Chaughule, P. Chavan, S. Ganjave, in Soil Analysis and Crop Fertility Prediction
(2018)
17. S.R. Pratibha et al., IOT based monitoring system in smart agriculture. Int. Conf. Recent. Adv.
Electron. Commun. Technol. 5(2), 81–84 (2017)
18. P. Sihombing, N.A. Karina, J.T. Tarigan, M.I. Syarif, Automated hydro-ponics nutrition plants
systems using arduino uno microcontroller based on android. J. Phys: Conf. Ser. 978(1), 12–14
(2018)
19. P.R. Harshani, M. Newlin Rajkumar, T. Maheshwari, in Monitoring the Nutrient Level for
Efficient Crop Productivity Using IOT (2017)
20. J.K. Patil et al., Advances in image processing for detection of plant diseases. J. Adv. Bioinform.
Appl. Res. 2(2), 135–141 (2011)
416 A. Singh and S. V. Budihal
21. A. Vibhute et al., Colour image processing approach for nitrogen estimation of vineyard. Int.
J. Agric. Sci. Res. 3(3), 189–196 (2013)
22. B. Patil, H. Panchal, S. Yadav, A. Singh, in Plant Monitoring Using Image Processing,
Raspberry Pi IoT (2017)
23. Y. Sun et al., Identification of nitrogen, phosphorus, and potassium deficiencies based on
temporal dynamics of leaf morphology and colour. Sustainability 10(3), 762–769 (2018)
24. A.P. Janwale, A survey on Nitrogen estimation techniques for plant using Image Processing
Techniques (2016)
25. R.F. Muñoz-Huerta, R.G. Guevara-Gonzalez, in A Review of Methods for Sensing the Nitrogen
Status in Plants (2013)
26. S. Mukherjee, S. Laskar, VIS-NIR-based optical sensor system for estimation of primary
nutrients in soil. J. Opt. 48(1), 87–103 (2019)
27. Mani et al., Estimating plant macronutrients using VNIR spectroradi-ometry. Pol. J. Environ.
Stud. 28(3), 1831–1837 (2019)
28. F.A. Azizan et al., Soil NPK variability mapping for harumanis mango grown in greenhouse
at perlis. Int. J. Adv. Sci., Eng. Inf. Technol. 9(2), 495–501 (2019)
29. Joseph V. Sinfield, Daniel Fagerman, Oliver colic, evaluation of sensing technologies for on-
the-go detection of macro-nutrients in cultivated soils. Comput. Electron. Agric. 70(1), 1–15
(2010)
30. H.J. Kim et al., Soil Macro nutrient sensing for precision agriculture. J. Environ. Monit. 1810–
1824 (2009)
31. V.I. Adamchuk et al., On-the-go soil sensors for precision agriculture. Comput. Electron. Agric.
44(1), 71–91 (2004)
32. A. Ehsani et. al., A NIR technique for rapid determination of soil mineral nitrogen. Precis.
Agric. 1385–2256 (1999)
33. J. Jeosadaque et al., A sensor for monitoring the volume of nutrient in a solid substrate-based
growth media by using electrochemical admittance spectroscopy. Sens. Actuators 87, 268–273
(2002)
34. B.R. Jahn et al., Mid-infrared spectroscopic determination of soil nitrate content. Bio Syst.
Eng. 94(4), 505–515 (2006)
35. I. Bogrekci, W.S. Lee, Effects of soil moisture content on absorbance spectra of sandy soils
in sensing phosphorus concentrations using UV-VIS-NIR spectroscopy. Trans. ASABE 49(4),
1175–1180 (2006)
36. A.M. Mouazen et al., On-line measurement of some selected soil properties using a VIS-NIR.
Sensor 93, 13–27 (2007)
37. W.S. Lee, I. Bogrekci, Portable Raman Sensor for Soil Nutrient Detection. US Patent:
US20070013908 (2007)
38. W.S. Lee et al., Sensing technologies for precision speciality crop production. Comput.
Electron. Agric. 74, 02–33 (2010)
39. G. Dooly et al., Highly selective optical fiber ammonia sensor for use in agriculture. Proc. Eur.
Sensors XXV, 1113–1116 (2011)
40. Y.S. Xiaoyu et al, MEMS Near-infrared reflectance absorption spectrometer for detection of
composition in soil. Int. J. Digit. Content Technol. Its Appl. (JDCTA) 7(1), 2013
41. V.B. Suneeta, et al., Facial Expression Recognition Using Supervised Learning. International
Conference On Computational Vision and Bio Inspired Computing. Springer, Cham (2019)
42. Yubing Wang et al., Soil pH value, organic matter and macronutrients contents prediction using
optical diffuse reflectance spectroscopy. Comput. Electron. Agric. 111, 69–77 (2015)
43. Y. Kulkarni et al., Primary nutrients determination in the soil using UV spectroscopy. Int. J.
Emerg. Eng. Res. Technol. 2, 198–204 (2014)
44. D.Q. Huo et al., Recent advances on optical detection methods and techniques for cell-based
micro fluidic systems. Chin. J. Anal. Chem. 38(9), 1357–1365 (2010)
45. N. Singh et al., D NPK measurement in soil and automatic soil fertilizer dispensing robot. Int.
J. Eng. Res. Technol. 3(7), 635–637 (2014)
Non-Invasive Techniques of Nutrient Detection in Plants 417
46. L.C. Gavade, Detection of N, P, K using fiber optic sensor and PIC controller. Inter Natl. J.
Eng. Sci. 13787 (2017)
47. D.V. Ramane et al., Detection of NPK nutrients of soil using fiber optic sensor. Int. J. Res.
Advent Technol. pp. 13–14, 2015
48. A. Rashid, LED based soil spectroscopy. Bul. Opt. 2016(3), 1–7 (2016)
Implementation of Cryptographic
Approaches in Proposed Secure
Framework in Cloud Environment
Manoj Tyagi, Manish Manoria, and Bharat Mishra
Abstract Through Internet, cloud computing shows extraordinary potential to give
on-demand services to its clients with higher suppleness in an economical manner.
Security is the primary need for this leading approach of computing ability. This
paper presents the analysis and implementation of some symmetric cryptographic
algorithms. It finds the best symmetric method and combines it with asymmetric
approach and shows the importance of hybrid encryption model for cloud security.
Moreover, this paper presents a framework by integrating CAPTCHA, Two Factor
Authentication (TFA), hybrid encryption approach, and separate server concepts
like authentication server with cryptographic server. It also suggests the use of
optimization for attaining availability and security strength.
Keywords Cloud computing ·Authentication ·Symmetric encryption ·
Asymmetric encryption ·Data confidentiality ·Integrity ·TFA ·CAPTCHA ·
Hybrid encryption
1 Introduction
In the last few years, online users and data demand have enormously grown, the
conventional computing infrastructure getting more expensive and more challenging
to maintain. Conventional computing is inappropriate for accessing data at anytime
and anyplace. To achieve this, it requires keeping the data on outside storage. Internet
usage is rapidly increasing which introduces new model for handling the quantity,
M. Tyagi (B)·B. Mishra
Mahatma Gandhi Chitrakoot Gramodaya Vishwavidyalaya, Chitrakoot, India
e-mail: manojtyagi80.bhopal@gmail.com
B. Mishra
e-mail: bharat.mgcgv@gmail.com
M. Manoria
Sagar Institute of Research Technology, Bhopal, India
e-mail: manishmanoria@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_36
419
420 M. Tyagi et al.
availability, and variety of data. This new model is cloud computing allowing resource
pooling, ubiquitous, convenient, on-demand access which are delivered by various
service provider interfaces [1]. Cloud computing allows its client to pay for only the
utilized time of resources. In this model, the clients request for the required resources,
then the cloud fulfills the clients’ request by providing such resources to the clients. It
gives enormous benefits to the enterprise as well as individual clients. By virtualiza-
tion technique, cloud computing allocates resources to its clients efficiently. Cloud
computing has characteristics like availability, manageability, and scalability. Addi-
tionally, it has various features such as multitenant, ubiquitous, expedient, elasticity,
on-demand service, stability and also it is economical [2].
2 Security Schemes
Security is an essential factor for cloud computing. End-users keep their data in the
cloud, remotely, so that users have no control over its data and don’t know about
the location of its stored data. Authentication and cryptographic schemes are used
to fulfill the security objectives. CAPTCHA can prevent Internet services from bots
through a simple test. The basic idea to design the CAPTCHA test is that a Human can
get success in it, but the computer program can fail [3]. Single-Factor-Authentication
(SFA) is simple to implement, and security is to depend on password strength. SFA
suffered if a hacker stole/break the password and access to the cloud through an iden-
tity check process. To overcome this problem, Multi-Factor-Authentication (MFA)
is an excellent option to protect the cloud from unauthorized access [4].
2.1 Symmetric Encryption
National Institute of Standards and Technology (NIST) has published DES which
suffered from weakness in both cipher-design and key-size where small key-size
invites brute-force attacks [5]. 3-DES is secure and the replacement of DES but
it is slow [6]. NIST announced the AES scheme which is widely accepted and
replaced the DES and 3-DES [7]. Famous cryptologists Bruce Schneier designed
a secure approach named Blowfish but it does not suit for large data due to its
small block size. Twofish is the successor of blowfish [8]. Lai and Massey intro-
duced a block cipher IDEA, which also provide good security [9]. Ross-Anderson,
Lar-Knudsen, and Eli-Biham designed block cipher named as serpent which is a
very secure approach but slow than AES [10]. Stafford Tavares and Carlisle Adams
also developed symmetric encryption approach CAST-128 and CAST-256 [11]. Ron
Rivest designed RC4 which is a renowned stream cipher known for its simplicity
and execution speed [12]. Ronald L. Rivest invented RC5 and RC6 where RC6 is
Implementation of Cryptographic Approaches in Proposed … 421
providing higher security, less round, and increased throughput [13]. Here symmetric
encryption is implemented in java on Intel core i5-8250U CPU 1.60–1.80 GHz with
8GBRAM.
2.2 Asymmetric Encryption and Hashing
Rivest, Shamir, and Adleman designed an asymmetric scheme RSA which is based
on prime number factorization. RSA is secure when using large prime number factor-
ization because of its hardness [14]. Koblitz and Miller independently developed an
elliptic curve based cryptosystem which is faster and needs less computing power.
A popular elliptic curve scheme is known as Elliptic-Curve-Integrated-Encryption-
Scheme (ECIES) [15]. The hash function generates a message digest to detect any
change in message to maintain integrity. Many hashing approaches are available like
SHA-3, SHA-2, SHA-1, MD5, where SHA approaches are better than MD5 due to
less secure behavior of MD5 [16] (Figs. 1and 2).
Fig. 1 Encryption time of symmetric algorithms except read/write time of file
422 M. Tyagi et al.
Fig. 2 Decryption time of symmetric algorithms except read/write time of file
3 Problem Definition
Lack of protection affects the acceptability of cloud. Practically, the computation
burden in cloud computing reduces its performance, and a single computation server
is not preferable in terms of security. Then, use two separate servers, one for authen-
tication operation and other for cryptography operation to reduce the burden and
improve security aspects [17]. CAPTCHA is accessible in a web-based application
to protect the system from bots and MFA is commonly accepted to verify the genuine
users. Generally, two factors are in practice for user verification in clouds which is
called TFA. By analysis of various symmetric key-methods, it is clear that AES is
better than other methods. It is fast as well as it offers excellent security.
Whereas, public-key cryptography like RSA is not practical to encrypt exten-
sive data due to its slow execution. This issue introduces an idea of hybrid encryp-
tion which combines the RSA and AES. The hybrid approach applies a symmetric
method AES to encrypt/decrypt the original message, and asymmetric method RSA
encrypts/decrypt the symmetric key of AES [18]. ECC is also another option of
asymmetric approach, and ECIES gives the complete integrated approach to secure
the cloud data. So the need for framework is required to achieve confidentiality,
authentication, and integrity as well as also it reduces the computation burden.
Implementation of Cryptographic Approaches in Proposed … 423
4 Proposed Framework
•Effective framework averts the data from various attacks in the cloud. Performance
of framework depends on techniques that are applied in a particular framework
for preventing the data. It shows that the selection of security techniques for
framework is very important. Various approaches are being applied for solving
the security issues like CAPTCHA, OTP, authentication, encryption, and integrity
schemes. Various techniques and methods are available for each approach, so
selection of effective techniques or methods is directly proportional to the security
of the system. Separate server enhances the security by reducing the chances of
server compromise situations and also reduces the overhead which improves the
speed. Proper resource utilization improves the availability and reduces the time
complexity of all running processes (Fig. 3).
•The user has data and wants to keep it on the cloud. First, it sends the verification
request which is verified by Authentication Server (AS) and allows the genuine
user to store the data. AS applied two factors, namely username–password and
OTP, to authenticate the user. CAPTCHA is also used to prevent the cloud from
bots attack.
•At the user side, encipher the original data by utilizing AES approach then use
an asymmetric approach like RSA or ECC to encipher the AES key. This process
is handled by the Cryptographic Server (CS). The user also generates the tag of
enciphered data using SHA-256.
•The user sends the data in an enciphered form through the Internet to the cloud
server, and keeps their data in the cloud.
•When the user wants to access the data from the cloud, then first it is verified
by the authentication server through CAPTCHA and the above two verification
factors. After successful authentication, the user accesses the enciphered data and
generates the tag again using SHA-256, and matches both tags. If tags are same,
then accept the data otherwise reject.
•At the user side, first, decipher the AES key by applying the asymmetric approach
and with the help of AES key interpret the original data.
•Strong password prevents dictionary attacks. OTP secures the system from
Keylogger and replay attacks. CAPTCHA averts from brute-force attack.
•Security algorithms are having two major issues in cloud, where one is time
complexity and the other is security. Time complexities are depending on server
and allotted resources at cloud, so efficient utilization of resources at cloud directly
improves the other issues like availability, time complexity of all running algo-
rithms including security algorithm, and access time. Nowadays, optimization
methods also play an imperative role in recourse allocation, server selection at
the cloud side. Bio-inspired optimization techniques like Cuckoo search, Honey
bee, Particle swarm, Bat, Flower pollination algorithms are widely researched in
the last few years for resource scheduling. Optimization algorithm like CMA-ES
is also giving good results to find the quality cipher in case of DES symmetric
encryption. Avalanche effect is applying to calculate the avalanche value which
424 M. Tyagi et al.
Fig. 3 Proposed framework with hybrid approach
measures the quality of cipher-text, where higher value presents the higher strength
of the cipher. Any change in plaintext or key generates a lot of change in cipher-
text. CMA-ES produces the many cipher-texts for the same data by applying a
little bit change in key, then find avalanche value. In this way, find the symmetric
key which provides the quality cipher with high avalanche value by CMA-ES
optimization [19]. So optimization improves efficiency by resource utilization as
well as enhances the security by providing an excellent key.
Implementation of Cryptographic Approaches in Proposed … 425
5 Conclusion
In the modern era, cloud computing services are top-rated and used by individ-
uals as well as organizations. A lot of users keep their data on the cloud which are
situated at a remote location. It creates issues regarding trust and security, which
are an essential requirement for cloud acceptability. This paper implements various
symmetric approaches in the cloud environment and proposes a framework including
CAPTCHA, TFA, separate computing server, and hybrid encryption. CAPTCHA
protects from bots, TFA applying two factors username-password and OTP to confirm
the genuine user. Two different computing-servers help to reduce the overhead
of computing at cloud as well as enhancing security aspects, hybrid encryption
combined symmetric approach AES with asymmetric approach RSA or ECC for
better security and speed. Bio-inspired optimization Cuckoo search approaches are
suggested for better resource utilization and safe server selection. CMA-ES is a good
option to optimize the ciphertext in case of DES symmetric encryption so, in this way,
this paper gives better implementation of various cryptographic methods in secure
framework for achieving privacy, confidentiality, and integrity in cloud. Moreover,
it protects from bots attack and reduces the burden of the server.
References
1. D. Zissis, D. Lekkas, Addressing Cloud Computing Security issues. J. Future Generation
Comput. Syst. Elsevier 28, 583–592 (2012)
2. S. Singh, Y.S. Jeong, J.H. Park, A survey on cloud computing security: issues, threats, and
solutions. J. Netw. Comput. Appl. (2016)
3. N. Roshanbin, J. Miller, A survey and analysis of current captcha approaches. J. Web Eng.
12(1&2), 001–040 (2013)
4. A. Ometov, S. Bezzateev, N. Makitalo, S. Andreev, T. Mikkonen, Y. Koucheryavy, Multi-factor
authentication: a survey. J. Cryptography, MDPI 1–31 (2018)
5. D. Coppersmith, The data encryption standard and its strength against attacks. J. IBM J. Res.
Develop. 38(3) (May, 1994)
6. D. Coppersmith, D.B. Johnson, S.M. Matyas, A proposed mode for triple-DES encryption. J.
IBMJ.Res.Develop.40(2) (March, 1996)
7. S. Almuhammadi, I. Al-Hejri, A comparative analysis of AES common modes of operation, in
30th Canadian Conference on Electrical and Computer Engineering (CCECE), IEEE (2017)
8. B. Schneier, J. Kelsey, D. Whiting, D. Wagner, C. Hall, N. Ferguson, Twofish: A 128-Bit Block
Cipher (1998)
9. E. Biham, O. Dunkelman, N. Keller, A. Shamir, New attacks on idea with at least 6 rounds. J.
Cryptology (2013)
10. H.M. Elkamchouchi, A.E. Takieldeen, M.A. Shawky, A modified serpent based algorithm for
image encryption, in 35th National Radio Science Conference, IEEE (2018)
11. F. Alaa Kadhim, G.H. Abdul-Majeed, Enhancement CAST block algorithm to encrypt big
data. in Annual Conference on New Trends in Information and Communications Technology
Applications-(NTICT’2017) (2017)
12. A. Mousa, A. Hamad, Evaluation of the RC4 algorithm for data encryption. Int. J. Comput.
Sci. Appl. 3(2) (2006)
13. R.L. Rivest, M.J.B. Robshaw, R. Sidney, Y.L. Yin, The RC6 Block Cipher, Version 1.1 (1998)
426 M. Tyagi et al.
14. A. Karakra, A. Alsadeh, A-RSA: augmented RSA, in SAI Computing Conference 2016, IEEE
(2016)
15. V.G. Martinez, L. Hernandez, A.Q. Dios, Security and practical considerations when imple-
menting the elliptic curve integrated encryption scheme. Taylor Francis 39, 244–269 (2015)
16. A.A. Alkandari, I.F. Al-Shaikhli, M.A. Alahmad, Cryptographic hash function: A high level
view, in International Conference on Informatics and Creative Multimedia, IEEE (2013)
17. F.F. Moghaddam, S.G. Moghaddam, S. Rouzbeh, S.K. Araghi, N.M. Alibeigi, S.D. Varnos-
faderani, A Scalable and Efficient User Authentication Scheme for Cloud Computing
Environments, IEEE (2014)
18. X. Li, J. Chen, D. Qin, W.Wan, Research and Realization based on hybrid encryption algorithm
of improved AES and ECC (ICALIP, IEEE, 2010)
19. M. Tyagi, M. Manoria, B. Mishra, A framework for data storage security in cloud, in Advances
in Data and Information Sciences, Lecture Notes in Networks and Systems (2017)
Home Automation With NoSQL
and Node-RED Through Message
Queuing Telemetry Transport
Naman Chauhan and Medhavi Malik
Abstract In the last decades, home automation technology has evolved and achieved
popularity. This paper discusses the current and emerging home automation technolo-
gies that will make life easier and swifter. Most automation systems use a smartphone
and a microcontroller. Current systems are mostly dependent on local Wi-Fi or Blue-
tooth connections for the transfer of data from the user to the microcontroller. In this
paper, wireless communication through mobile data or the Internet and controlling
the data through a centralized and secure NoSQL database on Google Firebase are
studied. The user will be connected to the system using a mobile application built for
both Android and iOS. Moreover, the advantages and drawbacks of different home
automation systems are discussed.
Keywords Home automation systems ·Firebase ·NoSQL ·MQTT ·Wi-Fi ·
Raspberry pi
1 Introduction
Today the technology is progressing rapidly to make the life of mankind easier,
efficient, and secure. The home automation system not just provides automated access
to the devices and appliances of your home to make your day easy but it also saves you
valuable time. Automation of home appliances can also be taken into consideration
for timed control of appliances for saving electricity. But with the advantages come
the drawbacks. Many systems of home automation are localized and thus cannot
be remotely accessed. Those systems which are able to provide you remote access
N. Chauhan ·M. Malik (B)
Department of Computer Science and Engineering, SRMIST, Delhi-NCR Campus, Ghaziabad,
Uttar Pradesh, India
e-mail: medhavimalik28@gmail.com
N. Chauhan
e-mail: chauhannaman98@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_37
427
428 N. Chauhan and M. Malik
come with a proprietary, dedicated device which act as the control center and often
they have low-level security for the database or high latency in toggling the devices
remotely.
This paper will discuss the various technologies which have been used to overcome
the snags in the previous systems. This home automation system has a cloud-based
NoSQL (Not only SQL) service which is being provided by Google Firebase. The
database can be accessed through the mobile application signed-in with the Google
account. This provides a personalized and secure way to store and access. Each
home has a dedicated server that shall handle all the requests and updates. This can
be a Raspberry Pi or any other minicomputer supporting IEEE 802.11 connectivity
[1]. Most of the automation uses the HTTPS protocol for communication but this
paper will discuss integrating the lightweight Message Queuing Telemetry Trans-
port (MQTT) for data transfer. The microcontroller being used is the ESP8266 by
Espressif Systems running on NodeMCU firmware. This further controls the relays
to switch on or off the appliances.
Different home automation systems are developed by different authors to auto-
matically on and off the appliances. We will discuss the design and development
of activation and control of the home automation system via the Android or iOS
application using a microcontroller. It predominantly concentrates on the control of
home appliances remotely when the person is far from the home.
2 Previous Technologies
2.1 Infrared
The most primitive and cheap method of automating a device is using the IR radiation.
In the electromagnetic spectrum, IR radiation lies between microwaves and visible
light. Consumer electronics like TV, security system, and short-range communica-
tions use this technology. It is simple and cheap but it comes with downsides. It can
only be used for line-of-sight communication and cannot penetrate walls. It can only
be used for short-range [2].
2.2 Radio Frequency
A pair of radio frequency modules (transmitter and receiver) can be used to auto-
mate the home. Common modules used to work on the frequency of 433 MHz or
modules like ZigBee which is a more powerful and long-range device operates on
radio bands of 868, 915, and 2.4 MHz [2]. It creates a wireless mesh network but
it has a disadvantage of having radio interference with other radio services in the
surroundings.
Home Automation with NoSQL and Node-RED through … 429
2.3 Bluetooth
Many consumer electronics are equipped with Bluetooth. At the speed of 1 Mbps,
Bluetooth can transfer data to about 10 m wirelessly [2]. As mobiles have Bluetooth,
this can be used for controlling appliances through mobile easily using an application.
But it has a snag as its operating range is not more than 10 m.
2.4 Global System of Mobile communication (GSM)
By using the cellular network GSM/GPRS module delivers 850/900/1800/1900 MHz
performance for voice, SMS, and data with low power consumption [2]. As cellular
networks have a wide range of coverage; it is possible to control home appliances
from very far off. Offering operations from remote regions and vast coverage, these
are very much reliable technologies for home automation. These modules require a
Subscriber Identification Module (SIM) to connect to a telecommunication service
provider. Also, the telecommunication provider charges rental fees and data charges
for SMS, voice, and data services. Thus, these are comparatively costlier as they
incur charges while operating.
2.5 Wireless Fidelity
Most Wi-Fi devices use 2.4 GHz frequency and implement frequency division multi-
plexing technology. Either personal computer or smartphone can connect with the
home Wi-Fi router which gives access to the local router’s network through which we
can communicate with other devices connected with the network [2]. These systems
have a server that handles the operations which provide better-centralized control as
compared to the above-discussed technologies. Most of the systems made on Wi-Fi
technology are usually localized to the router’s network only. If Internet access is
provided to the network, we can spread the reach of the system. This paper is all
about using this technology for expanding the reach of the automation system.
3 Proposed System
3.1 Overview
The proposed home automation uses the IEEE 802.11 Wireless Fidelity connec-
tivity at 2.4 GHz frequency to establish a connection between the ESP8266 and the
430 N. Chauhan and M. Malik
server which is actually a Raspberry Pi 3B running on Raspbian OS [1]. Debian-
based Raspbian operating system works mostly like an Ubuntu desktop but unlike a
computer, it is as small as a credit card [1,3]. For the server’s backend, Node-RED
developed by IBM is used which has a graphical flow editor. Node-RED makes a
Nodejs server without writing any JavaScript code. Wi-Fi router creates a network
in the home which is to be automated and server and the ESP8266 hubs connect to
it which further controls the relays.
3.2 Architecture
ESP8266 is connected to the server through the Wi-Fi network and communicates
using the Message Queuing Telemetry Transport protocol. It is a machine to machine
extremely lightweight connectivity protocol and has a very small code footprint that
requires less bandwidth and does decreases latency in the device toggling. The client
application is logged into the Google account to which the firebase is connected.
Now, the client application changes the state of an appliance in the app and the
update is sent to the firebase’s real-time database over the mobile Internet or any
Wi-Fi connection. On the server, Node-RED tracks the updates in the database on
firebase [4]. When any update takes place in the database, the payload is injected into
the flow and published to the specific topic via MQTT. The microcontroller which
has subscribed to that specific topic takes the command and changes the relay state
of the one having the required appliance attached to it [5].
Figure 1explains the whole architecture of the proposed system with a sequence
diagram. Server, ESP8266, and relays can be collectively known as a home hub.
There can be many home hubs in a single system and there can be many ESP8266
within a home hub which depends upon the requirements like the number of rooms or
number of appliances. Also, there can be many relays connected to a single ESP8266
depending upon the number of appliances to be controlled.
3.3 Working and Testing
When a button is pressed, the application sends an update to the firebase over the
Internet, be it mobile data, Wi-Fi, or any other means of connecting to the Internet.
Firebase authenticates the update request and then the update to the database. As
soon as the real-time database is updated, Node-RED running on the home server
takes the updates into the flow of the server and injects the payload to the other
node where the payload is published to the topic corresponding to the appliance state
changed via MQTT protocol [5].
When the subscribed ESP8266 gets the payload, it checks for the case which needs
to be executed and then changes the state of the pin required. Pin state defines the state
of the relay attached to it. Relays are the best way to control the appliances which
Home Automation with NoSQL and Node-RED through … 431
Fig. 1 Sequence diagram of the proposed automation system
make a bridge between the high potential alternating current and low potential direct
current. Figure 1having the sequence diagram explains this working in a graphical
mean.
A minimal working example has been tested by the black box testing technique.
Black box testing ensured that the system gave correct and desired output, given
various possible inputs. The average time taken by the system to switch on or off
the appliance when a button is clicked from the client application was calculated
which was further compared with the average time the system took to do the same in
reality while testing. It can be more or less depending upon the speed of the Internet
to which the mobile and server are connected.
4 Conclusion
4.1 Testing and Issue Fixing
While testing it was observed that when the server goes offline due to some power
failure or connectivity issue, database updates from the firebase do not reach the
server. If the buttons are pressed repeatedly from the application while the server
is offline, the firebase updates are not sent to the server but they are stored in a
432 N. Chauhan and M. Malik
queue. When the server comes online, all the updates are sent to the server from the
queue and the server publishes them further to the microcontroller. This leads to the
toggling of appliances when the server comes online. Repetitive switching on or off
appliances can damage them which is not desirable.
This issue can be solved by using the firebase on disconnect node which writes
data when Node-RED disconnects from the firebase servers. When the Node-RED
server goes offline, the value of server status on firebase is modified to false which can
be read by the mobile application. The mobile application can now get to know that
the server is offline and thus, it disables the buttons so that no repetitive toggling is
done. When the server comes online again, Node-RED sets the value of server status
on firebase to true. When the application reads the change, it enables the button for
the operations. Figure 2explains this feature through a sequence diagram.
Also, if the Internet of the system is down for a longer period due to some unfore-
seen reason, the server will no longer be able to get updates from the database.
Without the Internet, the system will not work as no updates can be made to the
database. This issue can be solved by manually overriding the automation system.
Figure 3shows the circuit diagram for the manual overriding for this issue fix.
This can be done by connecting the relays in parallel to the manual switches in the
home. If the relays keep the circuit open, the other parallel connection by switchboard
can close the circuit manually.
4.2 Advantages and Disadvantages
This system provides access to the home automation system from any corner of the
world. No matter how far the client application is sending updates to the firebase, as
long as the mobile phone is connected to the Internet, the home automation system
can be operated. The system maintains a NoSQL database of the home automation
system, unlike other previous systems that are secure with Google authentication.
The communication of data over the Internet is encrypted using Secure Socket Layer
(SSL) encryption. Moreover, the home server to microcontroller communication is
done using the Message Queuing Telemetry Transport which has a very lightweight
and small footprint.
Even though two major issues of this system have been addressed, the system
still has some snags which cannot be fixed. This system is totally dependent on the
Internet which means that if the Internet goes down for some reason, the system
fails to operate and manual overriding is done. Even though it works on Wi-Fi like
previously discussed systems that work on a local network only and does not depend
on the Internet connection, it cannot work in a similar manner. It requires some major
changes to be done on the home server to do so.
Home Automation with NoSQL and Node-RED through … 433
Fig. 2 Sequence diagram of server status updating to fix switch repetition issue
4.3 Future Prospects
The system has been experimentally proven to work with sample appliances attached
to it and the devices were controlled successfully through a wireless mobile phone.
The system still has a few disadvantages like every other system and these can be
fixed in the future.
Features like voice-controlled commands from the application or Google Assis-
tant’s integration through Google actions can also be done to make a personalized
434 N. Chauhan and M. Malik
Fig. 3 Circuit for manual
overriding
chatbot in the Google Assistant. Also, the home automation system can not just be
available for android or iOS devices but it also is accessed through a web client by
hosting a web application that updates the database just like a mobile application
does.
Home automation systems have the scope of several technological advancements
as new technologies are being introduced with time. This will take this system to
homes and it can be commercially available to common masses in the near future.
References
1. R. Prashant, S. Khizaruddin, R. Kotian, S. Lal, Raspberry Pi based home automation using Wi-
Fi, IoT, and android for live monitoring. Int. J. Comput. Sci. Trends and Technol. 5(2), 363–386
(April, 2017)
2. Types of wireless communication technology used in home automation. Electronics For You (19,
April 2018). Available: https://electronicsforu.com/resources/learn-electronics/wireless-techno
logy-types-home-automation. Accessed on 12 October 2019
3. M.P. Sathish, Dr. S.A.K. Jilani, D. Girish Kumar, Home automation through e-mail using
Raspberry Pi. Int. J. Adv. Res. Electron. Commun. Eng. 4(9), 2475–2480 (September, 2015)
4. Node-RED-contrib-firebase, Node-RED (February, 2018). Available: https://flows.nodered.org/
node/node-red-contrib-firebase. Accessed on June 2019
5. Documentation|MQTT v5.0, MQTT Version 5.0 (03, April 2019). Available: https://docs.oasis-
open.org/mqtt/mqtt/v5.0/os/mqtt-v5.0-os.html. Accessed on June 2019
Naïve Bayes Algorithm Based Match
Winner Prediction Model for T20 Cricket
Praffulla Kumar Dubey, Harshit Suri, and Saurabh Gupta
Abstract In the previous statistical analysis of sports, it is very difficult to predict the
winning team of the game/sports before the game starts. Nowadays, it has become
a challenge for all. There are a lot of indoor and outdoor games but Cricket is a
very popular game in every country over the world which is played between two
teams. Each team has 15 players in which only 11 selected players play a match. The
objective of this paper is to predict the winner of a cricket match through various
machine learning algorithms like logistic regression, support vector machines, k-NN,
Naive Bayes, and random forest classifiers applied on the past match performance
records and improve the accuracy of the model.
Keywords Logistic regression ·Naïve Bayes ·SVM ·K-NN ·Random forest
1 Introduction
In the last few years, we have contributed to the success in the field of sports with
the help of statistical modeling which has been used in sports for decades, and has
contributed to the success of sports. There are two most popular outdoor games in
the world, i.e., Cricket and football. Data analytics is easy in cricket as compared
to other games such as baseball, football, and basketball [1]. Various natural factors
influencing the game of cricket are huge media coverage, and a large betting market
has given sturdy incentives to model the game from different views [2]. However,
the complex rules governing a game like cricket, the ability of the team’s players on
P. K. Dubey (B)·H. Suri ·S. Gupta
Computer Science and Engineering, SRM Institute of Science and Technology, Delhi NCR
Campus, Ghaziabad, India
e-mail: praffullakrdubey@gmail.com
H. Suri
e-mail: suriharshit007@gmail.com
S. Gupta
e-mail: saurabh256837@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_38
435
436 P. K. Dubey et al.
a given day of the match, and their performance play an important role in influencing
the final outcome of a cricket match. This presents several challenges in predicting
game outcomes with accuracy. The game of cricket is played in all three formats—
Test match, ODI, and T20, but our research focuses on limited over matches, i.e.,
T20 (IPL) and ODI— the most popular format of cricket. In this paper, we use Naive
Bayes, random forest, and k-NN classifier algorithms for predicting the match’s
winner. The algorithms used in this research are explained in the model below.
1.1 Supervised Learning
The task of machine learning to teach a computer’s function that maps input to an
output based on input–output examples is known as supervised learning. It deduces a
function from the marked training data consisting of various training examples sets.
In this technique, each example is a pair which consists of an object as input object
(generally a vector) and a value of output as desired (also known as supervised signal).
The observed technique analyzes the training data and creates a deduced function
that can be used to map new instances. This scenario allows the learning algorithm to
uniquely identify class labels for unseen instances. The various supervised learning
algorithms are
1.1.1 Linear Regression
If the class and attribute data are numerals, a linear regression classifier is used for
classification. This is the primary statistical method. The approach in linear regression
is to derive the class expressions in terms of a linear combination of pre-established
weights and attributes [3]. It is of two types:
1. Simple Linear Regression
Q=cP +f
where
fis the y-intercept
cis the slope
Qis the dependent variable
Pis the independent variable
2. Multilinear Regression
Naïve Bayes Algorithm Based Match Winner Prediction Model … 437
Y=q0+q1r1+··· +qn+rn
where
q1is the regression coefficient
r1is the independent variable
Y is the dependent variable
1.1.2 K-Nearest Neighbor
In machine learning, the k-nearest neighbor (k-NN) algorithm is a non-parametric
technique that is applied for regression and classification. In these cases, k-closest
training example acts as input. The output depends on whether k-NN is applied for
regression or classification analysis. Distance measurements are used to determine
which examples of Kin the training data set are similar to the new inputs. For real-
valued input variables, Euclidean distance is the most common distance measure [4].
The distance of two data points can also be evaluated using Manhattan, Hamming,
Jaccard, Minkowski, Mahalanobis, and some other distances [1]. The k-NN algorithm
applies the closest neighboring proposal for classification. It takes a set of examples
from the training data that is divided into several groups or categories; training data
set is taken as an input and it is classified into unlabeled test data set examples into
one of those groups or categories. K-NN can also be termed as a sluggish learner
technique. K-NN models are never explicitly formed. This contributes to a faster
development training phase but a slower classification phase [5].
Algorithm
Classify (T, U, g)
for e from 1 to g;
Compute the distance (Ti,g)
End
Calculate set H containing indices for the k smallest distance (Ti,g).
return majority label {Uiwhere I belongs to H}
where training data is specified by T, class labels of T are specified by U, and an
unknown sample is specified by g.
1.1.3 Naive Bayes Classifier
Naive Bayes is one of the classifiers of the Bayesian type that uses predicting as
the likelihood of class membership and assigning a target class to investigate the
data instances. A Bayesian classifier is a statistical classifier that approximates the
probability of a given tuple belonging to a class. Naive Bayes is a curious learner
[1]. Various trends show that Naive Bayes is more accurate in comparison to other
more sophisticated and complex techniques. The Naive Bayes classifier makes an
438 P. K. Dubey et al.
assumption that each attribute possesses a distinct effect on class label, irrespective
of the other attribute values. Hence, this independency is called class-conditional
independence. These classifiers are based on Bayes’ theorem [4].
Bayes Theorem:
Let Ebe a data tuple and Fbe a class label
Let Ebelong to class F, then P(F|E)=P(E|F)P(F)/ P(E)
where
P(F|E) is the posterior probability of class Fgiven predictor E
P(F) is the prior probability of class
P(E|F) is the posterior probability of Egiven the class F
P(E) is the prior probability of the predictor
The classifier evaluates P(F|E) for each class Fi for a given tuple E. Then it will
predict that Ebelongs to the highest posterior probability class, conditioned on E.It
means Eis affiliated to the class Fi if and only if P(Fi|E)>P(Fj|E)for1≤j≤x,j
= i.
1.1.4 Random Forest Classifier
Random forest is an extremely powerful algorithm in machine learning and is an
ensemble method. Random forest classifiers create a set of discordant or inconsistent
trees that are disjoint and that choose the best class for creating a random forest [6].
Each of these trees has the same chance of sampling and is drawn randomly from a
set of possible trees. The random trees can be effectively created, by combining with
a collection of trees that leads to accurate models. The algorithm used is
1. Bagging
The random forest training algorithm applies a general technique of bootstrap
aggregation or bagging to tree learners.
With the given training set and response set bagging repeatedly (Btimes) selects
a random sample with replacement of the training set and fits into the samples:
For b=1, …, B
Sample, with replacement of the training examples
Train the regression or obtained classification tree
After training, predictions for unseen samples can be made by taking the average
of all the predictions from all the individual regression trees or by taking the majority
when classification trees are used.
1.1.5 Support Vector Machines
Isabelle Guyon, Bernhard Boser, and Vladimír Vapnik introduced the thorough idea
of support vectors in their work. SVMs are less prone to overfitting and are highly
Naïve Bayes Algorithm Based Match Winner Prediction Model … 439
accurate. Using non-linear mapping, SVM converts real data into a data which
contains data with high dimension. SVM is used for both classification and numer-
ical prediction. Then it looks for a hyperplane that is linearly optimal in this new
dimension that separates the tuples of different classes with each other. The tuples
belonging to the two classes can always be differentiated by a hyperplane when
properly mapped to a sufficiently high dimension. Using the support vectors and
the margins defined by the support vectors, the algorithm obtains this hyperplane.
The support vectors found by the algorithm gives a useful description of the learned
prediction model. The separating hyperplane is expressed in the following form: Q·
A+b=0, where Qdenotes the weight vector, Q={Q1,Q2,Q3,…,Qn}, the number
of attributes is specified by nand the scalar, often referred to as bias is specified by
b[4].
If we provide input with two attributes P1 and P2, and the training tuples are two
dimensional, (e.g., A=(A1,A2)), where A1and A2are the values of attributes P1
and P2, respectively. Then, the points occurring above the separated hyperplane are
classified to be belonging to the
Class P1:
Q·A+b>0
And any points occurring bottom to this separating hyperplane can be classified
to be belonging to the
Class P2:
Q·A+b<0
1.1.6 Logistic Regression
It is widely applied for problems in classification. It can be used to manage different
types of relationships as it uses a non-linear logarithmic transformation for the odd
ratio prediction. It is not necessary for logistic regression to have a linear relationship
between variables that are dependent and independent. In order to avoid under-fitting
and overfitting, we must not fail to include all the variables that are important. A
better technique for ensuring such practice is to estimate logistic regression by using
methods like gradual method. This leads to the requirement of big sized samples
as maximum possible estimates are least efficient than normal at least square at
low sized samples. There is no correlation between independent variables, i.e., no
multiplicity. However, the interaction effects of categorical variables can be included
in the analysis and model. If the value of the dependent variable is ordinary, it is
referred to as ordinal logical regression. If a dependent variable relies on multiple
classes, it is termed as multinomial logical regression [7].
The simple form is
440 P. K. Dubey et al.
L(q)=1/1+e−q
where
L(q) is the variable that increases with time and is a fraction between 0 and 1.
Variable qstands for time.
1.2 Unsupervised Learning
Unsupervised learning can be defined as a technique of machine learning where you
do not have to oversee the model. In fact, we must allow the model to process on its
own to predict information. It mainly deals with unlabeled data. These algorithms
allow us to execute more sophisticated processing tasks in comparison with super-
vised learning. Compared to other natural learning methods, unsupervised learning
can be highly unpredictable [8].
1.2.1 Partition Algorithms
The partition clustering algorithm gets one partition of data in place of structures as
clustering structures, such as a dendrogram created by techniques of hierarchy. The
problem accompanying in using a partial algorithm is the selection of the number of
output clusters that are required. These techniques have benefits in implementations
involving various big sets of the data where dendrogram construction is computa-
tionally prohibited. Divided or partitioned techniques typically create clusters by
optimizing the criterion function defined either globally (defined overall significant
patterns) or locally (on a small subset of patterns). The combinatorial search for a set
of possible labeling for the ideal criterion value is plainly prohibited from the calcula-
tion point of view. Therefore, in use, the partition algorithms are usually implemented
by running them multiple times with various initial states that are different from each
other, and after various runs, the most significant configuration that is obtained will
be final output clustering [9].
1.2.2 K-Means Algorithm
The K-means algorithm, possibly one of the firstly proposed cluster algorithms,
which has its base on a very basic and simple notion, provided the initial set of clusters,
assigning each point to one of them available, and then each cluster center is replaced
by a midpoint on the corresponding cluster. These 2 basic steps are repeated until
convergence. The cluster is assigned a point that is close in the Euclidean distance to
the point. Even though K-means are preferred as they have the benefit of being easily
put into various operations, it has 2 major disadvantages. The first one is that it can be
Naïve Bayes Algorithm Based Match Winner Prediction Model … 441
proved to be one of the extremely slow algorithms because at each step the distance
among each point and each cluster must be evaluated, which can be highly expensive
in the existence of a large data set. Secondly, the technique is indeed sensitive to
the initial clusters provided. But in recent years this trouble has been solved with
a certain degree of success [9]. At last, the algorithm is aimed at optimizing the
objective function which is known as the squared error function, and is given by
J(O)=(i=1to c)(j=1to ci)
ri−oj
2
here,
‘||ri−oj||’ is the Euclidean distance between xiand vj.
‘ci’ is the number of data points in ith cluster.
‘c’ is the number of cluster centers.
Steps for k-means clustering
•Let R={r1,r2,…,rn} be the sets of data points and O={o1,o2,…,on}bethe
set of centers
•Randomly select the cluster centers ‘c’
•Calculate the distance among each point of data and centers of clusters and assign
data points to cluster center whose distance is minimum and recalculate the new
center for the cluster by the equation:
Oi=(1/ci)(j=1to ci)Oi
•Now calculate the new distance between the data points and new center
•If no data point was reassigned the halt, else repeat the procedure [10].
1.2.3 Hierarchical Clustering Algorithms
An efficient algorithm of such kind is hierarchical clustering; it is applied in the
famous arithmetic-based software like MATLAB. It is an algorithm which belongs
to agglomerative algorithm having few changes based on the metric used to determine
the distances of clusters. For single points Euclidean distance is used. No criteria are
known as to which clustering distance should be used, and it seems to be sturdily
reliant on the data set [9].
1.3 Reinforcement Learning Algorithm
In reinforcement learning algorithm, the evaluation of learner’s views is used to set
the standards of correctness in the service of behavioral goals [11].
442 P. K. Dubey et al.
2 Methodology
Data science is a technique of processing and analyzing data, i.e., applied for a
precise intention. This technique is also beneficial in different project performance
predictions, i.e., useful in business analytics [12]. It contains six steps that are
•Data Import: we have obtained the relevant data set for our project, i.e.,
‘matches.csv’. It is a csv file and it contains six hundred and thirty-six rows
and eighteen columns. This data set has been imported and used for the analysis
as well as prediction.
•Data Cleaning: Data cleaning is a very time-consuming process. Most of the data
comes with missing parameters or duplicate values so it becomes important to
clean our data so that our data set contains only the required data. In our data set,
we found many columns that were useless as they were not required for prediction
so we dropped them and our data set also contained many columns that had null
values in them so we dropped them also. Many values were redundant and some
of the columns in the data set had no significance in the prediction process so
those columns were dropped during the cleaning process.
•Data Munging/Wrangling: It is the process of transforming and mapping data
from raw format into other forms of data that is more appropriate and valuable
for a variety of purposes such as analysis and visualization of the data present in
the data set [13].
•Data Visualization: In this, we try to show the various analysis of the data present
in the data set through various graphs or patterns. The analysis is shown with the
help of various graphs present in mat plot library of python.
•Data Modeling: The process of creating a descriptive diagram illustrating the
relationship between different types of information stored in a database is called
data models. These diagrams can be graphs which include joint plot, histogram,
scatter plot, bar-plot, etc.
•Cross-Validation: The process of training one set of data and testing it using
other data set is called cross-validation. We have used Naïve Bayes to train our
model and test our model to do prediction.
3 Proposed Algorithm
Step 1: Collect match Statistics of various countries T20 cricket (RD1,RD
2,….
RDn) where RD =Resource Data
Step 2: Data Preprocessing
1. Cleaning of data, i.e., clean of data
2. Normalization, i.e., normalize the values
3. Transformation
4. Integration of the data
Naïve Bayes Algorithm Based Match Winner Prediction Model … 443
Step 3: The data is to be split into Training data and Test data for training the
model
Step 4: Create a model using Naïve Bayes Algorithm
model =sc.Naive Bayes(x_train)
Step 5: Apply Test data on model
test =sc.fit_transform(test)
model.fit(x_train,y_train)
prediction =model.predict(x_test)
Step 6: Predictive outcome
1. Knowledge Discovery in terms of various graphs like histogram, scatter plot,
line graph, etc.
2. Performance accuracy of the model
3. Score prediction in terms of overs and score of the team.
4 Result
In our project, we have applied support vector machine, random forest, logistic
regression, k-NN, and Naïve Bayes algorithm and after training our model on the
data set we found that the accuracy of our model is approximate 71% for Naïve Bayes
Algorithm, 66% for k-NN, 64% for logistic regression, 63% for random forest, and
59% for Support Vector Machine. We found that using Naïve Bayes Algorithm we
get the maximum accuracy for our model therefore we can conclude that using Naïve
Bayes Algorithm we get more accurate results as compared to other algorithms used.
Figure 2shows the most succesful IPL team till the year 2017 (Figs. 2,3, and 4).
5 Conclusion and Future Scope
The important limitation in carrying out this project was the limited data set, which
I had at my end. The next step in the direction to improve the accuracy of prediction
problems at hand would be to test out the approaches and various techniques proposed
in this paper using a larger and more representative data set. Also, I would like
to extend the features like venue, pitch condition, and weather condition. If Deep
Neural network (Tensorflow, Thaeno, and Keras) comes into the implementation,
the accuracy will be even higher. Also, a similar model can be developed for other
sports like Football, Kabaddi, Olympics, Asian Games, and so on. We can further
increase the accuracy of our model by using unsupervised learning and reinforcement
learning.
444 P. K. Dubey et al.
Fig. 1 Methodology
Fig. 2 The most successful IPL team till 2017
Naïve Bayes Algorithm Based Match Winner Prediction Model … 445
Fig. 3 Performance of the model trained
Fig. 4 Accuracy of various classifiers applied
References
1. V. Purkayastha, Predicting outcome of cricket matches using classification learners. MSc
Research Project, 2017
2. M.G. Jhawar, V. Pudi, Predicting the outcome of ODI cricket matches: a team composition
based approach, in European Conference on Machine Learning and Practice of Knowledge
Discovery in Databases (ECML-PKDD 2016). Report No: IIIT/TR/2016-1
446 P. K. Dubey et al.
3. T. Singh, V. Singla, P. Bhatia, Score and wining prediction in cricket through data mining, in
International Conference on Soft Computing Techniques and Implementation (ICSCTI) (2015)
4. S. Kumar, S. Roy, Score prediction and player classification model in the game of cricket using
machine learning. Int. J. Sci. Eng. Res. 9(8) (2018). ISSN 2229-5518
5. https://cse.buffalo.edu/~jing/cse601/fa12/materials/classification_methods.pdf
6. H.A. Sulaiman, M.A. Othman, M.F.I. Othman, Y.A. Rahim, N.C. Pee, Advanced Computer and
Communication Engineering Technology: Proceedings of ICOCOE 2015, vol. 362 (Springer,
2015)
7. A. Nimmagadda, N.V. Kalyan, M. Venkatesh, N.N. Sai Teja, C.G. Raju, Cricket score and
wining prediction using data mining. Int. J. Adv. Res. Dev. 3(3) (2018)
8. https://www.guru99.com/unsupervised-machine-learning.html
9. S. Mythili, E. Madhiya, An analysis on clustering algorithm in data minings. Int. J. Comput.
Sci. Mob. Comput. (IJCSMC) 3(1) (2014)
10. https://sites.google.com/site/dataclusteringalgorithms/k-means-clustering-algorithm
11. R. Lokhande, P.M. Chwan, Live cricket score and wining prediction. Int. J. Trend Res. Dev.
5(1) (2018). ISSN: 2394-933
12. https://medium.com/@praffullakrdubey/understanding-data-science-pipeline-and-its-gen
eral-terms-5898c6927707
13. Wikipedia, https://en.wikipedia.org/wiki/Data_wrangling
Design of an Efficient Deep Neural
Network for Multi-level Classification
of Breast Cancer Histology Images
H. S. Laxmisagar and M. C. Hanumantharaju
Abstract Breast cancer is the second most prominent cause of death among women,
and it is a major health problem across the world over the past many years. In this
work, four hundred different biopsy images were collected from the Bio-imaging
2018 breast histology classification challenge by participating Grand Challenge
on breast cancer histology images (BACH) for multiclass breast cancer histology
image classification using deep learning technique. Computer-aided detection or
diagnosis (CAD) system plays an important role in the detection and to increase
death survival rate of women suffering from breast cancer, which can decrease death
rate among women. The main purpose of this paper is to develop a CAD system
which can be used to detect whether the input biopsy image fits one of the four
different classes like benign, in situ, Invasive and Malignant. We have incorporated
an efficient lightweight neural network such as MobileNet2.10ex instead of normal
convolution neural network for feature extraction and at top a fully connected deep
neural network is designed to do classification among different classes. The confu-
sion matrix is developed in order to draw the model accuracy of both techniques. It
is observed that MobleNet2.10ex is giving about 88.92 % training accuracy at 257
epochs and the corresponding loss value is 0.2249. In order to progress the perfor-
mance during pre-processing, a staining technique was used to remove staining from
a digital scan of histology microscopic image of fine-needle aspirate (FNA) slides and
feature invariant approach such as data augmentation technique was used to rotate
each image in pipeline about the axis 45° apart three times in order to increase dataset
and prediction accuracy. The model was trained till 280 epochs to evaluate training
and validation accuracy using Google Colab having strong GPU 2496 CUDA cores,
12 GB GDDR5 VRAM and 12.6 GB RAM.
Keywords Breast cancer ·Biopsy ·CAD ·Benign ·In situ ·Invasive ·FNA ·
MobileNet ·Google Colab ·GPU
H. S. Laxmisagar (B)·M. C. Hanumantharaju
BMS Institute of Technology and Management, Yelahanka, Bangalore 560064, India
e-mail: sagar8.hs@bmsit.in
M. C. Hanumantharaju
e-mail: mchanumantharaju@bmsit.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_40
447
448 H. S. Laxmisagar and M. C. Hanumantharaju
1 Introduction
Cancer continues to be one of the second [1] biggest health risks in all over the world.
There is no doubt that cancer today is a major threat to the society. Breast cancer is a
foremost basis of death and touches millions of lives every year. In primary finding
could benefit to intensification the endurance of several lives in [2] addition to saving
billions of dollars. Hence this area of research has been largely unexplored.
Breast cancer is the greatest vastly destructive cancer disease and a foremost health
[3] problem in women. Breast cancer for women is very serious and common cancer.
Digital Mammography is the preliminary method used to find breast cancer [4,5].In
filtering all the important features to support the clinical disease identification is a
challenging and time-consuming task.
Some of the authors have done a lot of research in the advance of computer-aided
detection (CAD) system. In 2009, an expert [6] system was developed for finding
breast cancer using association rules (AR) and neural network (NN). In 2017, Breast
lump [7] detection and classification in ultrasound images using surface analysis and
super-resolution approaches improves the performance of the system. Computer-
aided detection or diagnosis (CAD) systems [8] play an important role in the earlier
identification of breast cancer which can decrease the death rate among women
suffering from breast cancer In [9], machine learning techniques used to improve the
diagnostic accuracy of breast cancer from nuclear features of fine-needle aspirates.
To speed up the execution process [10] make use of high-performance hardware such
as GPU, CPU and RAM. One of the major problems in the analysis of breast cancer
microscopic image because of H&E stained so its appearance is variability [11]. It
is due to notable differences in light detectors, Optics and light sources used in the
scanners.
2 Breast Cancer Dataset
The microscopic image dataset contains 400 high-resolution (HR) uncompressed
images and each image has 2040 ×1536 pixels. It is ordered of hematoxylin and
eosin (H&E) stained histology microscopic images from the Bio-imaging breast
histology classification challenge 2018. We restrictthe scope of the project to just the
classification of histology microscopy [12,13] images. Each image class is labelled
with one of four categories such as benign, normal, in situ and invasive. The image
class that is labelled was accomplished by two expert pathologists to provide a
diagnostic from the image contents. The desired goal of the Bio-imaging challenge
is to perform classification between different classes of data.
The dataset has 400 different microscopic images, which are distributed as
follows:
1. Normal images: 100
2. Benign images: 100
Design of an Efficient Deep Neural Network … 449
3. In situ carcinoma images: 100
4. Invasive carcinoma images: 100
5. Histological images have RGB colour model with .tiff format
6. Microscopic image size: 2048 ×1536 pixels
7. Size of the pixel (Height ×width): 0.42 µm×0.42 µm
8. Type of label: image-wise.
3 Data Pre-processing
After obtaining the dataset, we pass the images through two steps of pre-processing
namely
1. Normalization
2. Data Augmentation.
Normalization is a pre-processing technique that uses staining technique to remove
staining for the biopsy image slides preparation. The HR images are transformed to
optical density (OD) using logarithmic [14,15] transformation. The singular value
decomposition (SVD) is then applied to the optical density tuples and the obtained
colour space transform is used to the unusual biopsy image. Then the histogram image
[16] is stretched so that it covers the lower 90 % of the data. Figure 1shows some
image samples before and after normalization. The staining allows the pathologist
to identify the type of cancer and its severity in the tissue under the microscope. But
computer does not require staining because using algorithm it can operate on pixel
wise. Computer does not require different colour for understanding.
Data augmentation is nothing but adding similar data to the existing dataset to
further substantiate the model. Further, a dataset is created from the normalized
images to rotate the image in pipeline. To avoid under fit due to low number of
samples, data augmentation technique using image rotation and flipping [16] is incor-
porated without altering the diagnosis and image quality to further increase [17]the
dataset. Once the data has been normalized and augmented, it is fed as input to the
CNN model. The model proceeds to learn the underlying features in the four classes
of images following which, it will be able to make predictions on images without
predefined labels.
Fig. 1 Normalization of images (before v/safter)
450 H. S. Laxmisagar and M. C. Hanumantharaju
(a) Original Image (b) Data augmentation-1 (c) Data augmentation-2
Fig. 2 Image augmentation (rotation about the axis)
Keras has this ImageDataGenerator class which lets the handlers to achieve image
augmentation in a very simple way. The ImageDataGenerator class has three different
approaches flow_from_directory(), flow_from_dataframe() and flow() to read the
images from folder having images and a big NumPy array.
The dataset whatever we are using is very small for our deep neural network
for training the model. To avoid it, we used different data augmentations on pre-
processed histology images to rotate about the axis, cropping and flipping of the
image in order to develop robust model and to prevent overfitting problem. Figure 2
depicts data augmentation of original image with respect to axis.
4 Training and Testing Phase of Proposed CAD System
Figures 3and 4represent the two phases such as training and testing phase of the
proposed CAD system. During pre-processing about 400 h (High-resolution) images
were converted to low-resolution (LR) images using image normalization. In that 352
LR images were used for training and remaining 48 LR images were used for testing.
Fig. 3 The training stage of the proposed CAD system
Design of an Efficient Deep Neural Network … 451
Fig. 4 The testing stage of the proposed CAD system
During training phase about 88 LR images were allocated for each class such as
benign, in situ, Invasive and Normal. During testing phase about 12 LR images were
reserved for each of the above class of data. Observe that during training phase the
class labels (benign, in situ, Invasive and Normal) of the input LR image sequences
are known. The difference between testing and training phase is that during testing
the known class label of the LR input image is fed into the trained model after
pre-processing to detect whether the patient belongs to one of the four categories.
4.1 Framework of Histology Image Classification
To perform convolution operation a MobileNet2.10ex convolutional neural network
approach is used, which is depth-wise separable convolution that includes two layers
one layer is used to perform depth-wise convolutions and another layer is used for
point-wise convolutions. At the output of each neuron, a ReLU activation function
is used to avoid dying of neuron. We can use depth-wise convolution layer to apply a
single filter for each input depth channel and a 1 ×1 convolution for [18] combining.
This process is continued and feature matrix is extracted which contains labelled data.
The feature data is given as input to deep neural network which contains four hidden
layers to perform classification between four different class of data (Fig. 5).
4.2 Architecture of MobileNet
The architecture is shown in Fig. 6. The depthwise separable convolutions divide
standard convolution into a depthwise convolution and a 1×1 pointwise convolution
[19]. The standard convolutional filters (Fig. 6a) are replaced by two layers first
one is depthwise convolution (Fig. 6b) and second one is pointwise convolution
(Fig. 6c). The depth-wise separable convolution uses two layers in that one layer
is used to perform filtering and another layer is used for combining. Computation
cost and the size of model can be reduced by using this factorization technique. The
Design of an Efficient Deep Neural Network … 453
MobileNet architecture is a lightweight convolutional neural network. It uses depth-
wise separable convolutions as it does a single convolution on each colour channel
instead of combining all three channels such as RGB and flattening it. MobileNet [20]
is a stack of the separable convolution modules which are composed of depth-wise
convolution and a 1 ×1 point-wise convolution.
Suppose in Fig. 6, we have five channels, then it is almost five DK×DKspatial
convolution. Point-wise convolution is the 1 ×1 convolution to change the dimension.
Operation cost =Depth-wise Separable Convolution Cost: Depth-wise +Point-
wise Convolution Cost
DK·DK·M·DF·DF+M·N·DF·DF
where ‘N’ represents the output channels, ‘DK’ and ‘DF’ indicates Kernel size and
Feature map size, ‘M’ represents the input channels.
The Standard Convolution Cost is:
DK·DK·M·N·DF·DF
Thus, the computation reduction is:
Depth-wise Separable Convolution Cost/Standard Convolution Cost
When DK×DKis 3 ×3, 8–9 times less computation can be achieved, but with
only small reduction in accuracy.
Depth-wise convolution can be preferred to decrease the complexity and size of
the model. The number of multiplications required is very large using conventional
convolution method. Suppose there are 512 5 ×5×3 kernels that change 8 ×8
times, it results in 512 ×3×5×5×8×8=2,457,600 multiplications.
In the depth-wise convolution, we have three 5 ×5×1 kernels that move 8 ×8
times. That’s 3 ×5×5×8×8=4800 multiplications. In the point-wise convolution,
suppose we have 512 1 ×1×3 kernels that transfer 8 ×8 times, i.e 512 ×1×
1×3×8×8=98,304 multiplications. Now we need to add them together, that’s
about 103,104 multiplications. It is observed that 103,104 is very less than 2,457,600.
With less multiplications, the deep neural network is able to process more data in a
shorter amount of time. In deep neural network, the speed and power consumption
depend on number of multiplications and accumulates unit, which directly measures
the number of multiplication and addition operations.
4.3 Model Training and Validation
4.3.1 Model Specifications
•Architecture—MobileNet2.10ex
•Epochs trained for—280
•Classifier used—Fully connected Neural network
454 H. S. Laxmisagar and M. C. Hanumantharaju
Fig. 7 Transfer learning strategies [21]
•Activation function—Soft Max
•Other layers—ReLU
•Loss function—Categorical Cross Entropy
•Batch Size—32
•Optimizer—Adam.
4.3.2 Transfer Learning
During the process of training the model, the concept of transfer learning is imple-
mented. In computer vision, transfer learning can be expressed using pre-trained
models. Consequently, due to this less multiplication is required for training such
models, it is very good practice to use models such as VGG, Inception, MobileNet
(Fig. 7).
4.4 Various Architectures of Convolution Neural Networks
4.4.1 MobileNetV2 (Apr. 2018)
MobileNetV2 architecture was developed in April 2018 [21] appropriate for
embedded and mobile-based vision applications in which there is deficiency of
computational power.
4.4.2 AlexNet (2012)
In 2012, AlexNet was developed. It is similar to MobileNet [22] architecture but has
more filters in each layer with stacked convolutional layers.
Design of an Efficient Deep Neural Network … 455
4.4.3 VGGNet (2014)
VGGNet was developed in the year 2014, and it is having sixteen convolutional
layers and is very analogous to AlexNet but has lots of filters [23]. Presently, it is one
of the best algorithm choices in the community to do research for extracting different
features from each image.
4.4.4 ResNet (2015)
At last, a new architecture called residual [24] neural network (ResNet) has lower
complexity compared to VGGNet. Using this technique, we can able to train a neural
network model with 152 layers.
5 Experiments and Results
Our dataset consists of total 400 h hematoxylin and eosin (H&E) stained images
categorized as benign, in situ, invasive and normal (total 100 images of each class).
Figure 8depicts the sample of four different classes of images. The required histology
microscopy image dataset is available at http://iciar2018-challenge.grand-challenge.
org/dataset/.
The dataset whatever we are using is very small for our deep neural network
for training the model. To avoid it, we used different data augmentations on pre-
processed histology images to rotate about the axis, cropping and flipping of the
image in order to develop robust model and to prevent overfitting problem.
In total, we have conducted 280 iterations of the model training procedure, with the
objective of obtaining the best possible results with the given technical constraints.
In order to progress the performance during pre-processing a staining technique
was used to remove staining from a digital scan of histology microscopic image
of fine-needle aspirate (FNA) slides and feature invariant approach such as data
augmentation technique was used to rotate each image in pipeline about the axis
45° apart three times in order to increase dataset and prediction accuracy. Below is
(a) Benign (b) In situ (C) Invasive (d) Normal
Fig. 8 Sample histology images of ICIAR 2018 challenge on Breast Cancer Histology [12](BACH)
456 H. S. Laxmisagar and M. C. Hanumantharaju
a graphical summary of the results obtained over the course of the model training
process.
5.1 Model 2.10ex
The model was trained till 280 epochs to evaluate training and validation accuracy
using Google Colab having strong GPU 2496 CUDA cores, 12 GB GDDR5 VRAM
and 12.6 GB RAM. Once the model is trained it is required to plot training and vali-
dation percentage accuracy related to the number of epochs used for deep learning.
Figure 9depicts the graphical illustration of training and validation accuracy which
is moving very close to each other and the gap between the two is minimum (Fig. 10).
Fig. 9 Depicts training and validation accuracy using MobileNet2.10ex
Design of an Efficient Deep Neural Network … 457
Fig. 10 Depicts training and validation loss using MobileNet2.10ex
Sample Code:
def plot_training(hist):
accr = hist.hist['accr']
accr=[ele*100 for ele in accr] # Precentage Value of Accuracy
value_accr = hist.hist['value_accr']
value_accr=[ele*100 for ele in value_accr] # Precentage Value of Acc
uracy
loss = hist.hist['loss']
value_loss = hist.hist['value_loss']
epochs = mention range (len(accr))
plt.grid(visible=True)
plt.plot(iteration, accr, 'g',label='Accr on Training Samples')
plt.plot(iteration, value_accr, 'b',label='Accr on Validation Samples')
plt.legend(loc='best')
plt.xlabel('Epochs')
plt.ylabel('Accuracy %')
plt.title('AccuracyTraining and Validation (MobileNet 2.10ex)')
plt.savefig('Training_and_ValidationAccuracy (MobileNet 2.10ex)
.png')
plt.figure()
plt.grid(visible=True)
plt.plot(iteration, loss, 'r',label='Loss on Training Samples')
plt.plot(iteration, val_loss, 'y',label='Loss on Validation Samples')
plt.xlabel('Iteration')
plt.legend(loc='best')
plt.title('Training and Validation loss (MobileNet 2.10ex)')
plt.savefig('Training_and_ValidationLoss (MobileNet 2.10ex).png'
)
plot_training(ModelEvolution)
6 Conclusion
The major work done in this paper is to do the classification between four different
classes of data such as normal, benign, in situ and invasive using his to patho-
logical images. The dataset, obtained from the ICIAR-BACH challenge 2018, was
augmented to create a more versatile dataset to feed into the convolutional neural
458 H. S. Laxmisagar and M. C. Hanumantharaju
network. In the feature extraction step, the deep CNN architecture of MobileNet is
used. The MobileNet was trained to differentiate among four classes of data and its
parameters were altered to classify microscopic medical images. The creation and
exploration of various architecture versions produced a high performing architecture
with a calculated accuracy of 87.5 %.
This version of the MobileNet architecture is very promising not only due to the
high accuracy obtained but also due to its resource-efficient architecture. This low-
cost model works out to be an exciting opportunity for industry-related projects in
the arena of breast cancer detection. In future, we can implement this lightweight
model using FPGA. For future scope, other architecture of deep neural networks
will be recommended to evaluate performance characteristics which include visual
geometry group (VggNet) and the residual network (ResNet) architecture.
References
1. D. West, P. Mangiameli, R. Rampal, V. West, Ensemble strategies for a medical diagnosis
decision support system: a breast cancer diagnosis application. Eur. J. Open. Res. 162, 532–551
(2005)
2. Deep learning for cancer diagnosis. https://software.intel.com/en-us/articles/deep-learning-for-
cancer-diagnosis-a-bright-future
3. D. Sun, M. Wang, A. Li, A multimodal deep neural network for human breast cancer prognosis
prediction by integrating multi-dimensional data (2018), pp. 1–10
4. S.-M. Choua, T.-S. Leeb, Y.E. Shaoc, I.-F. Chenb, Mining the breast cancer pattern using
artificial neural networks and multivariate adaptive regression splines. Exp. Syst. Appl. 27,
133–142 (2004)
5. B. Zheng, S.W. Yoon, S.S. Lam, Breast cancer diagnosis based on feature extraction using a
hybrid of K-means and support vector machine algorithms. Exp. Syst. Appl. 41, 1476–1482
(2013)
6. M. Karabatak, M.C. Ince, An expert system for detection of breast cancer based on association
rules and neural network. Exp. Syst. Appl. 36(2 Part 2), 3465–3469 (2009)
7. M. Abdel-Nasser, J. Melendez, A. Moreno, O.A. Omer, D. Puig, Breast tumor classification
in ultrasound images using texture analysis and super-resolution methods, in Engineering
Applications of Artificial Intelligence, Dec 2016 (2017), pp. 84–92
8. J. Tang, R.M. Rangayyan, J. Xu, I.E. El Naqa, Y. Yang, Computer-aided detection and diagnosis
of breast cancer with mammography: recent advances. IEEE Trans. Inf Technol. Biomed. 13(2),
236–251 (2009)
9. W.H. Wolberg, W.N. Street, O.L. Mangasarian, Machine learning techniques to diagnose breast
cancer from fine needle aspirates. Cancer Lett 77(2–3), 163–171 (1994)
10. P. Rajendra, A. Subbarao, G. Ramu, V. Brahmajirao, Prediction of drug solubility on parallel
computing architecture by support vector machines. Network Model Anal Health Inf Bioinform
7(1), 130 (2018). https://doi.org/10.1007/s13721-018-0174-0
11. M. Veta et al., Breast cancer histopathology image analysis: a review. IEEE Trans. Biomed.
Eng. 61(5), 1400–1411 (2014)
12. ICIAR Grand Challenge 2018 on Breast Cancer Histology Images. https://iciar2018-challenge.
grand-challenge.org/dataset/
13. P. Athology, A. Sethi, Fast GPU-enabled color normalization for digital pathology (2019)
14. X. Li, S. Member, K.N. Plataniotis. A complete color normalization approach to histo-
pathology images using color cues computed from saturation-weighted statistics, Feb
2015.https://doi.org/10.1109/TBME.2015.2405791
Design of an Efficient Deep Neural Network … 459
15. M. Macenko, M. Niethammer, J.S. Marron, D. Borland, J.T. Woosley,X. Guan„ C. Schmitt, N.E.
Thomas, A method for normalizing histology slides for quantitative analysis. 3 Statistics and
Operations/Research, 4 Lineberger Comprehensive Cancer Center, 5 Renaissance Computing
Institute, 6 Pathology and Laboratory Medicine, 7 Dermatology University of Nor. 1107–1110
(2009)
16. Y. Roohani, E. Kiss, Improving accuracy of nuclei segmentation by reducing histological image
variability (2018)
17. M.D. Bloice, C. Stocker, A. Holzinger, Augmentor: an image augmentation library for machine
learning (2017). arXiv preprint arXiv:1708.04680
18. D. Scherer, M. Andreas, S. Behnke, Evaluation of pooling operations in convolutional
architectures for object recognition, Sept 2010
19. Transfer learning using MobileNet and Keras. https://towardsdatascience.com/transfer-lea
rning-using-mobilenet-and-keras-c75daf7ff299. Last access 2018
20. A.G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H.
Adam, MobileNets: efficient convolutional neural networks for mobile vision applications
(2017). arXiv:1704.04861
21. M. Sandler, M. Zhu, A. Zhmoginov, C.V. Mar, MobileNetV2: inverted residuals and linear
bottlenecks. Transfer learning from pre-trained models: https://towardsdatascience.com/tra
nsfer-learning-from-pre-trained-models-f2393f124751.
22. Z. Alom, et al., The history began from AlexNet : a comprehensive survey on deep learnin
approaches
23. S. Liu, W. Deng, Very deep convolutional neural network based image classification using
small training sample size (2015)
24. Y. Chen, Q. Zhang, Y. Wu, B. Liu, M. Wang, Y. Lin, Fine-Tuning ResNet for Breast Cancer
Classification from Mammography (Springer, Singapore, 2019)
Autonomous and Adaptive Learning
Architecture Framework for Smart Cities
Saravanan Muthaiyah and Thein Oak Kyaw Zaw
Abstract The context of smart cities should really be anchored onto two key
attributes. First is the ability for a city to learn adaptively with the aid of machine
learning (ML) or artificial intelligence (AI) and second is the ability for a city to
sustain operations autonomously without any human intervention. While Internet
of Things (IoT) is seen as an enabler by making all devices connect to a network
that communicates with one another with minimal human interference; however,
critical problems such as sewer management, health, parking woes, traffic conges-
tion, pollution, waste management, and noise are not fully being addressed. In this
paper, we discussed the most recent literature for smart city initiatives across the
globe including the comprehensive Alcatel–Lucent study in 2011 and proposed an
overarching autonomous learning city baseline and target architecture with specific
functionalities for each layer.
Keywords IoT ·AI ·Machine learning ·Adaptive ·Autonomous ·Architecture ·
Smart city
1 Introduction
IoT has been drawing in enormous volumes of research as the world progresses
toward making everything connected. It is a future vision that everyday devices
will be equipped with microcontrollers, transceivers, and protocol stacks that will
communicate with each other making it integrated with the Internet [1]. Connected
devices will enable easier access to the sensors, actuators, and so on, which will create
many applications that will make life for the people easier. Currently, there are a wide
range of applications of IoT in the market for different areas such as medical aids,
S. Muthaiyah (B)·T. O. K. Zaw
Faculty of Management, Multimedia University, Cyberjaya, Malaysia
e-mail: saravanan.muthaiyah@mmu.edu.my
T. O. K. Zaw
e-mail: rahman93.iu@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_41
461
462 S. Muthaiyah and T. O. K. Zaw
traffic management, home automation, and many more [2]. However, each of the IoT
systems has its own mechanism and methods in realizing its objectives. This makes
integration for IoT applications into a system or control almost impossible.
Given this complexity, the driving force comes from users, technology partners,
and government to truly realize the Smart City vision [3]. Dameri [4] stated that the
main problem for the vague definition of a smart city is due to a bottom-up problem
where the ideas arise from the technology application itself toward urban problems as
shown by Fig. 1. The author stated that the main driver for the smart city was the birth
and development of the technology, especially ICT, which connects different digital
services. IoT, ML, and AI have the potential for solving urban problems such as crime,
pollution (e.g., carbon footprint), and congestion. Nevertheless, with stakeholders
having their own objectives without a unified vision or direction efforts for Smart
City initiatives seemed to be misaligned. Focus, vision, policy, and rules toward Smart
City that are vague is the main reason for the lack of a comprehensive framework.
The moment we can obtain the unified vision, a comprehensive definition can be
established and urban problems can be solved in a more effective manner toward
realizing a true Smart City. As no comprehensive definition exists for a Smart City,
one could argue that a city may not reach the status of smart even though most of
the devices if not all, are connected. One could also argue that IoT may not be the
answer in realizing it. Therefore, elements to achieve in a Smart City must have solid
guidelines or statements so that directions toward it can be realized. From literature
review and scenario stated above, a clear gap exists for a unified vision and definition
on what a Smart City truly should be and how should it be realized. In this paper,
we intend to narrow the gap by proposing a baseline and target architecture for an
autonomous and adaptive learning Smart City framework.
Fig. 1 Bottom-up Smart City development path
Autonomous and Adaptive Learning Architecture Framework … 463
2 Smart City Definitions
Giffinger et al. [5] stated that “A Smart City as a well-performing city built on the
‘smart’ combination of endowments and activities of self-decisive, independent and
aware citizens.” Caragliu et al. [6] highlighted that a city is smart “when investments
in human and social capital and traditional (transport) and modern (ICT) communi-
cation infrastructure fuel sustainable economic growth and a high quality of life, with
a wise management of natural resources, through participatory governance.” Both
the definitions differ from each other with the former discusses results and the latter
highlights components of achieving it. Rouse [7] on the other hand, stated that “a
smart city is a municipality that uses information and communication technologies
to increase operational efficiency, share information with the public and improve
both the quality of government services and citizen welfare.” It means information,
communication, and technology should be used by the city to increase the oper-
ational efficiency. When the situation happens, a city will be regarded as a smart
city. Again, the definition is obscure without any specific methods, technology, or
measurement to be used to achieve it. Techopedia [8,9] bears the same idea on
the Smart city. Both authors mentioned the usage of ICT in making a city smart to
enhance quality, performance, and address urbanization problems. Perhaps, the most
realistic definition is “a smart city is said to be a well-defined geographical area, in
which technologies such as ICT, logistics, energy and production cooperate to create
benefits for citizens in terms of their well-being” [4]. IoT offers a wide range of
capabilities for a Smart City in the context of solving core urban challenges and also
being very cost-effective at the same time [10]. Listed below are some of the main
points that advocate the need for smart solutions according to a study by Alcatel–
Lucent in 2011.1The study done across 52 cities around the globe indicate common
problems and issues of cities that require smart solutions, they are (1) energy, (2)
telecommunications, (3) traffic congestion, (4) intelligent community, (5) public util-
ities and lastly, (6) industry. In energy adoption, one of the key areas that is being
addressed is the Smart Grid [11]. The major contribution is for an efficient power
grid to support energy requirements for the city in terms of energy conservation
monitoring and quality of service. Telecommunications is needed to support home
automation smart buildings and public infrastructure for data acquisition using smart
sensors. Key areas also include home surveillance, lighting, panic button, and energy
conservation. Although there are research developments, implementation is still low
[12]. Traffic congestion is a norm that happens in cities and one of the proposed
solutions is public transportation. This problem has persisted after the emergence
of cheaper cars. Smart cities should offer a solution by providing monitoring of the
traffic and even controlling signal lights when needed. The monitoring also can be
possible by having vehicles equipped with GPS sensors [13].
Information from monitoring can greatly help authorities to build better and bigger
roads for places that have high traffic. Also, provide an early warning for potential
1https://www.tmcnet.com/tmc/whitepapers/documents/whitepapers/2013/7943-alcatel-lucent-get
ting-smart-smart-cities-recommendations-smart.pdf.
464 S. Muthaiyah and T. O. K. Zaw
traffic burst that could happen so that prevention can be taken earlier. Another aspect
is parking management. Limited parking spaces for cars and motorcycles is a norm
in the city (Table 1).
Smart parking, smart meters, and gazetted parking bays can utilize sensors and
intelligent displays [14]. Therefore, smart parking will be made available for cars and
motorcycles alike as well as reservation of parking bays ahead of time. Intelligent
community framework focuses on adding green spaces, recreation, which includes
noise as a part of a community.
Although, there are laws enforcing in making noise pollution lesser, it is still
rampant especially in city areas where there are many businesses located. Sensors
Tabl e 1 Alcatel–Lucent 2011 Smart City survey
No. City No. City
1Amsterdam (The Netherlands) 27 Malmö (Sweden)
2Ballarat (Australia) 28 Masdar (UAE)
3Besançon (France) 29 Moncton (Canada)
4Birmingham (U.K.) 30 Ottawa (Canada)
5Bottrop (Germany) 31 Paredes (PlanIT Valley, Portugal)
6Bristol (U.S.A.) 32 Pedra Branca (Brazil)
7Cape Town (South Africa) 33 Porto Alegre (Brazil)
8Chattanooga (U.S.A.) 34 Quebec City (Canada)
9 Cleveland (U.S.A.) 35 Recife (Brazil)
10 Copenhagen (Denmark) 36 Riverside (U.S.A.)
11 Curitiba (Brazil) 37 Rotterdam (The Netherlands)
12 Dakota County (U.S.A.) 38 Shanghai (China)
13 Dongtan (China) 39 Shenyang (China)
14 Dublin (Ireland) 40 Songdo (South Korea)
15 Dublin (U.S.A.) 41 Sopron (Hungary)
16 Eindhoven (The Netherlands) 42 Suwon (South Korea)
17 Gdansk (Poland) 43 Tallinn (Estonia)
18 Gold Coast City (Australia) 44 Taoy u a n (Tai w a n)
19 Gujarat international financial tech-city
(GIFT, India)
45 Tianjin Binhai (China)
20 Ipswich (Australia) 46 Toronto (Canada)
21 Issy-les-Moulineaux (France) 47 Trikala (Greece)
22 Jubail (Saudi Arabia) 48 Trondheim (Norway)
23 Kalundborg (Denmark) 49 Urumqi (China)
24 Lavasa (India) 50 Windsor-Essex (Canada)
25 Lyon (France) 51 Winnipeg (Canada)
26 Malaga (Spain) 52 Wuxi (China)
Autonomous and Adaptive Learning Architecture Framework … 465
Tabl e 2 ML/AI techniques
with use cases Machine learning (ML)
techniques
Use cases
Neural network Traffic woes and weather
Regression and clustering Healthcare
K-means nearest neighbor Waste management and home
surveillance
Anomaly detection Weather, pollution, carbon
footprint, and traffic
Support vector machine Weather and pollution
can be deployed for monitoring noise pollution [15]. The information gathered will
be provided to authorities so that timely action can be taken. Residents can report or
request for noise scanning (dB) smart devices if they feel that their area is getting too
noisy. In this area, the community plays a role as well. Public utilities include waste
management and carbon footprint per square foot calculator. Cities create problem
for waste storage, shortage of garbage landfills. Using intelligent containers that are
able to detect the level of waste, collectors can plan their routes more effectively.
This in return will reduce cost for collecting waste and improving the quality of
recycling [16] and also enable household waste tracking. the industry sector focuses
on environment management, which includes air quality. Air in urban areas is heavily
polluted. Smart technologies will be able to help to monitor air quality in crowded
places such as parks and fitness trails [17].
It will be able to monitor factory concentrated areas as well. Companies residing in
the polluted area can be imposed with carbon taxes so that efforts will be undertaken in
making the quality of air better in those places. Public can also be warned of affected
areas beforehand to live an environmentally friendly lifestyle. The technology can
also be used in buildings for rainwater harvesting and creating smart building index
as well. Although a city may differ in problems that it faced in its city, smart solutions
will be able to intelligently adjust to its requirements. However, IoT alone can only
be an enabler, what is needed is an overarching framework which is proposed in the
next section (Table 2).
3 Overarching Framework for Smart Cities
While we agree with the author, our contribution is the overarching framework
that includes specific layers, i.e., (1) Acquisition, (2) Data, (3) Business, and (4)
Application architectures that make a city smart. Therefore, after connecting all the
dots, we propose the framework (see Fig. 2) with emphasis on the framework being
autonomous and self-learning at the Business Architecture layer. These two items
were proposed because Smart should consist of the ability to reason and with very
little or no human input (i.e., independent or autonomous). Ability to think on its own
466 S. Muthaiyah and T. O. K. Zaw
Fig. 2 Smart City ML/AI use cases for smart solution
will require learning algorithms and the actions must be autonomous. Therefore, it is
clear that in order to be called smart, a system or device needs to be able to achieve
two things: (1) ability to think independently in difficult situations and (2) ability
of independent actions. With current IoT as the solution, both of the elements will
not be able to be fulfilled, as IoT does not possess learning algorithms for it to think
independently in difficult situations. IoT only possesses the first main component,
which is completely autonomous with minimum human interaction. Even so, it still
requires human intervention even though the interactions are minuscule. It will be
made possible by the combination of two technologies, which are IoT and artificial
intelligence (AI). Smart technology implies automatic computing principles like self-
configuration, self-healing, self-protection, and self-optimization [18]. Therefore, in
order to reach a Smart City status, IoT alone is not adequate and a combination of
IoT and AI is needed (Fig. 3).
Table 3illustrates each layer’s capabilities of the proposed framework. AI for
self-learning must perform tasks that will require intelligence to perform task with
machine learning reasoning capabilities. Douglas [19] stated that AI has four main
abilities. They are the ability to sense, converse, act, and learn. Although AI and
IoT as a solution is not the mainstream solution, the ideas and initiatives with both
technologies combined in realizing a Smart City are catching up.
Huawei in 2018 unveiled its artificial intelligence Smart City platform, which
utilizes IoT capabilities [20]. Roven [21] and Bailey and Coleman [22] highlighted
solutions with both the elements mentioned earlier.
Autonomous and Adaptive Learning Architecture Framework … 467
Fig. 3 Overarching autonomous learning framework
Tabl e 3 Layer functionality
Layers Functionality
Application architecture i. Present data for service allocation
ii. Respond to waste, traffic, carbon emission, healthcare, and
surveillance
iii. Mitigation responses and decision-making
Business architecture i. Data analytics—prescriptive and predictive
ii. Semantic analysis with aided reasoning capabilities
iii. Agent learning capabilities supported by autonomous features
Data architecture i. Data acquisition—multiple sources
ii. Data optimization and gleaning—from structured as well as
unstructured data
iii. Learning and training capabilities
Acquisition architecture i. Smart readers—feeder for data architecture
ii. Smart sensors—feeder for data architecture
iii. IoT sensors—feeder for data architecture
iv. Smart contracts for participative user network
4 Baseline to Target Architecture
Understanding baseline architecture (“As-Is”) is necessary for us to determine the
target architecture (“To-Be”) for Smart Cities. In order to be fully autonomous with
learning capabilities, our Business Architecture layer should be capable of execution
for decision without human intervention. Müller and Bostrom [23] estimated that AI
will have equal cognitive ability with humans by 2050 and the ability that will exceed
468 S. Muthaiyah and T. O. K. Zaw
human’s capability by 2075. A combination of different types of machine learning,
can be used to achieve this and they are (1) Supervised Learning, (2) Unsupervised
Learning (3) Reinforcement Learning, and (4) Semi-supervised learning. The transi-
tion strategy for Acquisition Architecture layer must include full deployment of smart
readers and IoT sensors coupled with smart contracts for decision-making. As for
the Data Architecture layer, data optimization and inclusion of learning algorithms
supported by machine learning will be the core priority. Metadata that are structured
and unstructured should be also considered and normalized. This will ensure that the
analytics done at the Business Architecture layer will provide accurate reasoning for
actions to be executed on-the-fly. For this, a proper taxonomy for definitions must
be in place with objects defined for each use case, e.g., surveillance, healthcare, CO2
emission, and traffic routing, sewer, and waste management. We recommend the Java
Agent Development Framework (JADE) which is based on the Foundation of Intel-
ligent Physical Agents (FIPA) to be used as reference (see Fig. 4). A specification
of conceptualization for different stakeholders can be developed with ontologies and
semantic processing will enable mediation among these ontologies. Agent softbots
will then fire services needed to be presented at the Application Architecture layer
for each use case discussed earlier (Fig. 5).
Fig. 4 Multi-agent system (MAS) for use cases
Autonomous and Adaptive Learning Architecture Framework … 469
Fig. 5 Baseline and target architecture
5 Conclusion
Due to population expansion our world will soon realize a total population of 10
billion by 2050. The enormous pressure on clean water, sanitation, waste manage-
ment, carbon footprint, noise, urban healthcare, and crime management will push
forward the need for smart solutions in the context of smart cities. The World Urban
Forum no. 9 (WUF 9) declaration summarized sustainability as the core driving force
for smart inclusion for cities. Urban governance, however, cannot be a one size fits
all approach. Core needs of a city must be understood, carefully planned out, and
deployed using technology. For this, a use case for cities must be addressed first in
their own context. After that a baseline and target architecture must be developed
and lastly proper maintenance is required. In this paper we have (1) proposed Smart
City use cases based on the Alcatel–Lucent study in 2011, (2) introduced the overar-
ching autonomous learning framework (OALF), (3) transition strategies for OALF,
and (4) a multi-agent approach as to how to deploy the use cases keeping in mind
autonomous and learning attributes for city sustainability and governance.
470 S. Muthaiyah and T. O. K. Zaw
References
1. L. Atzori, A. Iera, G. Morabito, The Internet of Things: a survey. Comput. Netw. 54(15),
2787–2805 (2010). https://doi.org/10.1016/j.comnet.2010.05.010
2. P. Bellavista, G. Cardone, A. Corradi, L. Foschini, Convergence of MANET and WSN in IoT
urban scenarios. IEEE Sens. J. 13(10), 3558–3567 (2013). https://doi.org/10.1109/jsen.2013.
2272099
3. H. Schaffers, N. Komninos, M. Pallot, B. Trousse, M. Nilsson, A. Oliveira, Smart cities and
the future internet: towards cooperation frameworks for open innovation. Future Internet 6656,
431–446 (2011). https://doi.org/10.1007/978-3-642-20898-0_31
4. R.P. Dameri, Searching for Smart City definition: a comprehensive proposal. Int. J. Comput.
Technol. 11(5), 2544–2551 (2012). https://doi.org/10.24297/ijct.v11i5.1142
5. R. Giffinger, C. Fertner, H. Kramar, E. Meijers, N. PIichler-Milhanovic, Smart Cities: Ranking
of European Medium-sized Cities (2007)
6. A. Caragliu, C. de Bo, P. Nijkamp, Smart cities in Europe, in 3rd Central European Conference
in Regional Science (2009)
7. M. Rouse, What is Smart City?—Definition from WhatIs.com (2017). Retrieved from: https://
internetofthingsagenda.techtarget.com/definition/smart-city. 24 Jul 2017
8. Techopedia, What is a Smart City?—Definition from Techopedia (2015). Retrieved from:
https://www.techopedia.com/definition/31494/smart-city. 30 Nov 2015
9. Gemalto, What is a Smart City? (2018). Retrieved from: https://www.gemalto.com/iot/inspired/
smart-cities. 18 Jan 2018
10. M. Dohler, I. Vilajosana, X. Vilajosana, J. Llosa, Smart Cities: an action plan, in Proceedings
of Barcelona Smart Cities Congress (2011), pp. 1–6
11. M. Yun, B. Yuxin, Research on the architecture and key technology of Internet of Things (IoT)
applied on smart grid, in 2010 International Conference on Advances in Energy Engineering
(2010). https://doi.org/10.1109/icaee.2010.5557611
12. A. Monacchi, D. Egarter, W. Elmenreich, Integrating households into the smart grid, in 2013
Workshop on Modeling and Simulation of Cyber-physical Energy Systems (MSCPES) (IEEE,
2013), pp. 1–6
13. X. Li, W.Shu, M. Li, H.Y. Huang, R.E. Luo, M.Y. Wu, Performance evaluation of vehicle-based
mobile sensor networks for traffic monitoring. IEEE Trans. Veh. Technol. 58(4), 1647–1653
(2009). https://doi.org/10.1109/tvt.2008.2005775
14. S. Lee, D. Yoon, A. Ghosh, Intelligent parking lot application using wireless sensor networks,
in 2008 International Symposium on Collaborative Technologies and Systems (2008). https://
doi.org/10.1109/cts.2008.4543911
15. N. Maisonneuve, M. Stevens, M.E. Niessen, P. Hanappe, L. Steels, Citizen noise pollution
monitoring, in Proceedings of the 10th Annual International Conferenceon Digital Government
Research: Social Networks: Making Connections between Citizens (2009), pp. 96–103
16. T. Nuortio, J. Kytöjoki, H. Niska, O. Bräysy, Improved route planning and scheduling of waste
collection and transport. Expert Syst. Appl. 30(2), 223–232 (2006). https://doi.org/10.1016/j.
eswa.2005.07.009
17. A.R. Al-Ali, I. Zualkernan, F.Aloul, A mobile GPRS-sensors array for air pollution monitoring.
IEEE Sens. J. 10(10), 1666–1671 (2010). https://doi.org/10.1109/jsen.2010.2045890
18. J. Sairamesh, A. Lee, L. Anania, Information cities. Commun. ACM 47(2), 28–31 (2004)
19. E. Douglas, Artificial Intelligence: Abilities & Expectations (2018). Retrieved from: https://
orzota.com/2018/02/08/artificial-intelligence-abilities/. 8 Feb 2018
20. C. Reichert, Huawei Unveils Artificial Intelligence Smart Cities Platform (2018).
Retrieved from: https://www.zdnet.com/article/huawei-unveils-artificial-intelligence-smart-cit
ies-platform/. 14 Nov 2018
21. A. Roven, AI in Smart Cities (2018). Retrieved from: https://medium.com/neuromation-blog/
ai-in-smart-cities-dfe2fa7d2829. 27 Nov 2018
Autonomous and Adaptive Learning Architecture Framework … 471
22. D.J. Bailey, Y. Coleman, Urban IoT and AI: How Can Cities Successfully Leverage This
Synergy? (2019). Retrieved from: https://aibusiness.com/future-cities-iot-aio/. 8 Jan 2019
23. V.C. Müller, N. Bostrom, Future progress in artificial intelligence: a survey of expert opinion.
Fundamental Issues of Artificial Intelligence, pp. 555–572 (Springer, Cham, 2016). https://doi.
org/10.1007/978-3-319-26485-1_33
A Predictive Analysis for Heart Disease
Using Machine Learning
V. Rajalakshmi, D. Sasikala, and A. Kala
Abstract Predictive analysis plays a major role in healthcare industry where fore-
casting the disease will reduce the risk that happen to patients. Statistics show that
cardiovascular diseases have increased the mortality rate in India. Machine learning
which is used in developing a predictive model for various domains is nowadays
applied in the field of medical diagnostics. Machine learning is playing an integral
role in predicting the presence or absence of heart diseases. Such predictions, if done
well in advance, can help the doctors to carry out the treatment for the patients and
mitigate their health risk. Biological samples such as blood or tissues are collected
from the human body to predict cardiovascular diseases. The proposed work is
focused on developing various machine learning predictive models using support
vector machine, decision tree, neural network and K-nearest neighbour for predic-
tion of heart disease. For this work Cleveland heart disease dataset is used which
consists of 14 attributes and 294 records. A comparative analysis on the prediction
models were carried out. From the results, it was found that support vector machine,
decision tree and KNN (k=15) classifiers yield better accuracy to predict heart
disease than the other models.
Keywords Cardiovascular disease ·Machine learning ·K-nearest neighbour ·
Decision tree ·Neural network
V. Rajalakshmi (B)·D. Sasikala ·A. Kala
Sri Venkateswara College of Engineering, Chennai, India
e-mail: vraji@svce.ac.in
D. Sasikala
e-mail: dsasikala@svce.ac.in
A. Kala
e-mail: akala@svce.ac.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_42
473
474 V. Rajalakshmi et al.
1 Introduction
There is a major shift in traditional healthcare system after the boom of Internet and
telecommunication. Today almost majority of the common people have Internet in
some devices which they use in their day-to-day life. In earlier days, healthcare system
means hospitals with doctors, caretakers, skilled lab technicians, etc. In traditional
systems, patient’s health is examined with the lab test or with devices. These devices
are operated with the help of doctors or people who are skilled in that field. The results
from the health records are reported to doctors and they assess these records with their
knowledge of expertise. These records are analyzed by the medical professionals and
they assess the patient’s health. In the nineteenth century, the industry started using
Internet and this made a big revolution in industrial sectors. Slowly this technology
has entered into the medical industry and aided the medical professionals in many
ways. In some countries, machine-assisted operations are performed on patients by
the doctors. Internet technologies are playing a major role in medical sectors. For
example, medical professional assistance is given through video conference; these
are due to technology evolution.
1.1 Motivation and Background
The mortality rate due to cardiovascular disease is high when compared to other
diseases which is pointed out by the statistical inferences. If these diseases are
predicted much earlier then it can save human life. Prediction accuracy is very vital
to prevent any judgmental errors. Machine learning technology is applied in many
scientific and medical fields wherein it has brought a new lease of life. Building a
good machine learning model to predict diseases can save human life. There are
many machine learning algorithms and each of them produces the results based on
the training and hypothesis it has built from the data. This paper is focused on finding
the suitable machine learning model for accurate prediction of heart disease from
the Cleveland dataset.
Cardiovascular or heart disease generally refers to a narrow/blockedvessel causing
cardiac arrest. There are many other conditions which may affect the heart muscle
which also results in blocking the blood flow from the heart to other parts of the body.
Cardiovascular disease can be prevented by diagnosing them earlier. A simple and
cost-effective procedure should help mankind rather than costlier ones. In conven-
tional procedure diagnosis of a patient to find a normal or abnormal condition is
assessed through the blood sample or tissue collected from them. Many works related
to this research were done using image processing by other researchers. Our aim is
to build a machine learning model using numerical clinical dataset. In the first phase,
Cleveland dataset is used in the proposed system to build a machine learning model
A Predictive Analysis for Heart Disease Using Machine Learning 475
using support vector machine where the prediction accuracy was good. Later other
machine learning models like decision tree, neural network and K-NN classifier were
built. These models’ accuracy of prediction for the Cleveland dataset was analyzed
to find the best model.
In many parts of the world, healthcare systems built by machine learning model
and artificial intelligence are being widely used and successfully accepted by the
medical professionals. But there is a need for an accurate model which would further
help the mankind. There are still researches going on in this field for the betterment
of patient-centric care. Many researches had been done by the scientific and research
community and their few works are presented in the next section.
2 Literature Survey
Alba et al. [1] proposed a model for segmenting highly abnormal hearts using statis-
tical shape model. Their focus was to accurately identify the abnormal hearts from the
cardiac image even though there is an abnormal reshaping due to other conditions like
pulmonary hypertension. They employed virtual remodelling transformation from
the image generated from the patient and the reference modelling which was obtained
from the boundary analysis. A reference statistical shape model from a sample of
normal subjects was built and then the original and pathological statistical model
was compared to find any abnormality.
This work [2] finds the heart failure by employing Complex Event Processing
along with the statistical approach. Here the patient health record is monitored and
the history of data is stored and compared with the current data, based on the refer-
ence, prediction was done. But patient-centric threshold is set and the values are
dynamically changed based on the health generated data. The project which Karan
Bhanot [3] presented uses machine learning algorithms like support vector machine,
random forest, decision tree and K nearest neighbour. Among these algorithms, K-
nearest neighbour algorithm has higher accuracy prediction rate when compared to
other classifiers.
Improving the accuracy of heart disease prediction at an early stage and reducing
theerrorrateinprediction[4] was the target of their work. Ensemble techniques such
as bagging and boosting reduce the error rate of weak classifiers. Another similar
work was done [5], in this work they tested their heart disease dataset across multiple
classifiers like support vector machine, Naïve Bayes classifier, logistic regression,
gradient boosting and random forest. They built a model using the classifier for
prediction.
For diagnosing [6] a heart disease, a hybrid model was built with seven classifiers,
three feature selection algorithms and a cross-validation method. They improved their
model accuracy by removing the irrelevant feature using feature selection method
and improved the accuracy of prediction. This model has improved the execution time
as well. In this work [7], they developed machine learning models using decision
tree, random forest, multivariate adaptive regression splines, tree model from genetic
476 V. Rajalakshmi et al.
Tabl e 1 Dataset description
and attributes Clinical features (attributes) Description
Age Age in years
sex Gender (1 =male; 0 =female)
cp Chest pain type
trestbps Resting blood pressure
chol Serum cholesterol in mg/dl
fbs Fasting blood sugar > 120 mg/dl
(1 =true; 0 =false)
restecg Resting electrocardiographic
results
thalach Maximum heart rate achieved
exang Exercise-induced angina (1 =
yes; 0 =no)
olpeak ST depression induced by
exercise relative to rest
slope The slope of the peak exercise
ST segment
ca Number of major vessels (0–3)
coloured by fluoroscopy
thal 3=normal; 6 =fixed defect; 7
=reversible defect
num Diagnosis of heart disease
algorithm and used Cleveland dataset [8] to predict the heart disease. Their study
proved that decision tree outperforms the other machine learning algorithms.
3 Dataset Description
The Cleveland database is the heart disease dataset which is popularly used by the
machine learning researchers in heart disease prediction. The dataset consists of 14
attributes, namely (Table 1):
The target field is ‘num’ which refers to the presence of heart disease in the patient.
‘0’ refers to absence of heart disease and ‘1’ refers to presence of heart disease. The
dataset consists of 294 records.
4 Proposed System
The objective of the work is to build a machine learning model to predict the heart
disease with high accuracy. Cleveland dataset is used in this work. In the first step,
A Predictive Analysis for Heart Disease Using Machine Learning 477
Fig. 1 Heart disease prediction architecture using machine learning
data cleaning is done where the missing data and the duplicate records are identified
and removed. The data is divided into training and test set for further model building.
The ratio of the training and test data split is 70:30. The model for prediction of heart
disease is built with support vector machine, decision tree, neural network and K-
nearest neighbour. The model is evaluated with the test data to identify the best model
that predicts the disease with least errors. Figure 1describes the architecture for heart
disease prediction.
4.1 Machine Learning Models
•Artificial Neural network (Perceptron)
Perceptron is a supervised linear classifier that classifies the data into two classes.
•Decision Tree
Decision Tree is a supervised classifier where the data is repeatedly split based
on a certain parameter.
•Support Vector Machine
Given a labelled training data, the support vector machine defines a separating
hyperplane, that categorizes the new data.
•K-Nearest Neighbour
Given an input, which consists of kclosest training data, KNN assigns the new
data to the class that exists in maximum among the K-nearest neighbours.
478 V. Rajalakshmi et al.
5 Implementation and Results
The prediction of the heart disease is done by implementing four models: the percep-
tron, the decision tree, support vector machine and K-nearest neighbour with kvalues
1, 5, 10 and 15. The models are implemented in R Studio.
Figure 2shows the decision tree classification and perceptron model built for the
given training set. The models are evaluated by computing various metrics. Let ‘a’
be true positive, ‘b’ be true negative, ‘c’ be false positive, ‘d’ be false negative and
‘n’ be the total number of samples. The metrics computed are as follows (Table 2):
Accuracy =(a+d)/n(1)
Precision =a/(a+c)(2)
Recall =a/(a+d)(3)
Fig. 2 Decision tree and perceptron model
Tabl e 2 Comparison of
results from ML models ML model Accuracy Precision Recall
Perceptron 78.5 83.3 60.6
Decision tree 84.5 87.8 60.6
SVM 84.5 86.3 61.9
K-NN (k=1) 77.3 82.9 60
K-NN (k=5) 79.7 83.6 61.2
K-NN (k=10) 78.5 84.7 59.1
K-NN (k=15) 85.71 88 61.1
A Predictive Analysis for Heart Disease Using Machine Learning 479
6 Conclusion
From the results, it was found that support vector machine, decision tree and KNN (k
=15) classifiers yield better accuracy to predict heart disease than the other models.
The work is planned to extend further by implementing the Committee of Machines
and deep learning models.
References
1. X. Alba, M. Pereanez, C. Hoogendoorn, A.J. Swift, J.M. Wild, A.F. Frangi, et al., An algorithm
for the segmentation of highly abnormal hearts using a generic statistical shape model. IEEE
Trans. Med. Imaging 35 (2015)
2. A. Mdhaffar, I.B. Rodriguez, K. Charfi, L. Abid, B. Freisleben, Complex event processing for
heart failure prediction. IEEE Trans. NanoBiosci. 16(8), 708–717 (2017)
3. https://towardsdatascience.com/predicting-presence-of-heart-diseases-using-machine-lea
rning-36f00f3edb2c
4. C. Beulah Chirstalin Latha, S. Carolin Jeeva, Improving the accuracy of prediction of heart
disease risk based on ensemble classification techniques. Inform. Med. (2019)
5. K.G. Dinesh, K. Arumugaraj, D. Santhosh, V. Mareeswar, Prediction of cardiovascular disease
using machine learning algorithms, in International Conference on Current Trends towards
Converging Technologies, pp. 1–7 (2018)
6. P. Li, M.H. Memon, S. Nazir, R. Shah, A hybrid intelligent system framework for the prediction
of heart disease using machine learning algorithms. Mob. Inf. Syst. (2018)
7. T. Sharma, S. Verma, Kavita, Prediction of heart disease using cleveland dataset. A machine
learning approach. Int. J. Rec. Res. Aspects 4, 17–21 (2017)
8. https://archive.ics.uci.edu/ml/machine-learning-databases/heart-disease/
Application of Data Mining Algorithms
for Tourism Industry
Promila Sharma, Uma Meena, and Girish Kumar Sharma
Abstract This paper presents different Data Mining Algorithms to be used to access
data patterns in Tourism Industry and their applications using Data Mining Tool like
WEKA which is a popular software, and useful patterns can be found efficiently. Data
Mining, as the name implies is the science by which hidden and interesting informa-
tion can be found from large datasets analogous to Gold Miners who extract precious
metal sieving through a lot of raw material. Normal day to day business operations
create huge voluminous data which is needed to be utilized properly to find informa-
tion to make efficient, effective, and prompt decisions otherwise businesses can not
sustain in competitive environment. There are some techniques and algorithms used
in data mining for determining interesting hidden patterns. data mining algorithms
are becoming popular day by day in various applications like Travel and Tourism.
Nowadays, a number of data mining applications are available for Public use. There-
fore, we have to judge which particular algorithm or application suits a problem
domain. The idea behind applying data mining is very straight forward, it is similar
to a human being becoming intelligent from learning by examples and experience.
Similarly, in data mining the behavior of given system (training dataset) is learned by
application, and certain rules are developed. Then these rules are used to evaluate the
behavior/result for the given situation (test dataset). The overall goal of this paper is
to classify some data mining algorithms and applications, draw a comparative study
that which one is better for tourism industry and under what circumstances.
P. Sharma (B)
Department of Computer Science, Mewar University, Chittorgarh, Rajasthan, India
e-mail: prom_sharma@yahoo.com
U. Meena
CSE Department, SRMIST, Delhi-NCR Campus, Modinagar, UP, India
e-mail: uma.b18@gmail.com
G. K. Sharma
Department of Computer Applications, Bhai Parmanand Institute of PG Studies (Under DTTE,
GNCT of Delhi), Delhi, India
e-mail: Gkps123@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_43
481
482 P. Sharma et al.
Keywords Travel and tourism industry ·Travel pattern ·Apriori algorithm ·Weka
tool ·Data mining ·Decision trees ·Association rules ·Classification ·Clustering
1 Introduction
Data mining is the latest technology and is becoming popular because of the
increasing use by companies who have a strong customer focus like retail, marketing,
tourism, etc. It provides great help to companies who can collect data about the
behavior of their potential customers and can discover hidden information within the
data that queries and reports cannot reveal properly.
Data mining best known as analysis part of the “Knowledge Discovery in
Databases” process, or KDD. It is an interdisciplinary part of computer science, is
a data-driven research which applies algorithms on computer and discovers patterns
in huge datasets. It sits at the intersection of statistics database systems, artificial
intelligence, and machine learning but with few differences. The databases provide
transactional and query reports as output whereas data mining provides analyt-
ical output, Statistics makes generalizations from observations whereas data mining
provides specific information, Machine intelligence deals with small-sized data and
data mining with huge datasets. The main objective to use data mining is that it can
extract information from large datasets and transform to readable form for further
uses. Data mining uses some techniques to extract interesting patterns from large
datasets [1].
1.1 Steps of Data Mining
Data mining or Knowledge Discovery in Databases (KDD) comprises the following
processing steps.
1. Data cleaning step or Preprocessing—This is the first step to remove noisy data
that is missing, duplicate inconsistent data from large datasets.
2. Data integration step—This step combines data from multiple sources.
3. Data selection step—All attributes are not related or needed to be included for
analysis task. So in this step only data relevant to a particular problem domain is
selected.
4. Data transformation—Data mining techniques require that data should be trans-
formed to consolidated or some standard normalized form before beginning data
mining.
5. Data mining application—It is the important phase in which various data mining
algorithms are applied on selected datasets to extract unknown useful patterns.
6. Pattern recognition—This is the phase where we see benefits of data mining. It
is used to find unknown and interesting patterns from datasets.
Application of Data Mining Algorithms for Tourism Industry 483
7. Knowledge interpretation or Presentation—Various visualization techniques can
be used to present mined patterns to the user and managers.
1.2 What are Data Mining Applications
A Human analyst manually cannot evaluate and analyze huge volume of data so
computer-based efficient and speedy software applications are required. Therefore,
data mining techniques are becoming popular for businesses and other fields as
medical, engineering, sports, marketing, etc. With the help of data mining tech-
niques, businesses can analyze their marketing and sales patterns, develop latest
digital marketing campaigns, and accurately predict customer behavior and profile.
Some of the applications of data mining are [1]:
1. Market Identification—Common features/characteristics of customers can be
found who bought similar products from company.
2. Customer churning—One can predict which customers had left your company
and had gone to the competitor.
3. Fraud detection—Through this fraudulent transactions/hacking can be detected.
4. Direct sales and marketing—In order to obtain higher response rate of customers,
it accesses which prospects of mailing lists can be changed.
5. Interactive/Online/Digital marketing—Through the use of website mining,
potential customers behavior and buying patterns can be known as who have
done online shopping.
6. Market/Product basket analysis—Through association rule mining algorithms
one can make judgment as which products or services customers had purchased
together, i.e., Bread and milk.
7. Trend/Growth rate analysis—Sales differences between previous year and
current year can be known to analyze trends or growth rate of products.
1.3 Data Mining Uses for Travel and Tourism Industry
Travel is known as the movement of people for different purposes between remote
geographical locations. There can be different modes of travel by foot, by bicycle,
auto, railway, ship, bus, aeroplane, or other means one way or round trip. Informa-
tion Technology has found great uses in travel and tourism industry for information
processing [2]. Advancements made in information technology have influenced the
services and facilities offered and these are efficiently delivered and promoted. It also
has affected the way how organizations are structured and communication channel
between customers and services providers has improved [3]. Travelers too are bene-
fited from Internet and communication technology. They can easily search places of
their interests and avail services accordingly.
Tourism in India is considered the largest service industry as both inbound and
outbound tourists have contributed to 6.23% to the national GDP and 8.78% of the
484 P. Sharma et al.
total employment. According to a survey report travel industry in India has generated
about $100 bn USD in 2018 and it is expected to increase to $275.5 bn USD in near
future with an annual growth rate of 9.4% (India Tourism Statistics) [4].
Because of increasing business competition, the situation of sustainability for
Indian travel and tourism industry has declined. Therefore, the industry needs to
increase their position in the market and keep control of their businesses. In order
to develop market and to predict potential customers, to plan best tour packages and
services, they should use data mining tools and techniques efficiently [5]. Modern
technology has had a great impact on tourism (Domestic and International Both)
when Internet and communication technologies became the most dominant channel.
Tourists’ expectations are related not only to tourism services, but also to technology.
For this to be successful, one must have domain knowledge of tourism as well as
how it can be implemented using data mining [6].
In the tourism industry, tour operator companies can formulate better marketing
strategies and maximize their profits by knowing guests’ behavior—where they have
come from, the money they have spent, and when and on what they have spent. In day
to day businesses, travel and tour companies accumulate large amounts of customer
profile data, which they can organize in proper format and integrate in databases.
Advanced technology as data mining can be used to identify important variables
and analyze the information hidden in datasets which guides marketing decision [7].
Nowadays businesses specially related to customer focus are raising information and
services system by using data mining techniques, for example, uses in hospitality
management and service industries like hotel management, tourism, ticket booking,
railways, Information technology has been used to assist and expedite the delivery
services [8]. Improved capacity management and operational efficiency, inventory
control of items, check availability of room information, increasing management
control and capability, creating consolidated or summary marketing, sales and opera-
tional reports, keeping track of customers and hotel guests who visit frequently, trans-
forming transactions from internal operations of management to human resources
are some of the results of IT technology uses [9]. Most of these items apply only to
hotels and accommodation managers and providers. The latest technologies like data
mining helps Hotel Industry to predict future trends and access customer transactions
which are useful in sales and marketing by making high-quality market research and
planning [10].
2 Introduction to Data Mining Algorithms/Strategies
A data mining algorithm can be defined as a set of some rules and heuristics that
can be used to create Data mining model from dataset. In order to develop a model,
the algorithm first analyzes datasets (Training dataset) and views specific unknown
patterns. Then it defines optimal parameters for the model by using output/result of
the training set. These parameters are applied in the remaining dataset (Test dataset)
to extract hidden information, i.e., actionable rules and statistics.
Application of Data Mining Algorithms for Tourism Industry 485
2.1 Popular Data Mining Algorithms
There are different types of techniques/algorithms used in data mining which depend
on the application domain and type of data recovery process that provide useful
answers to problems. These are listed below:
1. Classification algorithms also known as discrimination or pattern recogni-
tion which makes prediction of one or more discrete or class variables, based
on the other independent variables/attributes in the dataset. These come under
supervised learning.
2. Regression or Estimation algorithms also come under supervised learning.
This method produces numerical attributes as output based on other input vari-
ables/attributes. For example, prediction of profit or loss, depending on other
attributes in the dataset.
3. Segmentation or Clustering algorithms come under unsupervised learning.
These divide datasets into unlabelled groups or clusters of objects that have
similar properties. Examples are image segmentation, market segmentation,
outlier detection, etc.
4. Association Rule Mining algorithms are also known as Market basket analysis
or Affinity analysis which is unsupervised learning. In this, relationship between
different attributes in a dataset is found whose most common application is to
identify buying patterns of customers and accordingly design store layouts.
5. Sequence Detection algorithms. A time series is a set of observations collected
as sequences. This algorithm summarizes frequent sequences or time series in
data, such as a website path flow. These are important in bioinformatics, sales
forecasting, stock market analysis, etc.
3 Significance of the Problem
3.1 Problem Statement
To Access Data Patterns for Indian Tourism Industry.
This research work can provide the solutions to the following questions pertaining
to tourism sector as following:
1. How many and what age group Domestic/Indians travelers travel in Delhi or
other places of their choices per year? How much per year they spend on tours?
2. How many foreign tourists come to Delhi and what are the places of their interest,
purposes of travel, what age group people usually come and from which countries.
The tour packages they are offered and their likings.
3. What are the lean and peak seasons/timings during the year to travel?
4. What is the probability that certain tourists will respond to the upcoming
promotion for a holiday tour package?
486 P. Sharma et al.
5. What places and packages are expected to be in the priority list of customers
whether Domestic Travelers/NRI (Non-Resident Indians)/FTA (Foreign Travel
Arrivals)?
4 Research Methodology
Research methods may be defined as all those methods/techniques that are applied
by researchers for conduction of research. Research methods or techniques, thus,
refer to the processes and methods and tools, the researchers use in performing
research operations. For solving the above problems some research methods and
techniques are used for obtaining the desired result. Some tools and algorithms are
required for obtaining results. Main steps followed under the research methodology
are listed below.
4.1 Literature Review of Research Papers
In this paper, all previous work from journal literature and research papers were
studied. It gave insight into what type of work has been done by previous
scholars/researchers in the particular domain and by what methodologies/techniques
of research were followed, and what tools were used to carry out research. Then
more knowledge was gained about the problem.
4.2 Identify Tools
Then a tool was identified for carrying various data mining applications on tourism
dataset and the popular tool “WEKA” (Waikato Environment For Knowledge Anal-
ysis) was selected from all. It was found very useful and reliable. There are algorithms
built in the software for Classification, Regression Analysis, Clustering, Association
Rule Mining, and Visualization.
4.3 Defining Required Database Structure and Attributes
The structure and attributes were prepared for collecting Tourists Information, their
profile, places they visited, and others about their tour packages. Questionnaire
method was used for primary data collection.
Application of Data Mining Algorithms for Tourism Industry 487
4.4 Data Collection
The questionnaire was filled by visiting domestic tour operators in Delhi and a few
data were collected by emailing questionnaires to friends group who visited Delhi
and its excursions, nearby places of interest.
4.5 Data Pretreatment
The Database has 1050 records and 14 useful attributes are selected. First the Data
is prepared in Excel sheets and missing values are dealt with. Then raw data file is
saved with CSV extension. It was converted to arff file format because this is Weka’s
native file format used for data analysis.
4.6 Data Mining
After determining nature and workflow of Research Problem Database is organized
and data mining algorithms are applied using WEKA in order to generate results.
The following techniques are applied.
5 Analysis Using Arm (Association Rule Mining)
5.1 Conceptual Framework
ARM is a type of unsupervised learning algorithm. Its main objective is in transac-
tional database to describe the rules of data relationship. The set of items which are
purchased in a transaction is called its market basket such as shopping cart. Through
this one can predict the buying patterns of customers like which products are bought
together [11]. The knowledge gained from ARM analysis can be used in maximizing
sales and accordingly store layouts can be managed. For this data mining discipline
offers a useful strategy known as Association Rule Mining. It is a popular algorithm
credited to Agarwal, Imielinski and Swami in 1993 [12].
Let us assume that a transaction contains data items A and B. If any transaction
that contains “A” is also likely to contain “B” is expressed as the rule “A” =>“B”
(A ∩B=ϕ). This is an expression of Association rule. A is denoted as antecedent
or Hypothesis and B is called a consequent or result which follows B. They are
related as “If A then B.” In ARM two estimates Support S and Confidence C from
transaction datasets are calculated. Support is the percentage of items that contain
both “A” and “B”. It is described as P(A ∩B) where P is probability. Suppose there
488 P. Sharma et al.
are 100 transactions in which 80 transactions contain Item “A”, 70 contain item “B”
and 60 contain both “A” and “B”. Then Support S is 60%. The rule has Confidence C
as the percentage of items “B” is present in transaction containing “A”. It is denoted
y Conditional probability P(‘B’|‘A’). In the above example C is 60/80 that is 75%.
Support S is used to measure the frequency of an Association Rule and Confidence C
is used to measure the strength of the Association Rule. The Three important phases
of ARM are
1. First the set of Items that has significant impact on businesses are found.
2. In the second step, information from numerous transactions on dataset items
from many sources are combined.
3. Then Association Rules are generated from counts of significant items in
Transactions.
The Apriori algorithm (Bodon, Zaki, Agarwal, and Srikant) is the most widely
used tool for ARM. This algorithm uses previous knowledge of properties of data
items that appear frequently. Firstly, Apriori algorithm generates candidate itemsets
of a given size and then analyzes the dataset items if their counts are large.
In the first pass all singleton itemsets are generated and those are selected
whose support value is above the minimum specified. Remaining singleton items
are combined to form two member candidate itemsets. Support values of this candi-
date itemsets are again calculated and the next phase creates three candidate itemsets.
The process is iterative and it stops when all frequent itemsets are accounted for [13].
5.2 Discovering Travel Pattern Through ARM Using Weka
An ARM Algorithm cannot handle numeric attributes. Therefore, all the numeric
attributes have to be converted to discrete or categorical attributes. This is called
discretization and this can be accomplished by using Weka’s Internal discretization
tools.
The complete dataset of Indian tourists is refined/preprocessed and loaded in
Weka. 110 instances are selected as training set for analysis. Numeric attributes
as age, stay_in_days, no_persons are discretized. All other attributes except 8 are
removed which are useless for the analysis through ARM. The best 10 rules were
found supporting each other according to their dominant category. These are gener-
ated by taking default settings of Apriori algorithm in Weka. These are listed in
Fig. 1.
Application of Data Mining Algorithms for Tourism Industry 489
Fig. 1 Result of ARM in Weka
5.3 Best Rules Found Are
1. Type_Of_Tours =Adventure 17 ==> Age =‘(-inf-39.666667]’ 17 means there
are 17 adventure types of tours and they are taken by age group of people up to
40 yrs of age.
2. Type_Of_Tours =Adventure Travel_Arrangement =Tour_Operator 14 ==>
Age =‘(-inf- 39.666667]’ 14 means it follows above rule 1. In addition to that
all 14 people have traveled through tour operator.
3. Type_Of_Tours =Adventure Mode_Of_Travel =Private Car 14 ==> Age =
‘(-inf- 39.666667]’ 14 means there are 14 records of adventure types of tour
and have traveled through Private car taken by 14 people of age up to 40 yrs.
4. Sex =Female Age =‘(39.666667-57.333333]’ Mode_Of_Travel =Private Car
14 ==> Travel_Arrangement =Tour_Operator 14 means There are 14 females
who are in the age group 40–58 and have traveled by private car and tour is
arranged by tour operator.
5. Tour_Package =AgraFatehpurSikri 13 ==> Type_Of_Tours =Historical 13
Means there are 13 records of tour package as AgraFatehpurSikri is type of
Historical tour.
6. Occupation =Service Mode_Of_Travel =Private Car 21 ==> Stay_In_Days
=‘(-inf-3]’ means there are 21 persons in service who have traveled by private
car and stayed for 3 days.
7. Type_Of_Tours =Religious 20 ==> Religion =Hindu 19 means the type of
Religious tour taken by Hindu religious people.
8. Tour_Package =DelhiSightseeing 13 ==> Type_Of_Tours =PleasureLeisure
13 means there are 13 records in which tour package was Delhisightseeing and
type of tour is PleasureLeisure.
490 P. Sharma et al.
9. Tour_Package =GoldenTriangleJaipur Travel_Arrangement =Tour_Operator
12 ==> Type_Of_Tours =Historical 12 means there are 12 GoldenTriangle-
Jaipur tour package arranged by tour operator which is historical.
10. Sex =Male Tour_Package =GoldenTriangleJaipur 11 ==> Type_Of_Tours
=Historical 11 means 11 male persons have taken GoldenTriangleJaipur tour
package which is historical.
6 Analysis Using Classification
6.1 Conceptual Framework
The classification, discrimination, or pattern recognition algorithm comes the cate-
gory of supervised learning. The process evaluates a function and a relationship
between a dependent variable (Output/Discrete) and contingent independent vari-
ables (Input parameters) is established. In the classification problem an object is
identified and its belongingness to a given class is analyzed [14].
In Classification algorithms rules are found to organize given datasets into prede-
fined classes. A part of the historical data is chosen (Training data), algorithm rules
are generated, and then are tested on the remaining data (Test data). During Training
phase the model is trained or learned and the model is developed. Then the developed
model is run on Test data. The Traditional and well-accepted method of classifica-
tion is the induction of decision trees, which partitions the dataset to develop rules.
Another approach uses probability theory for classification and is known as Naïve
Bayes’s algorithm.
The other popular algorithms that are used for classification are Artificial Neural
Networks (ANN) and Super Vector Machines (SVM) [13].
6.2 Decision Tree for Classification Using Weka
Decision trees are used for classification and prediction tasks of data mining, and they
are the most popular techniques. A Decision Tree is a flowchart hierarchical tree-like
structure in which the condition is tested on each internal node attribute, each next
level branch node represents an outcome of the test, and finally leaf node holds a
class label. During the late 1970s and 1980s, J. Ross Quinlan, a researcher in machine
learning had developed Tree Algorithm known as ID3 (Iterative Dichotomiser). He
later on modified it and presented C4.5 which has become a benchmark to newer
supervised learning algorithms. It is the advanced version of ID3. In Weka it is
called J48. It has some additional features like missing values are taken care of,
continuous attributes are converted, pruning of decision trees is done internally to
avoid overfitting and the rule can be derived. A heuristic approach is used for pruning
the Tree which is based on the statistical significance of splits. Many ways are defined
Application of Data Mining Algorithms for Tourism Industry 491
Fig. 2 Decision tree algorithm run in Weka
to select the attributes in the nodes of decision tree and to arrange these in order.
Information gain or entropy measurement is one of the simpler ways [14]. Information
gain is the measurement of much information/answer to a specific question that is
provided. It is the measure of how much uncertainty there is in the information. For
n number of possibilities with Pithe probability of the event i, the Entropy H(X)is
given by
H(X)=−Σn=1pilog 2 pi
Figure 2is the decision tree shown by C4.5. The database is selected and loaded
in Weka. Tree algorithms J48 with default setting was run on 110 records as Training
set of Indian_tourist database which classifies travelers as they have stayed at which
hotel type as per their attributes as Age, Types Of Tours, Sex, etc. Similarly, other
decision trees can also be built according to the type of attributes one wants to classify.
6.3 Analysis of Result
The generated tree is as shown below:
J48 pruned tree:
Type_Of_Tours =Religious: Dharamshala (20.0/9.0)
Type_Of_Tours =Historical
|Sex=Male: 4Star (21.0/11.0)
|Sex=Female
||Age<=51: 3Star (5.0/1.0)
492 P. Sharma et al.
| | Age > 51: 4Star (2.0)
Type_Of_Tours =PleasureLeisure
|Age<=51: 4Star (19.0/10.0)
| Age > 51: 5Star (3.0/1.0)
Type_Of_Tours =Adventure
|Sex=Male
||Age<=26: 5Star (3.0/1.0)
| | Age > 26: 4Star (7.0/2.0)
|Sex=Female: 3Star (7.0/2.0)
Type_Of_Tours =Wildlife: 4Star (13.0/5.0)
Type_Of_Tours =HeritageCultural: 4Star (10.0/4.0)
Each line represents a node in the Tree. The Root node is the Attribute without
any vertical bar ‘|’ to its left. All Other lines that start with a ‘|’ is a child node of the
node listed above its line which is its parent node. Therefore, if there is a node with
one or more ‘|’ characters before the statement of the rule then it is the child node
of the node immediately preceding that line. Leaf nodes are denoted by a colon (:)
and the class is identified by text after colon (:). Node that generates a classification
such as Age <=51: 3Star (5.0/1.0) are followed by two numbers. First number 5
tells how many instances set are correctly classified by this node where age <=51
and 1 tells that this is incorrectly classified by algorithm.
7 Analysis Using Clustering
7.1 Conceptual Framework
Clustering comes under the category of unsupervised way of learning. Clustering
algorithm is different but the idea is similar to classification. The difference lies
that classes are not predefined. We can only imagine the benefits of grouping
customers and merchandise, giving us insight into which customer is likely to
buy which product. There are two approaches to performing clustering tasks—
Hierarchical and Non-Hierarchical. Based on implementation technique, hierar-
chical has either agglomerative method or divisive method to build clusters. Non-
hierarchical clustering is used to perform data segmentation/market segmentation.
Some measurement of dissimilarity between segments and the degree of segmenta-
tion is required to be done. In this, three kings of algorithms are there partitioning
methods, density-based, and probability-based methods.
Simple K-Means is the most popular and commonly used partitioning method of
non-hierarchical clustering. The centroid of a cluster is defined as the mean value
of points within the cluster. First of all, it randomly selects kvalue of the objects in
D, which represents a cluster mean or center value. Then for the remaining objects,
Euclidean distance between the object and the cluster mean is calculated and the
Application of Data Mining Algorithms for Tourism Industry 493
object is assigned to the cluster if it is most similar or the distance is less. The
K-means algorithm then iteratively assigns objects and improves the within cluster
variation [15].
7.2 Clustering Using Weka
In Weka, the SimpleKMeans algorithm is used for clustering which can automat-
ically handle mix types of attributes whether they are categorical or numerical
attributes. The algorithm does distance computations and also normalizes the numer-
ical attributes. It uses Euclidean distance measurement to calculate distances between
instances and clusters [13].
7.3 Analysis
Following is the result of SimpleKMeans Clustering on Indian_tourist database.
Training data of 110 records and 7 useful attributes were selected. The number of
clusters initially is taken as 2 so it formed 2 clusters with following statistics (Fig. 3).
This is very useful for the future planning of tour packages which are more
saleable. There is another group of 3 clusters which was taken by reducing the
number of attributes as 4. From the demonstration, it is clear that the SimpleKMeans
algorithm is a straightforward algorithm for finding clusters in datasets (Fig. 4).
Fig. 3 Clustering result in Weka
494 P. Sharma et al.
Fig. 4 Clustering result in Weka
8 Conclusion and Future Work
The Paper has presented applications of different data mining algorithms using
popular tool for analysis WEKA. Each algorithm has different objectives to clas-
sify and analyze datasets. The overall objective of the paper is to evaluate the use of
data mining algorithms for the use of the tourism industry. This work used data of
Indian travelers who come to the capital of India, Delhi from different parts of the
country to visit Delhi and its excursions and avail tour packages according to their
interests. There always remains the scope of improvements and so it is here also with
this research. The same analysis can be done for in fact any tourism datasets. The
inbound tour operators who cater to NRIs and Foreigner Travelers can also use these
applications to analyze different datasets. Using data mining tools, especially in the
field of tourism can improve quality and customer preferences can be known. For
this person should have domain knowledge of both sectors: Tourism as well as data
mining technology. This certainly would bring in more tourists both from within the
country and from abroad. In future we can evaluate more datasets of the tourism
industry and draw a comparative study like which one is more efficient to access
patterns or behavior of tourists/travelers and under which circumstances.
References
1. J. Han, M. Kamber, Data Mining and Techniques, 2nd ed. (Simon Fraser University, 2006)
2. P.J. Sheldon, The Tourism Information Technology (CAB International, Wallingford, 1997)
3. M. Olsen, D. Connolly, Tour. Anal. 4(1), 29–46 (1999)
Application of Data Mining Algorithms for Tourism Industry 495
4. D. Junping, Z. Min, T. Xuyana, The realization of distributed sampling association rule mining
algorithm in tourism, in 7th World Congress on Intelligent Control and Automation. WCICA
2008 (2008), pp. 183–187. https://doi.org/10.1109/wcica.2008.4592921
5. R. Omar, A.O. Md Tap, Z.S. Abdullah, Web usage mining: a review of recent works, in 2014
the 5th International IEEE Conference on Information and Communication Technology for the
Muslim World (ICT4M) (2014). https://doi.org/10.1109/ict4m.2014.7020638
6. M.B. Karathiya, R.S. Sakshi, D.S. Sakshi, D.R. Kathiriya, Data mining for travel and tourism.
J. Inf. Oper. Manag. 3(1) (2012). ISSN: 0976-7754 & E-ISSN: 0976-7762
7. M. Danubianu, V. Hapenciuc, Improving customer relationship management in hotel industry
by data mining techniques, in Proceeding of Competitiveness and Stability in the Knowledge-
Based Economy, CD, 30–31 May 2008, Craiova, Romania (2008), pp. 2444–2452. ISSN/ISBN:
978-606-510-162-3
8. Bartwal, Incredible India Loosing Sheen (McClatchy—Tribune Business News, Washington)
9. D. Buhalis, eTourism: Information Technologies for Strategic Tourism Management (Financial
Times Prentice Hall, 2002)
10. N. Padhy, P. Mishra, R. Panigrahi, The survey of data mining applications & future scope. Int.
J. Comput. Sci. Eng. Inf. Technol. (IJCSEIT) 2(3) (2012)
11. G.G. Emel, Akat, Prolifing a domestic tourism market by means of association rule mining.
Anatolia (2007)
12. P. Rani et al., Association rule mining in discovering travel pattern in passport data analysis.
(IJCSIT) Int. J. Comput. Sci. Inf. Technol. 5(4), 5015–5019 (2014)
13. A.B.M. Shawkat Ali, S.A. Wasimi, C. Learning, Data Mining Methods and Techniques (2009)
14. E. Irawan et al., Application of classification algorithm C4.5 in recommendations for natural
tourism development in district Simalungun. J. Phys.: Conf. Ser. 1255, 012078 (2019)
15. N. Iswandhani, M. Muhajir, K-means cluster analysis of tourist destination in special region
of Yogyakarta using spatial approach and social network analysis. J. Phys.: Conf. Ser. 974,
012033 (2018), International Conference on Mathematics
Real-Time Safety and Surveillance
System Using Facial Recognition
Mechanism
Sachi Pandey, Vikas Chouhan, Rajendra Prasad Mahapatra,
Devansh Chhettri, and Himanshu Sharma
Abstract Face recognition mechanisms can be used in many diverse applications
such as Security, Retail, Healthcare, and Marketing. The webcams are often used as
a security standard with the integration of the facial recognition technology which
allows authorized users to unlock objects such as Computer, Laptop, and Mobile.
Cameras have a major role in recognition technology to track human faces and keep
a track of the number of people in a shot or a certain location such as in an entrance or
public places. This technology can be further narrowed down to the recognition and
tracking of eyes. In the proposed work, we used an existing Viola–Jones algorithm as
a basis for face detection to create a surveillance mechanism by analyzing captured
data. Besides, the e-complaint registration feature has been achieved by sending
security breach email alerts to the nearby police station or registered users. The
proposed system uses cloud storage along with cryptographic techniques to achieve
confidentiality, availability and provides surveillance against any uncertain incidents.
Keywords Face recognition ·Surveillance ·Availability ·Home security ·
Intruder detection ·Cloud storage
S. Pandey (B)·V. Chouhan
Department of Computer Science and Engineering, Indian Institute of Technology, Roorkee, India
e-mail: sachipandey.231@gmail.com
V. Chouhan
e-mail: vikaschouhan.iitr@gmail.com
S. Pandey ·R. P. Mahapatra ·D. Chhettri ·H. Sharma
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
Delhi-NCR, India
e-mail: Mahapatra.rp@gmail.com
D. Chhettri
e-mail: chhettridevansh@gmail.com
H. Sharma
e-mail: himanshu.sharma0756@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_44
497
498 S. Pandey et al.
1 Introduction
Face Recognition Security is an important biometry technology that captured the
digital images and evaluated the patterns to identify a particular person in a digital
snapshot. Biometric facial recognition technology is a highly effective mechanism
which mainly anchors on the same recognizer that a human primarily uses to differen-
tiate between one person’s face respective to another. The primary goals are to distin-
guish the humans’ faces to individualize the identification with high accuracy and to
interpret the byzantine human-based optic approach. The very first semi-automated
facial recognition system involves the system user or administrator to discover facial
features that are related in a pattern such as eyes, nose, mouth, and ear. On the digital
snapshot, first, we evaluate the difference in ratios based on a common remark point,
subsequently which were analyzed the remarked information.
The proposed system also deploys as a home security system to detect an intruder
and securely control the door access. In the home environment, the intruder’s motions
are identified and recognition is accomplished by capturing the live video feed from
the web camera and framing. The web camera captures the current frame when the
intruder motion is detected where the framing process implements on the captured
video frames. The captured image is sent automatically to the registered email users
and the nearby police station control rooms for further immediate necessary actions.
The proposed technique is implemented in multiple phases. The first phase recog-
nizes the intruders based on human face from the live video streaming for intruder
detection. In this phase, we exploit the Paul Viola and Michael Jones Algorithm.
In the second phase, it recognizes the human face, and then it transforms into a
digital image snapshot. In the third phase, the recognized image is automatically
sent to the respective registered user email ID. The automatic email sending process
is executed persistently until the user closes the program manually. The performance
of the current system is more desirable as compared to existing methods [1,7,11].
The proposed mechanism provides a concealed and continuous time covert secu-
rity recording feature, which is triggered when the camera is activated. The hidden
recording will be initiated along with image processing that runs in the background
and will be stored in the client’s local storage, such that it can be viewed later by
the user for surveillance purposes. We stored the captured footage in ‘avi’ extension
form. The footage will be recorded within the time frame of the image snapshot. Thus
the recording turns out to be in faster motion than being normal. Better accuracy in
recognition is realized with the existing proposed method [3]. In the face recognition
system, a major challenge is encountered inside the multidimensional visual model
and a trending extent of research. This system utilizes for accessing the door. The
facial recognition incorporates several features extracted from the face image, recog-
nition, and classification process [8]. In our work, we utilize the on-demand resource
availability of cloud computing, especially data storage. We store and retrieve the
captured video anytime, anywhere in ubiquitously using any computing devices such
Real-Time Safety and Surveillance System … 499
as Mobile, Laptop, and Desktop [9]. Therefore, we store the recorded data on the
cloud data centers in encrypted form to achieve data confidentiality along with high
availability.
2 Related Work
The authors in [7] discuss that several approaches have been proposed by different
individuals where some of them are an emphasis on high-resolution image surveil-
lance and automation. They utilize the single pan-tilt-zoom camera that enables
the zoom property, which helps in-depth analysis of the face images. The authors
proposed a system for traffic surveillance in [6] based on the background subtrac-
tion method, which is implemented in the MATLAB environment. Mishra et al.
[10] proposed the detection of moving objects with the help of Eigen-object from
the captured frames. They perform the frame pre-processing and processing for
better detection accuracy and construct the video frames, respectively. They subtract
the background images from the present images pixel by pixel. The multiple video
camera surveillance systems are discussed in [5]; they emphasized on how to handle
the complicated challenges based on real-world constraints. However, detection and
tracking, along with recognition features, are achieved by the adequate use of
geometric complexities. There are several algorithms that exist, out of them Viola–
Jones algorithm [2] is more vigorous and speedy for the detection of the faces. Haar
feature is used by the authors in [12] for detecting regions such as the nose, mouth,
and eye by adapting the Adaboost algorithm.
The aforementioned literature provides image detection, surveillance, and concen-
trates on the accuracy, but they did not consider the confidentiality, availability, and
accessibility of the stored information. Hence there is a need to develop a mechanism
that provides the surveillance features along with the confidentiality, availability, and
accessibility.
3 Problem Analysis
A highly effective and secure facial recognition mechanism is required in a real-
time scenario to capture the digital images and evaluating the patterns. Therefore,
the proposed system provides secure facial recognition mechanisms and surveil-
lance against any uncertain incidents using cloud storage along with cryptographic
techniques to achieve confidentiality and availability.
The preservation of the data or information is becoming essential and difficult.
Nowadays, the surveillance cameras are commonly present in Offices, Airports,
ATMs, Universities, Banks, and in any locality with a security system. The facial
detection system recognizes or verifies an individual from a digital image. However,
this approach is typically useful for security and surveillance purpose. Even though
500 S. Pandey et al.
the facial detection process identifies a face automatically in an image and extracting
its features, then it identifies regardless of expressive behavior, a sign of aging,
poor lighting, transitions, and different poses. The extraction of these features is
a strenuous task [4]. The semi-automated recognition started in the late 90s, and
it has recently caught the attention of the scientific association. Many facial anal-
ysis techniques with facial modeling methods have extremely advanced in the last
decade. However, the loyalty of the facial detection system and various constraints
structure still present a great challenge to the scientific association [13]. Face recog-
nition grasps various preferences over other biometric methods. It is natural, non-
obstructive, and quite easy to use. In a research study of the face recognition security
system, we consider the rapport of six biometric techniques (eye, finger, signature,
hand, face, and voice) with machine-readable travel documents. Figure 1shows
that the facial feature attains the highest level of percentage compatibility. In this
study, parameters like the enrolment, renewal, machine specifications, and public
perception were considered.
Figure 1illustrates the basic comparison of machine-readable travel documents
(MRTD) compatibility in percentage unit scale along with six different biometric
elements. The enhanced attention of automated facial detection systems has been
obtained from a domain other than the scientific community which is widely
widespread due to enhancing public concerns for security and safety, especially
due to the many occurrences of terrorists around the world after September 11,
2001. However, automated face detection or identification system can be utilized
in widespread areas other than typically security-based applications (for instance,
access-control/verification systems, surveillance systems) and such as computer
customized computer–human based interaction.
Fig. 1 Illustrates the basic
comparison of
machine-readable travel
documents compatibility in
percentage unit scale along
with six different biometric
elements
Real-Time Safety and Surveillance System … 501
We exploit the widely used Viola–Jones Algorithm for our proposed face
recognition system. The algorithm performs the following tasks:
Haar Detection: In this phase, the algorithm is to perform the Haar detection.
However, every human face has some common structural integrity. This integrity
can be identical and identified using feature detection recognized as Haar Detection.
Haar detection is applied to detect the eye region, nose bridge, location of eye, mouth,
etc. Each feature detection outcome is an isolated value which is determined by the
difference between black pixels and white pixels.
Creating Integral Image: The image representation is also known as the integral
image that assesses rectangular characteristics in constant time, which provides them
a significant speed benefit over more complex alternative features [6]. An integral
image at a particular location (x,y) contains the sum of all left pixels (x,y) and above
pixels.
AdaBoost Training: In this phase, the machine learning system algorithm is
applied to many other diverse kinds of learning languages to improve the rendering
of the system. AdaBoost serves for the feature detection and trains the classifier.
Cascading Classifier: The concept is constructed over the recognition process
where the detector works every time through the same frame image. The classifier is
worked to be fast in demand to be implemented on low-power CPUs, like cameras
and mobiles. Major features are used and operated in a repository, and the rest of
them are aligned in a cascade container to save some storage space.
4 Proposed Mechanism
The proposed security scheme (i.e., Human Facial Recognition Security) constructs
up on the frame subtraction method. The main tactic is to build a static frame series
(background frame), and then compare every upcoming frame of the series to the
static frame in sequence to identify the motion detection.
The web camera continuously streaming the video and then applies the corre-
sponding operations to determine the intruder as shown in Fig. 2. First, we captured
the frame, and then it accomplishes the sequencing of the frames. In the intruder
investigation module, we considered two cases. In the first case, if there are no
changes in further subframes, then the frame sequence is compared in the video
stream on the basis of its first or background frame. In the second case, if there are
any changes detected in subframes, then we can determine and handle the variation
of subframes of the captured streaming video. Further, the safety operations are initi-
ated, the intruder detection is thoroughgoing as rapidly as frame changes and auto
alert system forward the intruder frame to the registered user via email.
The face detection steps are classified and illustrated in Fig. 3a and the surveillance
and alert strategy steps are illustrated in Fig. 3b.
In Fig. 4we present the methodology of the proposed facial recognition system.
An input device such as webcam, external cam, CCTV, etc. can be used to capture the
video. After providing input from the camera, the system will automatically discover
502 S. Pandey et al.
Fig. 2 Investigation of intruder
Fig. 3 a Face detection steps. bSurveillance and alert strategy
faces in real time which is based on the Viola–Jones Algorithm. Haar uses structural
integrity to detect identical features, an integral image that assesses rectangular char-
acteristics in constant time, which provides them a significant speed benefit over
more complex alternative features.
Real-Time Safety and Surveillance System … 503
Fig. 4 The methodology of the proposed recognition system
Adaboost training is used for feature detection and to train the classifier lastly
cascading the classifier. Recorded video is forwarded to the cloud storage in encrypted
form to ensure confidentiality. We used the Secure Hash Algorithm to perform
the Encryption and Decryption operations. Consequently, the intruder investigation
process marks the intruder frame. After detecting faces in real time, it then captures
the frame and converts it into the desired image in encrypted form; then the image is
saved in the desired folder of local storage and also uploads this frame to the cloud
storage. Further, the alert system is activated, and it would be sent to registered user
email through Simple Mail Transfer Protocol (SMTP) protocol. Meanwhile, a hidden
recording will be initiated with image processing in a background area and will be
stored in the local client storage, such that it can be viewed later by the registered
user for surveillance purposes. The AVI encoder or decoder is used to convert camera
input into video output in ‘avi’ extension. The AVI record player will capture whole
image frames entirely until the process is stopped manually. The security footage
will be recorded within the time frame of the image snapshot. There would be an
alert of intrusion recognition on the given User email ID. The whole process is auto-
mated and would start again until the user stops it manually. Hence the mail transfer
is executed through the SMTP Server, where the SMTP is an Internet standard for
electronic mail transmission. In the SMTP protocol process, data (i.e., a document,
image, and file) would be sent to the mail server by using open ports of that server.
Every server has default communication open ports. These open ports are the way
of communication to the desired server.
504 S. Pandey et al.
5 Result Analysis
In this section, we discuss the implementation that incorporates a real-time safety and
surveillance system. Our implementation setup uses MATLAB computing environ-
ment, which combines the visualization with computer intended vision, to process
the captured frames. We set the input device module priority to one that enables the
camera as an input device.
Basically, in the case of the video in constant frame sequence mode (constant
background frame rate), the algorithm is executed and attempts to search for human
face’s structural integrity.
The algorithm marks the face boundary and captures the frame immediately when
the human visage or face is detected or the background frame rate changes which
are shown in Fig. 5.
Figure 6a shows that the camera detects the multiple faces precisely, there-
fore capable of detecting various faces simultaneously. Thus it captures the frame
immediately and processes it to the next phase.
Figure 6b demonstrates the captured frame in the form of the email notification.
The captured frame is transformed into an image and is automatically mailed to the
registered user email. This sending process will be continuous in every 10 s of a time
interval or frame delay.
The recording operation is run in background (hidden security recording) that
stored the captured video in local client storage, which can be accessed and played
Fig. 5 Face detected
Real-Time Safety and Surveillance System … 505
Fig. 6 a Multiple face detection. bUser Email inbox
anytime. This security recording can be played with any media player which supports
‘avi’ extension. Moreover, cloud storage allows the accessibility and availability of
the stored captured video ubiquitously.
6 Conclusions and Future Work
In highly restricted security areas, the face identification security structure is very
beneficial in the detection of intruders. This system helps in minimizing human need
and error detection. The security scheme comprises two primary modules. One is
the hardware module (tangible) and the other is the software module (Intangible).
The Tangible module entails of any camera, whereas the Intangible module
comprises face detection and recognition algorithms. When an individual arrives
at the deployed zone, a sequence of snapshots is taken via the installed camera and
directed to the registered user as an alarm indication. This process goes on every
10 s of the time interval. The proposed procedure is entirely automated and does not
involve any manual labor to tackle the situation.
Robust automated facial or expression identification system can be applicable
or be beneficial as individual identification arrangements at the examination, travel
documentation, grocery stores, and safety, and also criminal tracking. Nowadays,
face detection systems work very fast under restricted circumstances, although all
working methods nowadays are much better with frontal mug-shot snapshots and
incessant illumination.
Near in future generations, the person identification system will be more secure,
reliable, fast, and easy to identify people in real time and in extremely less restricted
conditions. It can be done possibly with the help of auxiliary support of the video
recording behind the auto capturing image frame. A hidden recording process
506 S. Pandey et al.
executes behind the image processing which leads to a covert recording that simulta-
neously works with image and video, which would be initiated as soon as the system
process starts. Hence, as a result, it is a reliable hidden technology that acts as a
clandestine for the human identification system. Additionally, our system stored the
user data in an encrypted form to achieve confidentiality. Moreover, the cloud envi-
ronment increases accessibility and data availability in a ubiquitous manner using
computing devices such as Mobile, Laptop, and Desktop.
References
1. K. Bonsor, R. Johnson, How facial recognition systems work. HowStuffWorks. Com Np 4
(2001)
2. M.K. Dabhi, B.K. Pancholi, Face detection system based on Viola-Jones algorithm. Int. J. Sci.
Res. (IJSR) 5(4), 62–64 (2016)
3. N.T. Deshpande, S. Ravishankar, Face detection and recognition using Viola-Jones algorithm
and fusion of PCA and ANN. Adv. Comput. Sci. Technol. 10(5), 1173–1189 (2017)
4. P. Divyarajsinh, B. Mehta, Face recognition methods and applications. Int. J. Comput. Technol.
Appl. 4(1), 84–86 (2013)
5. J. Doring, H. Sprave, Multiple video camera surveillance system, U.S. Patent 10/371,192, 16
Oct 2003
6. C.H. Low, M.K. Lee, S.W. Khor, Frame based object detection—an application for traffic
monitoring, in 2010 Second International Conference on Computer Research and Development
(IEEE, 2010), pp. 322–325
7. N.A. Manap, G. Di Caterina, J. Soraghan, V. Sidharth, H. Yao, Smart surveillance system based
on stereo matching algorithms with IP and PTZ cameras, in 2010 3DTV-Conference: The True
Vision-Capture, Transmission and Display of 3D Video (IEEE, 2010), pp. 1–4
8. R. Manjunatha, R. Nagaraja, Home security system and door access control based on face
recognition. Int. Res. J. Eng. Technol. 4(3), 437–442 (2017)
9. P. Mell, T. Grance, The NIST definition of cloud computing (2019), http://dx.doi.org/10.6028/
NIST.SP.800-145. Accessed 12 May 2019
10. S.K. Mishra, K.S. Bhagat, Human motion detection and video surveillance using matlab. Int.
J. Emerg. Res. Manag. Technol. 5(5) (2016)
11. J. Pagliery, FBI launches a face recognition system. CNN.com (2014)
12. H.M. Phuong, L. Dung, T. de Souza-Daw, N.T. Dzung, T.M. Hoang, Extraction of human
facial features based on haar feature with Adaboost and image recognition techniques, in 2012
Fourth International Conference on Communications and Electronics (ICCE) (IEEE, 2012),
pp. 302–305
13. S. Tathe, A.S. Narote, S.P. Narote, Human face detection and recognition in videos, in
2016 International Conference on Advances in Computing, Communications and Informatics
(ICACCI) (IEEE, 2016), pp. 2200–2205
Liver Disease Prediction Using
an Ensemble Based Approach
B. Muruganantham, R. P. Mahapatra, Kriti Taparia, and Mukul Kumar
Abstract Lately, the usage of Information system and strategic tools in the domain
of medical science is constantly growing. The liver is an Exocrine Gland that helps
in the digestion process, especially for fats along with altering the pH of food due to
its alkaline nature. One of the most common signs for the majority of liver disease is
hyperbilirubinemia which is very difficult to identify at an early stage. While certain
diseases such as obstructive Jaundice and acute viral hepatitis present themselves
with an early rise of bilirubin along with yellowish discoloration of the skin, many
other diseases don’t usually present themselves with an early rise of serum bilirubin
or skin discoloration because of which sometimes liver disease are overlooked or
misdiagnosed in primary level. However, serum bilirubin is not the only way to
diagnose liver disease because it is not specific. The most specific way to diagnose
liver disease is by liver function test. With the help of the detection of the enzyme
level, we can identify and confirm the presence of liver disease and intensity of
liver damage when coupled with suitable imaging modalities like Ultrasound, CT
scan, or MRI scan. The dataset contains patient parameters such as Age, Sex, Total
Proteins, Alkaline Phosphatase, Alanine Phosphatase, Total Proteins, Total Bilirubin,
Albumin, Albumin and Globulin Ratio, Direct Bilirubin, and the Result. We are using
Binary Classification which is basically classifying the element of given set into two
given sets i.e., Patient suffering from Liver disease or not. We will try to use an
Ensemble Based Approach to find the best prediction accuracy.
B. Muruganantham (B)·K. Taparia ·M. Kumar
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
Kattankulathur, Chennai, Tamil Nadu 603203, India
e-mail: muruganb@srmist.edu.in
K. Taparia
e-mail: krititaparia@gmail.com
M. Kumar
e-mail: mukulsharma1998@gmail.com
R. P. Mahapatra
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
Ghaziabad, India
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_45
507
508 B. Muruganantham et al.
Keywords Liver disease ·Prediction ·Classification ·Accuracy
1 Introduction
The liver is the largest gland of the body. Liver Disease is a very tricky disease to
diagnose. It is very hard to diagnose it at an early stage based on its symptoms. It
is an Exocrine Gland. This organ is present in the right hypochondriac region of
the abdomen. It serves many important functions such as digestion, storing glucose,
metabolizing toxin. Fatty Liver Disease which occurs due to fat build-up in the liver
is also one of the major liver diseases. It is a very prevalent disease accounting for
more than 10 million cases per year in India. It usually causes no symptoms and
needs a test to determine if a person is suffering from the disease.
Artificial Intelligence (AI) is the ability of computer programs to think learn and
apply the knowledge it has gained from training in various fields. Nowadays artificial
intelligence is applied in almost every field of life and it consists of various parts
such as Machine Learning and Deep Learning which further divides into many parts.
Artificial intelligence proceeds in its course of action in such a way as to obtain goal
maximization at each stage.
Machine Learning, is a sub-domain of artificial Intelligence basically involves
techniques which help the machine experience to learn to take decisions and respond
in certain scenarios based on the prior information it has gained. It consists of two
types: Supervised Machine Learning and Unsupervised Machine Learning.
Supervised Machine Learning is a learning in which we train our machine by
providing labeled datasets that means data is already well labeled with correct
answers. This data is known as training data and the data which we use to test if
our model has learned and giving expected result is known as testing data. In this
type of learning, we first train our model using training data and after training, we
test our model using test data. Some examples of Supervised Machine Learning are
Linear regression, Random Forest, and Support Vector Machine
Unsupervised Machine Learning is self-supervised learning which does not
contain training dataset it finds its solution using previous unknown patterns available
in the dataset. It can also be further grouped into clustering and association.
Semi-Supervised Learning—Some data is labeled and most of the data is unla-
beled, i.e., combination of testing and training data. It is a mixture of both Supervised
and Unsupervised Machine Learning Techniques in the sense that testing and training
datasets are not explicitly specified but are together.
2 Related Work
Nahar and Ara [1] have used various algorithms related to decision tree such as
Random Forest, Random Tree, Decision Stump, J48, LMT, and Hoeffding tree for
Liver Disease Prediction Using an Ensemble Based Approach 509
the classification process and it has been compared on the performance measures
such as Mean Absolute Error, Precision, Accuracy, Recall, and Runtime and shows
that Decision Stump has high accuracy. Jacob et al. [2] have used SVM, Logistic
Regression, K-NN, and ANN with 10 input neurons showing that ANN is highly
efficient. Sontakke et al. [3] have proposed the use of genetic microarrays along with
neural networks for the prediction showing that the molecular biology approach can
strengthen the predictive power. Kefelegn et al. [4] have shown that Random Forest
algorithm is not suitable due to the problem of overfitting and has suggested the
utilization of oversampling methods to address this problem. Thirunavukkarasu et al.
[5] have shown that both Logistic Regression and K-NN generated equal accuracy but
in medical biology, sensitivity plays a key role in differentiating the performance.
Hamid et al. [6] have proposed the use of the abstention paradigm to ensure that
no result is generated when the model is not confident about its prediction so as
to avoid the wrong classification. Ayeldeen et al. [7] have shown that biochemical
markers have a huge impact on the positive prediction of different stages of liver
fibrosis and HA level increases with increase in the grade of fibrosis. Belavigi et al.
[8] have proposed that in comparison to Rprop and SAG on text dataset and CNN
on image dataset, CNN gives high accuracy. Kumar et al. [9] have used different
classification algorithms such as Random Forest, Naive Bayes, K-NN, K-Means,
and C 5.0 proposing that C.5.0 works best after adaptive boosting. Hashem et al.
[10] have used various inclusion and exclusion parameters to determine which data
can be used for the training of the model to ensure that irregular data patterns do not
affect the classification. Sontakke et al. [3] have proposed the usage of unsupervised
learning methods such as K-Means, Affinity Propagation, and DBSCAN rendering
Silhouette coefficient as the primary factor for judging the performance (Table 1).
3 Proposed Methodology and Overview of System
Architecture
This proposed work is to process the datasets into the training and testing part and
apply various techniques to find the best approach. The architecture of the proposed
work as shown in Fig. 1, shows the various modules of our proposed system.
3.1 Data Cleaning
Data Collection contains a collection of datasets of various patient data. The data
collection contains a collection of patients various parameters and details. It is basi-
cally the first step to collect as much data as possible. The model will benefit propor-
tionally to the size of dataset, therefore the larger the size of the dataset the more the
model learns from it.
510 B. Muruganantham et al.
Tabl e 1 Related work
Ref. No. Proposed objective of
work
Algorithms
used/architecture
developed
Dataset/input
parameters used
Applicability Performance measures
used
Limitations found
[1]Prediction of Liver
Disease with the help
of different techniques
of Decision Tree
1. J48
2. REP Tree
3. Random Tree
4. Decision Stump
5. LMT
6. Random forest
7. Hoeffding Tree
UCI Machine
Learning Repository
For the given
parameters, Decision
stump has proved to
provide the highest
accuracy as compared
to other algorithms
1. Accuracy
2. MAE
3. PRE
4. REC
Does not use
advanced decision
trees such as CART
[2]Machine learning
algorithms for the
detection of liver
diseases
1. Logistic Regression
2. K-NN
3. ANN
4. SVM
Indian Liver Patient
Dataset (ILPD) (UCI
Machine Learning
Repository)
A back propagation
neural network with
10 input neuron layers
serves as the model
with highest accuracy
1. Accuracy
2. Sensitivity
3. Precision
4. Specificity
The results obtained
can be varied based
on the number of
neurons in the input
layer
[3]Prediction of Liver
Disease using various
approaches of
Machine Learning
1. Back Propagation
2. Support Vector
Machines
Algorithm
Indian Liver Patient
Dataset (ILPD) (UCI
Machine Learning
Repository)
Microarray analysis
and comparison of
two algorithms
1. Accuracy
2. Sensitivity
3. Precision
4. Specificity
Natural factors, for
example, age, sexual
orientation, ethnicity
and diet keep on
influencing the
consistency of the
outcomes
[4] Usage of data mining
algorithms on
imbalanced data for
liver disease prediction
1. Logistic Regression
2. Random Forest
3. Auto Neural
Network
4. K-NN
Datasets of liver
patients from Andhra
Pradesh region of
India
Imbalanced data
handled with the help
of sampling method
1. Accuracy
2. Train rate
3. Train-ROC
Problem of
overfitting in random
forest algorithm due
to lack of data set
(continued)
Liver Disease Prediction Using an Ensemble Based Approach 511
Tabl e 1 (continued)
Ref. No. Proposed objective of
work
Algorithms
used/architecture
developed
Dataset/input
parameters used
Applicability Performance measures
used
Limitations found
[5]Usage of different
classification
algorithms for liver
disease prediction
1. Logistic Regression
2. Support Vector
Machines
3. K-Nearest Neighbor
Indian Liver Patient
Dataset (ILPD) (UCI
Machine Learning
Repository)
Given the set of input
parameters, Logistic
Regression and
K-Nearest Neighbor
have been shown to
have the highest
accuracy whereas
logistic regression
gives the highest
sensitivity
1. Accuracy
2. Sensitivity
3. Specificity
Does not consider the
precision of results
as a performance
measure for the
algorithms
[6]Automated detection
of liver abnormalities
with the use of
ultrasound images
1. Nearest Neighbor
(NN)
2. Support Vector
Machine (SVM)
3. Learning with
Abstention (LWA)
Dataset involves a
total of 99 liver
ultrasound images
comprising of 56
images of liver
abnormalities example
FLD and 43 images of
healthy individuals
Compares usage of
LWA with
conventional
classification
1. Specificity
2. Sensitivity
Needs to incorporate
working on large
independent test set
with elastography
data
(continued)
512 B. Muruganantham et al.
Tabl e 1 (continued)
Ref. No. Proposed objective of
work
Algorithms
used/architecture
developed
Dataset/input
parameters used
Applicability Performance measures
used
Limitations found
[7] Decision Tree
Approach for the
prediction of different
stages of Liver Fibrosis
1. Decision Tree dataset includes
laboratory tests
conducted and fibrosis
markers collected
from the Department
of Medical
Biochemistry and
Molecular Biology,
Faculty of Medicine,
Cairo University
The study shows that
HA level increases
with the advancement
in the stage of liver
fibrosis and is a
crucial marker for
prediction
Accuracy Other biomarkers
need to be considered
for significance of
calculation and also
increase the sample
population
[8]Comparison of
Prediction of Liver
Disease using Rprop,
SAG on text dataset
and CNN on image
dataset
1. Resilient Back
propagation Neural
Network (Rprop)
2. Stochastic Average
Gradient (SAG)
3. Convolutional
Neural Network
(CNN)
Indian Liver Patient
Dataset (ILPD) (UCI
Machine Learning
Repository)
Shows that the image
dataset used for CNN,
including the
ultrasound images of
the liver gives the best
accuracy
1. Accuracy
2. Sensitivity
3. Precision
4. Specificity
5. Misclassification
rate
CNN get issues of
low accuracy
(continued)
Liver Disease Prediction Using an Ensemble Based Approach 513
Tabl e 1 (continued)
Ref. No. Proposed objective of
work
Algorithms
used/architecture
developed
Dataset/input
parameters used
Applicability Performance measures
used
Limitations found
[9] Data mining
techniques for liver
disorder analysis and
diagnosis
1. K-Nearest Neighbor
Algorithm
2. K-Means
3. Naive Bayes
4. C5.0 Algorithm
5. Random Forest
UCI repository Shows that Random
Forest Algorithm
performs best when
no changes are
applied but after
adaptive boosting the
C5.0 algorithm
provides an
increasingly accurate
result
1. Accuracy
2. Recall
3. Precision
Different algorithms
work best with
respect to different
performance
measures used
[10] Prediction of advanced
liver fibrosis in chronic
hepatitis C patients
using various machine
learning approaches
1. Decision Tree
Learning
Algorithms
2. Genetic Algorithms
3. Particle Swarm
Optimization
Cohort data of 39,567
chronic hepatitis C
patients enrolled in
Egyptian National
Committee for the
Control of Viral
Hepatitis database in
National Treatment
Program of HCV
patients in Egypt
Approach to develop
the classification
models by combining
the clinical
information and
serum biomarkers
1. Accuracy 2.
Sensitivity
3. Specificity
4. ROC analysis
Low sensitivity of
the ADT model
514 B. Muruganantham et al.
Fig. 1 Architecture diagram of the proposed system
3.2 Data Preprocessing
Data Preprocessing is the next crucial step after Data collection. The real-world data
available is not fit for direct use as it contains a lot of noisy, incomplete and inconsis-
tent data so we must clean it before we can use it. Data preprocessing is a method to
convert the raw data into useful data for further processing. Data preprocessing help
us get better results when we apply classification algorithms to the data. The various
stages involved in Data Preprocessing are Data cleaning, Data integration, and Data
transformation.
3.2.1 Data Cleaning and Data Integration
Data Cleaning is the process that involves the removal of incomplete, corrupt, or
inaccurate data from the dataset for example if a patient does not contain some
important parameters so data cleaning is used. Data integration basically involves
combining the various data residing at different sources into a single unified view.
Liver Disease Prediction Using an Ensemble Based Approach 515
3.2.2 Data Transformation
Data Transformation is used to convert the image from one form or structure to
another.
3.3 Classification
As the name suggests Classification is the process of classifying. The main goal is
to predict the result accurately or predict the target goal accurately. A classification
model can be used in many real-life applications. It is the initial phase that uses
algorithms like Decision Tree, Random forest etc.
3.3.1 Decision Tree Algorithm
Decision Tree is one of the examples of supervised machine learning which is used
for classification problem. Ayeldeen et al. [7]. In this technique, we split the sample
into two or more homogeneous sets which are based on input variables which are
most significant and it also includes a node, branches, and leaf nodes. As shown in
Fig. 2, each depicts different things such as internal node shows test on attributes,
branch shows the outcome and leaf node holds a class label. The nodes expand on a
branch having Higher Information Gain.
Fig. 2 Decision tree algorithm
516 B. Muruganantham et al.
Fig. 3 Random forest algorithm
3.3.2 Random Forest Algorithm
The algorithm works in the following manner as shown in Fig. 3.
1. Given there are n cases in the preparation dataset, from these n cases, sub-tests
are picked indiscriminately with substitution. These irregular sub-tests browsed
the preparation dataset are utilized to assemble individual trees.
2. Expecting there are kfactors for input, a number mis picked with the end goal
that m<k[1]. mfactors are chosen haphazardly out of kfactors at every hub. The
split which is the best of these mfactors is picked to part the hub. The estimation
of mis kept unaltered while the forest is developed.
3. Each tree is developed as huge as conceivable without pruning.
4. The class of the new item is anticipated dependent on most of votes got from the
blend of all the choice trees.
3.3.3 Support Vector Machine Algorithm
SVM is an example of supervised machine learning algorithm that is used for
analyzing data that is required for classification and regression analysis [2]. These
models intently look like neural systems. Consider an Xdimensional dataset.
Utilizing SVM, the data can be plotted into an Xdimensional space. The preparation
information focuses are then plotted into kvarious areas depending on their names
as characterized by hyper-planes of Xvarious dimensions [3]. After the testing stage
Liver Disease Prediction Using an Ensemble Based Approach 517
is finished, the test focuses are plotted in a similar Ndimensioned plane. Contingent
upon which area the focuses are situated in, they are fittingly ordered in that region
SVM. The fundamental objective is to arrange all testing vectors into two classes so
it can recognize them and give its very own predictions.
4 Performance Evaluation
The comparison of the results has been done based on the performance measures
of prediction such as Accuracy, Sensitivity, Specificity, Precision, and Recall. We
will perform various evaluations to differentiate the techniques among each other.
The performance and its measures are compared statistically using different types of
plots to view their performance against all performance measures.
5 Conclusion
The objective of this paper is to compare different classification algorithms on the
dataset and compare their results based on different performance measures specified.
This comparison will help in identifying which algorithm works best with this type
of text dataset with the specified set of input parameters. This model can aid medical
experts in accurate liver disease prediction as it is not easily identifiable in its early
stages. A GUI will be incorporated for this model to provide a user-friendly interface
for all medical professionals to be able to use this tool for running predictions on the
patients. In future usage, this proposed system can be used for dynamic and more
complex datasets of a much larger volume and with more number of parameters
taken into consideration to make the model perform more accurately.
References
1. N. Nahar, F. Ara, Liver disease prediction by using different decision tree techniques. Int. J.
Data Min. Knowl. Manag. Process (IJDKP) 8(2) (2018)
2. J. Jacob, J. Chakkalakal Mathew, J. Mathew, E. Issac, Diagnosis of liver disease using machine
learning techniques. Int. Res. J. Eng. Technol. (2018)
3. S. Sontakke, J. Lohokare, R. Dani, Diagnosis of liver diseases using machine learning, in
International Conference on Emerging Trends & Innovation in ICT (ICEI) (2017)
4. S. Kefelegn, P. Kamat, Prediction and analysis of liver disorder diseases by using data mining
technique: survey. Int. J. Pure Appl. Math. (2018)
5. K. Thirunavukkarasu, A.S. Singh, Md. Irfan, A. Chowdhury, Prediction of liver disease using
classification algorithms, in 4th International Conference on Computing Communication and
Automation (ICCCA) (2018)
518 B. Muruganantham et al.
6. K. Hamid, A. Asif, W.A. Abbasi, D. Sabih, Machine learning with abstention for automated
liver disease diagnosis, in International Conference on Frontiers of Information Technology
(2017)
7. H. Ayeldeen, O. Shaker, G. Ayeldeen, K.M. Anwar, Prediction of liver fibrosis stages by
machine learning model: a decision tree approach, in Third World Conference on Complex
Systems (WCCS) (2015)
8. D.H. Belavigi, G.S. Veena, D. Harekal, Prediction of liver disease using Rprop, SAG and CNN.
Int. J. Innov. Technol. Expl. Eng. (IJITEE) 8(8) (2019). ISSN: 2278-3075
9. S. Kumar, S. Katyal, Effective analysis and diagnosis of liver disorder by data mining, in
Proceedings of the International Conference on Inventive Research in Computing Applications
(ICIRCA 2018)
10. S. Hashem, G. Esmat, W. Elakel, S. Habashy, S.A. Raouf, M. Elhefnawi, M. El-Adawy, M.
El Hefnawi1, Comparison of machine learning approaches for prediction of advanced liver
fibrosis in chronic hepatitis C patients, IEEE/ACM Trans. Comput. Biol. Bioinform. 15(3)
(2018)
Significance of Route Discovery Protocols
in Wireless Sensor Networks
Guntupalli Gangaprasad, Kottnana Janakiram,
and B. Seetha Ramanjaneyulu
Abstract Wireless Sensor Networks (WSN) have numerous applications in various
fields like military, navigation, agriculture, health care sector, etc., due to their data
sensing and transmission abilities. In spite of their wide range of applications in many
fields, finding appropriate methods for data gathering, routing, enhancing network
lifetime, and efficient communication between the sensor nodes have become the
issues of continuing research during the last two decades. Among them, finding
the best route from source to destination is one of the prime issues. Many routing
mechanisms are proposed in the literature and implemented, but there is no single
routing mechanism that addressed all the issues like throughput, latency, reliability,
and energy efficiency. This article provides a review of the routing techniques in
wireless sensor networks, along with the discussion of the issues they addressed.
Keywords Wireless sensor networks ·Routing protocols ·Network life time ·
LEACH ·PEGASIS
1 Introduction
A wireless sensor network comprises a large number of sensor nodes placed in the
field. The sensor nodes are of small size, limited battery, less computing power, and
small memory for data storing. These sensors sense the physical information like
temperature, pressure, humidity, moisture, etc., from the surroundings, and forward
it to the base station or sink node, through many intermediate nodes. At the base
G. Gangaprasad (B)·K. Janakiram (B)·B. S. Ramanjaneyulu (B)
Department of ECE, VFSTR University, Guntur, AP, India
e-mail: Ganga.prasad96@gmail.com
K. Janakiram
e-mail: Janakiram006@gmail.com
B. S. Ramanjaneyulu
e-mail: ramanbs@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_46
519
520 G. Gangaprasad et al.
Sensor
Nodes
Base Station
Wireless Communication Link
Target
Sensing
Field
Gateway
Internet
User
Sink
Fig. 1 Wireless sensor network architecture
station, the processing is performed on the data and appropriate action will be taken
like forwarding it to remote locations through devices like gateways [1] (Fig. 1).
The WSNs have gained a lot of importance in many fields of applications because
they can be deployed for many useful purposes in the real world. The applications
of these networks are mainly categorized into two types, namely monitoring appli-
cations and tracking applications [2,3]. The first category is used for inhouse and
outside environment monitoring, Seismic and structural monitoring, physical condi-
tion monitoring, and control monitoring, whereas the second category is used for
tracking the vehicles, human beings, animals, etc.
From the application point of view, WSNs are used in the broad categories of the
following:
•Water quality testing
•Environmental monitoring
•Military applications
•Disaster relief management
•Soil moisture monitoring
•Health care monitoring
•Body area networks
•Home applications, etc.
Though WSNs have many applications in different fields, they are still facing
many issues. Due to limited power available with the sensor nodes, they are unable
to stay alive for longer durations. There are a number of design issues and challenges
which degrade the performance of the WSNs. The most important are:
•There is no prior topology to sensor networks; once placed into the field they
setup themselves only. Human involvement to place them in a specific topology
may not be possible in many cases.
•These networks are infrastructure-less and hence distributed routing mechanism
is required.
Significance of Route Discovery Protocols … 521
•Sensor nodes depend on their battery power and there is no chance to recharge
the battery frequently. So energy efficiency is important for long-time survival.
•Designing of sensor node hardware that minimizes energy consumption is needed.
•Providing communication among all the nodes in the field is essential. Otherwise,
it will form isolated nodes, which is an indication of non-reachable nodes in WSN.
In sensor networks, the best routing method is expected to minimize energy
consumption at all nodes to improve the network lifetime. Along with energy savings,
the best routing protocol should also offer the best throughput, minimum latency, and
guaranteed delivery of data. Extensive research is going on to find efficient routing
techniques. So many routing protocols are already proposed in the literature, but no
single protocol solves all the issues. Designing efficient routing protocols for WSN
is a challenging task for researchers from the last two decades.
The main scope of this paper is to provide a survey on energy-efficient routing
protocols and their classification existed at present. Any route-finding protocol in
sensor networks falls under one of the categories like Network Structure, Commu-
nication model, Topology based, and Reliable Routing Scheme [4]. Among all the
above-stated methods Network Structure protocols offers the best energy efficiency
in the network because they implement cluster-based routing. They are also more
scalable compared to all traditional routing protocols.
The remaining part of the paper is framed in the following way. Challenging issues
of wireless sensor networks are presented in section-2. Information about existing
routing protocols of WSNs is presented in Sect. 3. In Sect. 4, a detailed discussion
of various routing protocols is carried out. The conclusion of the paper is presented
in Sect. 5.
2 Challenges of Designing Wireless Sensor Networks
WSNs have a wide range of applications in various fields, but at the same time, they
are plagued with many problems also. In this section, some important challenges of
WSN are presented. These are listed below.
Quality of service Reliability, availability, throughput, and latency are some of the
quality attributes that a network can offer. They are usually described in terms of the
metrics associated with them. Appropriate routing, clustering, and channel access
mechanisms will help in achieving the quality of service expected from the given
WSN.
Fault Tolerance When the nodes in WSN are damaged or run out of energy that leads
to permanent failure of communication between the nodes. So it is important that
such faults should be addressed immediately or precaution taken for their avoidance.
Also, the network needs to be employed with some pre-tolerant methods.
522 G. Gangaprasad et al.
Lifetime In WSNs, once nodes are placed, it is highly impossible to replace the
drained batteries. So, the nodes are considered with limited energy. But, in addition
to its own data transmissions, every sensor node in WSN also acts as a relay node to
forward the data of neighbor nodes to the destination or base station. So the possibility
of energy drain is high and hence achieving enhancement of lifetime in the network
is a challenging task.
Scalability Due to the requirement of WSN with a varied number of nodes for each
application, network scalability is important. In spite of it, effective communication
has to be maintained between the nodes.
Programmability The nodes employed in WSN are not only for sending the pre-
decided information to the sink. In addition to that, they may also need to perform the
tasks for changing requirements of a later time. To facilitate this, the nodes should be
programmable on-the-fly, after their deployment also. They also need to reconfigure
as per changing topologies so that the network will perform in an optimum manner.
So, the nodes in the network should be programmed to fulfill the above.
Maintainability In WSN, some nodes may get drained quickly due to their higher
quantum of work. It may lead to the failure of ongoing communication because
the route that passes through this drained node may disappear. In that scenario, the
network should automatically find new routes. It must also monitor its health and
respond to the disturbances in the ambiance.
Multi-hop wireless communication The sensors forward the sensed content to the
destination via multiple nodes. It makes energy consumption less compared to the
direct communication to sink from all the nodes.
Energy-Efficient operation Designing of each protocol in an energy-efficient
manner is the desired approach. For that, it needs to address many constraints.
Collaboration and in-network processing The sensed data from all sensors in
the network need not be sent in the same for the sink, because some redundant
data may exist. So network should ensure that data aggregation is done prior to
transmission of data to the sink, thereby duplicate or redundant data is reduced. To
do this collaboration among all the nodes is required.
3 Related Work
The energy-efficient routing protocols are implemented based on the network model
and application needs. Different factors are taken into consideration while designing
routing protocols like power, memory size, communication, routing, and network
lifetime. Among all, routing is an important issue for energy-efficient sensor networks
Significance of Route Discovery Protocols … 523
to improve the network lifetime. There are many proposals available in the litera-
ture for energy minimization protocols in WSN. In [1], the authors explained some
protocols belonging to the quality of service and their metrics in detail. In [2], the
authors provide details about the networking of wireless sensors and their issues in
detail. In [3], the authors describe various applications of WSNs and their issues
while designing real networks. In [5], energy-aware protocols belonging to the
QoS concept and their techniques in WMSNs are explained. In [4], the authors
provide well-defined methods for energy efficiency and internal classification with
their applications and limitations in detail. In [6], the authors made an extensive
survey of design issues and techniques for WSN. They explained the constraints of
sensor nodes and networks stack. Then, they discussed various applications of WSNs.
In [7], the authors presented various routing concepts in WSN. They describe the
routing concepts based on network structure into three types i.e., flat, hierarchical,
and position-based routing methods. And more ever, these protocols are categorized
into energy, multipath, negotiation, and QoS-type routing protocols. Different energy
minimization methods for sensor networks are provided in [8]. These methods are
designed based on data-centric, location-based, and hierarchical.
In [9], the authors present a classification of protocols based on broad-
cast/multicast problem is in WSN. The proposed minimum energy broad-
cast/multicast is aimed at reducing the total transmission energy usage at all sensor
devices. In [2], the authors present multiple applications and different aspects of
WSNs. These aspects are divided into the following types: underlying operating
system, a communication protocol stack, internal platform, services of network, and
placement issues. In [10], the authors presented details about the quantum of battery
energies used for various parts of the node. The hardware of the wireless sensor device
is divided into four parts: a subsystem for sensing, a computing subsystem including
microcontrollers and data storing, a unit for communication, and a battery unit. This
article explained the characteristics and advantages of energy consumption schemes
in detail. In [11] the author presented a categorization of route-finding methods and
issues in WSN. This work explained a number of route-finding methods briefly. In
[12] the important challenging design parameters of medium access control proto-
cols in WSN are presented. It describes many MAC protocols with their strengths
and weaknesses. It didn’t provide any energy-efficient routing protocols for network
lifetime improvement in WSN. In [13], the authors presented routing protocols for
WMSNs. They explained some design parameter issues of routing protocols for
WMSN and their issues.
4 Classification of Routing Methods in Sensor Networks
There are a number of routes discovering methods in wireless sensor networks.
Generally, these routing protocols are classified into the following types [4].
(A) Network structure.
524 G. Gangaprasad et al.
(B) Communication model.
(C) Topology type.
(D) Reliable routing.
(E) Nature of nodes.
(A) Network Structure
The network structure can be classified based on how the nodes are placed in the
network and connected to each other and to the sink. The protocols under this classi-
fication describe the way nodes are interconnected to provide the routing according
to the network structure. The models in this type are (a) Flat protocols, and (b)
Hierarchical protocols.
(a) Flat Protocols
In this category of protocols, all the nodes have the same level of connectivity in the
network. This type of network uses to minimize the overhead for communicating
nodes [4]. The protocols under this category are described below.
Proactive Protocols Here complete route-finding information is provided with the
help of routing tables at each node. Through the periodic exchange of data between
nodes, it builds a routing table at each node. In these protocols, the bandwidth require-
ment is high and extra battery power is required to maintain routing tables. Here a
good number of protocols have existed and one of them is explained below.
Wireless Routing Protocol (WRP)is a table-based method, which is implemented
by using the Bellman-ford algorithm [14]. In this method, each node focuses on
checking the route content received repeatedly from all neighbor nodes. These proto-
cols are not useful for very large scale WSN. These proactive protocols require more
bandwidth.
Reactive protocols This type of protocol works based on on-demand requests only.
Whenever a particular request is raised, that time only, the node will initiate the
route-finding process by sending/receiving route request/route reply information,
respectively. The below-mentioned protocol comes under this category.
The temporary ordered routing algorithm is distributed in nature and utilizes the
principle of link reversal with the loop-free mechanism. This protocol is implemented
with the help of topologies of the network to the minimization of communication
overhead for decreasing power usage. In this flat type of protocols, some more are
proposed for broadcasting, flooding, rumor routing, and energy-aware temporary
ordered routing algorithm. So all are used to minimize the energy and to improve the
network lifetime.
(b) Hierarchical Based Routing
In this routing protocol, the entire network is divided into different clusters, each
cluster maintaining one cluster head which is responsible for sending data to the
Significance of Route Discovery Protocols … 525
base station from all sensors in that cluster [4]. For energy efficiency purposes, hier-
archical routing is used. It uses two stages approach. The first stage is to make clusters
and elect cluster leaders based on some threshold value whereas the second stage
is about transferring data from cluster members to the sink through various cluster
leaders. The communication in hierarchical routing networks is through single-hop
communication and multi-hop communication. In the former case, the communi-
cation is between cluster members and cluster head whereas in the latter case the
communication is between cluster heads and the base station. Some of the advantages
of Hierarchical based routing are [15]:
•Reduction in transmission cost and energy saving.
•Improvement in network Scalability.
•Reduction in routing table size at each sensor node.
The protocols of this type are discussed below.
Low Energy Adaptive Clustering Hierarchy (LEACH) is implemented to mini-
mize energy consumption. In LEACH, cluster heads are selected based on threshold
value in different random rounds. For communicating data to the cluster leader,
respective TDMA time slots are provided. The aggregated data is sent to the sink
node in CDMA format to avoid the interference in the medium. This protocol is
implemented in two stages, namely the Setup stage, and Data transmission stage.
In the setup stage, cluster creation, and cluster head election are carried out. In data
transmission stage aggregation of data is performed. Aggregated data is transferred
to the destination through various cluster heads. Each sensor node provides various
random values between 0 and 1. If a sensor node gets a value less than the threshold
value, that node becomes a cluster leader for that round. If any node acted as cluster
head, it is not allowed to become a cluster head one more time for that round. The
formula for selection of cluster head is based on threshold value given by,
T(n)=
P
1−Pr∗mod1
pIf nεG(1)
where Pis the probability of selection of Cluster head, ris the existing round and
Gis the group of sensor nodes that have not been Cluster head in the current round.
By using this protocol maximum energy efficiency is achieved and network lifetime
is increased.
Power Efficient Gathering in Sensor Information Systems (PEGASIS) protocol
is a Hierarchical type. In this method, the near-optimal chain concept is used for
minimizing energy consumption in WSNs. It offers better performance than the
previous protocol. Here communication between the nodes in the network takes
place by communication between closest neighbor’s nodes only. The same nodes
participate in sending the information to the sink. This process is carried out in
different rounds and in each round, one node sends the data to the destination. After
completion of the round, a new round will begin. In PEGASIS protocol, nodes are
formed in chain structure to share the data per each round. The aim of this method is to
526 G. Gangaprasad et al.
minimize the battery power while forwarding the sensing information in each round
and to maintain uniform energy consumption at all nodes in the network. To get all
these advantages in PEGASIS, the following considerations need to be implemented
in the network.
•To enhance the lifetime of each sensor node in WSN, collaborative methods are
used.
•To optimize the bandwidth, the closest nodes should communicate with each other
(Figs. 2and 3).
In addition to the above, the following are some of the advanced routing protocols
which are aimed at improving the network lifetime in WSN.
•AZ-SEP: Advanced Zonal Stable Election Protocol is hybrid in nature. Here
sensor nodes directly communicate with the base station or sink node and
remaining sensor nodes participate in creating clustering. In this protocol, the
energy is distributed equally and improves network stability [16].
•Cross technology method: This protocol is implemented based on the usage
of various communication technologies between the sensor nodes. Generally
in WSN, ZIGBEE technology is preferred, but here both BLUETOOTH and
Fig. 2 LEACH protocol
Node
Sink
Cluster
Head
Fig. 3 PEGASIS protocol
Significance of Route Discovery Protocols … 527
ZIGBEE are used together for providing the best routing paths, thereby improving
the network performance [17].
•Based on Heterogeneous energy levels: By using energy levels in the heteroge-
neous network the following methods are proposed. These are [18]
–Circle Zone stable election protocol: A circle region is created in the center
of the network and provides random distribution around this circle. It is a
two-level energy protocol.
–Circle Zone stable election protocol with Helping nodes: It is similar to the
above but some helping nodes are placed in between the circle. This mechanism
improves the reduction of energy usage in the WSN. It is a three-level energy
protocol.
(B) Communication Model
This category is based on the working principle of the protocol to send the routing
information packets in the network. The classifications under this section are (a)
Query-type protocols and (b) Coherent and non-coherent type protocols.
•Query-TypeProtocols: Queries are generated in the network from the base station
for access to some data and upon receiving a query for a particular node send the
data to the sink node.
•Coherent and Non-Coherent Protocols: All sensed data in the network need
not be sent to the destination, because it contains some redundant information.
Redundancy needs to be eliminated from the data by the mechanism of data aggre-
gators. Incoherent type, the nodes perform some small processing and then transfer
it to the data aggregators whereas in non-coherent type the sensed information is
processed by some employed intermediate node for aggregation.
(C) Topology type Protocols: The name itself indicates that protocols in this cate-
gory have some structural or topology information present at sensor nodes in
the network. It contains effective route details of all nodes. These are divided
into two types
•Mobile Agent Type: In this type of protocol one node is completely allo-
cated to collect the sensed data from all the nodes in the network and it is
responsible to send the same data to the base station for further processing.
•Location Type: In the network, useful information resides in some regions
only for our requirement satisfaction. In this case, getting data to the
destination from that region needs to be done via some intermediate nodes.
(D) Reliable Routing type protocols: The routes in the WSNs are not always avail-
able due to node failures and battery drain etc. It leads to communication failure
in the network. These protocols provide alternative routes in those situations.
These are divided into
•Multipath-Based Protocols: They find alternate paths when a link failure
occurs or network is loaded with more packets.
528 G. Gangaprasad et al.
Tabl e 1 A basic comparison of various routing mechanisms
Parameter Network structure Communication
model
Topology type Reliable routing
Classification 1. Flat routing
2. Hierarchical
routing
1. Query type or
coherent based
2. Non-coherent
or negotiation
type
1. Location-based
2. Mobile
agent-based
Quality of service
(or)
Multipath based
Importance They find the path
details from the
way the sensor
nodes are
structured
The main
operation is based
on the manner the
protocol forwards
the packets in the
network
Network
maintains
topology
information
It maintains QoS
metrics like delay,
energy,
bandwidth,
services
Advantages Good for a small
network
Routing can be
done against sink
mobility also
Suitable for all
networks
Bandwidth can be
reduced
Effective routing
is possible
Quality of service
is provided to
ensure network
lifetime
•QoS-Based Protocols: It offers the quality of data at the base station.
(E) Nature of nodes. This classification is mainly based on the type of nodes
used in the network i.e., Homogeneous or Heterogeneous. If the sensor nodes
in the network are similar type, equal power supply levels, communication
bandwidth, and same storage capacities they belong to Homogeneous sensor
networks whereas counter to the above properties comes under Heterogeneous
sensor networks [19] (Table 1).
5 Conclusion
From the past two decades, WSNs are proposed for many application areas. In WSNs,
energy conservation is one of the main factors to describe the network efficiency and
performance, because nodes are equipped with low battery capacities. To suit a wide
range of applications using WSN, appropriate routing mechanisms are essential. In
this article, various routing techniques and their classification are provided along
with their energy consumptions. Here we mainly focused on Hierarchical routing
protocols, because among all the techniques they provide maximum reduction of
energy usage for many operations in the network.
Significance of Route Discovery Protocols … 529
References
1. M. Asif, S. Khan, R. Ahmad, M. Sohail, D. Singh, Quality of service of routing protocols in
wireless sensor networks: a review. IEEE Access 5, 1846–1871 (2017)
2. J. Yick, B. Mukherjee, D. Ghosal, Wireless sensor network survey. Comput. Netw. 52(12),
2292–2330 (2008)
3. Th. Arampatzis, J. Lygeros, S. Manesis, A survey of applications of wireless sensors and
wireless sensor networks, in Proceedings of the 2005 IEEE International Symposium on,
Mediterranean Conference on Control and Automation Intelligent Control, 2005 (IEEE, 2005)
4. N.A. Pantazis, S.A. Nikolidakis, D.D. Vergados, Energy-efficient routing protocols in WSN: a
survey com. IEEE Commun. Surv. Tutor. 15(2) (2013)
5. Samina Ehsan, Bechir Hamdaoui, A survey on energy-efficient routing techniques with QoS
assurances for wireless multimedia sensor networks. IEEE Commun. Surv. Tutor. 14(2), 265–
278 (2011)
6. I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E. Cayirci, Wireless sensor networks: a survey.
Comput. Netw. 38(4), 393–422 (2002)
7. J.N. Al-Karaki, A.E. Kamal, Routing techniques in wireless sensor networks: a survey. IEEE
Wirel. Commun. 11(6), 6–28 (2004)
8. K. Akkaya, M. Younis, A survey on routing protocols for wireless sensor networks. Ad Hoc
Netw. 3(3), 325–349 (2005)
9. S. Guo, O. Yang, Energy-aware multicasting in wireless ad hoc networks: a survey and
discussion. Comput. Commun. 30(9), 2129–2148 (2007)
10. G. Anastasi, M. Conti, M. Francesco, A. Passarella, Energy conservation in wireless sensor
networks: a survey. Ad Hoc Netw. 7(3), 537–568 (2009)
11. R.V. Biradar, V.C. Patil, S.R. Sawant, R.R. Mudholkar, Classification and comparison of routing
protocols in wireless sensor networks. Spec. Issue Ubiquitous Comput. Secur. Syst. 4(2), 704–
711 (209)
12. R. Yadav, S. Varma, N. Malaviya, A survey of MAC protocols for wireless sensor networks.
UbiCC J. 4(3), 827–833 (2009)
13. S. Ehsan, B. Hamdaoui, A survey on energy-efficient routing techniques with QoS assurances
for wireless multimedia sensor networks. IEEE Commun. Surv. Tutor. 14(2), 265–278 (2011)
14. A.S. Toor, A.K. Jain, A survey of routing protocols in Wireless Sensor Networks: IEEE Hier-
archical routing, in 2016 International Conference on Recent Advances and Innovations in
Engineering (INCREASE) (2016), pp. 1–6
15. S. Kellner, M. Pink, D. Meier, E. Blass, Towards a realistic energy model for wireless sensor
networks, in Proceedings of 5th Annual Conference on Wireless on Demand Network Systems
and Services, Garmisch (2008), pp. 97–100
16. F.A. Khan et al., Hybrid and multi-hop advanced zonal-stable election protocol for wireless
sensor networks. IEEE Access 7, 25334–25346 (2019)
17. D. Gao et al., RowBee: a routing protocol based on cross-technology communication for
energy-harvesting wireless sensor networks. IEEE Access 7, 40663–40673 (2019)
18. F.A. Mohamed et al. Energy-efficient circle zones stable election protocol with helper nodes
for heterogeneous WSNs. IET Wirel. Sensor Syst. 9(5), 313–322 (2019)
19. J. Yan, M. Zhou, Z. Ding, Recent advances in energy-efficient routing protocols for wireless
networks: a review. IEEE Access 4, 5673–5686 (2016)
In-Memory Computation for Real-Time
Face Recognition
Nikhil Kumar Gupta and Girijesh Singh
Abstract In the modern era, a facial recognition-based authentication system is
quite demanding. Facial recognition systems are widely used in many disciplines
like branches of biometric verification, video-based surveillance systems, interaction
between human and computers, access control systems and network security. The
traditional face recognition systems work well for detecting and recognizing faces
but when it comes to real-time operation, the system usually faces a problem of
drop in frame rate with an increase in the number of faces which might affect the
performance of the whole system. Due to the dropping frame rate, we might miss
some frames during the operation which in turn makes our whole system unreliable to
use in a real-time environment. In this paper, we propose a real-time face recognition
system that uses in-memory computation to keep frame rate constant during the entire
process even with an increase in the number of faces and hence makes the system
consistent and precise during its operation.
Keywords Face detection and recognition ·Attendance system
1 Introduction
Accuracy is one of the important aspects of any facial recognition system. Any
fault, error or inaccuracy can make the whole system inconsistent to use in a real
environment. Face recognition is comprised of various steps. From capturing a frame
to detecting faces and then recognizing it. A recognized face is matched against
the stored images in the database, and consequently, actions are performed. With
N. K. Gupta (B)·G. Singh
Primus Software Development India Pvt. Ltd, Noida, Uttar Pradesh, India
e-mail: activenikhilg@gmail.com
G. Singh
e-mail: singhgirijesh1996@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_47
531
532 N. K. Gupta and G. Singh
increase in computation capabilities of hardware, the algorithm based on neural
network are trained extensively to achieve human level accuracy in face recognition.
These algorithms are getting smarter day by day and being widely used in daily life
application.
There are enough algorithms available for face detection and recognition that
work very well for a variety of faces and are widely used in the face recognition-
based attendance system. These algorithms give outstanding results even for multiple
faces in a frame. But when it comes to video stream processing, it was seen that
these algorithms show a significant downfall in frame rate which might be a serious
drawback of the whole system as any drop in a frame can make whole system
unsuitable to use in real-time application.
Processing frames in real time in series created a problem of frame rate drop so
to address the issue, the proposed system uses in-memory computation to process
the frames. The execution flow of processing frames has also been changed, rather
than performing operation in series, a method of side by side operation is used, in
which, the system detect faces from an image, write into memory and at the same
time a recognition module read images from the memory and recognize the faces.
2 Related Work
In this section, different methodologies which are used to detect faces in the image
processing are described. The framework is majorly split into two parts: first being
face detection and the second being face alignment and recognition. The literature
review of both parts is described below.
2.1 Face Detection and Face Alignment
An image frame may consist of more than one face in it. To determine the count, size
and location of the faces, we have to find a definite common structure of a human
face consisting of major parts like nose, forehead, eyes, mouth and chin.
For face detection, Viola and Jones [1] proposed an algorithm that involves the
scanning of an image with a sub-window that can detect faces in an image. Three
major components of the Viola-Jones face detector are image integral, AdaBoost clas-
sifier learning and the attentional cascading for efficient computational resource allo-
cation. In this method, the image is rescaled to different sizes and then the algorithm
with a fixed size detector is executed on the rescaled images. It’s a time-consuming
process since it operates on multiple sizes of images.
To detect objects in the image, a Histogram of Oriented Gradients (HOG) feature
descriptor [2] is widely used. The picture is first split into sub-images or “cells”, then
a histogram of edge orientation is accumulated inside the cell. The algorithm then
counts the number of gradient orientations in a localized segment of an image. At
In-Memory Computation for Real-Time Face Recognition 533
last, the HOG descriptor uses the classification algorithm, Support Vector Machine
(SVM), to find out whether there is a face or not.
Several research results have shown a significant improvement in the performance
of face detection systems with the help of a neural network. A Deep Convolutional
Neural Network based model known as Deep Dense Face Detector (DDFD) (by
Farfade et al. [3]) which, irrespective of a wide range of orientations, could detect
faces easily. There is no need for bounding-box regression, segmentation or even
Support Vector Machine classifier when using this method. It was found using the
study that the method has the capability of detecting faces from various angles and
an assumption of having correlation among the distribution of positive examples and
negative examples in the training data and score of the suggested face detector.
Zhang et al. [4] introduced an algorithm that had the capability of detecting face
and alignment by making use of unified cascaded Convolutional Neural Networks
(CNN) with the help of multi-task learning. There are 3 stages in this framework:
firstly, it acquires the regression of bounding box, employing a CNN named Proposal
Network (P-Net). After calculating the estimated bounding box regression, non-
maximum suppression (NMS) is utilized which helps in ignoring the bounding boxes
which remarkably converge on each other. The qualifying applicants are further
passed through the Refine Network (R-Net) which results in rejecting a plethora
of erroneous candidates. It also performs NMS candidates merge and calibration
with the bounding box regression. The third and the final step involves describing
precise facial characteristics and outputting 5 facial landmark positions using the
Output Network (O-Net). If using a 2.60 GHz CPU, an approximate frame rate of 16
frames per second (fps) is achieved. The Nvidia Titan Black GPU has the capability
of providing a frame rate of up to 99 fps. The above-mentioned strategy delivered a
95% average precision in each stage complying against major benchmarks and had
a stellar real-time outcome.
2.2 Face Recognition
The utmost integral component of a face recognition system is to be able to distinguish
distinct faces by identifying the unique attributes of a face. The design of a face
recognition algorithm which stays robust even under ill-illuminated conditions is
highly significant. The conventional face recognition approaches which worked on
the basic principles of geometric features analysis [5], principal component analysis
(PCA) [6] and linear discriminant analysis [7] were able to attain a precise recognition
but only in certain conditions. CNN based facial recognition algorithms were able
to recognize faces with greater efficiency. A brief review of some of the state-of-art
methods is given below.
DeepFace [8], an algorithm for facial recognition which was proposed by Taigman
et al. combined 3D alignment and similarity metric with a deep neural network.
The inspiration for the network architecture came from the raw pixel RGB values.
Support Vector Regressor (SVR) extract 67 fiducial points, which was trained to
534 N. K. Gupta and G. Singh
predict point configurations from an image descriptor based on Local Binary Pattern
(LBP) histogram. DeepFace uses a single-core Intel 2.2 GHz CPU which takes 0.33 s
per image resulting in the system attaining 97.35% in precision. DeepID3 is another
method suggested by Sun et al. [9]. DeepID3 is composed of 2 deep network archi-
tectures, one is, stacked convolution and the other one is, which makes it more
accurate to perform facial recognition. DeepFace is an improvement of DeepID2+
method. DeepFace utilizes Joint identification-verification supervisory signals to
fully connected layers branching out from the pooling layer along with the last fully
connected layers. DeepID3 also uses the same method which helps the architecture
in better supervision of the early feature extraction process. DeepID3 uses a 10–15
non-linear feature extraction layer network which is significantly deeper compared
to DeepID2+. These layered networks helped DeepID3 achieve 99.53% accuracy for
face verification and 96% for face identification.
FaceNet [10] which is a state-of-art face recognition based on Deep convolutional
network was suggested by Schroff et al. This network is so well trained that it has
the capability to directly optimize the embedding of the input image. This network
learns directly from the pixel of a face image. Triplet loss which helps in minimizing
the distance of the embedding in the feature space between an anchor (reference face)
and a positive (the same person) and maximize the anchor and negative (a different
person) is used by FaceNet. It is believed that the faces of distinct people have large
distances whereas the face of the same person has small distances. The network is
trained in such a way that the squared L2-norm distances in the embedding space
directly correspond to face similarity. FaceNet is based on CNN which is trained
with the help of Stochastic Gradient Descent (SGD) with standard backpropagation
and AdaGrad. An accuracy of 99.63% was achieved using the above approach. The
system only uses 128 bytes per face and make it more suitable for running in the
embedded computer vision.
3 Proposed Framework
This part of the paper describes the architecture and framework used to accomplish
the task of real-time face recognition. There are two major parts: face detection and
face alignment and face recognition.
3.1 Face Detection and Face Alignment Module
Face detection which works as an input to the recognition system is the primary
and initial step in the framework. Face detection is performed with the help of Max-
Margin object detector (MMOD) [11]. Rather than performing sub-sampling this
method Optimizes overall sub-windows. Object Detection methods can be enhanced
in performance by MMOD if the learned parameters are linear. The model used here
In-Memory Computation for Real-Time Face Recognition 535
uses CNN based feature extractor that can detect faces in different variations and
illumination condition. A bounding box is generated around the detected face in the
input image, and is given as output.
Face alignment is performed using Ensemble of Regression Trees [12]. Out of
other available options like DeepFace by Facebook and Constrained Local Model
(CLM), Ensemble of Regression Trees was selected as it was trained by using MMOD
face detector and hence had the best accuracy among others. From a sporadic of
pixel intensities, this model directly estimates the positions of face’s landmark. An
Ensemble of Regression trees based 68-point face landmarking model was used and
affine transforms were applied on the obtained output to get a vertically aligned face
without changing its shape. The vertically aligned face image is required as this is
the type of input taken by the face recognition model to work.
3.2 Face Recognition and Classification Module
Face recognition is a critical part of the attendance process. In this paper, a ResNet
network was used for mapping the face to a 128D feature vector. For Image Recog-
nition we used a modified version of ResNet-34 network which is based on Deep
Residual Learning [13], with only 29 convolutional layers and the filters decreased
to half as compared to the original model in each of the layer. This network takes
150 * 150 px image of a face and embeds its features into a 128D vector. Pair-wise
hinge loss was used to train this network. It operates over a pair of matching set in a
mini-batch and comprise hard-negative mining at the mini-batch level. Any identities
fall in non-overlapping radius of 0.6 are projected.
For classifying the faces, KNN classifier was used. KNN classifies the new 128D
face feature vector to predict the identical face from previously learned feature
vectors. If no feature vector lies within the euclidean distance threshold of 0.6, the
obtained vector is marked as unidentified.
3.3 The Framework
The complete abstract of the framework is shown in Fig. 1. The process starts with
input captured from the camera. The feed from the camera was captured in the form
of video stream and was stored in a buffer memory from where the most recent frame
was kept in the memory and rest are dumped.
In the initial stage, the system picks a frame from the buffer memory to process.
The frame was given to the face detector model as input and the model returned an
array of coordinates of bounding boxes. The array gives the number of faces detected
in the image along with the location of each detected faces in the image. The data
received from the face detector model and the original image frame was encapsulated
and was saved into a secondary memory.
536 N. K. Gupta and G. Singh
Fig. 1 Process framework
The second stage, which runs parallelly with the first, picks the encapsulated data
from the secondary memory continuously in series with time and unpacks it to restore
the original image frame and the face detector data array. Each detected face in the
frame was given as an input to the face landmarking model which gives 68 coordinate
points, landmarking the facial features like eyes, nose, mouth, face boundary, and
certain other features. These coordinate points describe the alignment of the face and
hence used to perform affine transforms to align the face vertically without affecting
the shape of the face. The output obtained by this step was a vertically aligned face.
The vertically aligned image of the face is then given as input to the face recog-
nition model. This model returns a 128-dimensional feature vector for the input face
image. This 128-D feature vector is specific to a person and hence this data was
employed to classify the frame and determine the identity of an individual. To clas-
sify, a pre-trained KNN classifier was used. This takes the 128-D feature vector and
determines the person in the image.
To classify an unknown person, the euclidean distance threshold of 0.6 was consid-
ered. If there are no other learned (labelled) vector within the threshold of 0.6, then
the current vector was marked as unidentified. This step returns the best match of the
current vector from the pre-learned vectors, which defines the identity of the person
present in the image.
In-Memory Computation for Real-Time Face Recognition 537
Fig. 2 Working of recognition process
4 Experimental Setup and Results
4.1 Experimental Setup and Dataset
To accomplish the task of face detection and recognition, a Linux based system with
Intel i7-7700T processor with 4 cores and 8 threads and 16 GB of DDR4 System
RAM along with Nvidia 1060-Ti with 6 GB RAM and with 1280 Cuda cores was
used. The camera used to capture the frames was Logitech C920 webcam. To validate
the efficiency of the algorithm a custom dataset of a video stream with people passing
by the camera and number of people varying from 0 to 8 was created.
4.2 Model and Classifier
The models used in the system are mentioned in this part of the paper. For face
detection, a CNN feature-based Max-Margin Object Detector model was used with
pre-trained weights provided with the Dlib toolkit by Davis E. King.
The face landmarking task for face alignment was done by Ensemble of regression
trees model which was trained on ibug 300-W dataset [14], with pre-trained weights
provided by Davis E. King in Dlib toolkit.
The face recognition was done by a 29 layered ResNet network, a modified version
of ResNet-34 network [13], which was trained on 3 million faces (most of them from
face scrub dataset and VGG dataset). On LFW benchmark, the pre-trained model
available with Dlib toolkit shows 0.993833 as a mean error and 0.002727 as standard
deviation.
A K-nearest neighbour classifier (KNN) was used to classify the input face vector
on the basis of previously learned vectors from the database of known faces (Fig. 2).
4.3 Results
From Table 1and Fig. 3, it can be seen that there was a significant downfall in frame
rate with the increase of number of faces when we are running the algorithm without
538 N. K. Gupta and G. Singh
Tabl e 1 Framerate in CPU and GPU with/without in-memory computation
Number of faces in frame Frame rate
Without in-memory
computation
With in-memory computation
CPU GPU CPU GPU
15.24 20.05 5.03 20.26
23.25 17.39 5.02 20.20
32.31 15.07 5.01 20.12
41.82 12.28 4.99 20.01
51.39 10.01 4.97 19.96
60.97 8.12 4.94 19.89
70.76 7.05 4.92 19.82
80.62 5.16 4.88 19.79
0
5
10
15
20
25
12345678
Frame-rate
Number of Faces
Without In-Memory on CPU
Without In-Memory on GPU
With In-Memory on CPU
With In-Memory on GPU
Fig. 3 Framerate versus number of faces graph
in-memory computation on CPU and GPU. On the other hand, in the case of in-
memory computation, it was observed that there was a very minor fall in frame rate
on both CPU and GPU. This result encourages, to use in-memory computation on
the same hardware specification to achieve a better performance along with a better
result.
5 Conclusion
In this paper, a framework for real-time multiple face recognition systems was
proposed which uses in-memory processing to handle the drop of frame rate during
the processing of frames with an increase in the number of faces. This framework
shows a very consistent behaviour with no or very minor loss in frame rate during
the increase of the number of faces in a frame. This helps to achieve results even
In-Memory Computation for Real-Time Face Recognition 539
with limited resources. This framework suggests that better and efficient computa-
tion capabilities could be obtained with the same algorithms with no extra cost just
by changing the method of computation.
References
1. P. Viola, M.J. Jones, Robust real-time facedetection. Int. J. Comput. Vis. 57(2), 137–154 (2004)
2. N. Dalal, B. Triggs, Histograms of oriented gradients for human detection, in IEEE Computer
Society Conference on Computer Vision and Pattern Recognition, 2005. CVPR 2005,vol.1
(IEEE, 2005), pp. 886–893
3. S.S. Farfade, M.J. Saberian, L.-J. Li, Multi-view face detection using deep convolutional neural
networks, in Proceedings of the 5th ACM on International Conference on Multimedia Retrieval
(ACM, 2015), pp. 643–650
4. K. Zhang, Z. Zhang, Z. Li, Y. Qiao, Joint face detection and alignment using multitask cascaded
convolutional networks. IEEE Signal Process. Lett. 23(10), 1499–1503 (2016)
5. L. Wiskott, N. Krüger, N. Kuiger, C. Von Der Malsburg, Face recognition by elastic bunch
graph matching. IEEE Trans. Pattern Anal. Mach. Intell. 19(7), 775–779 (1997)
6. J. Yang, D. Zhang, A.F. Frangi, J.-Y. Yang, Twodimensional PCA: a new approach to
appearance-based face representation and recognition. IEEE Trans. Pattern Anal. Mach. Intell.
26(1), 131–137 (2004)
7. W. Zhao, A. Krishnaswamy, R. Chellappa, D.L. Swets, J. Weng, Discriminant analysis of
principal components for face recognition, in Face Recognition (Springer, Berlin, Heidelberg,
1998), pp. 73–85
8. Y. Taigman, M. Yang, M.A. Ranzato, L. Wolf, Deepface: closing the gap to human-level
performance in face verification, in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition (2014), pp. 1701–1708
9. Y. Sun, D. Liang, X. Wang, X. Tang, Deepid3: face recognition with very deep neural networks,
arXiv preprint arXiv:1502.00873 (2015)
10. F. Schroff, D. Kalenichenko, J. Philbin, Facenet: a unified embedding for face recognition and
clustering, in Proceedings of the IEEE Conferenceon Computer Vision and Pattern Recognition
(2015), pp. 815–823
11. D.E. King, Max-margin object detection, arXiv preprint arXiv:1502.00046 (2015)
12. V. Kazemi, J. Sullivan, One millisecond face alignment with an ensemble of regression trees,
in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2014),
pp. 1867–1874
13. K. He, X. Zhang, S. Ren, J. Sun, Deepresidual learning for image recognition, in Conference
on Computer Vision and Pattern Recognition (CVPR) (2016), pp. 770–778
14. C. Sagonas, G. Tzimiropoulos, S. Zafeiriou, M. Pantic, 300 Faces in-the-Wild Challenge:
the first facial landmark localization challenge, in Proceedings of IEEE International Confer-
ence on Computer Vision (ICCV-W), Workshop on 300 Faces in-the-Wild Challenge (300-W),
Sydney, Australia, 2013, pp. 397–403
A Novel Approach for Detection
of Basketball Using CFD Method
G. Simi Margarat, S. SivaSubramanian, and K. Ravikumar
Abstract Video analysis plays a vital role in detecting and tracking an object. In
a dynamic environment, such as sports or military area, it is difficult to detect and
track specified object. In this paper, we have presented a novel algorithm to detect a
basketball in the court. The objects are detected from video frames using different
types of realms like HSV space, time, frequency, histogram, etc. For ball detection,
Color Based and Frame Differencing (CFD) method is proposed and experimented.
The complexities arise in each stage such as preprocessing, segmentation, feature
extraction, background subtraction, and hole filling. Color quantization is applied
to minimize the number of distinct colors in the video. Frame differencing method
utilizes the difference in components between frames for ball detection. Morpholog-
ical operations on connected components are applied to increase the performance
and reduce the complexities.
Keywords Basketball tracking ·Ball detection ·Video capture ·Frame
difference ·Histograms
1 Introduction
The elemental steps are to detect, identify, and classify the object. The machine
based vision detection method segments the pixels of moving objects and thereafter
G. S. Margarat (B)
Department of Computer Science and Engineering, Bharath Institute of Higher Education and
Research, Chennai, Tamilnadu, India
e-mail: simimargaratphd@gmail.com
S. SivaSubramanian (B)
Dhanalakshmi College of Engineering, Chennai, Tamilnadu, India
e-mail: drsivatbm@gmail.com
K. Ravikumar (B)
Department of Computer Science and Engineering, Manonmanium Sundaranar University,
Thirunelveli, Tamilnadu, India
e-mail: ravikumarcsephd@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_48
541
542 G. S. Margarat et al.
tracks it. During these processes unnecessary interventions such as noise (the shadow
which is a connected set of background points are detected as relevant object) and
dynamic background curb the work. The video is read frame by frame. The frames
that contain the ball and ball similar object are retrieved based on the histogram
of the frames. The color palates run to around 224 combinations of RGB, the color
quantization algorithm is experimented to minimize the color palates and improve
the image quality. Further, implementation of edge detection methodology exhibits
robustness against the shadows, illuminations, repeated motions, etc. The Back-
ground Subtraction is appropriate for the moving object detection system to extort
the moving basketball from the static background. However, dynamic environment
changes such as shadows, illumination changes, camera noise, etc. make the ball
detection less accurate.
2 Literature Survey
Motion based recognition is the method which had been used to detect the object in
the array of an image. A huge number of images could be grouped to form a video and
by this method motion information had been extracted. Problem of holes had occurred
by employing two frame differencing. In order to overcome this problem, three frame
differencing and background subtraction had been used. Background subtraction had
been used in immobile background where dynamic background subtraction had been
used in mobile background for detecting objects. This would help to solve problem
on holes and could detect the objects in the mobile background. This literature had
presented an idea for solving the problem on holes in the mobile background [1]. A
motion based recognition method [1] solved the problem due to holes in two frames.
This would be overcome with the help of three frame differencing and background
subtraction. The dynamic background subtraction had been used to detect the object
in the video which has background changes. This could reduce the problem of holes
and had given the better detection of objects.
To track the objects, point tracking, kernel tracking, and silhouette tracking had
been used. The proposed work [2] tracks and detects the moving objects as their
existing system had failed to give the noiseless results in the narrow variations.
To extract the feature color, size, and motion had been extracted. Efficient motion
estimation algorithm had been used to extract the motion of feature. The same had
been used for other feature extractions.
This literature had presented an algorithm for motion detection and segmenta-
tion which was simple and robust [3]. This algorithm had worked on the basis of
background subtraction. A Differencing and Summing Technique (DST) had been
employed for motion detection and segmentation. This had reduced the difficulties in
computation. In addition to this, the method had detected and segmented the moving
objects in both indoor and outdoor environments with immobile background.
A Novel Approach for Detection of Basketball … 543
Reference [4] developed an algorithm for highlighting the levers to broadcast the
table tennis. To track the ball, the color, size, and position of the ball candidates were
attained using Bayesian decision framework. And from this framework, the position
of the ball can be identified by Kalman filter.
An efficient approach for detection of moving object under a static background [5]
is presented in the literature survey. Selection of maximum pixel intensity value of
the two frames had been calculated by segmenting the frames into non-overlapping
block. Reflections of objects captured in certain cases create haziness and hinder
the detection of targeted object. The joint probability density function (PDF) or the
scatter plot of two observed or mixed images is used for separation [6].
The object reconstruction algorithm for moving vehicle detection is based on three
frame differencing [7]. The algorithm implements Preset-Threshold-based Region
Filling (PTRF) method to reconstruct the moving object. The ORTFD has high-
accuracy in detecting dark moving vehicle object, and accelerates the detection speed.
To derive a good color model, a general optimization technique such as simu-
lated annealing is used. Color quantization [8] is regarded as a clustering problem
and is solved using fuzzy set [9]. The color palette is seen each as clusters of color
from which the suitable color with regard to the video frame can be derived. ViBE
(Visual Background Extractor) is foreground object segmentation in a dynamic back-
ground. As an extension to this ViBEalgorithm [10], the paper has included additional
parameter ‘blinking pixel’ to analyse the pixels of the inner border of the background
area.
3 Examination of Methodologies
3.1 Color Based Method
In this method, the features of video frames are extracted by implementing histogram
on it. The colors of histogram are compared against the frames. There are 16, 777, 21
color in Red, Green, Blue (RGB) slices, which increases the computational complex-
ities. The color quantization technique produces minimum number of color palettes
along with quality, is implemented to retrieve the best quality of the object frame.
The color based techniques have more provisions or features compared to geometric
cues and grayscale intensity. Generally, the color pixels are represented by three 8
bits each for Red, Green, and Blue. During the process of object segmentation, the
values are converted to an equivalent color model parameters. A comparison of these
color palette parameters is made to the video frame.
The most popular color models used are Hue, Lightness, Saturation (RGB, HSV,
HLS), Hue, Saturation, Intensity (HIS), and Y is the luma component and CB and CR
are the blue-difference and red-difference chroma components (YCbCr). To retrieve
intensity variations HSV and YCbCr color spaces are used [11]. The histogram gives
544 G. S. Margarat et al.
INPUT VIDEO
OBTAIN TEMPLATE BALL
RGB TO INDEXED COLOR
CALCULATE HISTOGRAM OF BALL
READ VIDEO FRAME1
OBTAIN REFRAINED FRAME1
END
Fig. 1 Block diagram showing the working of the color based algorithm
the summary of the data allocation in the video or the image. In this paper, it is with
regard to the video frames.
The resultant framework proceeds for detecting the ball automatically. Algorithm
has been designed to retrieve the group of ball objects from each frame and test their
locations in the consecutive frames (Fig. 1). Tracking operations are executed on the
ball objects by data associated algorithms. The foreground and background moving
objects are segmented by implementing the background subtraction technique.
3.2 Frame Differencing Method
This method is used for segmenting moving object in the frames. Two or more
successive frames are worked out, to get the desired result. The frame difference
method uses a very small time interval between successive frames and thresholding,
extracts the image region of the movement.
In a sequence, a morphological operation has been applied in order to bridge the
gaps and remove the interventions such as noise. After segmenting, morphological
operations [12] had been executed in order to bridge the gaps and eliminate the noise.
In order to differentiate the objects in motion in the background, an edge detection
technique had been exploited. The Prewitt technique [13] reduces the blurring effect
A Novel Approach for Detection of Basketball … 545
Reframed
Frame1
Reframed
Frame2
Apply
Theshold
Value
Morphological
Operation
Connectivity
Analysis
Frame
Difference
Fig. 2 Block diagram illustrates working of frame difference method
while detecting edges [14]. This technique uses a 3 ×3 gradient operator. A group
of ball candidates and the objects which had looked like balls were obtained in the
segmentation results and from that the ball candidates had been strained according
to their color, shape, and compaction trait of the ball range.
While pointing the ball candidates’ centroid location over time, the generation
of 2D candidate generation plot had been done. The computational morphological
operation includes erosion and dilation (Fig. 2). These take data as an image to be
eroded and a structuring element. The combined operations are applied to include or
exclude pixels from the boundaries of features to get a smoothened. Further, these
operations separate portions of features or touching features and to remove isolated
pixel noise from the image.
3.3 Proposed Method
The Color Based and Frame Differencing (CFD) algorithm is employed to detect the
basketball, tracks all the visual involved. The algorithm gives a clear map to track the
place and color of the high-density area which would be around the detectable ball.
The implementation of color based technique enables to track the object and object
similar identities. The color models which are around 224 that is 16 million color are
available for comparison. However to minimize these color models and to retrieve
the best quality object frame, the color quantization technique is implemented.
In CFD algorithm, Hue, Saturation, and intensity level of the frame is computed,
these diminish the noises (illuminations, shadows of the ball) from the frame. The
Hue, Saturation, Intensity Value (HSV) of each frame gives the hue and saturation
level. The object’s intensity level is analyzed and connectivity between the frames
is established. This establishment endorse the detection of the ball.
The edge detection algorithm using fuzzy sets requires threshold values to remove
the noisy edges from and around the ball in the frame. There can be strong and weak
edges, which needs different levels of focus by the edge detection [15,16]. Thus in
some cases, without the threshold value the noises can be edged out.
546 G. S. Margarat et al.
Further, the morphological operation for segmentation and the threshold value for
the detection of ball removes the unnecessary intervenes and noises. The morpho-
logical operation is a filtering measure. This blends the narrow breaks and cut-offs,
and bridges the gap in the contour. This operation results, the elimination of small
gaps between cut off segments and bridge the region. The morphological operation is
employed by erosion after dilation. The dilation process, in the proposed algorithm,
converts the ball pixel in one frame to other ball pixel, thus summing to find the
boundaries of the ball and closes the isolated background pixels. This is followed by
the erosion method. This gives the position of the basketball in the first frame.
Algorithm
Step 1: Input the Video and read the frames in order
Step 2: For every frame
Step 3: Convert RGB to HSV image
Step 4: Consecutive two frames are subtracted using Frame differencing
Step 5: Set threshold value and obtain binary frame.
Step 6: Apply morphological operation
Step 7: Display detected ball
Step 8: End
4 Conclusion
In this paper, the detection of basketball in a vision based domain is done. A range of
existing algorithms were also reviewed. Their performance and complexities were
discussed. The proposed method is based on color and frame difference algorithm. In
support of the proposed technique, color quantization, edge detection, and morpho-
logical operation are implemented. These reduce the illuminations and other basket-
ball similar objects from detection. On comparison, the proposed has a better perfor-
mance. The enhancement of this proposal is to design and experiment an algorithm to
identify the trajectory of the basketball in the video. The trajectory of the basketball
will be used to envisage the make or miss of the ball.
References
1. N. Srivastav, S.L. Agrwal, S.K. Gupta, S.R. Srivastava, B. Chacko, H. Sharma, Hybrid object
detection using improved three frame differencing and background subtraction, in 7th Inter-
national Conference on Cloud Computing, Data Science & Engineering—Confluence, 2017,
pp. 613–617
2. G. Sharmila Sujatha, V. Valli Kumari, An efficient motion based video object detec-
tion and tracking system, in International Conference on Innovative Trends in Electronics
Communications and Applications (ICIECA15), vol. 1, no. 13, 2015, pp. 85–97
A Novel Approach for Detection of Basketball … 547
3. G. Thapa, K. Sharma, M.K. Ghose, Moving object detection and segmentation using frame
differencing and summing technique. Int. J. Comput. Appl. 102(7), 20–25 (2014)
4. W. Chen, Y.J. Zhang, Tracking ball and players with applications to highlight ranking of
broadcasting table tennis video, in Proceedings of IMACS Multi Conference on Computational
Engineering in Systems Applications, 2006, pp. 1896–1903
5. S.S. Sengari, S. Mukhopadhyal, A novel method for moving object detection based on block
based frame differencing, in 3rd International Conference on Recent Advances in Information
Technology, 2016, pp. 1–6
6. M. Wajid, M. Sharma, Digital image separation algorithm based on joint PDF of mixed images.
Image Process. Commun. 20(1), 5–12 (2015)
7. Y. Lin, M. Fang, D. Shihong, An object reconstruction algorithm for moving vehicle detection
based on three-frame differencing, in 2015 IEEE 12th International Conference on Ubiquitous
Intelligence and Computing and 2015 IEEE 12th International Conference on Autonomic and
Trusted Computing and 2015 IEEE 15th International Conference on Scalable Computing and
Communications and Its Associated Workshops (UIC-ATC-ScalCom), 2015, pp. 1864–1868
8. G. Schaefer, Soft Computing-based colour quantization. EURASIP J. Image Video Process.
2014, 1–9 (2014)
9. P.A. Khaire, N.V. Thakur, A fuzzy set approach for edge detection. Int. Jo. Image Process.
(IJIP) 6(6), 403–412 (2012)
10. Tomasz Kryjak, Foreground object segmentation in dynamic background scenarios. Image
Process. Commun. 19(2–3), 25–36 (2015)
11. Giuliani Donatella, Color image segmentation based on principal component analysis with
application of firefly algorithm and Gaussian mixture model. Int. J. Image Process. 12(4),
101–112 (2018)
12. S. Behera, M.N. Mohanty, Land boundary detection of an island using improved morphological
operation. Int. J. Image Process. (IJIP) 6(6), 413–421 (2012)
13. R. Maini, J.S. Sohal, Performance evaluation of Prewitt edge detector for noisy images. GVIP
J. 6, 39–46 (2006)
14. E. Kermani, D. Asemani, A robust adaptive algorithm of moving object detection for video
surveillance. EURASIP J. Image Video Process. 2014, 1–9 (2014)
15. W. Yang, Analysis of sports image detection technology based on machine learning. EURASIP
J. Image Video Process. 2019, 1–8 (2019)
16. K. Aggarwal, M.S. Bhamrah, H.S. Ryait, Detection of cirrhosis through ultrasound imaging
by intensity difference technique. EURASIP J. Image Video Process. 2019, 1–10 (2019)
17. Z. Weng, Y. Guan, Action recognition using length-variable edge trajectory and spatio-temporal
motion skeleton descriptor. EURASIP J. Image Video Process. 2018, 1–15 (2018)
Ensemble Similarity Clustering Frame
work for Categorical Dataset Clustering
Using Swarm Intelligence
S. Karthick, N. Yuvaraj, P. Anitha Rajakumari, and R. Arshath Raja
Abstract Many attempts are made to resolve the clustering problem over cate-
gorical dataset through the process of clustering ensemble. However, the conven-
tional techniques failed to generate proper data portioning to cluster the data due to
the presence of incomplete information. Hence, this paper uses a modified cluster
ensemble similarity framework to cluster the relevant categorical datasets, and a
weight-based similarity approach has been proposed to work with type II full-
space and subspace ensembles that involve random-kand fixed-ktechniques. The
centroid selection is initiated using swarm-based bio-inspired optimization method
or swarm-based similarity ensemble clustering (WSE). Finally, Division Reminder
Hashing (DRH) technique is used to determine the weights of the cluster using
similarity measurement. This technique helped in eliminating the null and duplicate
datasets. The measurement for testing includes Adjusted Rand (AR), Normalized
Mutual Information (NMI), and Classification Accuracy (CA). The performance of
the proposed technique is tested over University of California Irvine (UCI) dataset,
and the results proved that the proposed technique provides better clustering using
Type-2 and Type-3 ensemble.
S. Karthick (B)·P. A. Rajakumari
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
DelhiNCR Campus, Ghaziabad, India
e-mail: karthick.usilai@gmail.com
P. A. Rajakumari
e-mail: sweetanip@gmail.com
N. Yuvaraj
Department of Computer Science and Engineering, St. Peter’s Institute of Higher Education and
Research, Chennai, Tamil Nadu, India
e-mail: yraj1989@gmail.com
R. A. Raja
Department of Electronics and Communication Engineering, B.S. Abdur Rahman Crescent
Institute of Science and Technology, Chennai, Tamil Nadu, India
e-mail: arshathraja.ru@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_49
549
550 S. Karthick et al.
Keywords Full-space ensemble ·Subspace ensembles ·Weight-based clustering ·
UCI data
1 Introduction
Data clustering is a learning method with inherent practical applications. Analysis
of clustered data is important for data analysis and data processing. Most clustering
methods divide the dataset into classes as per the data similarity and enable the class
element to have similar character and keep distance of the dataset cluster of different
classes as large as possible.
There are several popularly used clustering algorithms, which include k-means
[1] and PAM [2]. This facilitates the numerical data with its intrinsic properties in its
natural state. This measures the distance between vectors through distance metrics
[3,4]. The direct application is unfeasible to cluster categorical data with distinct
values and possess no order.
Several other clustering algorithms for the categorical data include uncertainty
clustering [6], Context-based Distance Learning [6], which is employed for real and
synthetic datasets. Other algorithms include concept-drifting categorical data [7].
The varied parameters in specific dataset, diverse, or same algorithms provide
distinct solutions. Consequently, it is difficult to determine the precise algorithm
for a specified dataset. This issue motivates to form cluster ensembles. In [8], there
exists an application that comprises reusing possibility of conventional knowledge
concerning the data and mining, wherein data partitions exist based on its locations.
Lately, clustering ensembles [9,10] gained popularity to overcome the issues of
clustering algorithms. It is apparent that clustering finds the patterns of any dataset.
No cross-validation is executed for tuning the parameters with inherent clustering. If
the user gets equipped without any guidelines, the clustering technique for specific
dataset is selected.
In [11], the relationships between clustering categorical data and clustering
ensemble are carried out. Categorical data clustering through maximal K-partite
cliques [12] determines the clusters in categorical datasets using K-partite maximal
cliques. The clustering ensemble process possesses two stages: population generation
with clustering partitions through resampling [13] attribute subspace [14] homoge-
nous algorithm [15] and others. Second stage integrates clustering results into a final
solution. Clustering ensemble uses supervised bagging and boosting algorithms. In
[16], an adaptive approach, ensemble individual partitions are generated sequen-
tially by subsamples of dataset. Also, clustering categorical dataset related issue is
analyzed and investigated in the proposed study. A suggested generic framework is
effectively applicable in all the other data types.
The paper proposes Weight-based Similarity Clustering Ensemble (WSE) method.
The weights of the each dataset point change during each iteration. A strategy is
adopted while weights are updated using similarity matrix. This is the base concept
for weight-based similarity method. The centroid values of the clusters are chosen
Ensemble Similarity Clustering Frame work for Categorical … 551
based on swarm behavior with respect to Fixed and Random-kclustering. The cluster
ensemble is generated through randomly chosen centroid values. The reduction is
maintained minimal for removing the irrelevant dataset from the clusters. This is
done using DRH technique with weight-based similarity function. This makes the
probability of Consensus Function to be applied with respect to RM. The new swarm
algorithm with weight-based ensemble framework has suggested deliberately for
engendering the precise and accurate measures in an economical manner.
Section 2provides the proposed method for clustering ensemble. Section 3eval-
uates the proposed clustering ensemble with conventional techniques. Section 4
concludes the paper.
2 Proposed Method
1. Weight-based Cluster Ensemble
A similarity method acquires λ(a), a∈Nr=1,2, …,r}fromrclustering algo-
rithms £(a). This forms a similarity matrix (m×m)M(a) and m—dataset size.
Once ensemble is applied, consensus similarity matrix is finally produced. This gets
transformed to the result for final clustering.
Basedontheresultsof£(a), similarity matrix M(a) is applied. The clustering
algorithms used here are crisp clustering, and the data point share goes to a single
class. The clustering results are expressed in two forms: λ(a) and M(a).
λ(a)=1×mvector,
where λ(a,i)—class label belonging to ith data point.
When the similarity matrix M(a) is generated, and if λ(a, m)=λ(a, n)orthe
mth and nth data lie in similar classes, then M(m,n)(a)=1, or else it is 0. Similarly,
if M(a)=binary matrix, then each data point pair occurs inside similar clusters.
Hence, for all similar matrices, raverage matrix amis calculated after the gener-
ation of consensus similarity matrix sm. Here, if smis binary, am=[0, 1]. If M(m,
n)=1, the possibility of mth and nth dataset lies in similar classes. A threshold of
0.5 is set. If M(m,n) > threshold, smis 1 or else smis 0. Then, binary consensus
definite similarity matrix is generated and consensus similarity matrix is generated,
and hence the final clustering result is obtained.
The weights are added with the similarity method, and clustering ensemble method
is proposed. Finally, the Weighted Similarity Clustering Ensemble (WSE) is formed
with weighted information which is shown in algorithm.
Swarm Algorithm
552 S. Karthick et al.
Input: Dataset D={x1,x2,…,xn};
Weight of individual data
Weight ={w1,w2,…,wn}
Base Cluster α(a), a∈Nα.
Process:
For a=1, …, m;
λ(a)=α(a)(D);
Derive m×msimilarity M(a)usingλ(a);
End
Evaluate and record λ;
M1=r
a=1M(a)
a=1
M(a);
For i=1, …, m
For j=1, …, m
If M1(i,j)==r
M1(i,j)=1.
Else M1(i,j)=0
End
End
Reduce wof each cluster points of same class over rprocess
For i=1, …, m
w=
w(i)
m
i−1w(i)
End
If final condition is satisfied
Clustering result is obtained;
else
Clustering is repeated with updated w.
Output:λ-clustering result
3 Results and Discussion
This section evaluates the proposed ensemble algorithm using weight cluster
ensemble or WSE framework. This employs several validity indices on real datasets.
Ensemble Similarity Clustering Frame work for Categorical … 553
Tabl e 1 Datasets
Dataset DN KAV
Zoo 16 101 736
Lymphography 18 148 459
Primary tumor 17 339 22 42
Data partition quality is created and evaluated based on categorical data ensemble
technique (Table 1).
The five ensemble types for evaluation are:
•Type-1,
•Type-2 Fixed-k,
•Type-2 Random-k,
•Type-3 Fixed-k, and
•Type-3 Random-k.
The k-modes clustering generates the clustering with proposed algorithm with
•Weighted Similarity Ensemble framework (WSE),
•Weight-based Cluster Ensemble (LCE),
•Cluster-based Similarity Partitioning Algorithm (CSPA),
•Hypergraph Partitioning Algorithm (HGPA).
Table 2shows the clustering accuracy for entire dataset in the Table 2.
Tabl e 2 Clustering accuracy
Dataset Ensemble type LCE HGPA WSE
Lymphography I0.743 0.66 0.7569
II-Fixed k0.796 0.709 0.812
II-Random k0.784 0.633 0.823
III-Fixed k0.782 0.692 0.795
III-Random k0.777 0.633 0.7925
Zoo I0.894 0.594 0.932
II-Fixed k0.921 0.836 0.941
II-Random k0.931 0.826 0.942
III-Fixed k0.941 0.831 0.954
III-Random k0.931 0.835 0.946
Primary tumor I0.445 0.293 0.689
II-Fixed k0.478 0.438 0.645
II-Random k0.496 0.428 0.694
III-Fixed k0.463 0.423 0.6945
III-Random k0.484 0.408 0.6936
554 S. Karthick et al.
Tabl e 3 Performance comparison
Ensemble type Methods CA NMI AR
B W B W B W
1WSE 187 28 146 53 164 53
LCE 170 35 137 65 149 61
HGPA 21 193 19 208 24 201
2–fixed kWSE 225 3228 5212 9
LCE 208 4204 9201 13
HGPA 64 80 93 92 93 91
2–random kWSE 235 2214 8219 8
LCE 209 4203 12 201 11
HGPA 67 121 41 139 49 132
3–fixed kWSE 219 5203 8197 9
LCE 197 7191 13 182 16
HGPA 71 82 56 101 65 101
3–random kWSE 227 3206 9196 9
LCE 203 6191 13 181 11
HGPA 51 117 38 134 63 131
Fig. 1 Type 2 ensemble:
fixed data clustering
Table 3shows significant performance of various ensemble types as per NMI, AR,
and CA evaluation indices. An enhanced effectiveness is proved in terms of suggested
weighted techniques in comparison with other conventional clustering method. The
Band Windicate better and worse performance than the others.
The parameter analysis presents a practical means and that allows the users for
employing the WSE at fullest range. Performance of the technique depends entirely
on decay factor, where DC ∈[0, 1] and that deploys the ascertained clustering
similarity. The relationship between DC ∈{0.1, 0.2, …, 0.9} and the clustering
performance values is presented in XY-axis.
Ensemble Similarity Clustering Frame work for Categorical … 555
The average score WSE value results of ensembles depend on DC. The Figs. 1,2,3,
and 4show the weighted average scores of proposed technique with the conventional
techniques. The WSE approach for the users ascertains consistent, high quality, and
best results as of DC values between 0.7 and 0.9. This is due to centroid calculation
through swarm intelligence and the clustering results for all DC.
Fig. 2 Type 2 ensemble:
random data clustering
Fig. 3 Type 3 ensemble:
fixed data clustering
Fig. 4 Type 3 ensemble:
random data clustering
556 S. Karthick et al.
4 Conclusion and Future Work
The weight-based ensemble clustering algorithm turns out to be an effective method
that enhances the performance of the categorical data clustering. The use of weight-
based similarity ensemble technique clusters the categorical data without empty
datasets. Here, the cluster ensemble calculates the centroid value using swarm intelli-
gence process. Then, the matrix transformation is carried out to transform the original
data matrix to RM. The duplicate entries are finally removed using DH method and
that uses graph partitioning method. The weight-based similarity is used to resolve the
issue while constructing a RM. The weight information is often changed due to limi-
tations in K-means and other conventional algorithms. This information is applied to
the forthcoming iteration, and the final score is calculated through similarity matrix.
The framework is substantiated and results prove its validity and performance of
clustering algorithm on different categorical datasets. The proposed framework can
be incorporated in other clustering algorithms, and the evidences prove the practical
possibility of ensemble technique.
References
1. J. Wu, Advances in K-means Clustering: A Data Mining Thinking (Springer, 2012)
2. L. Kaufman, P.J. Rousseeuw, Finding Groups in Data: An Introduction to Cluster Analysis
(Wiley, 2004)
3. A.S. Shirkhorshidi, S. Aghabozorgi, T.Y. Wah, A comparison study on similarity and
dissimilarity measures in clustering continuous data. Phys. Rev. Lett. 10(12) (2015)
4. L.O. Mafteiu-Scai, A new dissimilarity measure between feature-vectors. Int. J. Comput. Appl.
64(17), 39–44 (2013)
5. D. Parmar, T. Wu, J. Blackhurst, MMR: an algorithm for clustering categorical data using
rough set theory, in 25th International Conference on Conceptual Modeling, vol. 63, Issue No.
3, pp. 879–893 (2007)
6. D. Ienco, R.G. Pensa, R. Meo, Context-based distance learning for categorical data clustering,
in 8th International Symposium on Intelligent Data Analysis, pp. 83–94 (2009)
7. H.-L. Chen, M.-S. Chen, S.-C. Lin, Catching the trend: a framework for clustering concept-
drifting categorical data. IEEE Trans. Knowl. Data Eng. 21(5), 652–665 (2009)
8. N. Iam-On, T. Boongoen, S. Garrett, C. Price, A weight-based approach to the cluster ensemble
problem. IEEE Trans. Pattern Anal. Mach. Intell. 33(12), 2396–2409 (2011)
9. Sandro Vega-Pons, Jose Ruiz-Shulcloper, A survey of clustering ensemble algorithms. Int. J.
Pattern Recognit. Artif Intell. 25(3), 337–372 (2011)
10. T.Y. Li, Y. Chen, Fuzzy clustering ensemble with selection of number of clusters. J. Comput.
5(7), 112–1118 (2010)
11. Z.Y. He, X.F. Xu, S.C. Deng, A cluster ensemble method for clustering categorical data. Inf.
Fusion 6(2), 143–151 (2005)
12. M.J. Zaki, M. Peters, Clicks: mining subspace clusters in categorical data via k-partite maximal
cliques, in International Conference on Data Engineering, pp. 355–356 (2005)
13. J. Jia, X. Xiao, B. Liu, L. Jiao, Bagging-based spectral clustering ensemble selection. Pattern
Recogn. Lett. 32(10), 1456–1467 (2011)
14. M. Al-Razgan, C. Domeniconi, D. Barbara, Random subspace ensembles for clustering
categorical data. Supervised Unsupervised Ensemble Methods Appl. 31–48 (2008)
Ensemble Similarity Clustering Frame work for Categorical … 557
15. H.G. Ayad, M.S. Kamel, Cumulative voting consensus method for partitions with variable
number of clusters. IEEE Trans. Pattern Anal. Mach. Intell. 30(1), 160–173 (2008)
16. B. Minaei-Bidgoli, H.P.H. Alinejad-Rokny, H. Alizadeh, W.F. Punch, Effects of resampling
method and adaptation on clustering ensemble efficacy. Artif. Intell. Rev. 41(1), 27–48 (2014)
Multi-Focus Image Fusion Using
Conditional Generative Adversarial
Networks
A. Murugan, G. Arumugam, and D. Gobinath
Abstract Multi-focus image fusion is a task of generating a composite image from
the numerous limited depth of focus images of the identical clips. The aim of multi-
image fusion is to achieve a better quality image for better human visual perception
or better machine interpretation. Calculating the focus map to distinguish the focused
and unfocused pixels is the challenging task in image fusion algorithms. Numerous
ways of calculating focus map have been proposed in the last two decades. Very
recently, researchers show more interests on artificial intelligence and deep learning
based algorithms to identify the focused regions from the limited depth of field image.
Image fusion involves the following three subproblems: (1) Focus region detec-
tion, (2) Selecting the focus region based on confidence level, and (3) Final fusion.
Motivated by Conditional Generative Adversarial Network (cGAN), a novel way of
detecting the focused regions same as image-to-image translation has been proposed
here. GAN requires more number of images to train the network. Due to lacking of
limited depth of field images dataset, we synthesize a dataset based on PASCAL VOC
2012. We compare the results of the proposed deep learning based fusion methods
with the traditional fusion methods. Results prove that the proposed unsupervised
neural network based approach outperforms the other traditional methods.
Keywords Image fusion ·Deep learning and generative adversarial network
A. Murugan (B)·G. Arumugam ·D. Gobinath
Department of CSE, SRMIST, Kattankulathur Campus, Chennai, India
e-mail: murugana@srmist.edu.in
G. Arumugam
e-mail: arumugam@gmail.com
D. Gobinath
e-mail: gobinatd@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_50
559
560 A. Murugan et al.
1 Introduction
Acquiring an image with all the regions in focus is a confronting problem in the
current camera systems. Due to the shallow Depth Of Field (DOF), the partial regions
are in good focus and remaining area is in out of focus/blurred. An algorithm to
combine several incompletely focused pictures/images into a composite image is
called image fusion. This improves the visual perception of the image to human and
information to the machine. So far, many of the researchers have proposed several
algorithms to accomplish this task.
Shallow depth of focus makes images to blur in the region outside DOF. Very
small range which falls inside DOF gets fully focused. The algorithm to extract all
the focused pixels in the multiple shallow depth images and create a composite image
is multi-image fusion. It uses various camera parameters to combine the images.
Basic multi-focus fusion algorithms are classified into three levels based on the
complexity of the algorithm as low, mid, and high. Some authors have referred this
as pixel, feature, and characteristic level. But most of the pixel level algorithms work
either on spatial or on the transformed domain [1].
In the transform-based steerable image pyramid, the images are decomposed into
sub-bands, and the sub-bands are having the advantages of rotation and translation
invariant. Basic steerable image pyramid is a multiple orientation, self-inverting, and
multi-scale image decomposition [2]. In [3], Jing and Li apply wavelet transform to all
the input images. Then, for all the input images, the coefficient (wavelet approxima-
tion) along with detail coefficient is combined by applying different rules (fusion).
The final composite image is generated by applying inverse transform (wavelet).
Region-based wavelet coefficient is also used in [4]. The authors have implemented
two stages, where the textured and other regions are handled in the initial stage.
Later, in the second stage, spectral clustering method groups together in segmented
primitive regions. The DT-CWT coefficient was used in the texture process.
As compared against the transform-based methods, spatial methods do the fusion
task easily by processing the source images. Spatial methods are classified into three
subgroups, region, block, and pixel-based methods. Block-based approach has been
used in [5] as to decompose the input images into spatial blocks, and then the spatial
frequency is computed for each blocks in the source images.
Even though all of these handed craft recent algorithms performing better and
producing better fused image, choosing the right parameters and filter size manually
is a tedious task, and always producing some artifacts on the fused image and also
designing an ideal design is practically impossible due to some limitations or lack of
relevant knowledge. In [5], an already-trained Convolutional Neural Network (CNN)
based model is used to calculate the focus measure score map which basically has
all the focused pixels and non-focused pixels in the input image. And, a focus map
is calculated from the focus score map by taking average of the overlapping patches.
Then, the binary mask image is generated by applying the binarization to the focus
map; the pixel-wise weighted mean method is used to generate fused image from the
decision map.
Multi-Focus Image Fusion Using Conditional Generative … 561
Image Fusion
Fused Image
Image 2
Fig. 1 Basic image fusion
Inspired by generative models, Goodfellow [6] came up with the Generative
Adversarial Network (GAN). GAN has achieved success, and many algorithms have
been proposed on top of the generative modeling. Few of the algorithms are given
below: Image-to-Image translation [9], In-painting [7], and Text-to-Image transla-
tion [8]. As mentioned in [9], image fusion may be viewed as multiple images to
single image translation problem. We propose this N to 1 image translation using
conditional GAN [10].
The rest of the paper is organized as follows: related works in Sect. 2, proposed
method in Sect. 3, results and comparison in Sect. 4, and conclusion and future works
in Sect. 5(Fig. 1).
2 Related Works
(A) Generative Adversarial Network
Generative Adversarial Network (GAN) [6] is an effective generative model which
has two major components, Generator (G) and Discriminator (D). Generator takes in
random numbers and outputs an image. Discriminator accepts samples from training
data and G and outputs a scalar between 0 and 1. The D’s scalar output is an indicator
(probability) that the sample comes from training data. During adversarial training, G
is trained in such a way that it can generate a fake data and fool D. D’s responsibility
of distinguishing real and fake data is getting stronger, while the G’s output is more
close to real. Whenever D finds the differences between real and G’s output, G’s
parameter gets adjusted to generate a data more close to real. Eventually, G will have
the capability to generate a sample which cannot be distinguished by D.
(B) Conditional Generative Adversarial Network
Just feeding random noise and training the model in GAN is instable, and sometimes
it introduces artifacts in result image. To reduce the instability, Mirza et al. [10]
562 A. Murugan et al.
proposed an effective way of training GAN by adding an auxiliary input to both
Generator and Discriminator. This auxiliary input can be a label or data or information
from other model. This forms the GAN network with conditional input. By comparing
the MINIST results, authors prove that training GAN with this auxiliary (conditional)
input improves the Fused Image stability than vanilla GAN. The loss function with
auxiliary input is given below in (1):
minmaxC(G,D)=E,∼P(x,y)[log D(x,y)]GD
+E∼P(x), z∼(z)[log 1 −D(x,G(x,z))](1)
(C) Image-to-Image Translation
Image-to-image translation may be explained as the task of generating an image that
is translated from other image using generative models. The objective is to learn the
spatial map between a source and output grid. They do map a full resolution input
image into output grid. And also both input and output differ in surface appearance.
3 The Proposed Method
In this method, we construct a cGAN-based network model to generate a focus
measure map to distinguish focused pixels and non-focused pixels from the shallow
depth of field images. The focus map is generated with different labels for the focused
and non-focused regions. In the second step of focus map selection, we are selecting
the exact pixels and making the final selective focus map. General fusion map is
applied on the selective focus map to generate fusion image. The architecture of the
proposed method is shown in Fig. 2. The remaining part explains about architecture
and design.
Same like the vanilla GAN, this model also has two sub-networks/components, a
Generator (G) and a Discriminator (D). The input to the generator and discriminator
in this model is having multiple images and single image output. The multiple images
are being sent to network to extract the features, and in the feature concatenation layer,
the input features are concatenated and sent to discriminator. The network is being
forced to learn/train the same features for the multiple input images. The multiple
input features weights are shared between networks [11]. Since the tied weights,
the features are same for all the input images. And it is easy for concatenating the
different features. The proposed method detects the focused regions and generates
the focus measure map.
Accepts multiple input images and returns a focus measure map.
In Generator, the network accepts multiple images, and each image passes through
Encoder which has Convolution, Batch normalization, and Relu. To extract the high-
level features, Encoder output goes through the Resnet Blocks [12]. The Decoder
has deconvolution, batch normalization, and Relu to resize the concatenated feature
Multi-Focus Image Fusion Using Conditional Generative … 563
Fig. 2 Architecture of the proposed method. It has two sub-networks, “Generator (G) and
Discriminator (D)”
to the input images. Then the focus map is mapped through sigmoid operation. The
discriminator accepts both generated focus map and input images using cGAN’s
objective function to train the model.
The proposed algorithm is as follows:
1. The multi-focus images are passing through the GAN network and focus map is
generated. Each pixel is mapping to a focused region in the focus map image.
2. Selecting the focused pixel from the map and weighted average neighborhood
operation is applied to select the focused pixel.
564 A. Murugan et al.
Focus Map
Fused Image
Image Fusion (cGAN)
Fig. 3 Proposed algorithm flow
3. Final fusion rule is applied to the selected focus map, and fused image is
generated. Figure 3shows the flow of the proposed algorithm.
4 Results and Comparison
(A) Data Preparation
We have used the PASCAL Visual Object Classes 2012 (PASCAL VOC 2012) [13].
This dataset is having 2913 images with segmented labels with ground truth. For the
de-focused images, there is no large dataset available. To train the network, we would
need a large number of images with segmented labels. We synthetically created a
de-focused dataset using VOC 2012 images by applying Gaussian blur with different
sigma.
(B) Comparison
We compared the proposed method with the following three best conventional image
fusion techniques: (1) PCA-based method [14], (2) Spatial frequency [15], and
(3) Gray-level local variance [16]. The fused image generated by PCA method
is using weighted mean. The weights are calculated from the Eigen vector of the
corresponding Eigen values.
This method produces weak fused image as compared against other methods. The
spatial frequency is measuring the activity level of the each pixel in the source image.
Multi-Focus Image Fusion Using Conditional Generative … 565
Fig. 4 Subjective results, left to right: PCA, SF, all in focus, and proposed method
It is affecting the fusion quality due to blocking effect. The high level of weak features
in this method lacks in un-artifact fused image. The third method mainly relies on
the neighborhood size choosen for focus measurement. This leads to artifacts if the
proper neighborhood size is not chosen. The result of all the methods is given in
Fig. 4.
C. Quantitative Comparison
The quantitative comparison of stacked images is a tedious task as it involves ground
truth comparison. There are no proper quantitative measures proposed to measure
the fused image results. Comparing the mutual information may give us the proper
comparison. We used Normalized Mutual Information [17]; it is more powerful than
Mutual Information (MI). The definition of NMI is given below (2):
Q=2M(A,F)+H(A)+H(F)
M(B,F)(2)H(B)+H(F)(2)
5 Conclusion and Future Works
It is determined from the discussions above that there exist many common challenges
in typical activity level measuring techniques. Calculating the correct filter size isone
among the key challenges. Deep learning based image fusion research reduced the
hand crafting works and improved the quality of fusion for any real-time multi-focus
images. However, designing of neural network model for a specific fusion task is a
challenging task in deep learning based techniques. Another important drawback is
generating a dataset for model training (a sensible problem in existing deep learning
based image fusion is that the lack of large available dataset), lack of the applying of
domain information or lack of hardware, etc. In practice, these problems are having
close relations to one another and may be collectively thought-about. In this proposed
method, the features are extracted using the cGAN and there is no manual hand craft
involved in image fusion. And, the result shows that this method produces quality
fused image without manual work involved. The future work is to exploit the usage of
Generative Adversarial Network (GAN) to generate the large set of dataset without
manual labeling of segmented images.
566 A. Murugan et al.
References
1. G.A. Ratt, J. Aacute, A. V1ega, J.E. Murari, Image fusion: advances in the state of the art. Inf.
Fusion 8(2), 114–118 (2007)
2. Z. Liu, K. Tsukada, K. Hanasaki, Y.K. Ho, Y.P. Dai, Image fusion by using steerable pyramid.
Pattern Recogn. Lett. 22(9), 929–939 (2001)
3. J. Tian, L. Chen, Adaptive multi-focus image fusion using a wavelet based statistical sharpness
measure. Sig. Process. 92(9), 2137–2146 (2012)
4. J.J. Lewis, R.J. O’Callaghan, S.G. Nikolov, D.R. Bull, N. Canagarajah, Pixel- and region-based
image fusion with complex wavelets. Inf. Fusion 8(2), 119–130 (2007)
5. Y. Liu, X. Chen, H. Peng, Z. Wang, Multi- focus image fusion with a deep convolutional neural
network. Inf. Fusion 36, 191–207 (2017)
6. I. Goodfellow, NIPS 2016 tutorial: Generative adversarial networks. arXiv preprint arXiv:
1701.00160, 2016
7. R.A. Yeh, C. Chen, T.Y. Lim, A.G. Schwing, M. Hasegawajohnson, M.N. Do, Semantic image
inpainting with deep generative models, in Proceedings of IEEE Conference on Computer
Vision Pattern Recognition, pp. 6882–6890
8. S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, H. Lee, Generative adversarial text to
image synthesis. arXiv preprint arXiv:1605.05396 (2016)
9. P. Isola, J.Y. Zhu, T. Zhou, A.A. Efros, Image-to-image translation with conditional adversarial
networks, in Proceedings of IEEE Conference on Computer Vision And Pattern Recognition,
pp. 5967–5976 (2017)
10. M. Mirza, S. Osindero, Conditional generative adversarial nets. Comput. Sci. 2672–2680 (2014)
11. K.R. Prabhakar, V.S. Srikar, R.V. Babu, DeepFuse: A Deep Unsupervised Approach for
Exposure Fusion with Extreme Exposure Image Pairs, pp. 4724–4732
12. K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in Proceedings
of IEEE Conference on Computer Vision And Pattern Recognition, pp. 770–778 (2016)
13. M. Everingham, S.M.A. Eslami, L.V. Gool, C.K.I. Williams, J. Winn, A. Zisserman, The pascal
visual object classes challenge: a retrospective. Int. J. Comput. Vision 111(1), 98–136 (2015)
14. O. Rockinger, T. Fechner, Pixel-level image fusion: the case of image sequences. Proc SPIE
3374, 378–38829 (1998)
15. S. Li, T. James Kwok, Y. Wang, Combination of Images with diverse focuses using the spatial
frequency. Inf. Fusion 2(3), 169–176 (September 2001)
16. S. Pertuz et al., Generation of all-in-focus images by noise-robust selective fusion of limited
depth-of- field images. IEEE Trans. Image Process. 22(3), 1242–1251 (2013)
17. M. Hossny, S. Nahavandi, D. Creighton, Comments on ‘Information measure for performance
of image fusion. Electron. Lett. 44(18), 1066–1067 (2008)
Privacy Preservation Between Privacy
and Utility Using ECC-based PSO
Algorithm
N. Yuvaraj, R. Arshath Raja, and N. V. Kousik
Abstract In this paper, we propose a new privacy utility method, where privacy is
maintained by lightweight Elliptical Curve Cryptography (ECC) and utility is main-
tained through Particle Swarm Optimization (PSO) clustering. Initially, the datasets
are clustered using PSO, and then the privacy of clustered datasets is maintained
using ECC. The proposed method is experimented over medical datasets, and it is
compared with existing methods through several performance metrics like clustering
accuracy, F-measure, and data utility and privacy metrics. The evaluation shows that
proposed method obtains improved privacy preservation using clustering algorithm
than existing methods.
Keywords PSO clustering ·ECC ·Privacy and utility
1 Introduction
Data anonymization [1] is a major technique for preserving the data with privacy
in the field of data publishing or data mining. It involves anonymizing the data
records from linkage attacks [2] or probabilistic inference attack [3]. The process of
anonymization is achieved either through suppression of QID attributes or through
generalization. The former one leads to increased loss of data and the latter one is
an NP-hard problem. Hence, privacy preservation using data anonymization needs
N. Yuvaraj
Deputy Manager, Research And Development, ICT Academy, Chennai, Tamil Nadu, India
e-mail: yraj1989@gmail.com
R. Arshath Raja (B)
Senior Associate, Research And Development, ICT Academy, Chennai, Tamil Nadu, India
e-mail: arshathraja.ru@gmail.com
N. V. Kousik
Department of Computer Science and Engineering, Galgotias University, Greater Noida, Uttar
Pradesh, India
e-mail: nvkousik@galgotiasuniversity.edu.in;nvkousik@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_51
567
568 N. Yuvaraj et al.
proper improvement to avoid the constraints like increased data loss and NP-hard
problem.
On the other hand, privacy preservation process involves measuring the utility
of dataset and provides information on how the privacy-preserved data is useful to
the user. The sensitive information in the dataset is diminished such that the data
is removed, distorted, or transformed to obtain confidentiality [4]. The process of
finding the equilibrium between the utility and privacy of dataset remains obstinate
that leads to trade-off between them [5].
Specifically, in large datasets, the privacy and utility is a contradicting factor,
where either one should be sacrificed to achieve the other. If perfect privacy is
achieved without publishing the data, then perfect utility cannot be achieved since
original data is tend to be published without privacy. Hence, maintaining equilib-
rium between privacy and utility of datasets in e-health record datasets is of prime
importance.
The utilization of data anonymization in data publishing would be preferred to
obtain privacy [6], and clustering improves the data utility [2] in a fully distributed
database. Hence, in this paper, we propose an anonymization technique to preserve
the privacy of data in a fully distributed environment. Further, a clustering algorithm
is used to achieve the anonymization and provide resilient to probabilistic infer-
ence attack and linking attacks. Later, Hadoop distributed file system [7]isusedfor
distributing the privacy-preserved anonymized datasets.
The main contributions of the paper are given below:
1. A new privacy utility method is proposed using lightweight Elliptical Curve
Cryptography with PSO clustering.
2. The privacy is attained using lightweight ECC, and privatized data tends to get
published after the process of clustering.
3. The machine learning classification is carried out using PSO multi-label
clustering that reduces the errors in forming proper cluster.
2 Proposed Method
Initially, the privacy preservation is applied using lightweight ECC algorithm. After
the generation of privatized datasets, it is clustered using PSO clustering algorithm.
The clustering error obtained from PSO algorithm is used for the selection of cluster
with lowest clustering error. If the clustering error is lesser than the threshold level,
the data is published directly. On the other hand, the privacy parameters are adjusted,
and then the clustered instances are reclustered to obtain accurate clustered results.
Once the criteria are satisfied, the results are published.
The proposed method can perform data anonymization in order to provide
resilience against most common attacks, and then the data is sent to third party.
The third party access ensures privacy in distributed environment since the data is
anonymized in prior. Also, the data can be utilized to maintain utility without privacy
violation on a distributed environment.
Privacy Preservation Between Privacy and Utility Using ECC … 569
A. PSO Clustering Process:
Various techniques are used for achieving diversity by setting threshold values for
private attributes. On the other hand, it is very essential to categorize the private
attributes into different levels of confidentiality. It should further provide suitable
background information to assign threshold level for the private attributes. However,
skewing of data does not provide desired anonymization level and that satisfies t-
closeness constraints. The PSO data clustering [8] is carried out in this work to avoid
similarity attack. The similarity attack is handled using S-diversity principle.
The PSO clustering aims at dividing the entire class into various subclasses. The
algorithm is further divided into two major steps (i) partitioning and (ii) adjusting
step. The first step divides the data into multiple clusters, and the second step adjusts
the cluster size.
In order to further improve the uniformity of distributed clusters, we employ J48
and Naïve Bayes (NB) over PSO algorithm. This algorithm helps to generate the
clusters with uniformity such that the sensitive information is distributed uniformly.
The combination of J48 and NB with PSO finds the best neighbor in each cluster
group, and it assigns the values one at a time to the previous cluster. The proposed
method modifies the PSO algorithm to avoid the problem of skew in distributing
clusters of sensitive information using J48 and NB techniques. The nearest neighbors
are found using the J48 and NB algorithms over each cluster group with sensitive
information. Hence, the distribution of cluster is same in both the formed clusters
and original dataset. This avoids the probability of probabilistic inference attack in
data publishing.
The pseudocode of proposed J48 and NB with PSO is given in Algorithm 3.
Input: Privacy datasets
Output: Unequally clustered datasets with uniform distributions of sensitive
information
Step 1. The private dataset is sorted based on the values of sensitive attributes
Step 2. Then, the sorted input dataset is split into subgroups based on K,S—diver-
sity and Dmin—minimum distance between each itemset
Step 3. Check if the subgroup has the identical sensitive attributes values
Step 4. Else go to step 1, End
Step 5. Repeat the process for other subgroups or clusters
Step 6. Depending on the value of Kand S, we use two cases
Step 7. Case one, the value of Kis greater than S
(a) Cluster size =Dmin
(b) Distribute the records to all clusters, and it creates single element cluster
(c) Add nearest neighbor from the remaining cluster to the existing cluster
(d) Thus, the cluster formed has Sinstances with S-diverse private
information or sensitive information
(e) Go to 7(c)
(f) Remove the instances from clusters that are existing in each cluster
(g) Repeat the process
570 N. Yuvaraj et al.
Step 8. Find the centroid using ACO
Step 9. Find the instances in nearest neighbor from the clusters using Euclidean
distance estimation
Merge the single dataset with other subgroups.
3 Results and Discussion
The proposed method is implemented in Java environment on Eclipse IDE. The
Hadoop 1.2.1 is used to distribute the privacy-preserved dataset. Different classifiers
are executed to find the data utility of proposed algorithms using Weka 3.7.12. The
experiments are conducted to show the cluster size and total number of cluster formed
using the proposed method for measuring the data utility, privacy degree, execution
time, and scalability in HDFS. The experiments are conducted to compare the data
anonymization between the proposed and existing methods.
(A) Dataset
The proposed method is compared with conventional methods based on several exper-
iments conducted on the datasets. It includes adult dataset available [9], and its
description is given in Table 1. The experiments are further conducted on synthetic
data to determine the scalability proposed method. Two synthetic datasets are used
that consist of 200,000 and 500,000 records, respectively, each with ten attributes.
The experiments carried out with the given dataset are tested between the proposed
and existing Mondrian [11], Datafly [10], and Incognito [11] algorithms.
The experiments are conducted using Adult dataset. The proposed method is
evaluated in terms of discernibility cost and average equivalence class size. The
performance is compared in terms of F-measure and percentage of accurately clas-
sified instances. Finally, the proposed method is compared with existing methods in
terms of execution time.
The proposed system is tested in terms of two metrics, namely discernibility cost
and equivalence class size for determining the data utility loss of privacy-preserved
dataset. The lesser value of above metrics denotes the reduced information loss.
Tabl e 1 Discernibility cost (×108)
kIncognito Mondrian Datafly PSO
50 0.6 0.83 2.35 0.21
100 6.9 0.93 2.43 0.23
150 6.8 1.23 1.33 0.23
200 6.83 0.6 1.32 0.24
250 3.25 0.32 1.2 0.22
300 3.33 0.21 0.9 0.25
Privacy Preservation Between Privacy and Utility Using ECC … 571
Tabl e 2 Average values of equivalence class size
kIncognito Mondrian Datafly PSO
50 60 775 2
100 60 638 2
150 40 550 2
200 30 538 2
250 30 518 2
300 25 515 2
The Table 1provides discernibility cost based on the experiments conducted on
case one and case two datasets, respectively. The result shows that proposed method
obtains better performance in both the cases. The proposed method obtains reduced
discernibility cost than other algorithms, and it outperforms existing metrics in terms
of reduced discernibility cost.
The Table 2shows the results of average values of equivalence class size between
the proposed and existing algorithms, where the experiments are conducted using
anonymization toolbox. The average values of equivalence class size of proposed
method is lesser than existing algorithm in case of case one, and in case two, the
value is found higher for low k values, and it reduces gradually with increasing k
values. However, Mondrian algorithm obtains reduced equivalence class size than
proposed method.
Tables 3and 4show the percentage of correctly classified data instances and
F-measure for varying kvalues, say 5, 10, 25, and 30, where the experimentation
is carried out over case one. The result shows that performance of PSO algorithm
attains better performance than PSO algorithm and exhibits closer performance with
original datasets than other algorithm with Naïve Bayes and J48 classifier.
Further, the degree of privacy (Table 5) is experimented on both the cases using the
proposed algorithm. It is seen that attack links the value of individual QID attribute to
Tabl e 3 Accuracy of PSO clustering
kAlgorithm Incognito Mondrian Datafly PSO
5J48 84.2 84 84.24 84.05
NB 81.64 81.82 81.84 82.35
10 J48 84.2 83.96 84.24 84.05
NB 81.33 81.82 81.84 82.35
25 J48 84.2 83.99 84.28 84.05
NB 81.6 81.83 81.84 81.01
30 J48 84.2 84.04 84.26 84.24
NB 81.6 81.84 82 82.33
NB—Naïve Bayes; J48—J48 Classifier
572 N. Yuvaraj et al.
Tabl e 4 F-measure of clustered results
kAlgorithm Incognito Mondrian Datafly ACO
5J48 0.818 0.818 0.821 0.831
NB 0.809 0.809 0.81 0.818
10 J48 0.819 0.818 0.821 0.83
NB 0.81 0.809 0.81 0.818
25 J48 0.819 0.819 0.821 0.832
NB 0.809 0.809 0.81 0.805
30 J48 0.819 0.818 0.821 0.83
NB 0.809 0.809 0.81 0.805
NB—Naïve Bayes; J48—J48 Classifier
Tabl e 5 Results of privacy
degree Case kNB-ACO J48-ACO
One 23227 1423
53227 1423
10 3227 1423
15 3234 1365
25 3875 2476
Two 50 5 3
100 5 3
150 5 3
200 5 3
250 5 3
300 5 3
the value sensitive attribute with confidence level of P−1. In case one, the probability
of record linking is higher in ACO clustering algorithm than in KNN-ACO clustering
algorithm. The probability of record linking using case two is set to exhibit higher
linkage in ACO clustering algorithm than KNN-ACO clustered dataset. There exists
a 50% chance of record linking in case two using ACO clustering algorithm than
KNN-ACO clustering algorithm where it shows only 25% record linkage.
From the above three measures, we found that there exists a trade-off between
the data utility and data privacy. This is determined in terms of total number of
clusters formed using the proposed method, and this exhibits the trade-off. As the
total number of clusters increases, the data privacy tends to reduce with increased data
utility. Such reduction of data privacy with increased data utility is due to reduced
number of instances with increased clusters from the given datasets. This sets to form
reduced information loss; however, the level of privacy is low.
Privacy Preservation Between Privacy and Utility Using ECC … 573
4 Conclusions
In this paper, we propose data anonymization technique to provide equilibrium
between data utility and data privacy. The proposed data anonymization technique
using PSO clustering and ECC privacy-preserved datasets is experimented over adult
datasets. The evaluation metrics like discernibility cost and average equivalence class
size shows reduced value with improved performance in terms of data utility. Addi-
tionally, the data utility is measured in terms of accuracy and F-measure, and the result
shows that proposed method obtains improved performance than existing methods.
References
1. L. Sweeney, k-anonymity: a model for protecting privacy. Int. J. Uncertainty Fuzziness Knowl.
Based Syst. 10(05), 557–570 (2002)
2. W.Y. Lin, D.C. Yang, J.T. Wang, Privacy preserving data anonymization of spontaneous ADE
reporting system dataset. BMC Med. Inform. Decis. Mak. 16(1), 58 (2016)
3. J. Soria-Comas, J. Domingo-Ferrer, D. Sánchez, S. Martinez, t-closeness through microaggre-
tion: strict privacy with enhanced utility preservation. IEEE Trans. Knowl. Data Eng. 27(11),
3098–3110 (2015)
4. K. Mivule, C. Turner, S.Y. Ji, Towards a differential privacy and utility preserving machine
learning classifier. Procedia Comput. Sci. 12, 176–181 (2012)
5. R.C.W. Wong, A.W.C. Fu, K. Wang, J. Pei, Minimality attack in privacy preserving data
publishing, in Proceedings of the 33rd International Conference on Very Large Data Bases
(VLDB Endowment, 2007), pp. 543–554
6. K. Chen, L. Liu, Privacy-preserving multiparty collaborative mining with geometric data
perturbation. IEEE Trans. Parallel Distrib. Syst. 20(12), 1764–1776 (2009)
7. I. Polato, R. Ré, A. Goldman, F. Kon, A comprehensive view of Hadoop research—a systematic
literature review. J. Netw. Comput. Appl. 46, 1–25 (2014)
8. Y., Kao, K. Cheng, An PSO-based clustering algorithm, in International Workshop on PSO
and Swarm Intelligence (Springer, Berlin, Heidelberg, 2006), pp. 340–347
9. M. Lichman, UCI Machine Learning Repository, 2013. http://archive.ics.uci.edu/ml. Accessed
28 Aug 2018
10. L. Sweeney, Datafly: a system for providing anonymity in medical data, in Database Security
XI (Springer, Boston, MA, 1998), pp. 356–381
11. K. LeFevre, D.J. DeWitt, R. Ramakrishnan, Incognito: efficient full-domain k-anonymity, in
Proceedings of the 2005 ACM SIGMOD International Conference on Management of Data
(ACM, 2005), pp. 49–60
Predicting Energy Demands Constructed
on Ensemble of Classifiers
A. Daniel, B. Bharathi Kannan, N. Yuvaraj, and N. V. Kousik
Abstract Analysis of time series data and perfect future prediction are the most
stimulating tasks that the data analysts face in countless fields. Forecasting of energy
demands is very essential because both excess energy cost and delay lead to a signif-
icant reduction storing costs. In order to discover the uniformities in dynamic, non-
stationary data and time series prediction needs the use of models to be integrated
with multiple forecast models. The Ensemble learning model discovers the dynamic
patterns in energy time series data. The performance of two different Ensemble
learning techniques is compared against Bagging and stacking in forecasting energy
time series data. Stacking technique used in this paper combines different classifiers
like Multilayer Perceptron (MLP), Radial Basis Function (RBF), and Support Vector
Machine (SVM).
Keywords Predicting ·Ensemble ·Neural networks ·Stacking techniques ·
Dynamic pattern
1 Introduction
Time series analysis and perfect prediction play a significant role in various fields of
science like electricity demand predicting, financial, weather predicting, etc. Short-
term forecasting is the process of predicting events in short time periods such as days,
A. Daniel (B)·B. Bharathi Kannan ·N. V. Kousik
School of Computing Science and Engineering, Galgotias University, Greater Noida, Uttar
Pradesh, India
e-mail: a.daniel@galgotiasuniversity.edu.in
B. Bharathi Kannan
e-mail: bharathi.kannan@galgotiasuniversity.edu.in
N. V. Kousik
e-mail: nvkousik@galgotiasuniversity.edu.in;nvkousik@gmail.com
N. Yuvaraj
ICT Academy, Chennai, Tamil Nadu, India
e-mail: yraj1989@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981-15-5566- 4_52
575
576 A. Daniel et al.
weeks, and months. Medium-term forecasting involves prediction of future values
which can extend up to two years, whereas long-term forecasting problems can be
extended for many years [3].
Energy demand planning is an important factor for any country’s economic
growth. Predicting allows decision makers to make accurate conclusions and to get
competitive advantage. Improving the accuracy of the forecasting technique is the
major issue in time series analysis. Well-organized energy scheduling requires accu-
rate forecast of future energy demands. There are quite a lot of statistical methods for
time series prediction such as ARX, ARIMA, ARMA, GARCH, Box and Jenkins, and
Smoothing techniques. ML Techniques such as Artificial Neural Networks (ANN),
SVM, and Decision Trees are also widely used for time series prediction. The
forecasting accuracy of each model is different from each other.
The selection of a particular model is based on the accuracy needed the computa-
tional overhead incurred. In [2], it is reported that ANN provides better forecasting
accuracy than statistical techniques such as ARIMA and Multiple Linear Regres-
sion (MLR). But the computational complexity of ANN was higher over other two
techniques. Each forecasting method has its own advantage as well as disadvan-
tage in time series prediction. Ensemble methods improve the forecasting accuracy
by combining the hypothesis generated from individual base learners for the same
training data set [7]. Ensemble learning objective is to improve the performance of
classifier or predictor. Model Selection is an important task in ML. Accuracy depends
on type of classifier used and which realization of the classifier is to be chosen. For
example, in ANN, different initialization of parameters gives different outcomes.
An Ensemble model which consists of a group of learners is usually weaker
than an Ensemble. Ensemble learning is more advantageous because it can boost
weak learners to become strong learners which makes highly accurate forecast. Most
Ensemble methods such as Bagging and Boosting generate a homogeneous base
learner by making use of a single learning algorithm. There are also some other
methods such as Stacking and Voting which produce a heterogeneous learner using
multiple learning algorithms [7].
Two different steps can be used to build an Ensemble. The base learner generation
can be sequential or parallel while the base learner generation has an effect on the
future generation of learners. The base learners should be as accurate as possible
to get a more precise Ensemble. The accuracy of the learners is assessed by many
effective steps; for example, cross-validation, random sampling, stop testing, etc.
The advantages of various individual techniques are combined in Ensemble
learning, thereby improving the accuracy of predicted values. Accuracy is also based
on the Meta-learner to be chosen in the process of creating an Ensemble. This paper
compares the performance of the Ensemble with that of individual methods in the
prediction of energy demand.
Predicting Energy Demands Constructed on Ensemble of Classifiers 577
2 Literature Survey
From literature, a number of pattern recognition techniques are used for time series
forecasting. Those techniques include ARIMA and GARCH models, artificial neural
networks, fuzzy logic, genetic algorithms, and Support Vector Machines (SVM).
Ensemble of multiple classifiers can improve the prediction power rather than using a
single classifier. Combining the predicted results of a number of classifiers will signif-
icantly improve the accuracy of the prediction algorithm. Martínez-Álvarez proposed
the Pattern Sequence-based Forecasting (PSF) algorithm [3], which predicts the
future values of a time series based on pattern sequence similarity.
PSF produced a better prediction of energy time series compared to other well-
known techniques. Karin Kandananond [6] compared the performance of ANN
approach with that of ARIMA and Multiple Linear Regression (MLR) models. ANN
model outperforms MLR and ARMIA models in terms of accuracy. They reported
that even though accuracy of ANN is higher the computational overhead of ANN is
also higher compared to other models.
Wen and Vahan [1] proposed an algorithm Pattern Forecasting Ensemble Model
(PFEM) for day-ahead Energy Demand Forecasting. They combined various clus-
tering techniques like K-Means, SOM, Hierarchical Clustering, Fuzzy C Means,
K-Mediods, and applied forecasting models for the clustered output prediction of
the day ahead energy demand. The performance of PFEM was proved to be better
than other models.
Pasapitch and Nittaya [4] compared the performance of ARIMA and ARMA
models in forecasting household power consumption patterns. ARIMA model is
proved to be better for monthly and quarterly analysis, and ARMA model performs
better for daily and weekly analysis.
Reinaldo and Javier [8] forecasted day ahead electricity prices of Spain and
California Electrcity Markets with GARCH model.
Chitra [5] used an Ensemble of multiple classifiers such as Self-organizing Map,
K-Nearest Neighbors, and Radial Basis Function for time series prediction. The
performance of individual learners with that of Ensemble model is compared using
three different data sets. In this paper, the performances of two different Ensemble
learning techniques were compared by combining various classifiers such as RBF,
SVM, and MLP.
3 Methodology
Numerous types of classifiers are existing for pattern prediction. RBF, MLP, and
SVM are considered as the best methods for non-linear time series analysis and
forecast. Ensemble learning is one of the machine learning techniques in which
multiple learners are trained to solve a single problem. There are various Ensemble
methods: Bagging, Boosting, Stacking, and Voting.
578 A. Daniel et al.
Model Selection is a vital factor in supervised learning. Which model is best suited
for the given problem? There are two different ways to clarify this question: i) what
type of model is to be chosen among many competing models, such as MLP, SVM,
etc.; ii) given a particular prediction algorithm, which realization of this algorithm
is to be chosen, i.e., different initializations of SVMs give rise to different decision
boundaries, even if all other parameters are kept constant.
The most commonly used technique is to choose the algorithm which gives the
smallest error and high accuracy. But it is a flawed one; most of the techniques eval-
uating the accuracy such as cross-validation may be misleading because it provides
less accuracy for previously unseen data. So, Ensemble learning is the best alterna-
tive for this model selection problem. The following section describes two different
Ensemble learning techniques, Stacking and Bagging, respectively.
3.1 Stacking
The block diagram of Stacking is shown in the following Fig. 1.TheEnsemble
process is done with the following four steps: Data Processing, Applying Individual
Learner, Classification of the new data set with the Meta-learner, and Evaluation
of predicted output. It consists of three different classifiers: Radial Basis Function
Neural networks, MLP, and SVM. In the Stacking algorithm, a number of first-level
individual learners are generated from the training data set D by employing different
first-level learning algorithms L1, L2, …, LT.
Those individual learners are then combined by a second-level learner “L” which
is called as meta-learner. Ensemble learners are superior to single learners because
Fig. 1 Stacking of
classifiers
Predicting Energy Demands Constructed on Ensemble of Classifiers 579
of various reasons. The training data might not contain sufficient inputs for choosing
a single learner. The search process of the Ensemble learners is better than indi-
vidual learners. The hypothesis space might not have actual target function. Ensemble
learners are used whereever machine learning techniques can be applied.
Algorithm:
Input: Data Set D
First-level Learning algorithms RBF, MLP, and SVM.
Second-level Learning algorithm L.
1. Train the first-level individual learners h1, h2, and h3 by applying the first-level
learning algorithms RBF, MLP, and SVM to the original data set D.
2. Generate a new data set D’.
3. Apply individual learners h1, h2, and h3 to train the data set D.
4. Assign the result of first-level learning to the new data set D.
5. Train the second-level learner h’ by applying second-level learning algorithm to
the new data set D.
Output
Output of Stacking is the final predicted output of two-level Ensemble. The above
algorithm gives steps of the two-level Ensemble approaches with the set of base
learners and the Meta-learner. As a first step, the input data is preprocessed and
given to the set of base learners. Finally, the resultant data set is trained with the
Meta-learner. Hence, it uses the two-level heterogeneous Ensemble approach; the
generalization ability of the Ensemble is much higher than that of individual learners.
3.2 Bagging
Figure 2trains various base learners with the help of bootstrap samples obtained
from the training data set. Bootstrap sample is obtained from creating subsamples
from the original data set with replacement. The size of the bootstrap sample is same
as that of the training set. Bootstrap samples are generated from the given training
set as random samples. So there is a possibility of choosing the same tuples again.
For instance, the given data contains d tuples, and the data set is sampled d times,
thereby generating d bootstrap samples. Therefore, each tuple has a probability of
1/d to be chosen and has the probability of 1–1/d not to be chosen.
580 A. Daniel et al.
Fig. 2 Bagging of classifiers
Algorithm:
Input: Data Set D
Base Learning Algorithm L Number of rounds I
Process
For i =1,2, … I
1. Create Bootstrap sample Di from the data set D.
2. Train the base learner hi with the bootstrap
sample Di End
Output
Output of bagging is the average of predictions from the base learner. Like in
Voting Ensemble Learning, for numeric prediction, predictions of base learners are
averaged. For classification problems, Predictions of base learners are taken majority
voting.
Predicting Energy Demands Constructed on Ensemble of Classifiers 581
4 Results and Discussions
The data set was obtained from US department of Energy. The data set contains the
monthly energy consumption data. The data set is available at the following site:
http://www.eia.gov/totalenergy/data/monthly/index.cfm#consumption. The perfor-
mance of the above algorithms is evaluated with the following metrics.
Performance Metrics
The following are various metrics that are used for evaluating the performance of the
classifiers with that of Ensemble Model: Mean Absolute Error (MAE), Root Mean
Squared Error (RMSE), Mean Squared Error (MSE), and Mean Absolute Percentage
Error (MAPE).
A comparison of Prediction Accuracy, MAE, and RMSE of the algorithms RBF,
MLP, SVM, and Stacking and bagging is given below.
The data set is evaluated with hold-out evaluation technique, where the given data
is randomly partitioned in training and testing sets. Therefore, one-third of the data
is used as the test set. The data set consists of 490 records in which 327 records are
chosen as training set and remaining 163 records are used as test set (Table 1).
RBF Ensemble is the Ensemble obtained as a result of applying bagging to RBF
neural network. Similarly, SVM Ensemble and MLP Ensemble are the results of
bagging of SVM and MLP, respectively (Table 2).
The following chart shows the accuracy of the different algorithms in forecasting
of energy time series data (Figs. 3,4).
Tabl e 1 Performance comparison of classifiers with ensemble
Methods MAE RMSE MSE MAPE
RBF 0.2854 0.3783 0.1431 3.3788
MLP 0.395 0.499 0.249 4.8151
SVM 0.2153 0.2792 0.0779 2.6585
Stacking 0.2861 0.3477 0.1209 3.4222
Tabl e 2 Performance comparison of bagging of different classifiers
Classifiers MAE RMSE MSE MAPE
RBF 0.2989 0.3956 0.1565 3.5394
MLP 0.2643 0.3563 0.1269 3.1497
SVM 0.1864 0.2482 0.0616 2.2767
582 A. Daniel et al.
Fig. 3 Prediction accuracy
of classifiers in time series
forecasting
Fig. 4 Accuracy of
classifiers after bagging
5 Conclusion
The energy demand forecasting is very essential for optimization of energy resources
and application of green trends. The proposed Ensemble model can be applied for
prediction of both linear and non-linear time series data and can be used in wide
range of applications in time series forecasting. The experimental results show that
the Ensemble model outperforms all other individual learners.
References
1. S. Wen, B. Vahan, An Ensemble Model for Day-ahead Electricity Demand Time Series
Forecasting, e- Energy’13 (2013)
2. L. Felix, K. Wolfgang, Classification and clustering of electricity demand patterns in industrial
parks. Energies 5, 5215–5228 (2012)
3. A. Francisco, T. Alicia, Energy time series forecasting based on pattern sequence similarity.
IEEE Trans. Knowl. Data Eng. 23(8), 1230–1243 (2011)
Predicting Energy Demands Constructed on Ensemble of Classifiers 583
4. C. Pasapitch, K. Nittaya, Time series analysis of household electric consumption with ARIMA
and ARMA models, in Proceedings of the International Multi Conference of Engineers and
Computer Scientists (2013), pp. 295–300
5. A. Chitra, S. Uma, An ensemble model of multiple classifiers for time series prediction. Int. J.
Comput. Theor. Eng. 2(3) (2010)
6. K. Karin, Forecasting electricity demand in thailand with an artificial neural network approach.
Energies 4, 1246–125 (2011)
7. Z.-H. Zhou, Ensemble Methods: Foundations and Algorithms, 1st edn (CRC Press, 2012)
8. C.G. Reinaldo, C. Javier, A GARCH forecasting model to predict day-a head electricity prices.
IEEE Trans. Power Syst. 20(2), 867–874 (2005)
Use of RNN in Devangari Script
Madhuri Sharma and Medhavi Malik
Abstract In India, Devanagari script is used most prevalently. But recognizing the
handwritten documents of Devanagari Script is a challenging task due to variations
in writing, rendering, and cursive nature of handwriting. However, there has been a
lot of improvement in research in the field of recognizing Handwritten Documents.
In this, we will implement a framework that inputted the number of images gives an
output for recognized characters.
Keywords Devanagari script ·Natural language processing ·RNN
1 Introduction
Devangari script is used in Hindi and Sanskrit. Devanagari script contains 47 primary
characters; 14 are vowels and 33 are consonants. In HWR (Handwritten Recognition),
the model is trained for correct classification. For this, Supervised Learning, Unsuper-
vised Learning, and Reinforcement Algorithm are applied to train the model. Super-
vised learning is learned from Label data, while for unsupervised algorithm there
is no labeled data; so Clustering algorithm and K-Means algorithm are used. Rein-
forcement algorithm provides learning from previous experience. They can depict
syntactic data (e.g., grammatical form labeling, piecing, and parsing) or semantic
data (e.g., word-sense disambiguation, semantic job naming, named element extrac-
tion, and anaphora goals) [1]. There are no sufficient number of work on Indian
language character recognition although there are 12 major scripts in India [2].
M. Sharma (B)·M. Malik
SRMIST, Delhi-NCR Campus, Modinagar, India
e-mail: madhurisharma44@gmail.com
M. Malik
e-mail: medhavimalik28@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_53
585
586 M. Sharma and M. Malik
2 Identification of Handwritten Recognition (HWR)
After analyzing various word samples in Devanagari script, it has been found that, in
this script, generally words are written in non-cursive fashion which has prompted
us to use Recurrent Neural Network (RNN) models, as RNN models require pre-
segmented data [3]. To recognize handwritten words, it usually undergoes number
of stages. Following are the steps that can recognize a word:
•Process a word: This is the first step to process the handwritten words. First,
convert the image into Gray Scale image havingwhite background and black pixels
as text. Then, noise has to be removed since it will create a problem in the process
of segmentation. Various types of noise removal techniques are used: Smoothing,
Filtering, and Extraction. After this, normalization procedure is performed to
ensure the values lie in between −1 and 1. There are a number of different ways
for handwritten recognition style.
•Segmentation: Segmentation is in which text is divided into number of mean-
ingful units: Graphemes represent the lowest level characters; one or more
characters represent words, and sentences are made up of number of words (Fig. 1).
Word Segmentation is the process of dividing a text into number of tokens.
Sentence Segmentation is the process of breaking down into number of words sepa-
rated by boundaries in Devanagari Script (puran viram (|)). Both word and sentence
segmentations cannot be applied simultaneously and are independent of each other.
The inputted text is tokenized, and all the inconsistencies like spelling checking,
punctuation marking, grammar checking, etc., are removed. If there is any mistake
in grammar and spelling, it automatically corrects and gives the suggestions to the
user. The special words like numbers, date, time, abbreviations, and special symbols
(like: ) are identified [4].
But, there are some limitations that must be addressed in developing algorithms
for text segmentation which are described below:
Language Dependency: Written adaptations of language have a diverse set of
features.
Character Dependency: All texts are encoded in seven-bit ASCII-encoded set
which allowed only 128 characters and essential characters for writing English.
Extended character sets allow eight-bit encodings, but an eight-bit encoding allows
just 256 distinct characters. Still, there are tens of thousands of distinct characters
in all the writing systems. So, the main reason behind these characters in different
languages is currently encoded in a large number of overlapping character sets. The
HW
Document
Segmentation
Framework
Segments
Fig. 1 Segmentation
Use of RNN in Devangari Script 587
Unicode Standard, Version 12.0 specifies a single two-byte encoding system that
includes a range from U+090x to U+ 097x.
Application Dependency: These may include Morphological Analyzers, Part-of-
Speech taggers, Lexical Look-up Routines, or Parsers which are being used in its
process.
Corpus Dependency: This is due to the availability of large amount of corpus,
misspellings, erratic punctuation, and irregular white spaces.
•Feature Extraction: Feature extraction is the main component of the text recog-
nition. Various techniques are used in this process: LDA, PCA, and other various
algorithms to train the model.
Statistical Features: Statistical features include size, shape, and intensity. Various
statistical features are Area, Eccentricity, Orientation, Centroid, BoudingBox, Majo-
rAxisLength, MinorAxisLength, ConvexHull, ConvexImage, ConvexArea, Filled-
Image, and FilledArea. Various methods used are Moment, Zoning, Projection
Histogram, and Crossing and Distances.
Structural Features: Structural features are the geometrical features of the image
of the character. Some of the features are Vertical Line, Horizontal Line, End Points,
Number of Crossing Points, and Number of Contours. Various methods used in the
process are Geometrical and Topological Features, Coding, Graphs, and Trees.
Raw Features: Pixel values and deep learning like Convolution Neural Network
are used to learn features by themselves.
•Recognition Method: This step will assign a label from the given set of classes
or predict the labels using RNN. Various recognition methods are used in this
process: HMM (Hidden Markov Model), SVM (Support Vector Machine), CNN
(Convolution Neural Network), and MLP (Multilayer Perceptron) (Fig. 2).
3 Recurrent Neural Network (RNN)
RNN is a feedforward neural network which requires an internal memory for storing
the output at any stage since it is inputted back to the previous stage. RNN for every
input data performs the same function since it is recurrent in nature; however, input
is independent of each other. For decision, it depends on the previous input, current
input, and output from a node at a stage.
Recurrent nature of this helps in performing various tasks: Speech Recognition,
Content Handwritten Recognition, and Unsegmented task.
Current state is described as in (Fig. 3).
ct=f(ct−1,it)
Applying Activation Function:
ct=activation_ functionwpsct−1,wicit
588 M. Sharma and M. Malik
Fig. 2 Steps
Segmentation
Process a Word
Feature
Extraction
Recognition
Method
Fig. 3 RNN network
it
N
ct
Output:
ot=wcoct
4 Types of RNN
Different types of RNN are as follows:
Use of RNN in Devangari Script 589
One to One: It maps the fixed size of input to the fixed size of output, which does
not depend on the previous output. Example: Classifying an image, depending upon
the visual.
One to Many: It maps the fixed size as an input of data and provides the output
as a sequence of data. Example: Gives the textual description of an image.
Many to One: It gets the sequence of information as an input and provides a
fixed size output. Example: Automatically processes the data to detect positive and
negative opinions from a text.
Many to Many: It takes the sequence of information and process it recurrently,
produced an output a sequence of data Example: In Machine Translation, to convert
the one language to another.
5 Construct a Next Word Predictor
Main goal is to predict the next word; which can be used in many applications.
Pre-processing: First step is to make a corpus of unique words. But, it has to be
remembered that as the number of unique words in the corpus increased, complexity
also increased a lot.
It starts with the collection of how frequently each word co-occurs with one other
in a corpus.
Probabilistic Model: Each word is replaced with a frequently used word which
makes a vector of it. This vector makes a probabilistic model of it. The result of this
provides useful information for the next phase.
Model RNN Architecture: A Recurrent Neural Network (RNNs) is a model
which contains a self-connected hidden layer. This gives an efficient method for the
connections between neurons and makes a directed cycle.
It makes sequences of data to predict the favorable outcome. Suppose, we are
taking follwoing as an input: and it will predict the output. First, input
is provided as x0, and produces an output h0. For the next round with input x1,it
produces an output h1. At last, the output is got from ht, which demonstrates the next
word prediction.
The output particularly depends on the percentage of favorable words and predicts
the output:
590 M. Sharma and M. Malik
6 Conclusion
This paper demonstrates the development of a structure of Devangari Script using
RNN concept. This work will be extended toward online handwritten recognition
that will be more beneficial to the user. In the future, we would like to relax the
assumption of having pre-segmented images. We also plan to extend this work to
include other Indic scripts.
References
1. Medhavi Malik, Bersha Kumari, Madhuri Sharma, “Steps: To Make NLP More Enhance by
New Techniques”. Int. J. Eng. Adv. Technol. (IJEAT) 8(2S2), ISSN: 2249 – 8958 (2019)
2. U. Pal, B.B. Chaudhuri, Indian script character recognition: a survey. Patt. Recogn. 37(9),
1887–1899 (2004)
3. P. Keshri, P. Kumar, R. Ghosh, RNN based online handwritten word recognition in devanagari
script, in 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR)
(2018)
4. M. Sharma, M. Malik, N. Gupta, An efficient & learning approach of POS tagging using rule-
based for Devanagari script. Int. J. Innov. Technol. Exploring Eng. (IJITEE) 8(7C2). ISSN
2278-3075 (2019)
5. R.J. Milewski, V. Govindaraju, and A. Bhardwaj, Automatic recognition of handwritten medical
forms for search engines. IJDAR (2009)
6. N. Sankaran, A. Neelappa, C.V. Jawahar, Devanagari text recognition: a transcription based
formulation. ICDAR (2013)
7. J. Li, W. Monroe, A. Ritter, M. Galley, J. Gao, D. Jurafsky, Deep reinforcement learning for
dialogue generation. arXiv preprint arXiv: 1606.01541 (2016)
8. B. Shaw, S.K. Parui, M. Shridhar, Offline handwritten devanagari word recognition: a holistic
approach based on directional chain code feature and hmm, in International Conference on
Information Technology (ICIT) (2008)
9. R. Plamondon, S.N. Srihari, Online and off-line handwriting recognition: a comprehensive
survey. IEEE Trans. Pattern Anal. Mach. Intell. 22(1), 63–84 (2000)
10. P. Voigtlaender, P. Doetsch, H. Ney, Handwriting recognition with large multidimensional long
short-term memory recurrent neural networks, in 15th International Conference on Frontiers
in Handwriting Recognition (ICFHR) (2016)
11. B. Shaw, U. Bhattacharya, S.K. Parui, Combination of features for efficient recognition
of offline handwritten devanagari words, in 14th International Conference on Frontiers in
Handwriting Recognition (ICFHR) (2014)
12. R. Ishida, An introduction to indic scripts, in Proceedings of the 22nd International Unicode
Conference (2002)
Use of RNN in Devangari Script 591
13. U. Consortium et al., The Unicode Standard, Version 2.0 (Addison-Wesley Longman Publishing
Co., Inc., 1997)
14. N. Sankaran, C. Jawahar, Recognition of printed devanagari text using BLSTM neural network,
in 21st International Conference on Pattern Recognition (ICPR) (2012)
15. V. Bansal, R. Sinha, A complete OCR for printed hindi text in devanagari script, in 6th
International Conference on Document Analysis and Recognition (DAS) (2001)
16. R.J. Milewski, V. Govindaraju, A. Bhardwaj, Automatic recognition of handwritten medical
forms for search engines. Int. J. Doc. Anal. Recog. (IJDAR) (2009)
17. P. Liu, X. Qiu, X. Chen, S. Wu, X. Huang, Multi-timescale long short-term memory neural
network for modelling sentences and documents, in Proceedings of the Conference on
Empirical Methods in Natural Language Processing (2015)
18. A. Maas, R.E. Daly, P.T. Pham, D. Huang, A.Y. Ng, C. Potts, Learning word vectors for
sentiment analysis, in Proceedings of the ACL (2011), pp. 142–150
19. B. Pangand, L. Lee, Asentimental education: sentiment analysis using subjectivity summariza-
tion based on minimum cuts, in Proceedings of ACL (2004)
20. L.-C. Yu, J. Wang, K.R. Lai, X. Zhang, Refining word embeddings for sentiment analysis, in
Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing
(2017), pp. 545–550
21. L. He, K. Lee, M. Lewis, L. Zettlemoyer, Deep semantic role labeling: what works and what’s
next, in Proceedings of the Annual Meeting of the Association for Computational Linguistics
(2017)
22. D. Bahdanau, K. Cho, Y. Bengio, Neural machine translation by jointly learning to align and
translate. arXiv preprint arXiv:1409.0473 (2014)
23. A. Graves, Generating sequences with recurrent neural networks. arXiv preprint arXiv:
1308.0850 (2013)
24. Y. Kim, Convolutional neural networks for sentence classification, in Proceedings of the 2014
Conference on Empirical Methods in Natural Language Processing (EMNLP 2014), pp. 1746–
1751 (2014)
25. C. Dyer, M. Ballesteros, W. Ling, A. Matthews, N. A. Smith, Transition-based dependency
parsing with stack long short-term memory. arXiv preprint arXiv: 1505.08075 (2015)
Role of Data Science for Combating
the Problem of Loan Defaults Using
Tranquil-ART1NN Hybrid Deep
Learning Approach
Chandra Shaardha and Anna Alphy
Abstract Banking sector is facing a watershed with the problem of huge NPAs
being tackled with various measures such as IPB, Recapitalization to name a few.
Credit monitoring is indispensable to protect the lenders’ investment. An effective
monitoring system will protect and identify the loans which are to be red-flagged. The
resources of the bank manager to monitor the loans are limited. The parameters for
such surveillance are many. Some of them are available to the banker through systems
in place, while some others have to be instituted with the networking capabilities and
oversight of transactions of borrowers. With thousands of loans at various stages
of fulfillment, it would be a herculean and cumbersome task for the conventional
RDBMS and SQL systems to segregate the loans into categories that require better
attention and follow-up. This paper aims to present the possibilities with the banking
managers to make use of the advancements and also proposes Tranquil-ART1NN
approach using Artificial Intelligence (AI) to give advanced signals to bank offi-
cers regarding NPAs/Special Mention Accounts. Tranquil algorithm is meant for the
identification of an initial better neighborhood and the Intelligent Huddle Creation
algorithm is meant for creating clusters of borrowers based on their performance.
The clusters of borrowers are trained using ART1NN approach to identify the opti-
mized clusters of borrowers for early identification of NPAs. The quality of the
proposed system is compared with various traditional algorithms like Genetic clus-
tering (GC), K-means, Ant Colony Optimization (ACO), etc.,and the experimental
results show that our proposed system has a better performance in terms of purity
and DB Index, when compared to K-Means, ACO, and GC. The early detection of
default will enhance the profitability, liquidity of the banks. Technology will help
bankers to eliminate the gap between resources and workload available and will lead
to improved risk management and efficiency of the banks.
C. Shaardha (B)
Department of Management Studies, SRMIST, Delhi-NCR Campus, Ghaziabad, UP, India
e-mail: chandras@srmist.edu.in
A. Alphy
Department of Information Technology, SRMIST, Delhi-NCR Campus, Ghaziabad, UP, India
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_54
593
594 C. Shaardha and A. Alphy
Keywords Lending ·Credit monitoring ·Defaults ·NPAs ·Borrower profiles ·
Pheromones ·Neural network ·Adaptive resonance theory ·Ant colony
1 Introduction
“Money,” is the vitamin source for human life. “Fund,” another term of Money is
the essential source for any business development activities. In the banking scenario,
lending funds to an individual, organizations, and institutions is the major role to
play, to earn profit for the investors. When the borrower defaults in repayments, it
affects the profitability and liquidity of the banks.
Indian banking sector is under pressure to control the defaults of repayments and
dues from the borrowers. As per RBI guidelines, when the borrower fails to repay the
principal/interest for 90 days and when the loan account ceases to generate income,
the loan has to be classified/termed as “Non-Performing Assets” (NPAs). As on date,
the rising NPAs percentage gives increasing pressure to the banking sector. Gross
Non-Performing Assets (GNPAs) of Public Sector Banks (PSBs) have reduced by
Rs. 89,189 crores from Rs. 8,95,601 crores in March 2018 to Rs. 8,06,412 crores in
March 2019. The GNPAs as a percentage to overall loans remained at 9.3% as of 31
March 2019.
One of the major reasons for the increase in the NPAs percentage is the lending
design adapted to the banking sectors. The Nirav Modi’s Letter of Guarantee fraud
is the latest case in point to illustrate the faulty methods of lending by the banking
institutions.
RBI has taken various measures to curtail the NPAs percentage over years. But,
still it is very difficult to control as well to monitor the stressed assets. Credit rating
agencies have also initiated different methods to analyze the creditworthiness of the
borrower and submit the report to banking institutions before the disbursement of
loans. Even after the compliance of the credit rating agencies report, the borrower
makes default in repayment after availing the loans. The bankers are only giving
importance to analyze the borrower’s background before the disbursal of loans. Once
loans get disbursed, due to the workload/work pressure/manpower shortage, bankers
do not have much time to see or have an eye on the repayment of dues on time in the
loan accounts. Hence, it is important to report the issue of non-performance much in
advance before the loans convert to a stressed or Special Mention Account.
It is important to make certain the early recognition of loan repayments in the
borrower’s account. The bankers should be trained to identify the borrower’s accounts
with the help of symptoms they could see during the transactions. It becomes neces-
sary to have a proper mechanism to identify the factors that could avoid the loan
turning into NPAs. When the symptoms are seen in the borrower’s loan accounts, it
is the duty of the officer to make an honest assessment of that account and give a
proper alarm or follow-up to initiate the steps to identify the reason behind the cause.
Role of Data Science for Combating the Problem of Loan Defaults … 595
Generally, the steps are not taken on time due to various issues. It becomes manda-
tory for the bankers to incorporate the monitoring procedures with the risk manage-
ment process to ensure that the early warning symptoms are duly recognized and
followed up with due diligence to recover the dues without any default.
To identify the preventive measures of loans turning to be NPAs, RBI have given
the guidelines to monitor the loan accounts. The checkpoints are given below, where
bankers should be cautious about
1. Non-proper submission of (a) Stock statement, (b) Balance sheet, (c) P&L
accounts
2. Cheque Returns
3. Default in deferred payment guarantee
4. Default in payment of Letter of Credit
5. Continuous default in payment of Bank Guarantee
6. Return of Cheque/Bills discounted
7. Default in payment of Bills/Cheque discounted
8. Decline in Sales/Profit
9. Cash/Fund loss
10. Decline in Net worth
11. Incomplete documentation
12. Not fulfilling the terms and conditions
13. Special Mention Accounts I, II.
This paper tries to identify the possibilities of utilizing the latest technology of
the information systems with the help of Machine learning/Artificial Intelligence to
prompt/give a signal to bank officers in advance about the capabilities/capacity of
the borrower to make the repayment of loan due before it turns into NPAs/Special
Mention accounts.
2 Related Works
Author Naveenan [1] in his paper titled “Warning signals—A Tool to control NPAs in
banks” has strived to analyze various opinions of bank managers about the warning
signals identified by the Reserve Bank of India in predicting slippages and the resul-
tant NPAs. He has analyzed four types of warning signals, namely Financial, Opera-
tional, Banking, and Managerial. Under each category, the author has chosen different
parameters. For example, under the managerial category, he has narrowed on diver-
sion of funds, laxity in financial regime, poor interpersonal teamwork, miscalculated
risks, etc. Using non-parametric Friedman test, he concludes that ranking based on
means of warning signals in each category would be a valid approach.
HackerEarth [2] hosted a hypothetical problem of a bank facing huge bank
defaults. It challenged hackers to develop machine learning solutions to help the
bank. The challenge was attempted by more than 3000 ML enthusiasts across the
world. The solution of top performers with their statements is described in the site.
596 C. Shaardha and A. Alphy
The participants have used feature engineering to create new features such as to esti-
mate the number of periods when the loan was paid, ratio of insurance paid to the
total insurance due, etc., and used ML algorithms such as XGB.
Daietal.[3] introduced particle swarm chaos optimization mining algorithm
(PSCOMA). It used the strong global search ability of PSO and the strong local
search ability of chaos optimization for the process of web usage mining. It offered
a balancing between exploration and exploitation. Kennedy et al. [4] presented
particle swarm optimization (PSO) which is an evolutionary swarm intelligence-
based computational model PSO inspired by bird flocks. Here, each swarm repre-
sents a solution set. The swarms or particles fly through the solution space. Each
position of the particle in the problem space represents a solution. At each move,
a fitness function is evaluated to identify the closeness of particle solution to the
global optimal solution. Cui et al. [5] introduced a bio-inspired clustering model
called the Multiple Species Flocking clustering model (MSF). The flocking behavior
consisted of making decisions on movements based on certain rules that depended on
an action in response to the neighboring members and the environment it wit. In this
algorithm, each document vector was anticipated as a boid in 2D virtual space. Azzag
et al. [6] introduced an AntTree algorithm directly inspired by the ants’ self-assembly
behavior and its application to the unsupervised learning problem. This biologically
inspired clustering algorithm proceeded in a distributed way to build a tree-structured
organization of the data. Ramos et al. [7] introduced a clustering algorithm based
on corpse clustering. In certain species of ants (Messor Sancta) for cleaning up their
nest, they sort their larvae or form piles of corpses to form cemeteries.
3 Proposed Tranquil-ART1NN System
Data mining is the method to recognize patterns and to find liaisons to elucidate
problems via data analysis. It allows predicting future behavior. In the proposed
Tranquil-ART1NN system, we use simulated annealing approach to identify the
initial clusters of similar borrowers. The clusters obtained are optimized through the
ART1NN approach. It includes the following steps.
1. Preprocessing the banking data to identify borrower periods.
2. Pattern discovery based on data mining methods [8–12]
3. Generate borrower profiles from clusters
4. Trajectory of budding Borrower Profiles.
Figure 1describes the steps involved in web usage mining.
Role of Data Science for Combating the Problem of Loan Defaults … 597
Fig. 1 Steps involved in
early identification of NPA
using Tranquil-ART1NN
approach
Preprocessing
Banking
data Initial neighborhood
identification using Tranquil
algorithm
Cluster creation using Intelligent
Huddle_Creation Algorithm
Post Progressing & Cluster
Optimization using ART1NN
algorithm
3.1 Preprocessing the Banking Data to Extract Borrower
Periods
Data mining is the automatic discovery of borrower activity.
The first step in data mining processing is the preprocessing of input data available.
After availing of the loans, some borrowers make defaults in repaying loans install-
ments/interest. When they default in repayments for 90 days, then the accounts
become NPAs. To avoid or to identify or to alert, the bank managers have to follow
the parameters given below.
These are the guidelines by RBI which are mandatory to adhere to in the system of
banking. When the bank manager fails to identify the non-receipts of the statements
mentioned below, these are considered for creating binary borrower vector ‘ňi’
Non-proper submission of Stock statement. Balance sheet
P&L accounts Cheque Return
Default in Deferred Payment Guarantee—Default in payment of Letter of Credit
Continuous default in payments of Bank Guarantee Return of cheque discounted
Return of Bills discounted
Default in payment of Bills Discounted Default in payment of Cheque discounted
Decline in Sales and profit
Cash Loss Fund loss
Decline in Net Worth
Incomplete documentation
not full filling the terms and conditions Special Mention Account 1
Special Mention Account 2
ňij={1, if borrower did a valid submission 0, otherwise
Fo r t h e i th borrower and jth parameter.
598 C. Shaardha and A. Alphy
This binary input vector is given as input for the proposed data mining process to
identify the borrower who can fall into NPA category. This binary vector is given as
an input to the proposed tranquil algorithm.
3.2 Tranquil Algorithm for Identifying Initial Borrower
Segregation for Early Detection of NPA Assets
Initially, each borrower ňis placed randomly on the visualization panel. Visualiza-
tion panel is a two-dimensional plane represented by x–y coordinates. The x-axis
and y-axis values range from 0 to 1. To convey the similar borrowers nearer and
the dissimilar borrowers far apart in the visualization plane, the proposed Tranquil
algorithm based on simulated annealing concepts used in metals is used.
In annealing, a metal is heated to its melting point and then cooled back to solid
state. The eminence of the annealing process depends upon in what way the cooling
is performed. Slow cooling fallouts in low energy state with stable crystals, whereas
hasty cooling results in high energy state with imperfect crystals. The parameters
of thermodynamic annealing can be mapped into simulated annealing approach, in
which a state of the system epitomizes viable solutions, energy epitomizes outlays,
transformation of state epitomizes neighboring function, temperature epitomizes
control parameters, and the final state is represented by the frozen state.
The proposed Tranquil algorithm (Algorithm 1) uses a greedy heuristics allowing
the μs to travel from recent positions to the superlative neighboring solution.
μ=μ1,μ2, μnrepresents the activities of a borrower
Algorithm 1: Tranquil algorithm
Notations used:
Ʊ= borrower ƛ position on the visualization plane
ɫ = temperature length,
Ʊ *= borrowers updated locus on the visualization plane
Ui= ith borrower on the visualization panel
Pos (Ui)= position of borrower on the visualization plane
Ѝ= total number of borrowers on the x-y visualization
plane
I
nput: Uiwhere i=1 to Ѝ,Ʊ,ɫ,Ѝ
Output: Ʊ*
Engender a primary solution Ʊ 0ϵ Ʊ *= Ʊ 0// ϵ is the
element of operator
Engender primary temperature ɫ>0
Role of Data Science for Combating the Problem of Loan Defaults … 599
Repeat
For 1<= temp<= ɫ
Identify ƛjneighboring Ʊ* using Euclidean
distance
Ḓ= Sim (ƛi, ƛj)- Simƛi, ƛi+1) // where Sim (ƛi, ƛj) =
,
Where
If Ḓ<=0 //downhill move
Ʊ*= {ƛj,Pos (ƛj)}
Else
IF random (0, 1) < e^ (Ḓ ⁄T) then Ω*= {ƛj,
Pos (ƛj)}
End if
Increment i by one.
End for
temp=ϔ*temp//reduce temperature
Until bottom value of temperature or
looping doesn’t accept a new solution.
Return Ʊ*
The inputs to the algorithm Tranquil are borrower vector μ, their locus on the
visualization panel, and the temperature span. Initially, the primary solution ň0is
generated randomly and this is allocated as the final solution ň*. Now, the primary
temperature value is created. At this juncture, a new solution is bent by selecting
neighboring ňithat is analogous to the solution ň*. A budget function (Ḓ) is calculated
with a cosine similarity of input vectors with an idea of maximizing the cosine
similarity (Sim)ofň. Cosine similarity, which is able to handle the qualitative and
quantitative data between any two ňs, signifies the similarity between borrowers
that are mapped to that ň. It can also deal with high-dimensional sparse data. Then
the budget function is calculated and checked for the decrease in the same. The
budget functions are considered as the verve for simulated annealing. If the verve
is dwindled, the newfangled is accepted. Otherwise, we accept the newfangled state
with probability eˆ(/T). A linear temperature lessening with 0.8 ≤´
R≤0.99 is
used here. The T is adjusted in such a way that the number of iterations in higher
temperature are small and the larger number of iterations at the lower temperatures.
To make the final solution sovereign of the starting one, the initial temperature should
be high enough. There will be no uphill moves when the temperature is low.
600 C. Shaardha and A. Alphy
The Tranquil algorithm returns the ňs with their new position on the visualization
panel. Now the alike ňs lie nearer in the visualization panel. As the aloofness between
ňs on the visualization panel increases, their likeness decreases. The uphill moves are
supported to avoid local minima. As a result of the Tranquil algorithm, ňs converge
to a freezing profound state, where similar ňs are located nearby.
In this proposed method, when the temperature is high, bad moves are accepted
due to the difficulty in escaping from the neighborhood. When the temperature is
low, bad moves are rejected and the finest results are kept in final solutions.
The usage of Tranquil algorithm reduces the number of iterations thereby
decreasing the time to identify the default behaviors. The Tranquil algorithm identi-
fies the borrowers with similar loan payment behavior and keeps them as neighbors.
The ‘doubtful category’ borrowers will be grouped into the same cluster thereby
increasing the easy identification of malfeasance by the borrowers.
The borrowers with the updated positions are given as input to the Intelli-
gent_Huddle Creation algorithm. In this algorithm, each borrower is mapped to an
intelligent agent. The intelligent agents have the self-learning capability to identify
clusters of borrowers.
Algorithm 2: Intelligent_Huddle Creation Algorithm Notations used:
Notations used:
Pos(ƛ) = locus of agent on the visualization plane
dth=distance threshold.
I
nput: agent_ƛi, Pos (agent_ƛi)
Output: Clusters C1, C2… CN
Read the agent_ƛi
Assign the ƛ to cluster Ck
For all ƛjdo
If (Distance (agent_ƛi, .agent_ƛj)) < dth
<σth2// σk2is calculated using equation (1)
Assign agent_ƛjto Ck
Else
Assign agent_ƛjto Ck+1
End if
End for
and σk2
In Intelligent_Huddle Creation Algorithm (algorithm 2), Distance (agent_ňi,
agent_.ňj) represents the distance between agent_ňiand agent_ňjon the visualiza-
tion panel, σk2represents the mean squared error or average dissimilarity between
the cluster prototype and the data records. Mean squared error is calculated using
Eq. (1).
σ2
k=ΣS(i)eχxkd2
ki
d2
ki
(1)
Role of Data Science for Combating the Problem of Loan Defaults … 601
where S(i)represents the ith borrower, krepresents the set of borrowers assigned to
kth cluster, dki is the distance from S(i) to k. Initial clusters of borrowers are formed by
grouping the agent_ňs that lie within a distance threshold and whose mean squared
error lies within σth2into a cluster.
Thus, the ňs within a cluster represent similar borrowers. These clusters are given
as input for the ART1NN approach.
In the intelligent_HuddleCreation algorithm, borrowers having similar loan repay-
ment behavior will be grouped into the same cluster. The intelligent agents can iden-
tify borderline behavior such as ‘special mention’ category. It can also identify the
borrowers with genuine interest and they are grouped into the same cluster. The self-
learning, decision-making, and dynamic behavior of intelligent agents allow identi-
fying the borrowers who take multiple loans and regularly repaying all the loans but
not concentering on a single loan so that the borrower cannot be categorized as NPA.
To improve the quality of identification, the clusters obtained as the out of the
Intelligent_HuddleCreation algorithm are given as input to the ART1NN.
3.3 Postprogressing Using ART1 Neural Network Approach
In the postprogressing phase, the input is the clusters generated by Intelli-
gent_HuddleCreation algorithm. It gives better refined and optimized clusters as
output so that the slippage of NPA assets can be avoided.
ART1NN helps to cluster binary vectors. It is an unsupervised learning. The
algorithm finds a pattern and proceeds to examine the members to be included in the
cluster or for the creation of another cluster based on a distance threshold. This is
iterated to cover the entire dataset.
The ART1NN approach operates on three levels. The first level is the identification
phase. In this phase, the existing output classification is matched with the input vector.
The output of the neuron becomes ‘1’ for the best match and ‘0’ otherwise. This level
is called the recognition phase. The second phase is the Collation phase, in which
the input vector is compared with the collation layer vector. A reset function will be
applied if the degree of likeness is lower than the vigilance parameter. The third level
is the Exploration level. Here, the neural network will examine for the no reset and
good match. If this is obtained, then the network will consider as the classification
is over. Otherwise, the same process is repeated with the other stored patterns to
identify the best match.
Algorithm 3: ART1NN algorithm
602 C. Shaardha and A. Alphy
N
otations Used:
α= Number of attributes in the input vector
β=Maximum No. of Clusters that can be formed
μ=Vigilance Parameter
Wij=bottom-up weights.
Xij=top-down weights
Input: Clusters C1,C2,…CN
Output: Optimized Clusters C1,C2,…CM, M<N
Initialize the learning rate r as r>1, 0<=μ>1,
0<Wij (0)< and Xij(0)=1
Repeat
For every training input
Set activations of F1(a)=input vectors and F2=0
Sent input signal from F1(a) to F1(b) as sigi =yi
For each subdued F2 node
Xj= // Xj≠-1
End for
Repeat
Identify J for XJ≥Xj
Calculate activation on F1(b) as yi=sitJi
Until( < μ)
Wij(new)=
Xij =yi
End For
Until stopping condition is not true
The stopping condition of the above algorithm includes no weight change,
maximum number of iterations reached, or no reset is performed.
4 Performance Analysis
Figure 2shows the 2D visualization of the Visualization plane. Initially, borrowers
are placed randomly on the visualization plane.
Figure 3shows the output clusters generated by the Intelligent_HuddleCreation
algorithm: borrowers with similar behavior are grouped into the same cluster.
In Fig. 4, the refined and optimized clusters obtained after applying the ART1NN
approach is shown.
Here, quality is evaluated in terms of purity and DB Index.
Davies–Bouldin(DB) index [13] calculates the average similarity between each
cluster and its most similar one, in a clustering result. Let Clusiand Clusjbe two
Role of Data Science for Combating the Problem of Loan Defaults … 603
Fig. 2 Random placements of borrowers in visualization panel
Fig. 3 Intial clusters of borrowers using Intelligent Huddle_Creation algorithm
604 C. Shaardha and A. Alphy
Fig. 4 Optimized clusters of borrowers using ART1NN approach
clusters in a clustering result. i represents the measure of dispersion in a cluster Clusi
and dij is the cluster dissimilarity measure. To calculate the similarity measures
between the clusters (Sij), it has to satisfy the following conditions.
Sij >0
Sit =Sjt
If i=0 and j=0, then Sij =0
If j>kand dij =dkj dij =dkj, then Sij >Sik If j=kand dij <dkj, then
Sij >Sik
Sij =xi+xj
dij
(2)
DB index is defined by
DBnc=1
nc
Σnc
i=1si,(3)
where Si=max
i=1−noi= jSij,i=1,...nc.n
cis the number of clusters.
Since DB index represents the average similarity between each cluster and its
most similar one and it is desirable for the clusters to have the minimum possible
similarity to each other, a clustering that minimizes DB is more desirable (Fig. 5).
From the above figure, it is observed that DB Index is higher for Tranquil-
ART1NN approach, when compared to K-means, Genetic clustering, and Ant Colony
Optimization (Fig. 6).
Purity [14] is calculated using the equation given below
Role of Data Science for Combating the Problem of Loan Defaults … 605
Fig. 5 Quality comparison in terms of DB index
Fig. 6 Quality measure in terms of purity
purityi=max pij (4)
The overall purity of a clustering approach is given by
purity =
k
i=1
mi
mpurity j(5)
It is noticed from the above figure that purity is higher for Tranquil-ART1NN
approach when compared to K-means, Genetic clustering, and Ant Colony Opti-
mization
606 C. Shaardha and A. Alphy
5 Conclusion
The complex problem of NPA demands superior techniques for resolution. With
myriad parameters affecting the issue on the one hand and the rich experience of
the regulators and the industry to combat the NPA menace on the other, the meeting
ground can be none other than the ground of AI which throws up various tools to
apply the experience to the problem domain. An attempt has been made in this paper,
to apply the rigorous AI algorithms to this problem domain of predicting defaults.
References
1. Naveenan, R.V., Warning signals—a tool to control NPA in banks. Int. J. Adv. Res. Comput.
Sci. Manage. Stud. 4(7) (2016). Available at www.ijarcsms.com
2. ‘Bank Fears Loanliness’, Available at https://www.hackerearth.com/problem/machine-lea
rning/bank-fears-loanliness/
3. L. Dai, W. Wang, W. Shu, An efficient web usage mining approach using chaos optimization
and particle swarm optimization algorithm based on optimal feedback model. Math. Prob. Eng.
(2013). Article ID: 340480
4. J. Kennedy, R. Eberhart, Particle swarm optimization, n Proceedings of IEEE International
Conference on Neural Networks, vol. 4, 1995, pp. 1942––1948
5. X. Cui, T.E. Potok, A distributed agent implementation of multiple species flocking model
for document partitioning clustering model for document partitioning clustering. Coop. Inf.
Agents X 4149, 124–137 (2006)
6. H. Azzag, N. Monmarche, M. Slimane, G. Venturini, Anttree: a new model for clustering with
artificial ants, in 2003 Congress on Evolutionary Computation, 4 Dec 2003, pp. 2642–2647
7. V. Ramos, F. Muge, P. Pina, Self-organized data and image retrieval as a consequence of inter-
dynamic synergistic relationships in artificial ant colonies. Soft Comput. Syst. Des. Manag.
Appl. Frontiers Artif. Intell. Appl. 87, 500–509 (2002)
8. Cooley, R., Mobasher, B., Srivastava, J., Web mining: information and pattern discovery on
the world wide web, in Proceedings of Ninth IEEE International Conference on Tools with AI
(ICTAI ’97), 1997, pp. 558–567
9. Nasraoui, O., Krishnapuram, R., Joshi, A., Mining web access logs using a relational clustering
algorithm based on a robust estimator, in Proceedings of Eighth International World Wide Web
Conference (WWW ’99), 1999, pp. 40–41
10. O. Nasraoui, R. Krishnapuram, H. Frigui, A. Joshi, Extracting web borrower profiles using
relational competitive fuzzy clustering. Int. J. Artif. Intell. Tools 9(4), 509–526 (2000)
11. J. Srivastava,R. Cooley, M. Deshpande, P.N. Tan, Webusage mining: discovery and applications
of usage patterns from web data. SIGKDD Explor. 1(2), 1–12 (2000)
12. O. Nasraoui, M. Soliman, E. Saka, A. Badia, R. Germain, A web usage mining framework for
mining evolving borrower profiles in dynamic web sites. IEEE Trans. Knowl. Data Eng. 20(2),
202–215 (2008)
13. D.L. Davies, D.W. Bouldin, Cluster separation measure. IEEE Trans. Pattern Anal. Mach.
Intell. 1(2), 95–104 (1979)
14. P.H. Choua, P.H. Lib, K.K. Chenc, M.J. Wua, Integrating web mining and neural network for
personalized e-commerce automatic service. Expert Syst. Appl. 37, 2898–2910 (2010)
Optimized Multi-Walk Algorithm
for Test Case Reduction
U. Geetha, Sharmila Sankar, and M. Sandhya
Abstract Software testing is the most expensive process of the development and
maintenance phase of the software product. Whenever changes are made in the code,
it becomes very difficult to execute all the test cases in regression testing. Therefore,
that testing process needs some test case reduction and prioritization techniques to
improve the regression testing process. Although there are many existing techniques
for test case reduction as well as for prioritizations, this paper will take its focus on
the multi-level random walk algorithm which has been used for test case reduction.
In a multi-level random walk selection algorithm, test case selection for further
reduction is done randomly on every iteration that degrades the performance of the
testing process in terms of coverage and will also generate a situation for random
test case tie. To overcome this situation of random test case selection and handling
test case tie, a solution is being proposed in this paper which includes a combination
of optimized multi-level random walk and optimized algorithm (Genetic algorithm).
This proposed approach proves that the combination of both the techniques improves
the overall performance better than the existing multi-level random walk.
Keywords Test data generation ·Test reduction techniques ·Test case optimization
U. Geetha (B)
Department of Information Technology, B.S. Abdur Rahman crescent Institute of Science and
Technology, Chennai, India
e-mail: geetha_it_phd_17@crescent.education
S. Sankar ·M. Sandhya
Department of Computer Science and Engineering, B.S. Abdur Rahman crescent Institute of
Science and Technology, Chennai, India
e-mail: sharmilasankar@crescent.education
M. Sandhya
e-mail: sandhya@crescent.education
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_55
607
608 U. Geetha et al.
1 Introduction
Software testing is the process by which bugs and faults can be found in a software
product. It can also be done through the verification and validation process in the
early and late stages of software development. This can be done in functional and
non-functional development of the product. Regression testing is more important in
the maintenance phase of the software product that affects the cost of the testing
process. Still many issues are faced by regression testing in terms of resources, time,
and cost. So, to reduce the cost of regression testing, software testers introduced the
concept of prioritization of their test cases. Prioritization is done to find the useful and
representative set of test cases, by some measures, which are made to run on earlier
phases of the regression testing process. Another important goal of prioritization is to
increase a test suite’s rate of fault detection. An extensive review has been carried out
to find faults in the existing literature which says that the existing reduction technique
used in this paper faces some problems with test case tie due to the random selection
of test cases on every iteration, thus making the overall test suite complex. Therefore,
to improve the test suite and reduce the complexity by maintaining the coverage ratio,
some optimization techniques have been introduced. It considers two optimization
techniques which includes the combination of optimized multi-level random walk
and optimized algorithm (genetic algorithm). One of the most common algorithms
for test case selection is the random walk algorithm that uses local optima and
backbone test cases to simplify the original problem into small problems by removing
the shielded test cases that means test case reduction problem. On the other hand,
to improve the ordering of test cases and to reduce the prioritized test cases, an
optimization algorithm (genetic algorithm) is used which is very powerful and is a
widely used stochastic search process. Genetic algorithm (GA) is an evolutionary
algorithm based on natural selection. It is used to find approximate solutions for
optimization and search problems [4]. The aim of GA is to achieve better results
through selection, crossover, and mutation. Genetic processes are most effective for
optimizing test cases. Due to this, the proposed approach uses genetic and multi-walk
random algorithms for reduction problems.
2 Problem Description
Multi-level random walk is a software test case reduction technique that is taken
as a focus area in this research. It tries to find an optimal and refined solution for
the original problem instance [2,3]. But this algorithm still has some problems
associated with it. On each level, a search is being performed, selection of random
test cases is made, and this selection of random test cases increases the complexity
of the entire test suite. Moreover, re-execution of test cases will affect the regression
testing which is a very time-consuming and expensive process. Also, there will be a
situation when the random test cases will meet a test case tie at some point in time
Optimized Multi-Walk Algorithm for Test Case Reduction 609
due to which the statement coverage ratio will also get affected. To overcome all
these scenarios and to make the overall test suite more effective, a solution that is
preferred is the incorporation of optimization technique (genetic algorithm) with the
existing reduction and prioritization technique.
2.1 Existing Methodology
Test case reduction is more important for regression testing because the number
of test cases affects the cost of the regression testing process. In this situation, the
system needs effective test cases from the original test suite to check whether existing
product is getting affected by the modified one.
Multi-level random Walk algorithm is used in this paper for test case reduction
[5].
One of the most common algorithms for test case selection is the random walk
algorithm that uses local optima and backbone test cases to simplify the original
problem into small problems by removing the shielded test cases which means test
case reduction problem.
In each level, a random walk is made and an intersection or the common part is
locked and discarding those test cases which are not locked or unshielded. However,
this algorithm reduces the problem through random selection during the selec-
tion process that removes some effective test cases. To overcome the problem, the
proposed approach uses genetic and multi-walk algorithms for optimizing test cases
in spite of random selection [7,8].
This time, the solution obtained by the multi-level random walk is not much opti-
mized as the statement coverage ratio is not maintained properly. So some optimized
algorithm can be thought of invocation with this.
Let us take a look at how it works:
Initially, we will take a coverage matrix with some assigned weights and test cases
satisfying some of the statements at each level of the matrix.
(i) Initial Coverage Matrix
The steps given below are followed in the multi-walk algorithm for the reduction
process (Table 1).
Step 1: Selection of two random test cases, i.e., {T3, T7}
Step 2: Selection of these test cases is covering statements {S2, S6}
Step 3: These test cases and the statements that are taken into consideration will
become 0 and
will contribute to reduced level 1 matrix.
Step 4: Now, these covered statements and test cases will get a locally optimal
solution by finding
shielded test cases.
610 U. Geetha et al.
Tabl e 1 Initial Coverage Matrix
Weights 0.5 0.7 0.8 0.2 0.4 0.6 0.8
Test cases S1 S2 S3 S4 S5 S6 S7
T1 1 0 0 1 1 0 0
T2 0 0 1 0 1 1 1
T3 0 1 0 0 1 1 0
T4 1 1 0 1 0 0 1
T5 1 1 1 0 1 0 0
T6 0 0 0 1 1 1 1
T7 0 1 0 0 0 0 0
Step 5: Here, the shielded test case taken is {T6} which is shown in the reduction
level 1 matrix.
Step 6: Further,
the next random test cases selected are {T4, T5} covering statements
{S1, S3, S4, S5, S7}.
Step 7: Reductive level 1 will give a reduced and refined matrix but the solution
so obtained is
Not the expected optimal solution according to the statement weightage
covered and
Remaining test cases will not even contribute to any statements and test case
coverage (Table 2).
(ii) Reductive Level 1 Matrix
Backbone Test cases for initial coverage matrix: {T3, T7}
Shielded Test case: {T6}
Backbone test cases after reductive level 1 matrix: {T4, T5}
Therefore, after covering most of the test cases and statements, the entire matrix
is on the verge of
becoming 0 throughout and the optimal solution found by test cases {T3, T4, T5,
T7}.
Tabl e 2 Selection of Shielded Test Case
Test cases S1 S2 S3 S4 S5 S6 S7
T1 1 0 0 1 1 0 0
T2 0 0 1 0 1 0 1
T3 0 0 0 0 0 0 0
T4 1 0 0 1 0 0 1
T5 1 0 1 0 1 0 0
*T6 0 0 0 1 1 0 1
T7 0 0 0 0 0 0 0
Optimized Multi-Walk Algorithm for Test Case Reduction 611
(iii) Test Case Reduction Percentage: After performing the multi-level random
walk, the test case reduction percentage comes out to be 57% which has more
probability of getting enhanced and improvised.
(iv) Effectiveness of test cases in terms of statement weightage:
If we find the statement coverage and its effectiveness in terms of statement
weightage, it comes out to be: 4.7
T3 =1*0.6=0.6
T4 =1*0.5+1*0.3+1*0.2+1*0.8=1.8
T5 =0.5 +0.3 +0.8 +0.4 =2
T7 =0.3,
Total =0.6 +1.8 +2+0.3 =4.7
2.2 Proposed Approach
A comparison being performed proves that the GA is more agile and profound in
terms of coverage weightage w.r.t multi-level random walk. But to get the optimal
solution, a proposed approach is thought of which describes for a combination of
multi-level random walk and optimized algorithm (genetic algorithm). The proposed
approach of the combined algorithm is more effective and improvised than the
existing methodology.
GAs or Hereditary Algorithms are effective and broadly relevant stochastic pursuit
and advancement strategies in light of the thoughts of normal choice and character-
istic assessment. It chips away at a populace to the advancement issue. These issues
which either can’t be defined in correct or in an exact numerical shape may contain
uproarious or sporadic information or they basically can’t be settled by the customary
computational techniques [7,9].
(i) Algorithm # Combination of optimized multi walk algorithm and genetic
algorithm:
Input:
T=< T1…….Tn > set of test cases
S=< S1……..Sn > set of statements
R1—Reduction level 1
Output:
S1—Reduced test-suite
for i =1toR1
F=call genetic ()
bi =bi U F
Collect Refined and reduced test cases covered by F
612 U. Geetha et al.
and reduce instance
Find shielded test cases (SHi)
Discard SHi in T
End
Return (T =F)
Int[] genetic ()
//initial population &return value
For every Test case Ti in Tn and for every statement Si in Sn
If (Ti covering Si)
Mark =1;
A[i] =A[i] +w;
Else
Mark =0;
1. Select two test cases which has the highest weight and find the target
While (target not reached)
2. If (! target)
Crossover (T1 (2), T2 (6))
3. If (! target)
Mutation (Crossover (1))
4. if (! target)
Select highest fitness one and one from the existing suite
Repeat step 1
Else
return (Selected test cases)
End algo
(ii) Configuration setting for GA (Table 3).
(iii) Reduction process
1. Initial Population: The bits stored in the form of ‘0’ or ‘1’ in the coverage matrix
are the chromosomes here for the reduction techniques.
2. Fitness Values: The fitness values can be calculated using the associated test
cases by the formula F (t1) sn
s1wi∗coverage statement(Si) (Table 4).
For further manipulation, we will select those test cases which have the highest
weighted fitness value.
Optimized Multi-Walk Algorithm for Test Case Reduction 613
Tabl e 3 Experimental Setup
Genetic steps Parameters
Initial population Statement coverage information is used with respective chromosomes as ‘0’
(or) ‘1’
Fitness value Calculated by fitness value F(t1) sn
s1wi* coverage statement(Si)
Selection Tournament-based selection
Crossover 1st and 2nd of First fitness value and 4th and 5th of 2nd fitness value of the
test cases
Mutation 5th bit of crossover result (discrepant bit)
Termination New four test cases to be tested that should have 75% coverage value
Tabl e 4 Fitness Value
T1 T2 T3 T4 T5 T6 T7
1.1 2.6 0.6 1.8 2 2 0.3
3. Selection: Selection of the test cases is done based on Tournament-Based
Selection. Select any two test cases and perform XOR operation (Table 5).
4. Crossover:
Crossover can be performed on the test case 2 and 6 with some selected bits and again
the XOR operation is performed to get crossover value satisfying 75% coverage of
statements (Table 6).
5. Mutation:
Most bit of the output will get changed into 1 and, after performing mutation, it will
be on the verge of achieving 100% coverage on statements which will become 1
throughout after crossover. Now, suppose if 100% coverage on the statement is not
achieved, then it is needed to continue for all possible combinations. All combinations
will involve the fitness value with the highest weightage and a moment will come
Tabl e 5 Selection Output
T2 0 0 1 0 1 1 1
T6 0 0 0 1 1 1 1
XOR 0 0 1 1 1 1 1
Tabl e 6 Crossover Output
T2 0 1 1 0 1 1 1
T6 0 0 0 1 1 0 1
XOR 0 1 1 1 1 1 1
614 U. Geetha et al.
where all these combinations will contribute to achieve the 100% target. In this
example, we can take {T2, T6} as the backbone test case for reduction problem.
As we have performed multi-level random walk in the initial coverage matrix, it
is difficult to get the optimal solution by reduced test cases, so invocation of GA is
performed to compare the effectiveness genetic algorithm with multi-level random
walk.
We will take {T2, T6} as backbone test cases for reduced level 1 matrix which
will cover the respective coverage statements as depicted (Table 7).
Here, two shielded test cases possible for test cases T2 and T6, i.e., T1 and T4, but
we will go for T4 as it is covering more number of 1; moreover, T1 will ultimately
become optimal after further coverage with T4 as a shielded test case (Table 8).
Only one backbone is needed to be selected so no need to go for optimization
(Table 9).
Optimal solution:
As 1 is left uncovered in T5 test case, therefore, it is taken as the optimal solution:
{T5}
Reduction Test case in terms of weightage:
Backbone test case 1 ={T2, t6}
Backbone test case 2 ={t4}
Optimal solution ={t5}
Total =2.6 +2+1.8 +2=8.4
Tabl e 7 Reduction Level 1
Test cases S1 S2 S3 S4 S5 S6 S7
T1 1 0 0 1 1 0 0
T2 0 0 1 0 1 1 1
T3 0 0 0 0 0 1 0
T4 1 1 0 1 0 0 1
T5 1 1 1 0 1 0 0
T6 0 0 0 1 1 1 1
T7 0 1 0 0 0 0 0
Tabl e 8 Reduction Level 2
Test cases S1 S2 S3 S4 S5 S6 S7
T1 1 0 0 1 0 0 0
T2 0 0 0 0 0 0 0
T3 0 0 0 0 0 0 0
T4 1 1 0 1 0 0 0
T5 1 1 1 0 0 0 0
T6 0 0 0 0 0 0 0
T7 0 1 0 0 0 0 0
Optimized Multi-Walk Algorithm for Test Case Reduction 615
Tabl e 9 Reduction Level 3
Test cases S1 S2 S3 S4 S5 S6 S7
T1 0 0 0 0 0 0 0
T2 0 0 0 0 0 0 0
T3 0 0 0 0 0 0 0
T4 0 0 0 0 0 0 0
T5 0 0 1 0 0 0 0
T6 0 0 0 0 0 0 0
T7 0 0 0 0 0 0 0
2.6
4.6
6.4
8.4
0
2
4
6
8
10
T2-T3 T6-T4 T4-T5 T5-T7
opƟmized mulƟ walk
muliƟwalk
Graph 1. Multi-walk versus optimized multi-walk
Therefore, after covering most of the test cases and statements the entire matrix is
on the verge of becoming 0 throughout and the optimal solution found by test cases
{T2, T6, T4, T5}.
2.3 Performance Analysis
The following graph shows that the overall performance of multi-walk and optimized
multi-walk algorithms and optimized approach is better than multi-walk approach
(Graph 1).
2.4 Conclusion and Future Work
On taking a tour through this paper, we experienced that regression testing is a very
expensive form of testing which includes re-execution of all test cases, thus making
616 U. Geetha et al.
the overall test suite more cumbersome and leading to an increase in test cases. So to
enhance the regression testing, initially, a multi-level random walk is performed and a
common intersection point is found, but that walk being performed is not optimized.
Therefore, an optimized random walk is made which tries to decrease the redundancy
in the test suite and improves the overall test suite. When it comes to redundancy,
test case prioritization also plays a very important role. To get a more profound test
suite, one of the optimization techniques (genetic algorithm) is incorporated with
the existing multi-level random walk reduction technique. Thus, the combination
of both reduction and prioritization techniques with optimization technique (genetic
algorithm) is a good recommendation that is tried to be concluded in this paper.
In this paper, we have tried to overcome the random test case tie situation and
to decrease the redundancy in the test suite by using a combination of multi-level
random walk and optimization algorithms (genetic algorithm). But prioritization
can also work as a very good remedy for reducing the redundancy of test cases in
the test suite as it will provide a sequence and order to test cases, which can help
in simplifying the entire test suite. In future, a good prioritization technique can
be combined with an optimization technique. Moreover, some faults will also be
generated at the time of acceptance testing phase that can also be worked upon for
optimization using some reduced and prioritized optimization techniques.
References
1. M. Gligoric, L. Eloussi, D. Marinov, Practical regression test selection with dynamic file
dependencies, in Proceedings of International Symposium on Software Testing and Analysis,
Baltimore, MD, 12–17 July 2015, pp. 211–222
2. Hadoop, Apache Hadoop [Online] (2015). Available http://github.com/apache/hadoop/
graphs/commit-activity
3. S. Yoo, M. Harman, Regression testing minimization, selection and prioritization: a survey.
Softw. Test. Verif. Reliab. 22(2), 67–120 (2012)
4. M. Harman, S. A. Mansouri, Y. Zhang, Search-based software engineering: trends, techniques
and applications. ACM Comput. Surv. 45(1), 1–61 (2012)
5. A. Caprara, P. Toth, M. Fischetti, Algorithms for the set covering problem. Ann. Oper. Res.
98(1–4), 353–371 (2000)
6. R.M. Karp, Reducibility among combinatorial problems, in Complexity of Computer Computa-
tions, The IBM Research Symposia Series, ed. by R. Miller, J. Thatcher, J. Bohlinger (Springer,
New York, 1972), pp. 85–103
7. H. Hsu, A. Orso, MINTS: a general framework and tool for supporting test-suite minimization,
in Proceedings of 31st International Conference on Software Engineering, Vancouver, Canada,
16–24 May May 2009, pp. 419–429
8. J. Black, E. Melachrinoudis, D. R. Kaeli, Bi-criteria models for all-uses test suite reduction, in
Proceedings of 26th International Conference on Software Engineering, Edinburgh, Scotland,
23–28 May 2004, pp. 106–115
9. A. Gotlieb, D. Marijan, FLOWER: optimal test suite reduction as a network maximum flow,
in Proceedings of International Symposium on Software Testing and Analysis, San Jose, CA,
21–26 July 2014, pp. 171–180
10. D. Hao, L. Zhang, X. Wu, H. Mei, G. Rothermel, On-demand test suite reduction. in Proceedings
of 34th International Conference on SoftwareEngineering, Zurich, Switzerland, 2–9 June 2012,
pp. 738–748
Chennai Water Crisis—Data Analysis
Deepak Shankar, N. Aaftab Rehman, Sharmila Sankar, Aisha Banu,
and M. Sandhya
Abstract Chennai is a coastal city and is the capital city of the Indian state of
Tamil Nadu. It is densely populated and currently facing a severe drought as of
June 2019. Chennai mainly depends on reservoir water, i.e., metro water and its
groundwater resources to meet its water needs. The four main reservoirs of the city
are Red Hills, Chembarambakkam, Cholavaram, and Poondi. With a combined total
capacity of 11,057 Mcft, these act as the major source of water for the inhabitants of
the city. With the ever-increasing population and the added effects of climate change
which induces a decreasing rate of total rainfall, the situation of drought is likely
to increase over the upcoming years. This project aims to explain the severity of
drought in Chennai and how it can get worse in the upcoming years with the help of
data analysis to find the per capita water that is currently present for a civilian, and
how this has changed over the years due to various factors. This project also aims to
create awareness and provide scope for possible solutions and measures that can be
taken to mitigate the situation of drought in upcoming years.
Keywords Wat e r c r i s i s ·Reservoir ·Dataset ·Rainfall
1 Introduction
Chennai, one of the four major metropolitan cities of India, is hit with a never-
experienced before tremendous impact of drought, which is the highest so far
recorded in terms of water availability levels of the reservoirs: Red Hills, Cholavaram,
D. Shankar ·N. A. Rehman ·S. Sankar (B)·A. Banu ·M. Sandhya
Department of Computer Science & Engineering, B.S. A Crescent Institute of Science &
Technology, Chennai, India
e-mail: sharmilasankar@crescent.education
D. Shankar
e-mail: 01deepakshankar@gmail.com
N. A. Rehman
e-mail: aaftabrehman@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_56
617
618 S. Deepak et al.
Poondi, and Chembarambakkam which act as the potable source for the whole city
other than groundwater. The reason that can be accredited for this worse condition
of Chennai is its low monsoon rainfall record in the months between March and July
in the year 2019. Naturally, the city that is close to the equator might fall under the
list of places that experiences harsh climate and harsh temperature. Analysis and
records show that the rainfall record after 2016 has drastically lowered down and
almost less than 10 cm rainfall was totally recorded in the monsoon of 2019 which is
a catastrophic condition for any city. Chennai also witnessed two consecutive failures
of monsoons in 2018 and 2019. As a result, the city had to import water from other
states and had to circulate it to its residents to meet its water requirements. Again, the
climatic change could be reasoned as an effect of global warming. We have limited
ourselves to the city of Chennai which is now on a verge to turn into next Somalia
as its water resources are depleting. In this paper, we tried to cover up with all those
factors that have resulted in this condition of Chennai and we have calculated the per
capita water based on the details and data available only on the four major reservoirs
of the city.
2 Related Work
Many algorithms have been proposed for dealing with drought. We discuss some of
these in this section.
Agana et al. [1] discussed basically on drought conditions in California and how
big data analysis is used to build a real-time model for drought modeling and predic-
tion. Various predictions are made using diverse data sets such as weather data,
drought condition, climate sensor and satellite data, and water usage report’s dataset.
So, it is much more accurate in predicting the drought and its severity. It provides a
more realistic visualization and offers a real-time model for drought prediction which
describes the need for water conservation and drought awareness. Further, Anumalla
et al. [2] designed a method for groundwater monitoring systems based on a network
of wirelessly linked pressure sensors to measure the groundwater level. Gu et al. [3]
used the extended empirical orthogonal function along with a standardized precipi-
tation index to develop a dataset on spatial–temporal data. Jokhio et al. [4] presented
with an idea to innovate the wireless sensor networks and employ them in forecasting
systems to predict the drought and floods. Agana and Homaifar [5] have proposed
a method to predict the droughts in the future by using deep learning technology.
However, drought patterns are fluctuating in nature. Therefore, neural networks are
also employed to predict the patterns accurately and thereby on further application
of Standardized streamflow index (SSI) as an input in predicting this nonlinear time
series. Nivedha Deve et al. [6] talks about the agricultural drought and how it varies
change in temperature and rainfall levels with help of remote sensing which is used
to monitor the temperature and vegetation variation over the Trichy District over
the span of four years 2013–2016. The dataset used for calculating this factor is
Surface Temperature (LST) and Normalized Difference Vegetation Index (NDVI)
Chennai Water Crisis—Data Analysis 619
and the correlation between these two factors can help describing the severity of
drought and how it changed over the years. Rao and Chaudhari [7]discusseshow
soil moisture data can be used to calculate and predict the drought with minimum
accuracy. A comparison is done with estimated soil moisture data and rainfall data to
find the correlation between the two factors. It can be used to find out the drought and
flooding conditions and it is more suitable for bare fields with uniform conditions due
to a poor spatial resolution of AMSR-E. Kollukuduru Sravanthi and Rajan [8]talk
about identifying the spatial co-occurrence of droughts across various subdivisions
of India. The results show that similarities in the climate forcing factors have a role
in spatiotemporal correlation of events. De Azevedo et al. [9] analyzed the general
survey about the drought patterns across India using satellite remote sensing. The
project uses the Moderate Resolution Imaging Spectroradiometer (MODIS) to study
the stress on vegetation over India. Dynamics of monthly Normalized Difference
Vegetation Index (NDVI) and its relation to the climate conditions over a period of
10 years have been studied and presented in this paper. Liao et al. [10] discusses
monitoring the dryness or drought scale in grassland. The GDI was constructed by
integrating the canopy water content (CWC), soil moisture content, and precipita-
tion together with different weights. A detailed analysis is made on GDI regarding
temporal stats.
3 Algorithm
import and read the dataset into project.
import numpy, pandas, matplotlib libraries.
The dataset initially acquired is in the raw format. So, the collected data are first
cleaned using Pythonic Data Cleaning with Pandas and NumPy. Then the data are
filtered by choosing only the attributes that are necessary for plotting the graph.
Finally, the various attributes like ‘date’ are modified and converted into an appro-
priate format that is recognizable by Python compiler and can be meaningfully plotted
using Python Matplotlib and Seaborn Library (Fig. 1).
for each reservoir:
plot the graphs for various reservoir on a scale: Mcft every year.
plot the graph for rainfall at reservoirs on a scale: Mcft every year.
Water levels in each of the major reservoirs (y-axis) are plotted simultaneously
against the date (x-axis). This graph helps to visualize the water holding capacity
of all the reservoirs at each date of the year for a span of 15 years. These attributes
help to tell how the reservoir levels have changed over the years and whether any
effort was made to increase the capacity over the years. Also, the rainfall received at
the major reservoirs (y-axis) is plotted simultaneously against the date (x-axis). This
graph helps to visualize the total rainfall received at all the reservoirs on each date
620 S. Deepak et al.
Fig. 1 Architecture diagram
of the year for a span of 15 years. These attributes help in detecting fluctuations in
rainfall and analyze extreme scenarios such as floods.
total_water_level =Σ(water level at each reservoir)
total_rainfall =Σ(rainfall at each reservoir)
plot the graph for total_water_level and total_rainfall with scale in Mcft.
The total reservoir water level is calculated by finding the sum of water level at
all the reservoirs and the average rainfall level is also calculated. Finally, each factor
is individually plotted against the date. This helps to analyze the overall reservoir
levels for a given rainfall on any given day. This also helps to analyze the total water
that is actually made available to the public despite heavy rainfall.
read the population dataset
plot the graph of population growth taking a scale that measures years in x-axis
and population in y-axis.
The graph of population growth over a period of 15 years is plotted. This can help
analyzing the growth in population and how this factor is increasing at an accelerated
rate. This attribute is chosen because drought mainly occurs when water resource is
not able to meet the demand which is caused by increasing population.
(1) Per Capita Water =TotalReservoir Level
Population
plot the graph for per Capita Water taking years in x-plot and liters in y-plot.
Chennai Water Crisis—Data Analysis 621
The per capita water describes the water that is actually available to each person
through the reservoirs as metro water. This factor is calculated as it shows the severity
of drought over the given years. Thus, this factor is calculated and plotted against
the year. Analysis can give possible insights into remedies and solutions that can be
taken to mitigate the situation of drought in the upcoming years and thus safeguard
the future generation.
Finally, analyze the various graphs that are generated to arrive at possible
conclusions.
4Analysis
As inferred from Figs. 2and 3, The Redhills Reservoir has the highest capacity out of
holding the water and Cholavaram has the least capacity. The average total holding
capacity of all the reservoirs is around 4195.66 Mcft.
The graphs in Figs. 4and 5show how the water at individual reservoirs rise and fall
according to the rainfall pattern. Cholavaram seems to contribute the least with hater
holding capacity. So, it can meet only the needs of a small portion of the Chennai
public.
Fig. 2 Average water levels in reservoirs graph
Fig. 3 Overall Average
water availability in a year
(in terms of MCFT)
622 S. Deepak et al.
X-Axis: Year, Y-Axis: Water level
Fig. 4 Graph on water levels at various reservoirs
X-Axis: Year, Y-Axis: Water level
Fig. 5 Graph on rainfall levels at various reservoirs
During every year, rainfall peaks during November–January season which is seen
as the maxima at both the graphs. In Fig. 6, there’s a significant peak in the year
2005 and 2015 with a very steep rise and fall which marks two of the biggest floods
of Chennai history. A broader peak describes the gradual rise and fall of the water
level. This scenario is observed when there’s a low to medium rainfall. A steep and
sharp peak describes the scenario of a flood (Fig. 7).
Chennai Water Crisis—Data Analysis 623
Fig. 6 Graph on total water level in all reservoirs
Fig. 7 Graph on total rainfall received overall
Also, the steepness in these two regions indicates that the water holding capacity
of the reservoir is extremely low during excessive rainfall. So, a major amount of
water is wasted into the sea (Fig. 6).
On average, the groundwater table has gone down by 100% from earlier levels.
So, in locations where people got water at 150 feet, now they have to bore for 300
feet to even get a glimpse of water (Fig. 8). This is due to a significant reduction in the
number of water bodies, an increase in construction, and insufficient and defective
percolation pits (Rainwater harvesting).
From Fig. 9a, b, we infer that there is a steady rise in the population level. But the
capacity of the reservoirs remains the same. Hence, there’s an exponential decrease
in the per capita water from the reservoir. And as of June 2019, all four reservoirs ran
out of water, putting a large section of Chennai population underwater crisis. Water
624 S. Deepak et al.
Fig. 8 Article on average water level in Chennai [11]
Fig. 9 a Graph on population dynamics over the years. bGraph on per capita water over the years
crisis is dependent on multiple factors like rainfall, climate, population, changes in
land topography, and also groundwater replenishment level. But, many other projects
tend to miss out on many of these factors and solely blame on climate change. But
this project aims to relate and bring in all the possible factors influencing a drought
condition in a metropolitan city like Chennai.
5 Inference
As we can see, there is a variable increase and decrease of rainfall and water levels
in the reservoirs which makes it totally unfit to be completely dependent upon. As
inferred from the graphs plotted above, there are significant bars and plots above
representing the data of overall rainfall and water availability which were abundant
Chennai Water Crisis—Data Analysis 625
in the years 2004 and 2015 and has rapidly vanished in no time, which portrays the
irregular and careless behavior of its citizens and governing body to conserve water
and use it ideally. The study has also proved that the water resources in Chennai are
very scarce and the conventional source of water may no longer support its residents
to meet their needs on their basic amenities. Another graph also shows the study on
the population dynamics of the city and is a city of tedious growth the city has invited
millions of immigrants in the subsequent years after 2005.
The various factors that have resulted in drought condition as analyzed are:
•Insufficient rainfall due to climate change.
•Increase of population inducing an increase in the demand for the water resource.
•Reduction in the number of water bodies [12] by at least 50% due to illegal
encroachment and construction over existing water bodies.
•Water holding capacity during floods is extremely low. This can be attributed to
two factors:
– Insufficient de-silting done by the government. As we see, the reservoirs must
be properly de-silted every 5 years or so. But it’s not the case.
– The structure of the reservoirs is not that good. So, the officials fear for the
worst and release most of the water. If the structure was strong enough, this
wouldn’t be the case.
•Reckless mushrooming of bore-wells which has caused a reduction in ground-
water table by more than 100%
•The inefficient rainwater system:
– More than 40% of the houses don’t have a rainwater harvesting system present
[13].
– The existing rainwater harvesting pits are extremely old and not maintained
well. The rainwater percolation pits have to be cleaned at least once every five
years to avoid clogging. But not everyone does this. There are many modern
and much cost-effective methods present in recent times to implement this
model effectively. So if people step up and follow it, the problem could be
reduced several folds.
•Illegal sand mining:
– It is a very popular crime in southern states where the top layer of river sand
is mined and extracted illegally for the construction of sand. Due to this, the
land becomes arid, and water-retaining and holding capacity of the river basin
decreases severely. This also affects the groundwater table.
•Reckless irrigation methods:
– It is estimated that 80% of the freshwater in India is used for agriculture. Due to
improper implementation by the government, the farmers tend to over irrigate
the land. This is an extremely bad practice as it pollutes the excess water with
pesticides and fertilizers.
626 S. Deepak et al.
– Drip irrigation will prove to be a better alternative saving a major portion of
water.
•No system to Recycle and reuse wastewater:
– Almost all the water is let out into the sea even without being treated. If we
treat the water from households or from industries, more than 40% of pure
water can be recovered. But no proper infrastructure is present in the state due
to prejudice and carelessness of the government when it comes to conserving
essential resources.
6 Conclusion and Future Scope of Work
The project was based on the domain—Data Analysis under which we have analyzed
the water scarcity of Chennai that is on a verge to turn into a dry land sooner if not
acted upon. The study has detailed graphs on the water resources of the city mainly
concentrating on its water reservoirs and also has graphs on rainfall levels in those
areas and data on the population dynamics. The graphs on rainfall and water level
are far more terrorizing as they are decreasing at an alarming rate, if this goes on
there is no stopping of this land turning into barren. Different plots above show
that Chennai received adequate rainfalls before it fell short of rainfall in the years
preceding 2015, but due to the irregular management of water the city has been struck
with a drought-like condition to a level which was never experienced before.
However, the government has taken few steps to turn this city, the government
made it mandatory for all the houses which are built to install a rainwater harvesting
system and have also approved to provide small funding to build it. The government
is now taking up steps to build factories on the outskirts of Chennai.
The water resources of this city are drying up, the city with no rivers as a source
could be helped by its neighboring states by simply allowing the water rather stopping
it through dams. The condition of the city went to worse and if not supported by the
neighboring state the residents may face severe problems.
As observed, we have only provided this paper with the data on water level
and rainfall level of the major reservoirs of the city but this alone won’t suffice
to completely analyze the pattern of drought in Chennai. Additionally, people who
engage in this project in the future may take up additional parameters which are
naturally occurring like temperature, humidity, and many more. Also, we have not
covered up with all the sources of water to the city, the residents alternatively use
water from the ground, as these factors are mainly important and participate in the
overall water cycle.
Chennai Water Crisis—Data Analysis 627
References
1. N.A. Agana, A. Homaifar, A Deep Learning Based Approach for Long—Term Drought
Prediction, in SoutheastCon (2017)
2. S. Anumalla, B. Ramamurthy, D.C. Gosselin, M. Burbach, Ground water monitoring using
smart sensors, in IEEE International Conference on Electro Information Technology (2005)
3. X. Gu, N. Li, Study of droughts and floods predicting system based on spatial-temporal data
mining, in 6th International Conference on New Trends in Information Science, Service Science
and Data Mining (ISSDM2012) (2012)
4. S.H. Jokhio, I.A. Jokhio, F. Khan, S. Memon, Wireless Sensor Network Based
Flood/DroughtForecasting System, in IEEE SENSORS (2015)
5. N.A. Agana, A. Homaifar, A deep learning based approach for long-term drought prediction,
in Southeastcon-IEEE (2017)
6. S. Nivedha Deve, M. Jasmineniketha, P. Geetha, K.P. Soman, Agricultural drought analysis for
Thuraiyur Taluk of Tiruchirappali District using NDVI and land surface temperature data, in
11 th International Conference on Intelligent Systems and Control (ISCO) (2017)
7. Y.S. Rao, A.A. Chaudhari, Analysis of 7 years Aqua AMSR-E derived soil moisture data over
India, in 2009 IEEE International Geoscience and Remote Sensing Symposium
8. K. Sravanthi, K.S. Rajan, Mining spatial co-occurrence of drought events from climate data of
India, in 2010 IEEE International Conference on Data Mining Workshops
9. S.C. de Azevedo, R.P. Singh, E.A. da Silva, Assessing 2016 drought progression over India
using remote sensing data for the period 2006–2015, in 2017 IEEE International Geoscience
and Remote Sensing Symposium (IGARSS)
10. Z. Liao, B. He, X. Quan, X. Bai, C. Yin, X. Li, S. Qiu, Constructing a global grassland
drought index (GDI) product based on MODIS and ancillary data, in 2015 IEEE International
Geoscience and Remote Sensing Symposium (IGARSS)
11. Chennai Metro Water—Twitter Handle (@CHN_Metro_Water)
12. https://timesofindia.indiatimes.com/city/chennai/chennais-two-third-water-area-vanishes-in-
a-year/articleshow/68857592.cms
13. https://timesofindia.indiatimes.com/city/chennai/only-50-households-in-chennai-have-rain-
water-harvesting-structures-finds-corporation/articleshow/70205838.cms
Generalized Canonical Correlation
Based Bagging Ensembled Relevance
Vector Machine Classifier for Software
Quality Analysis
Noor Ayesha and N. G. Yethiraj
Abstract Early detection of defects helps to save cost and efforts. Few research
works have been designed in existing works for analyzing the quality of software
program using various machine learning techniques. However, the classification
performance of existing work was lower which reduces the defect detection accuracy.
In order to overcome the above existing issues, the Generalized Canonical Correla-
tion Analysis based Bagging Ensembled Relevance Vector Machine Classification
(GCCA-BERVMC) model is proposed. The GCCA-BERVMC model considers the
number of source code lines from software program dataset as input. After that, the
generalized Canonical Correlation Analysis (gCCA) algorithm designed in GCCA-
BERVMC model selects the relevant features (i.e., code metrics) in order to improve
the quality of software program. By applying Bagging Ensembled Relevance Vector
Machine Classification (BERVMC) algorithm, the classification of program files is
performed in GCCA-BERVMC model. The boosting algorithm creates ‘n’ number
of weak learners to classify the input source code lines as normal or defected by
analyzing the source codes and chosen metrics. After that, the weak learner’s results
are combined into strong classifier by using the majority votes. This helps for GCCA-
BERVMC model to enhance the accuracy of defect prediction for software quality
analysis with a lower amount of time. Experimental evaluation of GCCA-BERVMC
model is conducted using metrics such as defect detection accuracy, false positive
rate, and time complexity with respect to various software code sizes. The experi-
mental result shows that the GCCA-BERVMC model is able to increase the defect
detection accuracy and also minimizes the amount of time required for software
quality analysis when compared to state-of-the-art works.
N. Ayesha (B)
Department of Computer Science, Bharathiar University, Coimbatore, India
e-mail: razia.ayesha.bi@gmail.com
N. G. Yethiraj
Department of Computer Science, Maharani’s Science College for Women, Bengaluru, India
e-mail: rajsbmjce@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_57
629
630 N. Ayesha and N. G. Yethiraj
Keywords Bagging Ensembled Relevance Vector Machine Classification ·
Generalized canonical correlation ·Majority vote ·Software defect ·Strong
classifier and weak learner
1 Introduction
Software testing is the procedure of discovering faults in software while executing
it. The results of the testing are employed to identify and correct faults. Software
defect prediction finds where faults occur in source code. The results from the defect
prediction can be utilized to optimize testing and get better software quality. Software
Defect Prediction impacts quality and has attained significant popularity in the last
few years [1,2]. Therefore, a novel GCCA-BERVMC model is introduced in this
research work by using the generalized Canonical Correlation Analysis (gCCA) algo-
rithm and Bagging Ensembled Relevance Vector Machine Classification (BERVMC)
algorithm.
GCCA-BERVMC model is developed in this research work. The main contribu-
tions of GCCA-BERVMC model are described in below
•To enhance the software defect detection performance through classification
when compared to state-of-the-art works, GCCA-BERVMC model is introduced
using generalized Canonical Correlation Analysis (gCCA) algorithm and Bagging
Ensembled Relevance Vector Machine Classification (BERVMC) algorithm.
•To improve the software metric selection performance with a minimal amount of
time when compared to conventional works, generalized Canonical Correlation
Analysis (gCCA) algorithm is employed in GCCA-BERVMC model. The gCCA
algorithm determines the associations between each input software metric ‘μi’
and objective function to significantly choose the related metrics.
•To increase the defect prediction accuracy when compared to traditional works,
BERVMC algorithm is applied in GCCA-BERVMC model. BERVMC algorithm
formulates a strong classifier with aiming at classifying each source code line as
normal or defect with lower error for effective software quality analysis.
The remaining article structure is formulated as follows. In Sect. 2, the proposed
GCCA-BERVMC model is explained with the aid of architecture diagram. In Sect. 3,
experimental settings are presented, and the comparative result of GCCA-BERVMC
model is discussed in Sect. 4. Section 5portrays the conclusion of the paper.
Generalized Canonical Correlation Based Bagging … 631
2 Generalized Canonical Correlation Analysis Based
Bagging Ensembled Relevance Vector Machine
Classification Model
Software metrics (features or attributes) are collected during the software develop-
ment cycle. Metric selection is one of the most important preprocessing steps in
the process of building defect prediction models and may improve the final predic-
tion result. To improve the effectiveness and quality of software development and to
predict defects in software, various machine learning techniques were designed in
conventional works [3,4]. But, the classification performance of conventional work
was not sufficient to achieve better defect detection accuracy [5]. Therefore, Gener-
alized Canonical Correlation Analysis based Bagging Ensembled Relevance Vector
Machine Classification (GCCA-BERVMC) model is designed in this research work.
The architecture diagram of GCCA-BERVMC model is presented in Fig. 1.
Figure 1depicts the overall processes of GCCA-BERVMC model. As demon-
strated in the below figure, GCCA-BERVMC model contains two main processes
namely metrics selection and classification [6]. Initially, GCCA-BERVMC model
gets a number of source code lines from an input software program dataset. During
the metrics selection process, GCCA-BERVMC model discovers the most related
software metrics to test the quality of input software program. During the classifica-
tion process, GCCA-BERVMC model predicts the error in a given program by clas-
sifying each source code line as normal or defect [7]. From that, GCCA-BERVMC
model improves the software quality analysis performance when compared to state-
of-the-art works. The detailed processes of GCCA-BERVMC model are explained
in the below subsections.
Software Program Dataset
N
umber of source code lines
Generalized Canonical Correlation Anal
y
sis Choose relevant
Bagging Ensembled Classification Classify each program
Enhanced software quality Prediction
Fig. 1 Architecture diagram of GCCA-BERVMC model for software quality analysis
632 N. Ayesha and N. G. Yethiraj
2.1 Generalized Canonical Correlation Analysis
The selection of software metrics is significant for predicting the software quality
[8]. Selecting such metrics is not feasible because of limited project resources, partic-
ularly if the number of available metrics is large. Therefore, generalized Canonical
Correlation Analysis (gCCA) algorithm is designed in GCCA-BERVMC model in
order to choose only a more significant metrics to enhance the quality of software
program. The gCCA is designed to improve the feature selection (i.e., software
metrics) performance with minimal amount of time in order to efficiently improve
the software quality analysis. The process involved in gCCA is depicted in Fig. 2.
Figure 2shows the flow process of gCCA algorithm to perform metrics selec-
tion during the software quality analysis process. As depicted in figure, generalized
canonical correlation analysis is used in gCCA algorithm in order to compute the
associations between each input software metric ‘μi’ and objective function ‘oi’,
i.e., software quality analysis. Let us consider two features ‘μi’ and ‘oi’, and there
are correlations among the information. The gCCA algorithm at first finds linear
combinations between software metric ‘μi’ and objective function ‘oi’ which have a
maximum correlation with each other. From that, generalized canonical correlation
analysis is mathematically estimated as below
Cr →μioi=Cov(μi,oi)(1)
From the above mathematical Eq. (1), the covariance matrix is calculated by using
information of software metric ‘μi’ and objective function‘oi’. In gCCA algorithm,
the covariance between input software metric ‘μi’ and objective function ‘oi’is
mathematically determined as below
Number of software metrics
Determine correlation between metric and
ob
j
ective function
If ‘ ’ is higher
Metric is selected as
relevant
Yes
No
Metric is removed
asirrelevant
Fig. 2 Flow processes of gCCA algorithm
Generalized Canonical Correlation Based Bagging … 633
Cv(μi,oi)=(μi−¯μ)(oi−¯o)
n(2)
From the above mathematical formula (2), ‘μi’ signifies the values of software
metric and ‘oi’ denotes the values of objective function. Here, ‘ ¯μ’ indicates the mean
of the s oftware metric, ‘ ¯o’ refers the mean of objective function, and ‘n’ represents a
total number of the software metrics considered for testing input software program.
With assistance of the Eqs. (1) and (2), gCCA algorithm estimates the relationship
between the software metric ‘μi’ and objective function ‘oi’. Thus, gCCA algorithm
selects the software metrics with higher correlation value as best in order to enhance
the performance of software quality prediction with lower time complexity.
The algorithmic processes of gCCA algorithm are explained below.
Algorithm 1 Generalized Canonical Correlation Analysis
//Generalized Canonical Correlation Analysis Algorithm
Input: large number of software metrics ‘{μ1,μ
2,...,μ
n}’;
Output: Select relevant code metrics for improving the software quality
Step 1: Begin
Step 2: For each input software metric ‘μi’
Step 3: Determine covariance ‘Cv(μi,oi)’using(1)
Step 4: Evaluate the canonical correlation ‘Cr using (2)
Step 5: If (Cr is higher) then
Step 6: Select software metric ‘μi’ as relevant
Step 7: Else
Step 8: Remove software metric ‘μi’ as irrelevant
Step 9: End If
Step 10: End For
Step 11: End
Algorithm 1 describes the step by step processes of generalized canonical corre-
lation analysis. By using the above algorithmic process, gCCA algorithm efficiently
chooses the features that are more significant for increasing software quality with a
lower amount of time consumption. Hence, gCCA algorithm attains enhanced feature
selection performance for effective software quality management as compared to
existing works. As a result, GCCA-BERVMC model provides better metrics selection
performance with a lower time complexity as compared to state-of-the-art works.
2.2 Bagging Ensembled Relevance Vector Machine
Classification
Bagging Ensembled Relevance Vector Machine Classification (BERVMC) algorithm
is designed in GCCA-BERVMC model with aiming at improving the performance
634 N. Ayesha and N. G. Yethiraj
Apply votes for each weak
learner result
Create stron
g
classifie
r
Classify each source code line as normal or defect
Number of Source Code Lines with Selected
Metrics
Fig. 3 Block diagram of bagging ensembled relevance vector machine classification algorithm
of software quality analysis. On the contrary to conventional works, BERVMC algo-
rithm is proposed by using the Relevance Vector Machine and bagging ensemble
algorithm. The designed BERVMC algorithm takes a Relevance Vector Machine
classification as a weak learner to classify software code as normal or defect using
chosen software metrics. The flow processes of BERVMC algorithm are presented
in Fig. 3.
Figure 3shows the flow processes of BERVMC algorithm. As demonstrated in
the above figure, BERVMC algorithm at first takes number of source code lines
with selected metrics as input. Then, BERVMC algorithm designs ‘n’ number of
weak learner result for each source code. Let us consider an input software dataset
that includes of many source code lines denoted as ‘{Sc1,Sc2,Sc3,...,Scn}’. Here,
‘n’ point outs the total number of source code lines in the input dataset. In the
BERVMC algorithm, each source code is trained using weak learner, i.e., relevance
vector machine classification model to classify the source code as normal or defect.
The weak learner is represented as ‘{(Sc1,γ
1),(Sc2,γ
2),...,(Scn,γ
n)}’ where ‘Sci’
indicates a set of training samples (i.e., input source code lines) and ‘γi’ refers to the
output (prediction result). From that, weak relevance vector machine classification
result is mathematically obtained using below
z(Sci,w)=
n
i=1
wiK(Sc,Sci)+w0(3)
Generalized Canonical Correlation Based Bagging … 635
From the mathematical Eqs. (3), ‘K(Sc,Sci)’ symbolizes the kernel function,
‘wi’ refers to the weight of the ‘ith’ kernel function ‘w=[w0,w1,...,wn]T’, and
‘w0’ denotes the bias. In weak learner, Gaussian kernel function is applied in order to
map ‘z(Sci,w)’ to (0, 1) for two class classification, i.e., normal class or defect class
because, the weak learner returns output as ‘0’ or ‘1’. For minimizing the over-fitting
of weak learner due to excessive support vectors utilized, all weight vectors fulfill a
zero-mean Gaussian prior distribution with help of below expression,
P(w,a)=
n
i=0
n(wi|0,a−1
i)=
n
i=0
ai
√zπexp−aiw2
i
2(4)
From the above mathematical formula (4), ‘a=[a0,a1,a2,...,an]T’ signifies
the hyper-parameter vector which determines the prior distribution of weight vector
‘w’ and controls the degree of weight diverges from its zero-mean. For a given prior
probability distribution and the likelihood distribution, Bayes’ rule is applied in weak
learner to find out the posterior probability of models ‘w’ and ‘a’ with aid of below
equation,
P(w,a|t)=P(w|t,a)P(a|t)(5)
Subsequently, approximation procedure is employed in weak learner with aid
of Laplace’s method which identifies maximum hyper-parameter vector ‘a’using
below,
wnew
MP =wi+w,(6)
anew
MP =1−aiij
w2
MP
(7)
From the above mathematical representation (6) and (7), optimal hyper-parameter
vector ‘a’ is identified in weak learner to classify each input source code line as
normal or defect. In weak learner, the result ‘γi=0’ represents that a source code
line is normal, whereas ‘γi=1’ symbolizes that the source code line is defect.
To further enhance the accuracy of weak learner when compared to conventional
works, bagging algorithm called bootstrap aggregation is utilized in GCCA-BEC
model. Let consider a BERVMC algorithm constructs ‘n’ number of weak learner
results for each source code in bootstrap samples. Subsequently, BERVMC algorithm
all weak learner results into a strong classifier with help of below equation,
w(Sci)=w1(Sci)+w2(Sci)+···+wn(Sci)(8)
Next, BERVMC algorithm applies vote ‘δi’ for each weak learner results ‘w(Sci)’
using below
636 N. Ayesha and N. G. Yethiraj
δi→
n
i=1
w(Sci)(9)
Consequently, the majority vote of all weak learner results is employed to formu-
late a strong classifier for efficiently categorizing each source code line as normal or
defect. Accordingly, strong classifier result is mathematically performed as follows:
A=arg max
nδ(w(Sci)) (10)
From the above mathematical formulation (10), ‘A’ designates the final strong
classifier result, whereas ‘arg max
nδ’ signifies majority votes of weak learner results.
The strong classifier helps for BERVMC algorithm to correctly classify each source
code line as normal or defect based on selected software metrics with a minimal
time when compared to state-of-the-art works. The algorithmic steps of BERVMC
are explained below.
Algorithm 2 Bagging Ensembled Relevance Vector Machine Classification
//Bagging Ensembled Relevance Vector Machine Classification Algorithm
Input: Software Program Dataset; Number of Program Files
‘pf1,pf2,pf3,..., pfn’; Number of Source Code ‘Sc1,Sc2,Sc3,...,Scn’
Output: accurately classify source codes as normal or defect
Step 1: Begin
Step 2: For each input program files ‘ pfi’
Step 3: For each source code line ‘Sci’
Step 4: Design bootstrap samples
Step 5: Obtain ‘n’ number of weak learner results
Step 6: Aggregate all weak learner results using (8)
Step 7: Apply voting scheme using (9)
Step 8: Design strong classifier by considering majority voting results
using (10)
Step 9: Exactly Classify source code line as normal or defect
Step 10: End For
Step 11: End For
Step 12: End
BERVMC algorithm as presented in the above algorithmic process, initially gets a
number of source code lines as input and then produces a number of bootstrap training
samples. For each source code line in bootstrap training samples, next BERVMC
algorithm generates ‘n’ number of weak learner results.
Then, BERVMC algorithm aggregates the results of all weak learners together
and then applies voting scheme. Finally, BERVMC algorithm considers majority
votes of all weak learners to design strong classifier that correctly classifies each
source code line as normal or defect according to selected software metrics with a
lower amount of time consumption when compared to conventional works. Thus,
Generalized Canonical Correlation Based Bagging … 637
GCCA-BERVMC model attains better software quality analysis performance when
compared to conventional works.
3 Experimental Settings
In order to determine the performance of the proposed, GCCA-BERVMC model is
implemented in Java Language using schoolmate dataset. The data set using open
source projects is obtained from https://sourceforge.net/. This dataset contains 66
PHP programs for elementary, middle, and high schools. The GCCA-BERVMC
model takes different sizes of software code as input in the range of 10–100 KB.
Then, GCCA-BERVMC model classifies each source code line as normal or defect
in order to increase software quality. The performance of GCCA-BERVMC model is
estimated in terms of defect detection accuracy, time complexity, and false positive
rate with respect to different software code sizes. The experimental result of GCCA-
BERVMC model is compared with weighted majority voting techniques [9] and
neural network based classification technique [10].
4 Results
In this section, the comparative result of GCCA-BERVMC model is discussed.
The performance of GCCA-BERVMC model is compared with weighted majority
voting techniques [9] and neural network based classification technique [10], respec-
tively. The performance of KPFS-SBPRVM technique is determined along with the
following metrics with the help of tables and graphs.
4.1 Performance Measure of Defect Detection Accuracy
In GCCA-BERVMC model, Defect Detection Accuracy ‘DDA’ is evaluated as the
ratio of number of source code lines that are accurately identified as normal or defect
through classification to the total number of source code lines taken as input. The
defect detection accuracy is calculated using below
DDA =nCD
SCn∗100 (11)
From the above mathematical representation (11), accuracy of software defect
prediction is determined. Here, ‘SCn’ point outs the total number of source code
638 N. Ayesha and N. G. Yethiraj
lines and ‘nCD’ indicates number of source code lines that are properly detected. The
defect detection accuracy is evaluated in terms of percentage (%).
Sample Calculations
•Proposed GCCA-BERVMC model:DDA=45
50 ∗100 =90%
•Existing weighted majority voting techniques:DA=40
50 ∗100 =80%
•Existing neural network based classification technique:DDA=38
50 ∗100 =
76%.
The experimental result analysis of defect detection accuracy obtained during
the processes of software quality analysis using three methods namely the proposed
GCCA-BERVMC model and conventional weighted majority voting techniques [9]
and neural network based classification technique [10] is demonstrated in Table 1.
Tabl e 1 Tabulation result of defect detection accuracy
Software code size
(KB)
Defect detection accuracy (%)
GCCA-BERVMC
model
Weighted majority
voting techniques
Neural network based
classification technique
10 90 80 76
20 91 82 78
30 89 82 77
40 89 81 79
50 88 80 79
60 90 81 77
70 91 79 77
80 92 78 75
90 95 77 74
100 91 77 73
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
Defect Detection Accuracy
(%)
Software Code Size
(KB)
GCCA-BERVMC
model
weighted majority
voting techniques
neural network based
classification
technique
Fig. 4 Graphical result analysis of defect detection accuracy versus different software code sizes
Generalized Canonical Correlation Based Bagging … 639
Figure 4presents the impact of software defect detection accuracy along with
diverse number of software code size in the range of 10–100 KB using the proposed
GCCA-BERVMC model and traditional Ensemble weighted majority voting tech-
niques [9] and neural network based classification technique [10]. As demonstrated in
the above graphical representation, the proposed GCCA-BERVMC model achieves
higher defect detection accuracy with increasing size of software code as input
when compared to existing weighted majority voting techniques [9] and neural
network based classification technique [10]. This is because of application of gener-
alized Canonical Correlation Analysis (gCCA) algorithm and Bagging Ensembled
Relevance Vector Machine Classification (BERVMC) algorithm in the proposed
GCCA-BERVMC model on the contrary to state-of-the-art works.
By using the concepts of gCCA algorithm, the proposed GCCA-BERVMC model
selects the features that are more relevant in order to enhance software quality. Further
with application of BERVMC algorithm, the proposed GCCA-BERVMC model
designs strong classifier that accurately classifies each source code line as normal
or defect based on selected software metrics with higher accuracy when compared
to conventional works. Hence, the proposed GCCA-BERVMC model enhances the
ratio of number of source code lines that are accurately identified as normal or defect
when compared to other works. As a result, the proposed GCCA-BERVMC model
increases the software defect detection accuracy by 14 and 19 % when compared
to weighted majority voting techniques [9] and neural network based classification
technique [10].
4.2 Performance Measure of Time Complexity
In GCCA-BERVMC model, time complexity ‘TC’ determines the amount of time
needed to identify the defects in given software program. The time complexity is
mathematically estimated using below
TC=n∗Time(DDSSC )(12)
From the above mathematical Eq. (12), the time complexity of software defect
prediction is calculated. Here, ‘n’ refers to the number of source code lines and
‘T(DDSSC )’ symbolizes the time utilized for detecting the defect in a single source
code line. The time complexity is evaluated in terms of milliseconds (ms).
Sample calculations
•Proposed GCCA-BERVMC model:TC=50 ∗0.2=10 ms
•Existing weighted majority voting techniques:TC=50 ∗0.3=15 ms
•Existing neural network based classification technique:TC=50 ∗0.33 =
17 ms.
640 N. Ayesha and N. G. Yethiraj
Tabl e 2 Tabulation result of time complexity
Software code size
(KB)
Time complexity (ms)
GCCA-BERVMC
model
Weighted majority
voting techniques
Neural network based
classification technique
10 10 15 17
20 12 16 19
30 14 18 20
40 19 26 26
50 25 31 32
60 29 40 41
70 33 44 46
80 35 46 48
90 41 51 54
100 45 57 59
The tabulation result analysis of time complexity involved during the processes
of software quality analysis using three methods namely the proposed GCCA-
BERVMC model and traditional weighted majority voting techniques [9] and neural
network based classification technique [10] is depicted in Table 2.
Figure 5illustrates the impact of time complexity of software quality analysis
based on different number of software code sizes in the range of 10 KB to 100 KB
using the proposed GCCA-BERVMC model and state-of-the-art weighted majority
voting techniques [9] and neural network based classification technique [10]. As
presented in the above graphical figure, proposed GCCA-BERVMC model attains
lower time complexity to predict the software defect with increasing size of software
code as input when compared to existing weighted majority voting techniques [9]
and neural network based classification technique [10]. This is owing to applica-
tion of generalized Canonical Correlation Analysis (gCCA) algorithm and Bagging
Ensembled Relevance Vector Machine Classification (BERVMC) algorithm in the
proposed GCCA-BERVMC model on the contrary to existing works.
With the help of gCCA algorithmic process, the proposed GCCA-BERVMC
model picks the key software metrics with minimal amount of time consumption to
0
20
40
60
10 20 30 40 50 60 70 80 90 100
Time complexity
(ms)
Software Code Size…
GCCA-BERVMC
model
weighted majority
voting techniques
neural network based
classification technique
Fig. 5 Graphical result analysis of time complexity versus different software code sizes
Generalized Canonical Correlation Based Bagging … 641
improve software quality. In addition with application of BERVMC algorithm, the
proposed GCCA-BERVMC model generates strong classification result and thereby
perfectly classifies each source code line as normal or defect using chosen soft-
ware metrics with lower amount of time consumption when compared to existing
works. Therefore, the proposed GCCA-BERVMC model minimizes the amount of
time needed to discover the defects in given software program when compared to
other works. Accordingly, the proposed GCCA-BERVMC model reduces the time
complexity by 24 and 29 % when compared to conventional weighted majority voting
techniques [9] and neural network based classification technique [10].
4.3 Performance Measure of False Positive Rate
In GCCA-BERVMC model, False Positive Rate ‘FPR’ measures the ratio of number
of source code lines that are wrongly detected as normal or defect to the total number
of source code lines taken as input. The false positive rate is determined as follows.
FPR =nWD
SCn∗100 (13)
From the above mathematical expression (13), false positive rate of software defect
prediction is calculated. Here, ‘SCn’ indicates the total number of source code lines,
whereas ‘nWD’ denotes number of source code lines that are mistakenly detected.
The false positive rate is estimated in terms of percentage (%).
Sample Calculations
•Proposed GCCA-BERVMC model: FPR =5
50 ∗100 =10%
•Existing weighted majority voting techniques: FPR =10
50 ∗100 =20FPR%
•Existing neural network based classification technique: FPR =12
50 ∗100 =
24%.
The performance result analysis of false positive rate of software defect predic-
tion during the processes of software quality analysis using three methods namely
the proposed GCCA-BERVMC model and state-of-the-art weighted majority voting
techniques [9] and neural network based classification technique [10]isshownin
Table 3.
Figure 6demonstrates the impact of false positive rate of software defect predic-
tion according to diverse number of software code size in the range of 10–100 KB
using the proposed GCCA-BERVMC model and traditional weighted majority voting
techniques [9] and neural network based classification technique [10]. As shown in
the above graphical figure, the proposed GCCA-BERVMC model obtains minimal
false positive rate to correctly detect the software defects with increasing size of
software code as input when compared to conventional weighted majority voting
techniques [9] and neural network based classification technique [10]. This is due
642 N. Ayesha and N. G. Yethiraj
Tabl e 3 Tabulation result of false positive rate
Software code size
(KB)
Falsepositiverate(%)
GCCA-BERVMC
model
Weighted majority
voting techniques
Neural network based
classification technique
10 10 20 24
20 918 22
30 11 18 23
40 12 19 21
50 11 18 21
60 10 19 23
70 921 23
80 822 25
90 523 26
100 923 27
0
5
10
15
20
25
30
10 20 30 40 50 60 70 80 90 100
False positive rate (%)
Software Code Size
(KB)
GCCA-BERVMC
model
weighted majority
voting techniques
neural network
based classification
technique
Fig. 6 Graphical result analysis of false positive rate versus different software code sizes
to application of generalized Canonical Correlation Analysis (gCCA) algorithm and
Bagging Ensembled Relevance Vector Machine Classification (BERVMC) algorithm
in the proposed GCCA-BERVMC model on the contrary to state-of-the-art works.
With the support of gCCA algorithmic steps, the proposed GCCA-BERVMC
model finds the most important software metrics to accurately perform software
quality analysis process. As well with application of BERVMC algorithm, the
proposed GCCA-BERVMC model categorizes each source code line as normal or
defect considering selected software metrics with minimal misclassification error
when compared to conventional works. For that reason, the proposed GCCA-
BERVMC model reduces the ratio of number of source code lines that are imperfectly
detected as normal or defect when compared to other works. Thus, the proposed
GCCA-BERVMC model decreases the false positive rate of software defect predic-
tion by 52 % and 59 % when compared to conventional weighted majority voting
techniques [9] and neural network based classification technique [10].
Generalized Canonical Correlation Based Bagging … 643
5 Conclusion
The GCCA-BERVMC model is proposed with the intention of increasing the perfor-
mance of software defect prediction via performing classification with a minimal
error rate. The objective of GCCA-BERVMC model is obtained with the applica-
tion of generalized Canonical Correlation Analysis (gCCA) algorithm and Bagging
Ensembled Relevance Vector Machine Classification (BERVMC) algorithm the
contrary to traditional works. The proposed GCCA-BERVMC model increases the
ratio of number of source code lines that are correctly predicted as normal or defect
by using BERVMC algorithm when compared to conventional works. As well, the
proposed GCCA-BERVMC model minimizes the time utilized to predict software
defect with the support of gCCA and BERVMC algorithmic process when compared
to other state-of-the-art works.
In addition, the proposed GCCA-BERVMC model minimizes the ratio of number
of source code lines wrongly predicted for software quality analysis when compared
to other existing works. Therefore, proposed GCCA-BERVMC model gives better
software quality analysis performance in terms of defect detection accuracy, time
complexity, and false positive rate when compared to conventional works. The
experimental result shows that the proposed GCCA-BERVMC model presents
better software defect prediction performance with an enhancement of accuracy and
minimization of time when compared to state-of-the-art works.
References
1. H.B. Yadav, D.K. Yadav, Early software reliability analysis using reliability relevant software
metrics. Int. J. Syst. Assur. Eng. Manag. 8, 2097–2108 (2017)
2. S. Chatterjee, B. Maji, A new fuzzy rule based algorithm for estimating software faults in early
phase of development. Soft. Comput. 20(10), 4023–4035 (2016)
3. A. Okutan, O.T. Yıldız, Software defect prediction using Bayesian networks. Empir. Softw.
Eng. 19(1), 154–181 (2014)
4. S. Huda, K. Liu, M. Abdelrazek, A. Ibrahim, S. Alyahya, H. Al-Dossari, An ensemble over-
sampling model for class imbalance problem in software defect prediction. IEEE Access 6,
24184–24195 (2018)
5. T. Wang, Z. Zhang, X. Jing, L. Zhang, Multiple kernel ensemble learning for software defect
prediction. Autom. Softw. Eng. 23(4), 569–590 (2016)
6. H. Tong, B. Liu, S. Wang, Software defect prediction using stacked denoising autoencoders
and two-stage ensemble learning. Inf. Softw. Technol. 96, 94–111 (2018)
7. Z. Xu, S. Li, J. Xu, J. Liu, X. Luo, Y. Zhang, T. Zhang, J. Keung, Y. Tang, LDFR: learning
deep feature representation for software defect prediction. J. Syst. Softw. 158 (2019)
8. C.-Y. Huang, T.-Y. Kuo, Queuing-theory-based models for software reliability analysis and
management. IEEE Trans. Emerg. Top. Comput. 5(4), 540–550 (2017)
9. S. Moustafa, M.Y. El Nainay, N. El Makky, M.S. Abougabal, Software bug prediction using
weighted majority voting techniques. Alexandria Eng. J. 57(4), 2763–2774 (2018)
10. R. Jayanthi, L. Florence, Software defect prediction techniques using metrics based on neural
network classifier. Cluster Comput. 22, 77–88 (2019)
A Deep Learning Approach Against
Botnet Attacks to Reduce
the Interference Problem of IoT
Pramathesh Majumdar, Archana Singh, Ayushi Pandey,
and Pratibha Chaudhary
Abstract Today we are witnessing a world where hacking into a user’s computer
using tiny bots or intercepting a group of interconnected devices is no more impos-
sible. These tiny bots are called botnets which are a group of malicious codes that
can hamper the whole security system without the knowledge of the user. As Internet
of Things (IoT) is emerging rapidly, the interconnected devices are susceptible to
breach as one affected device can hamper the whole network. The security threat
remains as botnet attacks increase their presence to the interconnected devices. In
this work, we are implementing Restricted Boltzmann Machine (RBM) algorithm
of deep learning approach on the CTU-13 dataset to train the algorithm about the
botnet attack patterns in IoT and to prevent the botnet attacks on IoT devices, thus
reducing the interference problem in the network.
Keywords Bitnets ·DeBot ·Interference problem ·Internet of Things (IOT) ·
Security
1 Introduction
Today we are living in an era of advanced technology where sneaking into a remote
system is not impossible anymore. Malwares have become so advanced today that
they even don’t need to write anything into the system files. The concept of dropping
a malicious file into a target system is long gone now. Malwares are now able to
use legitimate inbuilt applications to access the mainframe of the system. These
P. Ma j u m d ar ·A. Singh ·P. Chaudhary
Amity University, Noida, India
e-mail: Pramathesh@outlook.com
A. Singh
e-mail: asingh27@amity.edu
A. Pandey (B)
SRM IST, Ghaziabad, India
e-mail: Pandeyayushi185@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_58
645
646 P. Majumdar et al.
advanced class of malwares are called file-less malwares. The concept of botnets is
somewhat like file-less malwares. Botnets don’t need to drop a malicious file into
the target system. Their main goal is to hamper as much devices in the network
as possible without leaving any trace of the source. In IoT, where every device
is connected, botnet attacks have proven themselves to be an alarming challenge.
Until the IoT concerns improve their security mechanisms, botnet attacks have the
potential to become the new weapon of future cyberattacks. Since the last decade,
the statistics of the botnet attacks have increased rapidly. The more technology is
becoming advanced, the more advanced botnets are becoming.
Present technology of botnet detection mechanisms like signature or flow-based
error detection is providing sound results when there is a small-scale botnet attack.
In larger scale, they are not meeting the desired output. Nowadays we are witnessing
the rise of artificial intelligence-based botnets which can learn the pattern of the
users’ usage mechanism and deploy themselves accordingly. Therefore, there is an
unavoidable need of learning-based mechanisms to be implemented in the field of
botnet attack prevention.
We are implementing a deep learning approach (Restricted Boltzmann Machine)
to train the algorithm more accurately with a higher accuracy level about the attack
patterns of the botnets. The dataset used in this work is the CTU-13 dataset of
botnet attacks which shows several botnet attacks patterns and their feature-set and
parameters.
This paper proposes a solution to the detection of botnet activity within consumer
IoT devices and networks. A novel detection algorithm was developed based on deep
learning mechanism. Detection was performed at the packet level with Wireshark by
creating a fake network using ApateDNS.
2 Literature Review
Many researches have been done on providing an efficient solution against botnet
attacks. Torres et al. proposed a methodology to compare several Recurrent Neural
Network (RNN) models and the efficiency is analyzed to behavior of the traffic
by designing it as a sequence of states that changes [1]. The focus of this work
is to analyze the behavioral characteristics of the botnets. This work analyses the
RNN methodologies because of the two main reasons—unbalanced network traffic
and length of the data sequence. To perform the proposed concept, a K-fold cross-
validation and test was conducted on unseen data extracted from different botnet
samples. From the results, it was found that the RNN model is capable of efficiently
differentiating the botnet attacks with a high detection rate. But it is not as much
efficient when using imbalanced network traffic.
The ongoing development of the Internet of Things (IoT) has brought about an
ascent in IoT-based DDoS assaults. McDermott et al. exhibit an answer for the iden-
tification of botnet action inside purchaser IoT gadgets and systems [2]. A tale use of
A Deep Learning Approach Against Botnet Attacks … 647
deep learning is utilized to build up an identification model dependent on a Bidirec-
tional Long Short-Term Memory based Recurrent Neural Network (BLSTM-RNN).
This paper shows that although the bidirectional methodology adds overhead to every
age and expands preparing time, it demonstrates to be a superior dynamic model after
some time. A named dataset was created as a component of this examination and is
accessible upon solicitation.
Yan et al. worked on the new patterns and attributes of DDoS assaults in distributed
computing and gave a far-reaching overview of characterizes systems against DDoS
assaults utilizing SDN [3]. Moreover, the examinations about propelling DDoS
assaults on SDN, just as the techniques against DDoS assaults in SDN were analyzed.
Apparently, the repudiate or connection among SDN and DDoS assaults has not been
all around tended to in past works. This work shows how to utilize SDN’s focal points
to crush DDoS assaults in distributed computing conditions and how to avert SDN
itself from turning into a casualty of DDoS assaults, which are significant for the
smooth advancement of SDN-based cloud without the diversion of DDoS assaults.
Botnets comprise an essential risk to Internet security. The capacity to precisely
recognize botnet traffic from non-botnet traffic can help essentially in moderating
pernicious botnets. Roosmalen et al. presented a novel way to deal with botnet
discovery that applies profound learning on streams of TCP/UDP/IP-parcels [4].
As of late, botnets have turned out to be one of the significant dangers to data secu-
rity since they have been continually developing in both size and advancement. A few
botnet identifications measures, for example, honeynet-based and Intrusion Detec-
tion System (IDS)-based, have been proposed. Nonetheless, IDS-based arrangements
that utilization marks appear to be insufficient on the grounds that ongoing botnets
are outfitted with advanced code update and avoidance strategies. A few investiga-
tions have appeared unusual as botnet recognition strategies are more successful than
mark-based techniques since oddity-based botnet discovery strategies don’t require
pre-constructed botnet marks and henceforth they have the capacity to recognize
new or obscure botnets. Toward this path, Hoang X. and Nguyen Q. proposed a
botnet discovery model dependent on AI utilizing Domain Name Service question
information and assessed its viability utilizing famous AI strategies [5]. Exploratory
outcomes demonstrate that AI calculations can be utilized successfully in botnet
recognition and the irregular timberland calculation creates the best by and large
identification precision of over 90%.
Since the past decade, botnet has risen as an intense risk to digital security by
giving its ability of bargaining billions of PCs and making them do the unlawful
work. There are various existing ways by which botnet can be recognized. A thorough
review of the current strategies is likewise expressed in this paper. Because of the
inclusion of immense measure of information, recognition of botnet utilizing AI
calculations is in tremendous pattern. Mathur et al. proposed a method to utilize
AI to prepare classifiers by a system stream dataset. From that point, the prepared
classifiers were connected on the gathered information to assess the outcomes [6].
Investigation of system stream information is utilized as a technique for location since
it doesn’t rely on the bundle content henceforth giving resistance toward the most
recent type of encryption and confusion utilized by assailants to shroud their bots.
648 P. Majumdar et al.
Results are unmistakably demonstrating that the proposed technique can separate
the typical traffic and the bot-traffic with a high precision and low false positive rate.
Also, pretty much every sort of botnet can be distinguished utilizing the proposed
model.
In the age of the Internet of Things (IoT), a gigantic measure of detecting gadgets
gathers as well as produces different tangible information after some time for a
wide scope of fields and applications. Considering the idea of the application, these
gadgets will result in huge or quick/constant information streams. Applying investi-
gation over such information streams to find new data, anticipate future experiences,
and settle on control choices is a significant procedure that makes IoT a commend-
able worldview for organizations and a personal satisfaction improving innovation.
Mohammadi et al. proposed an analysis outline on utilizing a class of cutting-edge
AI strategies, Deep Learning (DL), to encourage the examination and learning in the
IoT area [7]. The procedure is performed by articulating IoT information qualities
and recognizing two noteworthy medicines for IoT information from an AI view-
point, IoT huge information examination, and IoT spilling information investigation.
Additionally, why profound learning is a promising way to deal with accomplish the
ideal examination in these sorts of information and applications were talked about.
Kudugunta and Ferrara proposed a system utilizing streak against 2D parodying
assault to be utilized as a face liveness identification component [8]. Alongside the
upgrade of separation among genuine just as ill-conceived clients, there is mini-
mization of the impacts caused because of the natural components present because
of the nearness of glimmer. A picture that incorporates streak and the one that does
exclude streak is taken from the subject here. To catch the data of two pictures,
present inside the model, four surfaces just as 2D structure descriptors are used that
have low computational unpredictability.
Wireless devices are the largest part of the Internet of Things (IoT) units. Radiated,
such as specific absorption rate, and spurious related issues are regulated by the
standards bodies. The functionality performance will greatly influence IoT system
stability, user experience, and cost. Radio receiver desensitization and coexistence
aspects of IoT devices are discussed in work done by Yihong Qi et al. [9].
Measurement technologies are briefly reviewed for transmit/receive mode only
of single-input and single-output, and multiple-input multiple-output devices, to
provide a means of device performance evaluation and electromagnetic interference
troubleshooting. An understanding of the electromagnetic compatibility aspects of
the IoT and its measurement methods can help ensure the best user experience for
ubiquitous wireless systems.
Bakshi et al. presented central thoughts of EMIT by portraying the worldwide
impedance insights as far as single-gadget task and creates control rate allotment tech-
niques to ensure low-postpone high-unwavering quality execution [10]. The future
Internet of Things (IoT) systems are required to be made from a huge populace of
ease gadgets discussing progressively with passages or neighboring gadgets to convey
little packages of postponing delicate information. To help the high-force and brief
requests of these developing systems, we propose an Efficient MAC worldview for
IoT (EMIT). The worldview sidesteps the high overhead and coordination expenses
A Deep Learning Approach Against Botnet Attacks … 649
of existing MAC arrangements by utilizing an impedance averaging methodology
that enables clients to share their assets all the while. As opposed to the prevalent
obstruction stifling methodologies, EMIT abuses the thick and dynamic nature of
IoT systems to diminish the spatio-fleeting fluctuation of impedance to accomplish
low-postponement and high-unwavering quality in administration.
The Internet of Things (IoT) imagines inescapable, associated, and shrewd hubs
cooperating self-rulingly while offering a wide range of administrations. Wide appro-
priation, transparency, and generally high preparing intensity of IoT objects made
them a perfect focus for digital assaults. Additionally, the same number of IoT hubs
is gathering and preparing private data, they are turning into a goldmine of infor-
mation for pernicious entertainers. Along these lines, security and explicitly the
capacity to recognize traded off hubs, together with gathering and protecting confir-
mations of an assault or malevolent exercises rise as a need in effective arrangement
of IoT systems. In this paper, we initially present existing significant security and
legal sciences challenges inside IoT area and after that quickly talk about papers
distributed in this exceptional issue focusing on distinguished difficulties.
Deep learning is a promising methodology for separating exact data from crude
sensor information from IoT gadgets sent in complex situations. Due to its multilayer
structure, profound learning is additionally fitting for the edge processing condition
[11]. Consequently, in this article, we initially bring profound learning for IoTs into
the edge figuring condition. Since existing edge hubs have restricted handling ability,
we likewise structure a novel offloading technique to advance the presentation of IoT
profound learning applications with edge processing. In the presentation assessment,
we test the exhibition of executing various profound learning undertakings in an edge
registering condition with our technique. The assessment results demonstrate that our
strategy outflanks other advancement arrangements on profound learning for IoT.
Kasprzyk et al. performed an investigation of digital dangers, with specific accen-
tuation on the dangers coming about because of botnet movement [12]. Botnets
are the most widely recognized sorts of dangers and frequently saw as essential
regarding national security. Their grouping and strategies for spreading are the reason
for making the internet model including the nearness of various kinds of digital
dangers. A well-structured internet model empowers to build an exploratory situ-
ation that takes into consideration the investigation of botnet attributes, testing its
protection from different occasions and recreation of the spread and advancement.
For this reason, committed stages with capacities and practical attributes to meet
these prerequisites have been proposed.
The expansion of IoT gadgets that can be more effectively bargained than work-
stations has prompted an expansion in IoT-based botnet assaults. To alleviate this
danger, there is a requirement for new techniques that recognize assaults propelled
from traded off IoT gadgets and that separate among hours-and milliseconds-long
IoT-based assaults. Meidan et al. proposed a novel system based on abnormality
discovery strategy for the IoT called N-BaIoT that extricates conduct previews of
the system and uses profound autoencoders to recognize bizarre system traffic from
traded off IoT gadgets [].
650 P. Majumdar et al.
3 Problem Statement
Many technologies have been developed over the years to cope with the increasing
threat of the cybersecurity, but the ration of cybercrime is increasing instead of
decreasing. Heuristic-based tools use rules to examine suspicious codes and classify
them as malware. This approach is limited, however, due to the fact that it relies on
the sequence of repeated code that is indicative of malicious intent. Hence, in this
work, we are presenting a view on the combined approach of static and dynamic
analyses with tools based on real-time extraction (Fig. 1).
3.1 Botnets in Internet of Things
A botnet is a network of interconnected small group of computers which runs on one
or more bots. Their main target is to deploy botnet framework into the target system’s
mainframe to gain access to the system’s administrative privileges. Once this step is
done, they use their framework to manipulate the users’ behavior and intervention to
the system processes. For example, the Torii botnet attack (September 20, 2016) was
one of the most severe botnet attacks which took over the IoT network. The “Bot-
master” uses the C&C server to gain access to the victim system and transfer the
information via IP trafficking through the botnet’s framework. Botnet is an abbrevi-
ation which was taken from “Robot” and “Network” (Bot +Net). The main motive
behind a botnet attack is to collapse as much IoT devices as possible and to gain
access to the privileged information of a target system. The effects of botnets are
Fig. 1 Existing problem of botnet attack on IoT network
A Deep Learning Approach Against Botnet Attacks … 651
increasing in a rapid manner which was anticipated in the early ages. Statistics show
that most of cyberattacks somehow involve in botnet implementations.
3.2 Work-Flow Model
The above work-flow model shows the step by step process of the proposed concept.
They are as follows (Fig. 2):
i. At first, we need to collect the required botnet sample to update the dataset
repository. Here, we used the CTU-13 dataset.
ii. A secure lab environment was created to ensure the safety of the public DNS
and internal resources. Oracle Virtual Machine was installed using Windows
10 operating system to run the botnet sample. ApateDNS tool was used for this
purpose, which creates a fake network to restrict the botnet from realizing that
it is running inside a virtual machine.
iii. Analyze the botnet sample and record its severity level to determine the number
of epochs to run.
Fig. 2 Work-flow model of
the proposed concept
652 P. Majumdar et al.
iv. In the next step, the analyzed result was combined using the CTU-13 dataset to
update the data repository.
v. Based on the updated record, the number of epochs was determined.
vi. Now, in this step, two experiments were performed using different epoch
numbers. The first one used epoch number less than the pre-determined number
and the second one used the pre-determined epoch number. This step was
performed to cross-verify the accuracy of the epoch selection and accuracy
level.
vii. Compare both the results from the previous step and determine the best result
among them. The model is then ready and trained to detect botnet samples using
the updated repository.
3.3 Deep Learning Against Botnets
Now we are leaving in a world of IoT where almost every device is interconnected to
ease our lives. But the same mechanisms can and are being used against the security of
the cyberspace. As we know, botnets target the most vulnerable devices and connec-
tions in an IoT network to spread into other devices, the question remains that how
we are going to stop this attempt cybersecurity invasion. Malwares, ransomwares and
botnets are becoming advanced faster than we can improve our defense against them.
AI-enabled malwares, ransomwares, botnets are not new today. This threatens cyber-
security and raises the question that are we secure enough and advanced to defend
ourselves in case in future any more devastating security threats raise concern.
Deep Learning (DL) mechanism is an advanced extension of the Machine
Learning (ML) mechanism which includes the following main types—Boltzmann
Machine (BM), Neural Network (NN), and Stacked Auto Encoders. In this work, we
are comparing machine learning approaches with the deep learning Neural Network
Sequential algorithm to find out whether machine learning or deep learning does
more accurate job in learning the attack patterns accurately.
Figure 3shows the implementation of Restricted Boltzmann Machine (RBM) on
the CTU-13 dataset in our secure lab environment by creating a imitated network
using ApateDNS tool to restrict the botnet sample from gathering information about
the virtual environment. ApateDNS starts a Virtual Private Network (VPN) inside
the Virtual Machine to limit the botnet’s access to the public DNS.
As we can see from the result, the accuracy of the algorithm is 64.88% when we
are using epoch =5. Epoch is the count of iterations that are implied on the dataset
to train the algorithm in a recursive manner. The more the count of epoch, the better
the accuracy level is.
Count of epoch ∞Accuracy level
A Deep Learning Approach Against Botnet Attacks … 653
Fig. 3 Implementation of Restricted Boltzmann Machine (RBM) using epoch =5
Using the CTU-13 dataset, we extracted the features from the dataset and applied
Restricted Boltzmann Machine (RBM) algorithm on it.
We adjusted the number of epochs in different phases (e.g., epoch =1, 2, 3 and
4) and found that the model gives underfitting results till epoch =9. In epoch =10,
the model provides optimal result with an accuracy of 95.13%. Again, in epoch =
11, the model becomes overfitted. Hence, we selected number of iterations of epoch
=10 (Fig. 4).
4 Conclusion and Future Scope
In this paper, we worked on a deep learning-based botnet detection algorithm, which
trains the IoT devices few ways for future research have been distinguished. To show
the capacity of our created model to identify new varieties of botnets, a changed
adaptation of the Torii sample will be utilized in the next phase of the work, to
produce a second combined dataset and will be looked at against existing mark and
stream-based oddity location strategies.
This work will be extended using a sample analysis of Torii botnet and combining
the parameter set extracted with the CTU-13 dataset. A novel comparison will be
done in the next phase on the dataset between machine learning approach and deep
learning approach to find out which algorithm provides the best result and creating
a model which can predict the attacks.
The following figure shows the implementation result of the RBM model on the
CTU-13 dataset with epoch =10 (0–9). This model provides an optimal result with
a minimum processing time. Also, this model analyses the dataset and its patterns
more accurately than other phases performed.
654 P. Majumdar et al.
Fig. 4 Implementation of Restricted Boltzmann Machine (RBM) using epoch =1
References
1. C.D. McDermott, F. Majdani, A.V. Petrovski, Botnet detection in the internet of things using
deep learning approaches, in 2018 International Joint Conference on Neural Networks(IJCNN)
(IEEE, 2018), pp. 1–8
2. Q. Yan, F.R. Yu, Q. Gong, J. Li, Software-defined networking (SDN) and distributed denial of
service (DDoS) attacks in cloud computing environments: a survey, some research issues, and
challenges. IEEE Commun. Surv. Tutor. 18(1), 602–622 (2018)
3. P. Torres, C. Catania, S. Garcia, C.G. Garino, An analysis of Recurrent Neural Networks for
Botnet detection behavior, in 2016 IEEE biennial congress of Argentina (ARGENCON) (IEEE,
2016), pp. 1–6
4. J.V. Roosmalen, H.P.E. Vranken, M.C.J.D. van Eekelen, H.H. Haddad, Applying deep learning
on packet flows for botnet detection, in SAC18: The 17th edition of the Computer Security Track
at the 33rd ACM Symposium on Applied Computing, 9–13 April 2018, Pau, France, ed. by H.H.
Haddad (ACM, New York, 2018), pp. 1629–1637
5. H. Zin, C. Kim, M. Wu, S. Kim, Avoidance of channel interference in polygonal IoT networks.
Concurr. Comput.: Pract. Exp. 29(11), e4060 (2017)
6. X. Hoang, Q. Nguyen, Botnet detection based on machine learning techniques using DNS
query data. Future Internet 10(5), 43 (2018)
7. L. Mathur, M. Raheja, P. Ahlawat, Botnet detection via mining of network traffic flow. Procedia
Comput. Sci. 132, 1668–1677 (2018)
8. S. Homayoun, M. Ahmadzadeh, S. Hashemi, A. Dehghantanha, R. Khayami, BoTShark: a deep
learning approach for botnet traffic detection. Cyber Threat Intelligence (2018), pp. 137–153
9. M. Mohammadi, A. Al-Fuqaha, S. Sorour, M. Guizani, Deep learning for IoT big data and
streaming analytics: a survey. IEEE Commun. Surv. Tutor. 20(4), 2923–2960 (2018)
10. S. Kudugunta, E. Ferrara, Deep neural networks for bot detection. Inf. Sci. 467, 312–322 (2018)
A Deep Learning Approach Against Botnet Attacks … 655
11. A. Bakshi, L. Chen, K. Srinivasan, C.E. Koksal,A. Eryilmaz, EMIT: an efficient MAC paradigm
for the Internet of Things, in IEEE INFOCOM 2016—The 35th Annual IEEE International
Conference on Computer Communications, April (IEEE, 2016), pp. 1–9
12. M. Conti, A. Dehghantanha, K. Franke, S. Watson, Internet of Things Security and Forensics:
Challenges and Opportunities (2018)
13. H. Li, K. Ota, M. Dong, Learning IoT in edge: deep learning for the internet of things with
edge computing. IEEE Netw. 32(1), 96–101 (2018)
Predicting Movie Success Using
Regression Techniques
Faaez Razeen, Sharmila Sankar, W. Aisha Banu, and Sandhya Magesh
Abstract Hollywood is the largest and most profitable movie industry in the world.
In 2018 alone, it generated a massive global box office of over $42 billion. A single
production company with multiple movies may benefit greatly from knowing which
movies are likely to succeed—it would help them focus their resources on the required
advertisement and promotion campaigns. Furthermore, theaters would get a prefer-
ence on which movies to run for a longer duration based on its success rate. Large-
scale investments come with large risks. Using machine learning to predict revenues
may help investors mitigate these risks. The algorithms in this paper aim to recognize
historical patterns in the movie industry to try and predict the success of upcoming
movies using a variety of machine learning algorithms. The success metric used
is the box office, i.e., the commercial success of a film in terms of overall money
earned. The results show that it is indeed possible to predict revenue with a consid-
erable amount of accuracy, with better results than a majority of the papers that were
reviewed.
Keywords Regression ·Machine learning ·Movie revenue
1 Introduction
The movie industry is one of the most lucrative industries in the world. Including
the global box office and home entertainment revenue, the global film industry is
worth $136 billion as of 2018. Marketing and advertising costs take up to 50%
of the entire budget allotted for a movie. For a big chunk of the budget, spending
needs to be done wisely. Here is where machine learning comes to play. A movie
production company often has many movies and TV shows under production at the
F. Razeen (B)·S. Sankar ·W. A . B a n u ·S. Magesh
B.S. Abdur Rahman Crescent Institute of Science and Technology, Chennai, India
e-mail: razeenfaaez@gmail.com
S. Sankar
e-mail: sharmilasankar@crescent.education
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_59
657
658 F. Razeen et al.
same time. Since advertising is a huge expense, it would help the company to know
which movie to put their money in. This is better supported by the fact that not all
movies become blockbusters, i.e., movies that are of great commercial success. In
this paper, the movies used are of only U.S. release in hopes of better performance
due to homogeneity.
Successful prediction of a movie’s success is complicated, as it depends on
multiple factors. Success here can be defined in two ways—the revenue it earns
or a quantification of how well it is received, i.e., user or critic score. Revenue is a
much better definition of success in this case, as the amount of money a movie earns
compared to its budget is a more rewardable measure of success.
The data used in this paper was a combination of data taken from IMDb (Internet
Movie Database), TMDb (The Movie Database), and Facebook. Due to the fact that
this paper uses metadata about a movie, predicting the revenue before it releases
was not possible. Other papers which had done this used data like Twitter tweet rate,
number of theaters the movie was going to be released in, number of Wikipedia edits
at the time before release, etc. These were possible to retrieve before release. This
paper uses Facebook data. The disadvantage here is that Twitter is more representative
of the hype surrounding a movie. Someone excited about a movie who uses both
Facebook and Twitter will obviously be using Twitter to talk about it due to the fact
that Twitter tweets are meant to be real-time and to share trending information, while
Facebook is more about connecting with friends, sharing photos and videos, etc.
The approach presented in the paper mainly focuses on regression algorithms. This
paper aims to see if using machine learning algorithms can help predict movie revenue
accurately. It also aims to find the features that are most important in determining a
movie’s revenue.
2 Literature Review
Quader et al. [1] used multi-class classification with five classes to predict whether
a movie would be a “flop” or a “blockbuster,” along with three other categories in
the middle. Sentiment analysis was done by calculating the sentiment score for each
review of a movie using Microsoft Azure’s Text Analytics API and multiplying it by
the total number of reviews. Using a variety of classification techniques, they found
that IMDb vote count and number of screens were the best predictors. They have used
two types of predictions: one, an exact match which rewards correct classifications,
and the other being a “one-away” prediction, where the predicted class is considered
correct if it is one away from the actual class. Considering the fact that there are only
five classes, this method of classification is not indicative of actual accuracy. The
inaccuracy is evident by the fact that using one-away prediction gave accuracies in
the high 80s, while using normal prediction gave accuracies hovering around 50%.
Such a difference is significant and should not be considered for real usage scenarios.
Yoo e t a l. [ 2] found that budget had the most correlation (~0.63) with revenue.
They have used numerical and textual features and a combination of both, “sentiment”
Predicting Movie Success Using Regression Techniques 659
features, which are taken from movie reviews and given a positive or negative weight
based on a subjectivity lexicon obtained from the University of Pittsburgh Opinion-
Finder project, although it only leads to a minuscule increase in accuracy. They have
theorized that this was due to two factors: the sentiment is already captured by the
“rating” feature, and the number of people who watched the movie matter more than
the number of people who think the movie is good. When using linear regression,
they managed to achieve a correlation of 0.7479 between the predicted revenue and
the actual revenue, which they consider not accurate enough to be used in practice.
Vr et al. [3] had used movies released in the USA and in the English language, in
hopes of getting higher accuracies due to the fact that reviews are written in English.
They implemented Linear Regression, Logistic Regression, and SVM Regression,
and had accuracies of 51, 42, and 39%. The supposed cause for this was the small
size of their dataset (1050 samples). The overall low accuracies would not prove
useful in practice.
Liu et al. [4] had used purchase intention based on tweets to find out movie revenue.
They found that the best correlated factor to box office revenue was how many people
are willing to see the movie. They took tweets from Twitter and split them based on
purchase intention. They used this to predict a movie’s first week income and eventual
revenue. They used Linear Regression and Support Vector Machine as their models.
Their best results are an adjusted 2value of 0.95 for predicting first week revenue,
and 0.74 for predicting eventual gross. Their experiment verified their assumption
that purchase intention is a better indicator than both popularity of the movie and
sentiment analysis.
Jernbäcker et al. [5] have used classification to predict success of movies, using
both revenue and rating as metrics. They have taken inflation into account in order
to reduce the number of possible outcomes from millions to a little more than a
hundred. They have used three classification algorithms: decision trees, support
vector machines, and variants of nearest neighbor classifiers. Although the accu-
racies for predicting rating were on par with other papers, the highest accuracy they
achieved on predicting the gross revenue was 15%.
Asur and Huberman [6] have shown how social media can be used to predict
future outcomes. They have used tweets referring to movies prior to their release to
predict the box office revenue generated by a movie in its opening weekend. Their
research shows that the tweet rate (the average number of tweets per hour) of any
movie is highly indicative of the success of a movie’s opening week revenue. Using
only the average tweet rate, they got an 2value of 0.80. Additionally, they have used
seven variables, created using time series values of the tweet rate for the seven days
before release, each corresponding to a particular day, along with another variable
which contains the number of theaters the movies were released in. Using sentiment
analysis to calculate the ratio of positive tweets to negative tweets yielded a slight
improvement in performance; however, these sentiments were not as important as
the tweet rates themselves. Maximum prediction accuracy was achieved using both
tweet rates time series and theater count, with an 2value of 0.97 in predicting opening
weekend gross. They have demonstrated that a simple linear regression model with
660 F. Razeen et al.
only tweet rate as a feature can perform better than artificial money markets like
Hollywood Stock Exchange (HSX).
Afzal [7] used various classification algorithms and managed to achieve an accu-
racy of 84% using logistic regression. However, they have not mentioned the number
of output classes, which is important in gauging the efficacy of the algorithm.
Classification would not prove useful in practice if a low number of classes were
used.
Pal et al. [8] have used a combination of movie metadata and sentiment anal-
ysis to predict movie revenue. They extracted sentiment scores from movie reviews
and applied various regression techniques with and without the sentiment scores to
see if they actually helped improve regression accuracy. The sentiment scores were
calculated using Naive Bayes and VADER (Valence Aware Dictionary for Sentiment
Reasoning). Using the RMSE (Root Mean Square Error) scoring metric did not seem
appropriate for this problem due to the fact that RMSE is very sensitive to outliers.
Thus, they decided to convert their regression problem into a classification problem.
A leeway of 15% was given, which means that if the prediction lies in this range,
it is counted as an accurate prediction. The highest accuracies were achieved using
Decision Tree Regression and Random Forest Regression. Although the inclusion
of sentiment scores improved accuracy and decreased RMSE, the improvement was
minor and deemed low enough to be not significant. Additionally, because they have
converted their regression problem into a classification one, there is bound to be a
loss in accuracy when used in real-time applications.
Mestyán et al. [9] have used data from Wikipedia to predict movie box office
success. To estimate the popularity of a given movie, they measured four types of
activity on a page: number of views, number of users, number of edits, and collabora-
tive rigor (quality of thoroughness of the article). They also included another feature
which they considered essential: number of theaters. A multivariate linear regression
model was used to predict revenue. Using all five features, the R2value achieved was
0.77.
3 Data Preprocessing and Feature Engineering
3.1 Preprocessing
Two separate datasets were merged using an inner join on the IMDB ID column.
Since they were merged, there were columns with the same attributes. Columns
with either lesser information or which contained a lesser degree of relevancy were
discarded. Columns which played no statistically significant role in prediction were
also removed.
The merged dataset had two of the same columns showing movie revenue.
However, the difference between them was found to be 43%. The cause of this
was because one dataset contained revenue for the region of USA, while another
Predicting Movie Success Using Regression Techniques 661
had global revenue. The column with the global revenue was dropped in favor of
USA revenue in order to keep the scope of the project homogeneous. The samples
used were only chosen if the country of production was USA and the language was
English. Other samples were dropped. Samples containing at least one null value
were also dropped. The table containing all the features and their descriptions are
shown below (Table 1).
Tabl e 1 Dataset features and
their description Feature name Description
vote_average TMDB (The Movie DataBase)
average vote rating
vote_count TMDB total vote count
num_critic_for_reviews Number of critic reviews on IMDb
(Internet Movie Database)
num_user_for_reviews Number of user reviews on IMDb
director_facebook_likes Number of likes on Facebook for
the director of the movie
actor_1_facebook_likes Number of likes on Facebook for
actor 1 of the movie
actor_2_facebook_likes Number of likes on Facebook for
actor 2 of the movie
actor_3_facebook_likes Number of likes on Facebook for
actor 3 of the movie
cast_total_facebook_likes Number of likes on Facebook for
the entire cast
movie_facebook_likes Number of likes on Facebook for
the page of the movie
revenue Box office earnings of the movie
inside U.S.A. (in dollars)
num_voted_users Number of people who voted for
the movie
facenumber_in_poster The number of actors featured on
the movie poster
duration Duration of the movie in minutes
budget Total budget allocated for the
movie in dollars
imdb_score IMDb score of the movie
release_year Year of release
662 F. Razeen et al.
3.2 Feature Engineering
Feature Engineering is a process of creating new features by utilizing available
features and domain knowledge in order to improve the performance of machine
learning models. Three features were engineered.
Genre. The revenue across different genres was also studied on. Upon visual
analysis, it was found that Action and Adventure movies on an average earned more
revenue when compared to other genres. Thus, a new column was created, which
contains a 1 if the genre of the movie is either action/adventure, else contains a 0.
Top Director. The top ten directors sorted in descending order by average revenue
per movie were used to create a new column, called “top_director,” which contains
a 1 if the director of that particular movie is a top director, else contains a 0.
Collection. If a movie was part of a series, this column contains a value of 1. If
the movie was a standalone movie, this column contains a 0.
4 Feature Selection
Feature selection is the process of selecting only a set of useful features from the
available lot. It is used to prevent overfitting, a condition where an algorithm captures
too much noise of the data, rendering it inefficient on data it has not seen before. The
removed features usually have little to no effect on prediction accuracy.
4.1 Numerical Attributes
Since only regression techniques are used, all nominal features were either dropped
or converted to categorical features.
4.2 Backward Elimination
Backward Elimination is an iterative process starting with all candidate variables,
and in each iteration, deleting the variable whose loss gives the most statistically
insignificant deterioration of the model fit, and repeating this process until no further
variables can be deleted without a statistically significant loss of fit.
In this paper, the probability value (p-value) was used. According to Wasser-
stein and Lazar [10], for a given statistical model, the p-value represents the prob-
ability that, when the null hypothesis is true, the statistical summary (such as the
sample mean difference between two groups) would be equal to or more extreme
Predicting Movie Success Using Regression Techniques 663
than the actual observed results. The smaller the p-value, the greater the statistical
incompatibility of the data with the null hypothesis.
In simpler terms, if the p-value of a certain feature is 0.05, then it means that there
is a 5% chance that the results obtained were due to pure chance rather than due to
the statistical features of the data. The industry standard for p-value is 0.05 or 5%,
and the same is used here. In each iteration of calculating the p-values, the feature
with the highest p-value is removed until all features have p-values <0.05.
5 Regression Models
All algorithms were carried out using the Scikit-learn library [11] in Python. The
efficacy of each model used would be judged based on four criteria.
•2, a value which measures goodness of fit of the model. 2=1 indicates a perfect
fit.
•Mean Absolute Error (MAE) measures the difference between two variables: the
actual value (yi) and the predicted value (yi). For easy interpretation of MAE, the
features are normalized between 0 and 1 before applying algorithms.
MAE =1
n
n
i=1
yi−ˆyi
•Root Mean Squared Error (RMSE) is the square root of the Mean Squared Error
(MSE). RMSE is a better metric compared to MAE, as it is sensitive to outliers.
RMSE =
1
n
n
i=1
yi−ˆyi
2
•Correlation between the predicted revenue and actual revenue. From the final
subset of features, vote_count was found to be the highest correlated feature
with revenue, with a Pearson correlation coefficient of 0.7515. This is used as a
baseline—all models used in this paper should be able to achieve a correlation
equal to or greater than 0.7515 (Fig. 1).
All metrics except correlation were cross validated over ten sets. Cross validation
is a resampling procedure where each subset of the data is used both as a training set
and as a testing set. Doing this helps prevent the model from overfitting on the data.
664 F. Razeen et al.
Fig. 1 Correlation between vote_count and revenue
5.1 Linear Regression
Multiple linear regression is a technique that uses multiple explanatory variables
to predict the outcome of a single response variable through modeling the linear
relationship between them. It is represented by the equation below:
y=β0+β1x1+··· +βnxn
where yi=dependent variable, xi=independent variables, β0=y-intercept, and βn
=slope coefficients.
5.2 Support Vector Regression
A Support Vector Machine is a classifier that aims to find the optimal hyperplane (the
separation line between the data classes with the error threshold value epsilon) by
maximizing the margin (the boundary between classes and that which has the most
distance between the nearest data point and the hyperplane). In this paper, a linear
kernel was used.
Predicting Movie Success Using Regression Techniques 665
5.3 Decision Tree Regression
A decision tree is a supervised classification model that predicts by learning decision
rules from features. It breaks down data into smaller subsets by making a decision
based on asking a series of questions (the answers are either True or False), until the
model gets confident enough to make a prediction. The end result is a tree, where
the leaf nodes are the decisions.
The questions asked at each node to determine the split are different for classifi-
cation and regression. For regression, the algorithm will first pick a value, and split
the data into two subsets. For each subset, it calculates the MSE (Mean Squared
Error). The tree chooses the value with the smallest MSE value. After training, the
algorithm runs it through the tree until it reaches a leaf node. The final prediction is
the average of the value of the dependent variable in that leaf node. The usage of a
single decision tree gave the worst results. While decision trees are supposed to be
robust against collinearity, they did not perform better than linear regression.
5.4 Random Forest Regression
Random forest is an ensemble method, which means that it combines predictions
from multiple machine learning algorithms, in this case, decision trees. The problem
with decision trees is that they are very sensitive to training data and carry a big
risk of overfitting. They also tend to find the local optima, as once they have made
a decision, they cannot go back. This was evident by the fact that the 2value and
correlation varied in each iteration of running the algorithm. Random forest contains
multiple decision trees running in parallel, and in the end, averages the results of
multiple predictions. Random forests with 100 trees were found to have the best
results among all algorithms used in this paper.
5.5 Ridge Regression
Ridge regression L2 uses regularization, which is a method used to avoid over-
fitting by penalizing high-valued regression coefficients through shrinkage, where
extreme values are shrunk toward a certain value. Particularly, in L2 regularization,
the coefficients are penalized toward the square of the magnitude of the coefficients.
Ridge regression is a technique used to mitigate multicollinearity in linear regres-
sion. Multicollinearity is where one feature in a multiple regression model can be
predicted from other features due to a positive linear relationship. This causes inac-
curate estimates of the regression coefficients. While the fulfillment of the multi-
collinearity assumption is not necessary for prediction and is only necessary for
inference, using Ridge Regression decreased performance. The cause of this was not
666 F. Razeen et al.
Fig. 2 Correlation heatmap of features indicating the presence of multicollinearity
clear. The features themselves were confirmed to contain multicollinearity, indicated
by the correlation heatmap which is shown in Fig. 2.
5.6 Lasso Regression
Similar to ridge regression, lasso regression shrinks all coefficients toward a value,
in this case, the absolute value of the magnitude of coefficients. This is called L1
regularization, and can sometimes lead to elimination of some coefficients. Lasso
regression had similar performance to Ridge regression.
Predicting Movie Success Using Regression Techniques 667
6 Results
6.1 Comparison of Models
Table 2compares the performance of different regression algorithms based on certain
metrics. The usage of a single decision tree was found to be the least effective algo-
rithm for this scenario. The best performing algorithm was Random Forest, while
Linear, Ridge, and Lasso fell shortly behind. While the former three algorithms
seem to be similarly performing in terms of correlation and 2, the error metrics are
much higher in Ridge and Lasso compared to Random Forest and Linear Regres-
sion. Overall, Random Forest is the most suitable algorithm, followed by Linear
Regression (Figs. 3and 4).
Tabl e 2 Aggregation of results
Regression model R2MAE RMSE Correlation
Linear 0.7299 0.0331 0.0509 0.8584
Support vector 0.6893 0.0408 0.2337 0.8532
Decision tree 0.6893 0.0408 0.2337 0.7626
Random forest 0.7356 0.0326 0.051 0.8941
Ridge 0.6893 0.0408 0.2337 0.8614
Lasso 0.6893 0.0408 0.2337 0.8614
Fig. 3 Comparison of correlation and R2between algorithms
668 F. Razeen et al.
Fig. 4 Comparison of error metrics between algorithms
6.2 The Most Important Features
To find out the most important features in determining the success of a movie, random
forest regressor was used. In each individual tree, the decision is made based on the
MSE (Mean Squared Error). When training individual trees, the degree of how each
feature decreases the MSE can be averaged. The features are then ranked accordingly
in ascending order. The following bar plot shows that vote_count and budget are the
most important in determining the revenue of a movie. Similarly, vote_count was
the best predictor in [1] and budget was the best predictor in [2]. This shows that
people’s opinion of a movie is more important in determining the success of a movie
than the budget it was allotted (Fig. 5).
The bar plot also shows that, out of the feature-engineered columns, the
belongs_to_collection column had more importance compared to the other engi-
neered columns. The other two columns (action_or_adventure and top_director) had
very less importance but still contributed to a slight increase performance, proven
when they were removed, and the performance metrics were measured again. Next
to vote_count, the budget of a movie is the most indicative of a movie’s success.
The next most important features are all social media type features. They tell us how
many people voted for the movie, how many people left a review for the movie, and
how many liked the movie on Facebook. Overall, it can be inferred that people’s
opinion of a movie matters most in determining a movie’s success. Furthermore, a
movie’s genre does not have any role in predicting the revenue, nor does the success
of the director.
Predicting Movie Success Using Regression Techniques 669
Fig. 5 Feature importances using random forest regressor
7 Conclusion and Future Work
The aim of this paper was to use regression techniques in order to predict movie
revenue using movie metadata. Compared to other papers that were reviewed, the 2
values were higher, except in one case [6], where Twitter data was used. While it is
well researched that using real-time trend-based social media data like Twitter tweets
can have excellent results, most other papers using movie metadata and non-hype
based social media data like data from Facebook had lower performing algorithms
when compared to the results found in this paper.
Error metrics could not be compared as other papers did not mention them.
The best algorithm used was Random Forest, which had an 2value of 0.89. Most
papers had implemented classification algorithms, which are unsuitable for predicting
numerical values, as they have a window within which the actual value lies. If the
size of this window is too big, results obtained may vary from actual results.
Overall, the methods presented in this paper show that movie revenue can in fact
be predicted successfully, although some might argue that the accuracy is not enough
for real life use. People’s opinion of a movie matters the most in determining the
success of the movie. The measure of opinion here is based on the number of likes,
votes, and reviews on social media websites.
For future work, an aggregation of trend-based social media data like Twitter,
social media data from Facebook, and data from IMDb could be used to improve the
prediction results vastly, to the point of deploying these algorithms for real-time use.
670 F. Razeen et al.
References
1. N. Quader, M. Gani, D. Chaki, M. Ali, A machine learning approach to predict movie box-
office success, in 2017 20th International Conference of Computer and Information Technology
(ICCIT) (2018) (online). https://doi.org/10.1109/iccitechn.2017.8281839
2. S. Yoo, R.K. Kanter, D.C. Cummings, A. Maas, Predicting Movie Revenue from IMDb Data
(2011)
3. N. Vr, M. Pranav, P.B.S. Babu, A. Lijiya, Predicting movie success based on IMDB data. Int.
J. Bus. Intell. 03, 365–368 (2014) (online). https://doi.org/10.20894/ijbi.105.003.002.004
4. T. Liu, X. Ding, Y. Chen, C. Haochen, M. Guo, Predicting movie Box-office revenues by
exploiting large-scale social media content. Multimed. Tools Appl. 75, 1–20 (2014) (online).
https://doi.org/10.1007/s11042-014-2270-1
5. C. Jernbäcker, P. Shahrivar, Predicting Movie Success Using Machine Learning Techniques
(Stockholm, Sweden, 2017)
6. S. Asur, B. Huberman, Predicting the future with social media, in Proceedings—2010
IEEE/WIC/ACM International Conference on Web Intelligence, WI 2010, vol. 1 (2010) (online).
https://doi.org/10.1109/WI-IAT.2010.63
7. H. Afzal, Prediction of movies popularity using machine learning techniques. Int. J. Comput.
Sci. Netw. Secur. 16, 127–131 (2016)
8. L. Pal, L. H. Bui, R. Mody, Predicting Box Office Success: Do Critical Reviews Really Matter?
9. M. Mestyán, T Yasseri, J. Kertész, Early prediction of movie box office success based on
Wikipedia activity big data. PLoS ONE 8, e71226 (2013) (online). https://doi.org/10.1371/jou
rnal.pone.0071226
10. R.L. Wasserstein, N.A. Lazar, The ASA statement on p-values: context, process, and purpose.
Am. Stat. 70(2), 129–133 (2016) (online). https://doi.org/10.1080/00031305.2016.1154108
11. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P.
Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher,
M. Perrot, E. Duchesnay, Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12,
2825–2830 (2011)
Vehicle Recognition Using CNN
V. K. Divyavarshini, Nithyasri Govind, Amrita Vasudevan,
G. Chamundeeswari, and S. Prasanna Bharathi
Abstract We are designing the software with the help of Convolution Neural
Network (CNN) with 25 layers of max pooling. The proposed work is achieved
by python or MATLAB with Support Vehicle Machine (SVM) with HOG. The SVM
is used first to recognize the vehicle and then CNN is used to split or classify them
into domains and keyframes of a specific video. This automation can be used in toll
gates, highways, heavy traffic surveillance, etc.
Keywords Convolution neural networks ·Max pooling ·Support vector machine ·
Deep learning ·Machine learning
1 Introduction
In modern years, we have observed fascinating growth in fields related to artifi-
cial intelligence like automatic vehicles, vision of computers, machine learning and
deep learning. As a newcomer, it is very difficult to remain updated due to the
drastic changes that occur in this field. Till date there are no journal and conference
V. K. Divyavarshini (B)·N. Govind ·A. Vasudevan (B)·S. Prasanna Bharathi
Department of ECE, SRM Institute of Science and Technology, Vaadapalani, Chennai 600026,
India
e-mail: divyavarshini13@gmail.com
A. Vasudevan
e-mail: vasudevanritu@gmail.com
N. Govind
e-mail: nithyasrigovind23@gmail.com
S. Prasanna Bharathi
e-mail: prasanns@srmist.edu.in
G. Chamundeeswari
Saveetha School of Engineering, SIMATS, Chennai, India
e-mail: easwarig@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_60
671
672 V. K. Divyavarshini et al.
papers which deal with generalized study of automatic vehicles using large amount
of datasets, but for specific problem related papers are available.
This project provides the software which can be used accurately to solve such
problems. Here as it turns out, deep neural networks are outperforming the approach.
Categorization of vehicles manually is strenuous; hence by automating this work will
enhance the efficiency to the user. So, for the proposed problem, we have designed a
system using machine learning by developing a software for identifying each vehicle
separately in an image which is captured in a high definition camera. In real time,
taking photographs are a tedious process, in order to overcome we are using a video
which splits into images which we normally call that as ‘keyframes’.
Each keyframes can be splitted from minutes to nanoseconds, for example, if we
captured a video of one minute of moving vehicles in a pedestrian, we can split the
image from 1 s to 1 ns. Here, we are using three main algorithms for developing
this software which is CNN ‘Convolution Neural Network’, SVM ‘Support Vector
Machine’, HOG ‘histogram of oriented gradients’. Earlier, the classification of vehi-
cles was done by observing and analyzing a single vehicle at a time. This in turn
reduces the accuracy of the classification.
Every year, new algorithms/models keep on outperforming the previous ones. The
latest state of the art software system for object detection was just released last week
by Facebook AI team. The software is called Detectron that incorporates numerous
research projects for object detection and is powered by the Caffe2 deep learning
framework. Now, this is made easier through data science and the people who work
with these tools are called data scientist. In order to figure out the complicated
analytical problems, data science can be used which is a unique combination of
both datasets and algorithms. Datasets is a collection of information that has been
accumulated for a very long time in a specific location; this information can be used
for research purposes.
This data can be much learned in the mining process [1]. We can build with the data
available through advanced algorithms. The main purpose of this project is to detect
and classify the objects in an image using image processing of machine learning
for example vehicles. The sole purpose of this project is to help the government in
regulating the traffic issues that is being by the commoners. This can be achieved
through MATLAB, though there are different tools available in the market. The most
popular tool used for image processing recognition is python 3.7.0 (latest version).
The method by which this classification has to be done is simplified in this project
to make it user friendly and easy for execution. Mainly, with the renaissance of deep
learning, convolutional neural networks have automated this task while significantly
boosting performance.
The basic requirements to develop this software are
Hardware:
•Laptop
•High definition camera.
Vehicle Recognition Using CNN 673
1.1 Software Used
MATLAB—The tool which we are executing is MATLAB. It merges a desktop
environment which is adaptive for the repetitive analysis and design processes with a
programming language that disclose matrix and array mathematics directly. Within
few lines of MATLAB code, we can frame deep learning models without being an
experienced person. MATLAB can run models up to 7×faster than tensor flow and
up to 4.5×faster than caffe2. It uses the functions and tools to envision the common
results and debug deep learning models.
1.2 Proposed Work
Here is a short introduction into traffic surveillance cameras for vehicle detection.
Vision-based algorithms for vehicular detection can be divided into three cate-
gories: movement-based approaches, hand-made practical approaches and CNN-
based approaches. Motional solutions include subtracting frames and subtracting
the history. Motion-based methods include subtraction of frames and subtraction of
history. Image subtraction computers for the identification of the movement target
the variations in two or three series images. Frame subtraction has a simple calcula-
tion and a dynamic background, but is not ideal to move too quickly or too slowly.
Handcrafted feature-based approaches include Histogram of Oriented Gradients.
CNN methods have shown a rich power of representation and achieved promising
results. CNN uses object proposal to prepare CNN for detection tasks by selective
searching. In the framework of CNN, ALEX-Net and Fast CNN speed up the gener-
ation of regional proposal on the map; computer only needs these approaches once.
Instead of selective search, Faster CNN uses the region proposal network, then it can
train end to end and improve speed and accuracy.
2 Algorithms
The breakthrough for this project is done by three main algorithms:
•CNN—Convolution Neural Networks
•SVM—Support Vector Machine
•HOG—Histogram of Oriented Gradients.
674 V. K. Divyavarshini et al.
2.1 Convolution Neural Network
The convolution neural networks have tiny fragments as their input images and these
are called s receptive fields (Fig. 1). This type of feature extraction was inspired
by the visual mechanism in living organisms, where cells in the visual cortex are
perceptive to small regions of the visual field. There are two specific types of layers
in CNNs: the convolution layer and the pooling layer. In the convolution layer, the
images are convoluted by divergent convolution filters via shifting the receptive field
gradually (Fig. 1). The convolution filters contribute the identical parameters in every
small portion of the image, largely condensing the representation of hyperparameters
in the model. The pooling layer, taking leverage of the ‘stationary’ property of the
images, takes the mean, the max or other data of the countenance at discrete positions
in the feature maps, thus reducing the variance and capturing necessary features. In
the above table, the convolution layer fields are represented by coloured blocks.
The input patch is expressed by the letter ‘a’, and it is actually multiplied with the
matrices of the pooling layer. The outcome of this matrix is represented by aij,cij and
kij at line I and column J. Convolution layers operate by passing on the input images
to the next layer by applying a convolution [2]. The function of the convolution is
to imitate a similar response of a respective neuron to a visual stimuli. Each field
is taken concentrated individually and the data from each neuron is processed. It is
not practical to use can architecture on images which learn the features as well as
classify the images which is done by a fully connected neural network.
As the number of neurons that are required is very high, even in a shallow archi-
tecture, because the pixels in a very large input image which is relevant are very high.
Fig. 1 CNN
Vehicle Recognition Using CNN 675
For example, a small image which is actually connected with the layers will even-
tually have a size of 100 ×100 and its weight will be almost equal to 1000 neurons
which are present in the second layer. A solution is put forward by the convolution
operation by decreasing the free parameters. This allows the network to function
deeper with the help of fewer parameters. For example, regardless of the size of the
image, tiling regions present in an image is of size 5 ×5. They share their weights
equally and the learnable parameters required are only. In this method, the problem
of vanishing or exploding of gradients is resolved by training traditional multi-layer
neural networks with many layers. This method is called sing back propagation.
3 Model-Based on Vehicle Recognition Tensor
3.1 Tensor Flow
Today, Google’s TensorFlow is the world’s best-known deep learning library. Google
products use machine learning to enhance search engine, translation, underline
picture and suggestions in all of their products.
Google users may experience a more rapid and streamlined search with AI to give
a specific example. Google makes a recommendation on what the next word might be
when the user types the keyword in a search bar. Google would like to use machine
education, so that users get the best experience from their massive datasets. Machine
learning is used by three different groups: researchers, data scientists, programmers.
They can all work together and improve their effectiveness using the same toolset.
Google has not only data; it’s the largest machine in the world, so Tensor Flow has
been designed for scale. TensorFlow is a Google brain group library built to expe-
dite machine learning and groundbreaking research on neural networks. It has been
designed for running on several CPUs or GPUs and handheld devices, and is bundled
in several languages such as Python, C++ and Java.
3.1.1 Pooling
A pooling layer is another section of Knits main action to successively decrease the
structural size of the representation which will thereby decrease the computation
and the number of parameters which is necessary in the network. The pooling layer
operates only on a single feature map independently [3]. Both global and local pooling
layers can be included in the convolutional networks, by this the outputs of a neuron
cluster are combined in one layer with a neuron in another layer. One example
is average pooling, which takes the average value from each of a cluster of neurons
at the prior layer. And another example is max pooling which makes use of the
676 V. K. Divyavarshini et al.
maximum value from each of a cluster of neurons at the prior layer. The most common
approach used in pooling is max pooling.
3.2 Support Vector Machine
In the support vector machine the machine learning has a very good role in classifying
the images into different domains. It is a classifier defined by a hyperplane. The
algorithm learns to classify the new images when it is trained by a labelled dataset.
For instance, if an image is allowed for classification, this algorithm will divide itself
into a set of hyperplanes or a single hyperplane in an infinite or high-dimensional
space, which is then used to perform various tasks like outliers detection.
The tuning parameters in SVM classifier are
•Kernel
•Regularization parameter
•Gamma
•Margin.
By varying these parameters, we can achieve considerable accuracy in reasonable
amount of time.
3.2.1 SVM Classifier
Figuring out the (soft-margin) SVM classifier in minimum expression is
1
n
n
i=L
max(0.1−yi(w,xi−b))+λω
The problem of quadratic programme involves by selecting small value of λyields
(Fig. 2) and that’s the reason behind the full focus to the classifier first.
Fig. 2 CNN flow diagram
Vehicle Recognition Using CNN 677
3.2.2 Kernel Classifier
The nonlinear distribution rule which coincides with linear distribution rule for the
transformed data point ϕ(xi)is given by kernel function kwhich satisfies k(xi,xi)=
ϕ(xi),ϕ
(xi).
The classification vector ¯ωin transformed space gives
¯ω=
n
i=L
Ciyiϕ(xi)
where the Ciis obtained by explaining the problem.
3.2.3 Linear SVM
The training dataset of n points are taken into consideration
(xi,yi),...,(xn,yn)
where yican be −1 or 1, that indicates the group the point where xibelongs. Every
ρdimensional real vector has xi. The ‘maximum-margin hyperplane’ has to be
calculated that splits the group of points xifor which yi=1 and from the group
yi=−1 is taken and the distance between the nearest point xiand the hyperplane
from any one of the group is zoomed.
3.2.4 Nonlinear Classification
In Fig. 3authentic maximum-margin hyperplane algorithm known as linear classifier
is shown in a graph with x axis as ‘x1’ and y axis as ‘x2’. It is distributed by employing
the kernel trick to maximum-margin hyperplanes. The emerging algorithm is gener-
ally similar, but the dot product is restored by a nonlinear kernel function. Hence for
a particular area with the help of the algorithm, the maximum margin of the hyper-
plane is estimated. Even though a hyperplane is the classifier and changed feature
area, it might be the nonlinear original input, but the changed area will have higher
dimensions. Engaging in higher dimensional area can be an advantage as it reduces
rationalization error in SVM without effecting the performance of the algorithm
(Fig 4).
678 V. K. Divyavarshini et al.
Fig. 3 Maximum-margin
hyperplane and margins for
an SVM trained with
samples from two classes.
Samples on the margin are
called the support vectors
Fig. 4 Nonlinear classifier
3.3 Histogram of Oriented Gradients
In image processing and computer vision, a component called Histogram of Oriented
Gradients (HOG) is used which is a feature descriptor used for detection of object.
The function of a feature descriptor is to delete the extraneous information and
obtain the relevant content by compressing the image or image patch and provide an
indication. The method calculates the localized occurrences of gradient orientation
in an image. This method is same as that of changes descriptors, and shape contexts,
but varies in the fact that it is evaluated on a thick grid which is of equally spaced
cells and makes use of overlapping [4] local contrast generalization to achieve better
accuracy.
To calculate this the following are used:
•Pre-processing
•Calculate the Gradient Images
•Calculation of Histogram of Gradients in 8 ×8 cells
Vehicle Recognition Using CNN 679
Fig. 5 Pre-processing
•16 ×16 Block Normalization
•Calculate the HOG feature vector.
3.3.1 Pre-processing
For pedestrian detection HOG feature descriptor is used and 64 ×128 image is
applied for the estimation (Fig. 5). The size of the image can vary. Generally, each
patch is monitored at different image locations at multiple scales. The only limitation
is that there are fixed aspect ratios for the patches that are being monitored. To explain
this point, a patch of size 100 ×200 is selected for determining our HOG feature
descriptor. The image is then cropped and resized to 64 ×128; the image is retrieved
from the patch. The gamma correction is also done as a pre-processing step, but the
results obtained are minimum thus we are avoiding this step.
3.3.2 Calculate the Gradient Images
The HOG descriptor can be calculated using the vertical and horizontal gradients.
Figure 6shows the gradient images. This is quickly accomplished with the filtered
images of the following kernels
g=g2
x+g2
y
θ=arctan gy
gx
The vertical lines represent the x-gradient and the horizontal for y-gradient lines.
When there is change in the intensity, the magnitude of the gradient changes. At
680 V. K. Divyavarshini et al.
Fig. 6 Gradient image
each pixel, the gradient has vector and magnitude. For coloured images, it helps to
identify the gradients of the three channels (as shown in Fig. 6). The magnitude of
the gradient depends upon the magnitude of each pixel inside the gradient.
3.3.3 Calculate Histogram of Gradients in 8 ×8 Cells
Here, the picture is segregated as 8 ×8 cells and the histogram is found in each cell.
The main reason to use the feature descriptor is to get a patch of image which in turn
gives a compact representation of the gradients. Each 8 ×8 image patch has 8 ×8×
3 which is equal to 192-pixel range.
The gradient contains vector and magnitude per pixel in which the dimensions
have 8 ×8×2 which is 128 in number, these 128 pixels of magnitude and vector
representation are denoted by 9-bin histogram gradient which is kept at an array of 9
numbers. Now, by determining a histogram in an 8 ×8 patch will make the gradient
into less noise. Each gradient will have some noise, but this type of representation
will make the gradient into less disturbance (Figs. 7and 8).
The histogram has 9 bins parallel to angles 0°, 20°, 40°, …, 160°. Now, the
gradients look like 8 ×8 in an image.
The image mentioned above as Fig. 8shows the gradient and the direction of the
gradient with its length showing the magnitude. The direction of arrows directs to
the change in intensity and the magnitude shows how big the direction change is.
From Fig. 8, we are able to observe some raw numbers displaying with one
minimum difference inside an 8 ×8 cells—the angles are between 0 and 180° than
0 and 360°.
They are called ‘unsigned’ gradients. Else, the same category is obtained by the
180 degree opposite and gradient arrow. Now, the final step is to generate 8 ×8 cells
of histogram of gradients. The histogram has 9 bins parallel to angles 0°, 20°, 40°,
… 160°.
3.3.4 16 ×16 Block Normalization
Previously, image was generated using histogram based on the gradient. The lighting
is the main part of the gradient orientation (Fig. 9). The darker image is segregated
by pixels first, all the pixel values are divided by 2 so that the gradient magnitude
Vehicle Recognition Using CNN 681
Fig. 7 Fully layered image
Fig. 8 8×8 cell image
682 V. K. Divyavarshini et al.
Fig. 9 16 ×16 block
normalization
becomes half. Now, automatically, histogram value becomes half. Basically, wedon’t
want any lightning differences in the captured image, to rectify that we can normalize
our descriptor of the histogram.
Colour vector is of RGB type [128, 64, 32]. L2 norm is the length of the vector.
By dividing each element by 146.64 and the resultant normalized vector [0.87, 0.43,
0.22]. Now, consider another vector which has twice the value than the first one. So,
Vehicle Recognition Using CNN 683
the equation becomes 2x [128, 64, 32] =[256, 128, 64]. The normalized value of
the second vector is [0.87, 0.43, 0.22], which is exactly same as the first vector.
3.3.5 To Calculate the HOG Feature Vector
For calculating the complete image patch, we have to multiply 36 ×1 vector with
one giant vector [5] which forms a final feature vector. And to calculate that
1. Totally, there are 105 positions which 7 in horizontal and 15 in vertical positions.
2. So, in each 16 ×16 block, we have 36 ×1 vector (Fig. 10) which internally calcu-
lated, we get 36 horizontals and 105 vertical vectors of totally 3780-dimensional
vector.
4 Data Science
To expand the software, the primary concept of data science strategies is needed. It is
known as ‘data product’. It is an advantage for every data scientist. More precisely, a
‘data product’ is a biggest credit which makes use of the input and processes each and
every data to get logically and algorithmically generated outputs. The most common
example is an engine, which analyze the user data and personalize itself based on the
data which has been received previously. There are examples of these data products:
•The recommendation engines of the Amazon website which shows the suggested
items in one corner of the page which the engine suggests based on the interest
of the user by some algorithms.
•The next is Netflix which suggests movies for us.
4.1 Data Scientist
The interest and the training of data is the capital contribution for the scientist. The
people who deal with data with science are known as ‘Data Scientist’. A common
personality character of data scientists, they are with high curiosity and very deep
thinkers. Being analytical is the part of Data science like exploring new things,
different methods of solution for a single problem, asking questions for every single
problem thrown at them. Data science is multidisciplinary.
684 V. K. Divyavarshini et al.
Fig. 10 36 ×1 patch image
4.2 Analytics
If a person thinks quantitatively, then that thinker is analytical thinker by nature.
Materialistically, ‘science of analysis’ defines analytics. A person who connects both
science and statistics together with some math, technology and business awareness
Vehicle Recognition Using CNN 685
is said to be Data Scientist. They basically work with raw data and analyze it and
finally build a product out of it. Though data scientist analyses the data but not as
deep as analyst. Their job is to interact with the data scientist and develop a new
database for further research purpose. They mainly try to get the insights of the raw
data which they acquire from data scientist.
5 Detailed Description Using a Flow Chart
Step 1: A road swamped with vehicles is selected and a video of few seconds
is recorded. The video can be recorded for about few nanoseconds or
milliseconds.
Step 2: Now the images in the video are splitted into few keyframes for few
seconds or milliseconds. As the number of keyframes increases the accuracy
increases.
Step 3: From the keyframes, the object is recognized as vehicle by using SVM then
the data is forwarded to CNN.
Step 4: The data forwarded to CNN was nothing but the different keyframes. These
keyframes are recognized as vehicle by SVM and categorized with the help
of CNN. The classification can be done based upon the user need.
Step 5: These keyframes which were recognized and categorized are now attached
back and the original video is retrieved back.
Step 6: Now the final video is played in which the vehicles are categorized properly.
The following task of training a software with an input data set as well as the
expected output data is a method that is followed in machine learning. In order to
complete this, the training process has to undergo various stages.
•Construction of software.
•Training the datasets.
•Testing the images collected from the datasets.
•Evaluation of the processed image.
686 V. K. Divyavarshini et al.
6 Training Model
Training data is provided by images of vehicles (Fig. 11). In order to detect a vehicle
on the image, we need to identify feature(s) which uniquely represent a car. We could
try using simple template matching or relaying on colour features but these methods
are not robust enough when it comes to changing perspectives and shapes of the
object. Here are two samples.
In order to have a robust feature set and increase our accuracy rate, we will be
using Histogram of Oriented Gradients (HOG). This feature description is much
more resilient to the dynamics of the traffic. MATLAB library provides us with the
necessary API for calculating HOG feature. Here Y Cr Cb colour space and all its
channels are used as inputs for HOG features extraction (Fig. 12). Here’s a sample
of vehicle and nonvehicle image with HOG features from the same images as above:
In order to detect the vehicle based on our feature set, we would need a prediction
model. Linear Support Vector Machines (Linear SVMs) is an administered learning
model that will be able to classify whether something is a car or not after we train it.
In order to improve the accuracy of the final output, we will be trying to find
multiple hits for the object of interest in the similar area. This approach is equivalent
Vehicle Recognition Using CNN 687
Fig. 11 Training model from the dataset
Fig. 12 Differentiation in HOG identified vehicle and nonvehicle
to creating a heat map. The next step is threshold which needs to be met in order for
a specific hit count from the heat map to be accepted as a detected vehicle (Fig. 13).
7 Testing Model and Output
Here, the CNN process is tested with a real time video in the swamped roads of
vehicles in the busy morning time. The file here is saved as ‘test.mp4’ (Fig. 14).
688 V. K. Divyavarshini et al.
Fig. 13 Stereotypic identification by any machine learning
Fig. 14 Output of test.mp4
The performance of the pipeline is not great and can be improved. Deep Neural
Network approach would have better performance numbers. Since this car detection
approach is based on camera it’s prone to usual challenges with this kind of sensor
(bad visibility, reflections, etc.).
8 Advantage and Disadvantage
The coding is less complicated when compared to other software. The software is
effectively used in areas such as toll gates in classifying vehicles without human
intervention.
It can also be used for statistical analysis of the vehicles.
The software is smart enough to find the vehicle in less distorted image or keyframe
of specific video. The CNN is basically slow in every software regardless of whether
its python or MATLAB. The training process is slow in CPU when compared to
GPU.
Vehicle Recognition Using CNN 689
9 Acknowledgement
This software can be effectively implemented not only in the field of automobiles
but also in various other fields like health care, entertainment, electronic, security,
merchandise, etc.
9.1 Smart Parking
This system can be used for smart vehicle parking where the number of vehicles
enters the parking plot. If the number of parking slots in a particular level is stored
then when the vehicles entered exceeds that number, an indication will be shown.
9.2 Health Care
This software has been able to provide a new method for generating genetic data. This
recent advancement in technology will open new doors in the field of biotechnology
and help us to solve the mysteries of human genetic structure. Genomic sequence
analysis structure can be achieved using deep learning algorithm [6].
•Colorization of b/w frames.
•Adding some background music to no sound movies.
•Classification of objects in a picture.
•Generation of caption in an image.
•Speech recognition like Google assistant and Amazon echo.
•Face recognition like Google photos, Microsoft how-old.
10 Conclusion
The vehicle categorization is a very promising access since it incorporates the
advantage of noticeable types of information. The efficiency that is acquired so
far in this project is 67 % (Fig. 15), but the efficiency can be further increased by
using more data sets and training the software with them.
Vehicles Expected results (%) Arrived result (%)
Car 100 60
Auto 100 10
Motorbike 100 30
690 V. K. Divyavarshini et al.
Fig. 15 Accuracy of the output
References
1. B. Selbes, M. Sert, Multimodal vehicle type classification using convolutionalneural network and
statistical representations of MFCC, in 2017 14th IEEE International Conference on Advanced
Video and Signal Based Surveillance (AVSS) (2017)
2. C. Cao, F. Liu, H. Tan, D. Song, W. Shu, W. Li, Y. Zhou, X. Bo, Z. Xie, Deep learning and its
applications in biomedicine. Genomics Proteomics Bioinform. (2018)
3. Proceedings of the International Conference on ISMAC in Computational Vision and Bio-
Engineering 2018 (ISMAC-CVB). (Springer Nature, 2019)
4. Z.-S. Huang, C.-C. Chuang, C.W. Tao, M.-Y. Hsieh, C.-X. Zhang, C.-W. Chang, IOS-based
people detection of multiobject detection system, in 2016 Joint 8th International Conference on
Soft Computing and Intelligent Systems (SCIS) and 17th International Symposium on Advanced
Intelligent Systems (ISIS) (2016)
5. D. Prasanna, M. Prabhakar, An efficient human tracking system using Haar-like and hog feature
extraction. Cluster Comput. (2018)
6. J. Ravikumar, A.C. Ramachandra, K.B. Raja, K.R. Venugopal, Convolution based face recog-
nition using DWT and HOG, in 2018 International Conference on Intelligent Informatics and
Biomedical Sciences (ICIIBMS) (2018)
GLCM and GLRLM Based Texture
Analysis: Application to Brain Cancer
Diagnosis Using Histopathology Images
Vaishali Durgamahanthi, J. Anita Christaline, and A. Shirly Edward
Abstract Objective This work focuses on design and development of an auto-
mated diagnostic system using brain cancer histopathology images. Background
Detection and classification of cancerous tissues are one of the challenging tasks in
histopathology image analysis. The features based on cell morphology, cell distribu-
tion, and randomness in growth and placements have been considered as important
aspects of the cancer diagnosis. Results of the conventional methods are subjective
and dependent on the skills of histopathologists. The computer-assisted diagnosis
(CAD) offers the diagnostic results based on standard algorithms and standard test
database. Material and Methods This work shows the contribution of gray-level co-
occurrence matrix (GLCM) based Haralick features and gray-level run length matrix
(GLRLM) features in analysis and classification of Brain histopathology images. The
primary focus of this study is to analyze the complex random field of cancer images
with co-occurrence matrix features, run length matrix features, and also the fusion of
both features. These features are used for classifying healthy and malignant tissue.
The SVM classifier with RBF kernel has been used for classification. The Confusion
matrix is formed showing true positive (TP), true negative (TN), false positive (FP),
false negative (FN) information. The classifier performance has been described with
sensitivity, specificity, precision, accuracy, and F-score measures. Results Outcome
of this work shows that GLCM-based and GLRLM-based features offer excellent
discriminating features for statistical study of histopathology images and can be
useful for cancer detection. Overall accuracy improvement is seen by fusing the
GLCM and GLRLM based texture features. Conclusion The work describes an inno-
vative way of using GLCM and GLRLM based textural features to extract underlying
information in brain cancer imagery.
V. Durgamahanthi (B)·J. Anita Christaline ·A. Shirly Edward
Department of Electronics & Communication Engineering, SRM University, Vadapalani,
Chennai, Tamil Nadu, India
e-mail: vaishalb@srmist.edu.in
J. Anita Christaline
e-mail: anitaj@srmist.edu.in
A. Shirly Edward
e-mail: edwards@srmist.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_61
691
692 V. Durgamahanthi et al.
Keywords Computer-assisted diagnostics (CAD) ·Gray-level co-occurrence
matrix (GLCM) ·Gray-level run length matrix (GLRLM) ·Support vector
machine (SVM)
1 Introduction
1.1 Background and Motivation
For many decades, researchers have been working to understand cancer, its causes,
and its various development stages so that they can offer the best solution and arrest
its development. The main cause of cancer is uncontrolled mitosis of human cells
caused by a mutation in a cell’s DNA. This reaction triggers production of more and
more damaged cells. The accumulation of these cells forms a malignant tumor in
later stages [1,2]. As the entire mitosis process has become exceptionally random,
cancer detection is a challenging task. Histopathology is the study of microscopic
cells and tissue structures, cells growth and its placement forming which is identified
as a perticular textured pattern in a diseased cases. The texture is a main feature
of cancer histopathological image [3]. According to WHO reports, 80% of biopsy
samples studied for cancer diagnosis have been found to be benign. The processing of
biopsy samples has become quicker and efficient with the development of whole slide
imaging (WSI) or digital pathology. The introduction of fast scanning devices and
the advancement in computer technology have replaced the subjective judgments
of a histologist with an accurate objective judgment with computers’ assistance.
The vital aim of cancer research is to study its spreading activity (metastasis) and
to design the best treatment to confine it to a small area of the human body. In
a computer-assisted diagnosis (CAD), researchers have developed many algorithms
for detection of symptoms as prognosis tool and detection of disease-specific features
as a diagnosis tool. Early detection of the disease leads to accurate treatment and
also increases survival rate.
1.2 Our Contribution
This research is aimed to study brain histology images and classify these images
into normal and malignant, based on textural attributes in cancer biopsy samples
[4,5]. Our aim is to use the Haralick texture features which are based on gray-
level co-occurrence matrix (GLCM) and gray-level run length matrix (GLRLM)
features to classify histopathology images into the normal and malignant sample
using support vector machine (SVM) classifier [5,6]. This work also describes the
fusion of extracted features to achieve better accuracy in diagnosis. The performance
of SVM classifier has been checked by describing confusion matrix. Comparison of
GLCM and GLRLM Based Texture Analysis … 693
GLCM and GLRLM based classification with model-based approach has also been
discussed.
The article is organized as follows. Section 2gives an overview of related work
done in the past. Section 3describes materials and methods; it describes basics of
GLCM and Haralick texture features and GLRLM features. The brief description
on TMAD is also given. It also describes classification with performance evaluation
measures. Section 4describes experimentation and results. In Sect. 5,wedrawa
conclusion.
2 Histopathology Image Analysis: An Overview
Advancement in imaging modalities and computer resources has opened the new
doors for the research to analyze disease-specific image features. Several researchers
in the field of clinical pathology and image processing have acknowledged the impor-
tance of digital pathology and histopathology images. The work has been performed
to improve algorithms involved in computer-assisted diagnoses (CAD), mainly image
enhancement, underlying information extraction in terms of model parameters, and
classification for disease prognosis and diagnosis [7–9]. Histopathology image anal-
ysis has been classified into two parts, object-level and spatial-level feature analysis.
Object-level features are further characterized into (i) Size and shape, (ii) Radiometric
and densitometric, (iii) Textural features, and (iv) Chromatin-specific features. In
these image analysis techniques, irregularities in cell distribution and morpholo-
gies have been used to detect the malignancy and also the level of malignancy
[10–13]. To avoid interobserver and interobserver variation in results, researchers
have suggested quantitative analysis of tissue and cell structures. Gunduz et al. have
considered cell graph and have suggested how to quantify organizational charac-
teristics using cell graph for cancer diagnosis. O. Sertel et al. have proposed the
method to grade tissue abnormalities based on the amount of cytological components
present in a tissue sample. He has performed texture analysis based on gray-level
co-occurrence matrix [14,15]. Our previous work describes the use of autoregres-
sive parameters for classification of brain cancer histopathology images considering
model-based image analysis [16,17]. The lymphocytic infiltration has been consid-
ered as a key component in cancer prognosis and diagnosis through which one can
study the interactions of the tumor with immune systems. This study has further
been used for setting therapeutic targets [18]. Multiresolution analysis in wavelet
sub-band has been proved as a revolutionary tool to analyze complex and hetero-
geneous random field of histology images. Many researchers have addressed the
heterogeneity problem of cancer tissue in wavelet transform domain [19,20]. The
selection of correct classifier is one of the key steps while handling histology image
database. Several researchers have worked on different classification approaches
to achieve better classification accuracy. Several machine learning methods have
been established for classification; they are neural network based multilayer percep-
tron (MLP), k-nearest neighborhood (KNN), a logistic regression method, fuzzy
694 V. Durgamahanthi et al.
systems, support vector machine (SVM), etc. Considering the amount of database in
pathology imagery, it is suggested by many researchers to use multiclassifier systems
for a better result [20–29]. This work has been focused on texture-based feature
analysis. The texture is a connected set of a pixel which shows repetitive occurrence.
Texture provides information about the intensities in the form of smoothness, coarse-
ness, and regularity. Textural feature analysis primarily considers GLCM features,
GLRLM features, fractal features, multiresolutional features in wavelet sub-bands,
entropy-based features, and chromatin-specific features. These features form a base
for analysis of textural region seen in cancer images [11,12]. Graph run length matrix
(GRLM) is the major tool to study tissue organization. The analysis of random field
image texture for diagnosis purpose has been presented by AkifBurak et al. [30].
They have proposed a new homogeneity measure to describe the tissue components.
The study of arrangement and cell structure through GRLM has been undertaken by
Kannan et al. [31] Morphological approach and GLCM approach in texture feature
analysis of histology images have been well described by Omar Al-kadi [32]. Even
though the researchers have been working with the histopathology image analysis
for past few decades, we found that there is a very limited work on brain cancer
histopathology. This work proposes a novel concept of using statistics-based texture
features to understand abnormalities in histopathology samples of brain cancer.
3 Material (Data) and Methods
3.1 Image Database
This work has been focused on brain tissues for brain cancer detection using
histopathology samples. The tissue database has been taken from the Stanford
tissue microarray database (TMAD) and tissue microarray virtual slides available on
Biomax.us. These are the public resources providing reference images for researchers
and histopathologists. This database consists of stained tissue images from Tissue
Microarray experiments as shown in Fig. 1[33,34]. The tissue samples have been
stained to enhance the visibility of the structures on the slide. The tissue sections have
been dyed with standard stains such as Hematoxylin and Eosin (H & E) and Immuno-
histochemical (IHC). H & E stained slides enhances the spatial nuclei features like
size, shape, texture, spatial arrangements, tubules, and stromal, etc. [35]. The color
related and intensity changes were found enhanced in IHC-stained tissue slides. [36]
Once the tissues are stained, the imaging of stained samples has been performed by
whole slide scanner with adjustable resolution based on histological and cytological
requirements. For the study of nuclei deformities in cytology, resolution is set to
400X, and for histological analysis, it is below 40X, as entire tissue structure has
been examined in histology image.
GLCM and GLRLM Based Texture Analysis … 695
Fig. 1 a Tissue microarray core. bTMA dataset (TMAD)
4 Methodology
This work has been carried out to quantify deformities in cell, regional lymph
node metastasis, distant metastasis, and growth factor for cancer diagnosis using
GLCM-based Haralick features and GLRLM-based textural features. It consists of
three major stages: image enhancement, feature extraction, and analyzing extracted
features using SVM classifier.
4.1 Image Enhancement
The preprocessing and segmentation are the two major stages in enhancement. To
reduce noise component arising from image acquisition and staining process, it is
necessary to preprocess an image prior to feature extraction. The noise removal and
improving overall quality of an image is the main objective of preprocessing. This step
also involves segmentation of key areas for analysis and diagnosis. Depending upon
cell-based and tissue-based studies, segmentation of cell has been performed from
a segmented tissue sample. Cell-based study considers region-based and boundary-
based segmentations. But when there is a need to consider global features alone, such
as first order, second-order statistical features like mean, variance, autocorrelation,
etc., segmentation is not required. Researchers have proved that calculating global
features can help to diagnose presence and absence of the disease [8].
4.2 Feature Extraction
In this step, the entire image has been represented by extracted features. The spatial
independency or spatial dependency can be seen depending on the cell-based features
696 V. Durgamahanthi et al.
or the tissue-based features, respectively [7]. Tissue texture determines variations in
tissue/cell organization. In cancer diagnosis, interpretations of these changes are
significant. The gray-level run length matrix (GLRLM), gray-level co-occurrence
matrix (GLCM), transform-based methods, and statistical model based study have
been considered as widely used methods for texture image analysis.
4.3 Haralick Feature
Textural features extracted in this work have been based on the hypothesis that textural
information on an image ‘I’ represents overall spatial association among the different
gray levels of the image. In other terms, texture information is effectively given by
a set of gray-level spatial co-occurrence matrix (GLCM) [5]. GLCM is computed
for various angular relationships and distances between neighboring pixels carrying
aspecific set of gray levels. So while constructing GLCM for texture representation,
there have been three main parameters: the quantification of the image gray levels,
the displacement, and angular orientation. Sample structure of GLCM is given in
Fig. 2where Nis the total number of gray levels existing in an image and probability
P(i,j) is the probability of finding a pixel with gray-level value ‘i’ together with a
pixel gray-level value ‘j’. An example of the gray-level image and corresponding
calculated GLCM is given in Fig. 3.
Haralick defined textural features using GLCM. These features are related to
specific textural property such as contrast, correlation, homogeneity, absence or pres-
ence of aspecific pattern in an image, and also quantify complexities and gray-level
transitions. Features have been defined from F1 to F14 and shown in APPENDIX
A. They are F1: Angular second moment (ASM): It represents energy contents in
the image. Indirectly, it gives the measure of uniformity. F2: Contrast: This is the
measure of local intensity variation in the histological sample. F3: Correlation: It
measures linear dependencies of gray levels of two neighboring pixels. In other
words, it is a measure of deformities in the textural image. F4: Variance: The vari-
ance is a measure of how far the gray levels have been spread from their mean value
Fig. 2 GLCM matrix format
GLCM and GLRLM Based Texture Analysis … 697
A
B
0 1 2 3
003 4 2
121 3 4
203 2 1
321 0 3
C
0 1 2 3 4
001030
110310
203011
331102
400120
Fig. 3 a Sample image. bIntensity values. cGLCM at angle 0o
in the given sample. It is one of the measures of the probability distribution. F5:
Inverse difference moment (IDM): It is the degree of homogeneity. In heterogeneous
image, this measure will be small. This is one of the important measures for cancer
diagnosis. F6, F7, and F8: Sum of Average, Sum of Variance, and Sum of Entropy
are the average, variance, and entropy of the normalized gray tone image in the
spatial domain. F9: Sum of Entropy is the measure of randomness in the image. F10:
Measure of redundancy or information lying in the image which is further used for
coding while compression. F11: Difference variance is another measure of random-
ness. F12, F13, and F14: a measure of the correlation in a more precise form [5]. This
work considers thirteen Haralick features as they can be made invariant to translations
and rotations. As mentioned above, they have the ability to understand something
instinctively in given images of biopsy samples, for example, coarse versus smooth,
directionality of the pattern, image complexity, etc. Gray-level run length matrix
(GLRLM), P(i,j/8) gives information about pixel gray level ‘i’ present ‘j’ times
consecutively in direction ‘8.’ GLRLM can be generated at 0°, 45°, 90°, and 135°
like GLCM. The rows of GLRLM show gray levels in the image, and the columns
signify length of runs, with the entries corresponding to the number of runs of the
given length in the given image. Thus, if we observed a large number of neighboring
pixels of the same gray level, the texture is a coarse texture, and if we find a small
number of neighboring pixels of the same gray level, the texture is a fine texture.
GLRLM calculated for ‘8’=0°isshowninFig.4. We have considered seven statis-
tical features derived by Galloway [37] and Chu et al. [38]. They are the short-run
emphasis (SRE), long-run emphasis (LRE), gray-level non-uniformity (GLN), run
length non-uniformity (RLN), run percentage (RP), low gray-level run emphasis
(LGLRE), and high gray-level run emphasis (HGLRE). The detailed definitions are
given in APPENDIX B.
4.3.1 Classification
In this phase, we allocate one among the two classes (Malignant, Normal) to the
sample after extracting GLCM and GLRLM based texture features. This work uses
support vector machine (SVM) classifier to separate malignant tissues from a healthy
698 V. Durgamahanthi et al.
A B
0 1 2 3
003 4 2
121 3 4
203 2 1
321 0 3
C
00Run
Lengths
Gray
Level
1 2 3 4
03000
13000
23000
34000
42000
Fig. 4 a Sample image. bIntensity values. cGLRLM at angle 0o
one. In support vector machine (SVM), several hyperplanes are constructed to sepa-
rate dataset into two classes. We need to find the optimal hyperplane which works
as the linear decision boundary with maximum margins between the vectors of two
classes. This margin has been optimally computed by critical components of the
training data called support vectors. This optimal hyperplane can be described as
WTx+b=0xεRd(1)
SVM classifier is able to classify miscellaneous data by varying kernel func-
tions. These kernels are linear, polynomial function, radial basis function (RBF),
and sigmoid functions. This work considers the RBF kernel function.
4.4 Testing Classifier Performance
A confusion matrix as shown in Fig. 5reveals the information about actual and
predicted outcomes. The performance of classifier is commonly evaluated using
the information described in this matrix. Sensitivity (Recall), Specificity, Accuracy,
(a) Definitions
Predicted
Negative Positive
Actual Negative TN FP
Positive FN TP
Predicted
Negative Positive
Actual Negative 1.00 0.00
Positive 0.14 0.86
(b)Confusion Matrix Standard form (c)Observed Confusion Matrix with a fusion of GLCM and
GLRLM Data Set
TP True Positive Sick person correctly identified as sick.
TN True Negative Healthy person correctly identified as healthy
FP False Positive Healthy person incorrectly identified as sick
FN False Negative Sick person incorrectly identified as healthy
Fig. 5 Confusion matrix
GLCM and GLRLM Based Texture Analysis … 699
Tabl e 1 Statistical measures Sensitivity/recall TP/(TP +FN)
Specificity TN/(TN +FP)
Accuracy (TN +TP)/(TN +TP +FN +FP)
Precision TP/(TP +FP)
F-Scores 2 * (Precision * Recall)/(Precision +Recall)
Precision, F-Score, etc., are most chosen statistical measures to evaluate the perfor-
mance of diagnostics test, and in this work, it is higher order statistical features
(GLCM and GLRLM features) with SVM Classifier. Each measure is used to describe
how good and reliable the test is. Sensitivity/Recall identifies how much worthy the
test is at detecting a positive disease, while specificity estimates how likely patients
without disease to be correctly ruled out. Accuracy gives how good the test is to
identify as well as ruling out disease. The disease prevalence is a statistical term
that depends on population of the disease and always affects the accuracy measure.
Precision or confidence is the measure of predicted positive cases that are truly posi-
tives. F-Score is another measure of tests accuracies based on precision and recall.
Sometimes, “F-score” is referred as balanced mean or weighted average mean. The
confusion matrix consists of terms such as true positive (TP), true negative (TN),
false positive (FP), and false negative (FN) which are associated with diagnostic tests
as shown in Fig. 5a, 5b, and 5c. The statistical measures calculated with confusion
matrix have been listed in Table 1. In this study, fusion of GLCM and GLRLM
features has been considered to evaluate the performance of SVM classifier.
5 Results and Discussions
5.1 Experimentation
In this study, we have considered 78 image samples of astrocytomas, 118 image
samples of glioblastomas, and 54 image samples of normal (54) tissue. This work has
been carried out in three phases: preprocessing, features extraction, and classification.
In the first part, we cropped the image of size 51 ×512 pixels before preprocessing.
After preprocessing, we have calculated thirteen Haralick Texture features based on
gray-level co-occurrence matrix and seven features based on gray-level run length
matrix. Table 2and Table 3show Haralick and GLRLM based features calculated
for malignant and healthy brain tissues. We have formed three feature datasets using
GLCM, GLRLM, and fusion of GLCM and GLRLM features. The SVM classifier
using the k-fold technique, with ‘k’ value set to 5 and number of iteration set to 100
has been considered for classification (Diagnosis). Initially, individual GLCM and
GLRLM feature sets have been considered for classification. In the next step, we
700 V. Durgamahanthi et al.
Tabl e 2 Haralick features for
normal and malignant
samples
Haralick texture feature Normal sample Malignant sample
F1: Angular second
moment
0.7594 0.11875
F2: Contrast 0.1385 0.63417
F3: Correlation 347.0757 1415.6
F4: Variance 35.2915 31.747
F5: Inverse difference
moment
0.9453 0.75925
F6: Sum of average 11.8893 11.119
F7: Sum of variance 128.2208 86.779
F8: Sum of entropy 0.5841 1.9885
F9: Entropy 0.6874 2.4946
F10: Difference variance 0.0984 0.048007
F11: Difference entropy 0.3681 0.87048
F12: Information measure
of correlation
– 0.2079 – 0.23777
F13: Information measure
of correlation
0.3840 0.69993
Tabl e 3 GLRLM features GLRLM texture feature Normal sample Malignant sample
SRE 0.7199 0.8268
LRE 4.0671 2.1502
GLN 2193.4157 1627.6164
RP 0.6149 0.7688
RLN 4840.3810 8034.9178
LGRE 0.0108 0.0170
HGRE 132.8836 92.9851
have considered the fusion of both the datasets to see the improvement in classifi-
cation accuracy. To test the performance of diagnostic system, i.e., SVM classifier,
we constructed the confusion matrix as shown in Fig. 5a. It consists of true detec-
tion cancerous patients(TP), true elimination of non cancerous patients (TN), and
false detection of non-cancerous patients (FP), false elimination cancerous patients
(FN) terms. Using confusion matrix, we have calculated statistical measures like
sensitivity, specificity, accuracy, precision, and F-score.
5.2 Results
In this proposed work on brain cancer diagnosis, in the first part, GLCM feature and
GLRLM feature datasets have been separately given to classifier, and classification
GLCM and GLRLM Based Texture Analysis … 701
Tabl e 4 SVM classifier Dataset % classification accuracy
GLCM 87.33
GLRLM 91.14
GLCM +GLRLM 93.00
Tabl e 5 Classifier
performance based on
confusion matrix
Test parameters GLCM +GLRLM AR (4) model features
TP 86.00 87.50
TN 01.00 81.25
FP 00.00 18.75
FN 14.00 13.50
Sensitivity/recall 100.00 87.50
Specificity 100.00 81.25
Accuracy 93.00 84.50
Precision 87.79 82.35
F-Score 92.66 84.84
accuracies have been calculated. In the next step, fusion of thirteen GLCM and seven
GLRLM features has been performed, and this fused dataset has been tested with
SVM classifier. As shown in Table 4, the classification accuracy with GLCM-based
Haralick features has been observed as 87.33% and with GLRLM as 91.14%. The
improvement in classification accuracy has been observed with the fusion of GLCM
and GLRLM features which shows 93.00% classification accuracy as described in
Table 4. As described in Table 5, the numerical value of sensitivity is 100%, which
represents the probability of identifying the person who is suffering from the disease;
similarly, specificity value is 100%, which represents ability of the classifier to detect
the person who is not suffering from the disease. The classification accuracy is 93%
which represents the overall effectiveness of the classifier for true positive results in
the total population. Similarly, F-score (92.66%) value reveals the harmonic mean
of precision and recall and gives a better evaluation of performance test.
5.3 Discussions
In brain cancer detection, this work reveals the use of GLCM-based Haralick Textural
features and GLRLM features. In textural feature analysis for cancer diagnosis,
several methods have been described opting for various methods [39]. We believe
that in the texture-based study of brain cancer histopathology imagery this work
presents a novel approach where the fusion of thirteen Haralick Textures features
and seven run length features have been considered for cancer diagnosis. As shown
in Table 2, Haralick features such as ASM, correlation, entropy, sum of entropy, a
702 V. Durgamahanthi et al.
difference of entropy, and information measures of correlation have proved to be
good biomarkers to significantly differentiate between malignant and normal tissue.
Similarly, as shown in Table 3, GLRLM features such as long run emphasis (LRE),
gray-level non-uniformity (GLN), run length non-uniformity (RLN), and high gray-
level run emphasis (HGLRE) terms have been proved as biomarkers. This work
also gives performance evaluation of SVM classifier with a fusion of GLCM and
GLRLM features. Table 5displays major statistical measures for checking classifier
performance with fusion of dataset. We have compared this result with model-based
study [17]. It shows that textural analysis of histopathology images with GLCM
and GLRLM features offers better results than fourth-order autoregressive model
based features. The performance evaluation of SVM classifier checks two major
requirements; first is the confirmation about the correct predictions of positive and
negative labels, and second is the ability of the classifier to avoid failures. There
are a few limitations to the proposed method. The issues related to heterogeneity
and long-range pixel dependency found in cancer pathology images have not been
resolved in this work. We propose some future tasks such as Markov random field
modeling (MRF) and radial basis function (RBF) modeling which can be considered
while interacting with nonlinear heterogeneous textures in histopathology imaging
[40]. Similarly, long-range dependency issues can be addressed if the same work is
attempted in transformed domain by taking Fourier or Wavelet transforms. There
have been three more performance measures which can be considered in evaluating
diagnostic tests performance; these are Youden’s Index (γ), Likelihoods (ρ+and
ρ−), and discriminant power (DP). The Youden’s index evaluates the ability of the
diagnostic test to avoid failure. While likelihood evaluates classifier performance on
the finer scale, high value of positive likelihood and low value of negative likelihood
represent better classification algorithm. The discriminant power (DP) is the power
of classifier to distinguish between positive and negative labels [41]. Analyzing the
diagnostic performance of classifier algorithms on above measures will be considered
as our future tasks.
6 Conclusion
This work presents a textural analysis of brain cancer histopathology images using
GLCM-based Haralick features and Gray-level run length matrix based features.
Survey of texture analysis for histopathology images and its applications in cancer
diagnostics has also been carried out. The estimated texture-based features are excep-
tional discriminating capacity that provides the basis for statistical analysis. These
features can be used to quantify and classify tissue deformations in cancer diag-
nostics. The work also shows the SVM classifier performance based on standard
statistical measures. It also reveals fusion of GLCM and GLRLM features forms
excellent dataset which provides better diagnosis. As a future work, one can extract
more complex features from histopathology image using transform domain approach.
Similarly, one can also consider model-based approach for analyzing heterogeneity,
GLCM and GLRLM Based Texture Analysis … 703
and generate vector data of textural features so as to capture more deformities in
architecture and hence key regions in histology samples.
Acknowledgements The authors would like to thank S.R.M. University, Vadapalani, Chennai for
their continued support and encouragement during this research work.
Conflicts of Interest The authors declare that there are no conflicts of interest regarding the
publication of this article.
Appendix 1: GLCM Haralick Features
Haralick texture feature Standard formulae
F1: Angular second moment F1 =Ng−1
i−01Ng−1
j−0{p(i,j)}2
F2: Contrast F2 =
Ng−1
n=0n∗nNg−1
i−01Ng−1
j−0{p(i,j)}{
i−j}=n
F3: Correlation F3 =
Ng
i=11Ng
j=1{(i−μx)( j−μy)p(i,j)}/σ xσy
F4: Variance F4 =Ng
i=11Ng
j=1(i−μ)2p(i,j)
F5: Inverse difference moment F5 =Ng
i=11Ng
j=1{p(i,j)}/1+(i−j)2
F6: Sum of AVERAGE F6 =2Ng
k=2kpx+y(k)
F7: Sum of Variance F7 =2Ng
k=2(k−F6)2px+y(k)
F8: Sum of entropy F8 =2Ng
k=2px+y(k)logpx+y(k)
F9: Entropy F9 =Ng
i=11Ng
j=1{p(i,j)}log{p(i,j)}
F10: Difference variance F10 =
Ng−1
k=0k−Ng−1
l=0|px−y(l)lpx−y(l)2
px−y(k)
F11: Difference entropy F11 =Ng−1
l=0px−y(k)logpx−y(k)
F12: Information measure of correlation F12 =F9−HXY!
max(HX,HY)
F13: Information measure of correlation F13 =(1−exp(−2(HXY2 −F9)))1/2
F14: Maximal correlation coefficient F14 =Q(i,j)=k1p(i,k)p(j,k)/pi(i)pk(k)
704 V. Durgamahanthi et al.
Appendix 2: GLRLM Features
GLRLM features Standard formulae
SRE SRE =1
nrM
i=11N
j=1{p(i,j)}1
j2=1
nrN
j=1{p(j)}1
j2
LRE LRE =1
nrM
i=11N
j=1{p(i,j)}∗j2=1
nrM
j=1{p(j)}∗j2
GLN GLN =1
nrM
i=11N
j=1{p(i,j)}∗2=1
nrM
j=1pg(j)∗2
RLN RLN =1
nrN
i=11M
j=1{p(i,j)}∗2=1
nrN
j=1{pr(i)}∗2
RP RP =nr
np
LGLRE LGRE =1
nrM
i=11N
j=1{p(i,j)}∗1
j2=1
nrN
j=1pg(i)∗1
j2
HGLRE HGRE =1
nrM
i=11N
j=1{p(i,j)}∗j2=1
nrN
j=1pg(j)∗j2
References
1. D.C. Cires
,an, A. Giusti, L.M. Gambardella, J. Schmidhuber, Mitosis detection in breast cancer
histology images with deep neural networks, medical image computing and computer-assisted
intervention, in MICCAI 2013. 16th International Conference (Nagoya, Japan)
2. Glioblastomas and Malignant Astrocytomas, American Brain Tumor Association, Publications
and Services (2014)
3. C.H. Chen, L.F. Pau, P.S.P. Wang, The Handbook of Pattern Recognition and Computer Vision,
2nd edn. (World Scientific Publishing Co., 1998), pp. 207–248
4. R. Chellappa, Stochastic models in image analysis and processing, Ph.D. thesis, Purdue
University, 1981
5. R.M. Haralick, K. Shanmugam, I. Dinstein Texture features for image classification. IEEE
Trans. Syst. Man Cybern. SMC-3(6), 610–621
6. A. Baraldi, F. Parmiggiani, An investigation of textural characteristics associated with gray
level co-occurrence matrix statistical parameters. IEEE Trans. Geosci. Remote Sens. 33(2),
293–304 (1995)
7. C. Demir, B. Yener, Automated cancer diagnosis based on histopathological images: a system-
atic survey, Technical Report, Rensselaer Polytechnic Institute, Department of Computer
Science (TR-05–09) (2009)
8. M.N. Gurcan, L. E. Boucheron, A. Madabhushi, N.M. Rajpoot, B. Yener, Histopathological
image analysis: a review. IEEE Rev. Biomed. Eng. 2, 147–171 (2009)
9. S. Bhattacherjee, J. Mukherjee, S. Nag, I.K. Maitra, S.K. Bandopadhyay, Review on
histopathological slide analysis using digital microscopy. Int. J. Adv. Sci. Technol. 62, 65–96
(2014)
10. O. Sertel, J. Kong, U. Catayurek, G. Lozansski, J. Saltz, M. Gurcan, Histopathological image
analysis using a model based intermediate representation of color texture, follicular lymphoma
grading. J. Sig. Process. Syst. 55, 169–183 (2009)
11. L.E. Boucheron, Object and spatial level quantitive analysis of multispectral histopathology
images for detection and characterization of cancer. Ph.D. thesis, 2008
12. J.C. Caicedo, Features for histology images. Image Segmentation, pp. 1–9 (2009)
GLCM and GLRLM Based Texture Analysis … 705
13. A.D. Belsare, P.O. Brown, C.A. Ball, M.M. Mushrif, Histopathological image analysis using
image processing techniques: an overview. Sig. Image Process. Int. J. (SIPIJ) 3(4), 22–31
(2012)
14. M.K.B. Sims, A. Murray, Image analysis can be used to detect spatial changes in the
histopathology of pancreatic tumors. Phys. Mach. Biol. 48, 183–191 (2003)
15. C. Gunduz, B. Yener, S. Gultekin, The cell graphs of cancer. Bioinformatics 20, 145–151 (2004)
16. D. Vaishali, R. Ramesh, J. Anita Christine, Histopathology image analysis and classification
for cancer detection using 2D Autoregressive model. Int. Rev. Comput. Softw. 10(2), 182–188
(2015)
17. D. Vaishali, R. Ramesh, J. Anita Christaline, Autoregressive modeling: application to mitosis
detection in brain cancer histopathology. Int. J. Biomed. Eng. Technol. 20(2), 179–194 (2016)
18. A.N. Basavanthy, S. Ganesan, S. Agner, J.P. Monaco, M.D. Feldman, J.E. Tomaszewski, G.
Bhanot, A. Madabhushi, Computerized image- based detection and grading of lymphocytic
infiltration in breast cancer histopathology. IEEE Trans. Biomed. Eng. 57(3), 642–653 (2010)
19. S.G. Mallat, A theory for multiresolution signal decomposition: the wavelet representation.
IEEE Trans. Pattern Anal. 11, 674–693 (1989)
20. B. Weyn, G. Van de Wouwer, A. Van Daele, P. Scheunders, D. Van Dyck, E. Van Marck,
W. Jacob, Automated breast tumor diagnosis and grading based on wavelet chromatin texture
description. Cytometry 33, 32–40 (1998)
21. M.N. Gurcan, T. Pan, H. Shimada, J. Saltz, Image analysis for neuroblastoma classification:
segmentation of cell nuclei. In: Proceedings of the 28th IEEE EMBS Annual International
Conference New York City, USA, 30 Aug–Sept 2006
22. P. Spyridonos, P. Ravazoula, D. Cavouras, K. Berberidis, G. Nikiforidis, Computer-based
grading of haematoxylin-eosin stained tissue sections of urinary bladder carcinomas. Med.
Inform. Internet Med. 26, 179–190 (2001)
23. A.N. Esgiar, R.N.G. Naguib, B.S. Sharif, M.K. Bennett, A. Murray, Microscopic image anal-
ysis for quantitative measurement and feature identification of normal and cancerous colonic
mucosa. IEEE Trans. Inf. Technol. B. 2, 197–203 (1998)
24. P.W. Hamilton, P.H. Bartels, D. Thompson, N.H. Anderson, R. Montironi, Automated location
of dysplastic fields in colorectal histology using texture image analysis. J. Pathol. 182, 68–75
(1997)
25. M. Wiltgen, A. Gerger,J. Smolle, Tissue counter analysis of benign common nevi and malignant
melanoma. Int. J. Med. Inform. 69, 17–28 (2003)
26. S. Rajeev Srivastava, S. Sharma, S.K. Singh, Design, analysis and classifier evaluation for
CAD tool for breast cancer detection from digital mammograms. Int. J. Biomed. Eng. Technol.
13(1), 270–300 (2013)
27. M. Muthu Rama Krishnan, M. Pal, S. Bomminayuni, C. Chakraborty, R.R. Paul, J. Chatterjee,
A. Roy, Automated classification of cells in the sub-epithelial connective tissue of oral sub-
mucous fibrosis—an SVM-based approach. J. Comput. Biol. Med. 39, 1096–1104 (2009)
28. S. Doyle, M. Feldman, J. Tomaszewski, A. Madabhushi, A Boosted Bayesian Multi-Resolution
Classifier for prostate cancer Detection from digitized needle biopsies. Trans. Biomed. Eng.
59(5), 2005–2018 (2005)
29. M. Murat Dundar, S. Badve, G. Bilgin, V. Raykar, R. Jain, O. Sertel, M.N. Gurcan, Comput-
erized classification of intraductal breast lesions using histopathological images. IEEE Trans.
Biomed. Eng. 58(7), 1977–1984 (2011)
30. A. Baurak, M. Kanemir, C. Sokmensure, C. Demir, Object oriented texture analysis for unsu-
pervised segmentation of biopsy images for cancer detection. Pattern Recogn. 42, 1104–1112
(2009)
31. K.P. Kannan, A. Ananthakumarari, Texture analysis of histopathology images to identify
anomalous region. Int. J. Manag. IT Eng. 2(8), 1–10 (2012)
32. O. Al-Kadi, Texture measures combination for improved meningioma classification of
histopathology images. J. Pattern Recogn. 43, 2043–2053 (2010)
33. Stanford Tissue Microarray Database. https://tma.im/
34. www.biomax.us/tissue-arrays/Brain/
706 V. Durgamahanthi et al.
35. M. Macenko, M. Niethammer, J.S. Marron, D. Borland, J.T. Woosley, X. Guan, C. Schmitt,
N.E. Thomas, A method for normalizing histology slides for quantitative analyses. In: IEEE
International Symposium on Biomedical Image Processing (2009), pp. 1107–1110
36. R.A. Walker, Quantification of immunohistochemistry Issues concerning methods, utility and
semiquantitative assessment. Histopathology 49(4), 406–410 (2006)
37. M.M. Galloway, Texture analysis using Gray level run lengths. Comput. Graph. Image Process.
4, 172–179 (1975)
38. A. Chu, C.M. Sega, J.F. Greenleaf, Use of gray value distribution of run lengths for texture
analysis. Pattern Recogn. Lett. 11, 415–420 (1990)
39. D. Yan, P. Wang, M. Linden, B. Knudsen, T. Randolph, Statistical methods for tissue array
images-algorithmic scoring and co-training. Anal. Appl. Stat. 6(3), 1280–1305 (2012)
40. H. Irshad, A. Veillard, L. Roux, D. Racoceanu, Methods for nuclei detection, segmenta-
tion, and classification in digital histopathology: a review—current status and future potential
methodological review. IEEE Rev. Biomed. Eng. 7, 97–114 (2014)
41. M. Sokolova, N. Japkowicz, S. Szakowicz, Beyond accuracy, F-Score and ROC: a family
of discriminant measures for performance evaluation, in AI2006: Advances in Artificial
Intelligence (2006)
Resource Management in Wireless IoT
Using Gray Wolf Optimisation
Framework
S. Karthick and N. Gomathi
Abstract With advancement in Wireless Internet of Things (IoT), the connectivity
between the devices increases, and hence the effective resource management is
considered challenging. In this paper, we use a scheduler that adopts Gray Wolf
Optimization (GWO) model to learn the allocation of resources flexibly to the users.
This model acts as an adequate model to allocate resources on large-scale systems
and it does not avoids the setbacks like slow convergence, analytical complexity, and
increase overhead due to multiple IoT inputs. The comparison between the proposed
and existing resource management techniques shows that the GWO offers improved
resources management efficiency than the existing methods.
Keywords Gray wolf optimization ·Resource management ·Wireless IoT
1 Introduction
In the light of the various applications and the advances throughout wired networking
technology [1], interest for the Internet of Things (IoT) infrastructure has increased
rapidly. In the Internet of Things, the word stuff is a piece of equipment that has
the capacity to monitor, run, store, or cycle. Such apps have unique features, i.e.,
limited memory, lower battery life, and weak processing power [2]. The IoT has
a strong potential to influence the course of our lives. The IoT provides various
applications from home automation to healthcare systems, which allow cost-effective
communication between objects and devices [3] to improve industry and society.
Therefore, nearly 50 billion IoT devices are predicted to operate by 2050. Various
S. Karthick (B)
Research Scholar, Department of Computer Science and Engineering, Vel Tech Rangarajan Dr.
Sagunthala R&D Institute of Science and Technology, Avadi, Chennai, India
e-mail: karthick.usilai@gmail.com
N. Gomathi
Department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D
Institute of Science and Technology, Avadi, Chennai, India
e-mail: gomathin@veltechuniv.edu.in
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_62
707
708 S. Karthick and N. Gomathi
systems have been built to support IoT sensors or actuators leading to the proliferation
of IoT networks over the last decade. Likewise, a range of OSs have been established
slowly for running these tiny sensors [4].
IoT systems have minimal capacity and power and in some cases require real-time
capability. In addition, heterogeneous equipment and appropriate networking and
protection frameworks should be enabled. Efficient communication and maintenance
of this large number of devices is one of the research community’s major design
priorities. In an IoT framework, the fundamental problem of science is the organized
and regulated use of available resources. An IoT Resource Management Mechanism
is essentially configured to effectively meet IoT system specifications [5,6].
A variety of strategies for resource availability [7–11], such as a limited resource
base, and dispersed and complex ecosystems, are being implemented to meet the
challenges raised by the IoT lower end devices [12]. The main responsibility for the
effective operation of the computer is an IoT operating system. Various OSs also
provided various solutions to satisfy the resource requirements of low-end devices.
Various systems are provided by the various OSs to ensure that the sensing nodes
are working properly.
Among several OSs suggested for the use in a resource-starved network of low-
end devices, low-end IoT systems typically run on a limited energy tank. It is there-
fore extremely important to provide an energy-efficient OS [13–15]. Such low-end
machines relay sensory data using a protocol for connectivity. The messaging proto-
cols will minimize as much energy as possible in order to be energy-efficient. Energy
efficiency is required for transport layer protocols, MAC-layer, and network layer
[16].
For their sensing processes, IoT devices require computing capabilities. An OS
needs an efficient method and resource control system to accommodate the restricted
processing capacity and memory. IoT systems are powered by batteries and are
primarily used in remote areas. Therefore, energy management is very critical for an
OS. The primary goal of an IoT device is to provide sensing activity for the further
handling of sensed data at the base station. Efficient resources and connectivity must
also be used with respect to wireless architecture, signal processing, reception of
data, transmitting, and radio sleep/wake cycle. Load, index and file device data from
IoT OS. Therefore, in IoT OSs, it is very desirable to provide an effective, robust,
and adequate file system.
The main aim of the study is to design a scheduler that adopts GWO model to
learn the allocation of resources flexibly to the users. This model acts as an adequate
model to allocate resources on large-scale systems and it does not avoid the setbacks
like slow convergence, analytical complexity, and increase overhead due to multiple
IoT inputs.
Resource Management in Wireless IoT Using Gray Wolf … 709
2 Proposed Method
GWO is a system of swarm intelligence and a social intelligence with gray wolves,
hunting, and leadership. There is a common social hierarchy in every group of gray
wolves which control supremacy and strength. The order of gray wolves is shown in
Fig. 1.
Wolves kills and consumes all the pack, is the mighty dog. Alpha, which leads
the whole group to pursue, run, and eat, is the toughest dog. When the wolves are
not in the pack, ill, or injured the best beta wolf follows.
Delta and omega are stronger than beta and alpha and are less dominant. As Fig. 2
demonstrates the GWO algorithm is based mainly on social understanding. The
search for gray wolves is another motivation. A number of effective steps are taken
by gray wolves to target, encircle, threaten, and strike. This helps you to monitor
major bearings. The logical measures are as follows.
Mainly the alpha wants to eat, sleep, wake up, etc. The alpha’s choices are based on
the pack. Yet also there is a certain egalitarian mentality, where the other wolves are
led by an alpha. During sessions, the whole pack recognizes the leader by protecting
the legs. The Alpha Werewolf, as the pack will follow instructions, is also recog-
nized as the alpha werewolf. The pack will matte only the alpha wolves. To notice,
alpha is not necessarily the most important part of the package but the best part
of the product administration. This showed that a kit is much more essential for
organization, consistency, and power.
Beta is the second stage in the system of gray wolves. The betas are wolves who
help the alpha to determine or box otherwise. The wolf of the beta can be male or
female. If one of the wolves is dead or very aged, it is probably the best alpha choice.
The beta wolf would adhere to alpha, but the other lower wolves should also be in
charge. He is the role of an alpha counselor and a discipliner in the package. Beta
enhances the entire package of alpha commands.
Fig. 1 Social hierarchy of
wolves
710 S. Karthick and N. Gomathi
Fig. 2 Architecture of the proposed system
In the lowest ranking, the Gray Wolf is alpha. The omega is job of each wolf
and always has to sell all other powerful wolves. It’s the last wolf to hunt. The
omega may not be an important person, but if the omega is misplaced, it experiences
internal difficulties. That is because the omega of aggression and anger resides among
all dogs. The whole set is completed and the domination system is maintained. The
omega is in some instances also the babies in the bag.
It is considered a subordinate if an alpha, beta, or omega is not a bear. Alphas
and betas are to show the delta, but they overpower the omega. Scouts will track
the boundaries of the territories and alert the pack if there is a danger. It protects
and guarantees the health of the packaging. The elders are the alpha or beta dogs.
Hunter’s tends to help both alphas and betas while catching predators and feeding
Resource Management in Wireless IoT Using Gray Wolf … 711
the pouch. Ultimately, it is the duty of the guardians to take care of the poor, sick,
and wounded wolves in the bunch.
The algorithm is given below and the flowchart is given in Fig. 2.
Step 1. The social hierarchy helps GWO to save the best solutions it has achieved
so far over several iterations.
Step 2. The circular mechanism defines a circular district around solutions that can
be extended as a hypersphere to higher dimensions.
Step 3. The random parameters Aand Chelp candidate solutions to find the optimal
resources for the IoT device to transmit the information with different
random radius in hypersphere.
Step 4. The hunting method proposed enables candidates to find the likely position
of the prey.
Step 5. The adaptive values of aand Aare ensured for exploration and exploitation.
Step 6. GWO can smoothly transfer between exploration and exploitation with the
adaptive values of parameters aand A.
Step 7. With decreasing A, half of the iterations are used for the exploration (|A|
≥1) and remaining for exploitation (|A|<1).
Step 8. The main parameters of GWO are adjusted finally, i.e., aand C.
3 Results and Discussions
In this section, the proposed method is compared with Radio Resource Allocation
(RRA) method in [17] in terms of energy consumption and delay. The result shows
that the proposed method achieves reduced energy consumption in both uplink and
downlink for the data packet size of 100 and 32 bytes transmitted over the IoT
network, which can be evident from Fig. 3.InFig.4, the result shows that the
10
9
8
7
6
5
4
3
2
1
0
GWO - Downlink
RRA [13] - Downlink
GWO - Uplink
RRA [13] - Uplink
100 bytes 32 bytes
Data packet Size (bytes)
Energy Consumption
Fig. 3 Energy consumption (µJ)
712 S. Karthick and N. Gomathi
600
500
GWO - Downlink
RRA [13] - Downlink
GWO - Uplink
RRA [13] - Uplink
400
300
200
100
0100 bytes 32 bytes
Data packet Size (bytes)
Delay (ms)
Fig. 4 Delay (ms)
proposed method achieves reduced delay in both uplink and downlink for the data
packet size of 100 and 32 bytes transmitted over the IoT network.
4 Conclusions
In this paper, GWO model is adopted in WirelessIoT environment to learn the
resources allocation flexibly to the IoT devices. The GWO model allocates well the
required resources, avoids slow convergence, analytical complexity, and increases
overhead due to multiple IoT inputs. The results show that the proposed GWO offers
improved resources management efficiency in terms of increased throughput and
reduced power consumption than the existing methods.
References
1. M. Weiser, The computer for the 21st century. IEEE Pervasive Comput. 1(1), 19–25 (2002)
2. D. Miorandi, S. Sicari, F. De Pellegrini, I. Chlamtac, Internet of things: vision, applications
and research challenges. Ad Hoc Netw. 10(7), 1497–1516 (2012)
3. I. Yaqoob, E. Ahmed, I.A.T. Hashem, A.I.A. Ahmed, A. Gani, M. Imran, M. Guizani, Internet
of things architecture: recent advances, taxonomy, requirements, and open challenges. IEEE
Wirel. Commun. 24(3), 10–16 (2017)
4. O. Hahm, E. Baccelli, H. Petersen, N. Tsiftes, Operating systems for low-end devices in the
internet of things: a survey. IEEE Internet Things J. 3(5), 720–734 (2015)
5. E. Baccelli, O. Hahm, M. Günes, M. Wählisch, T. Schmidt, OS for the IoT-goals, challenges,
and solutions (2013)
6. N. Yuvaraj, R. Raja, C. Dhas, Analysis on improving the response time with PIDSARSA-RAL
in clowdflows mining platform. EAI Endorsed Trans. Energy Web 5(20) (2018)
Resource Management in Wireless IoT Using Gray Wolf … 713
7. J. Santos, T. Wauters, B. Volckaert, F. De Turck, Resource provisioning for IoT applica-
tion services in smart cities, in 2017 13th International Conference on Network and Service
Management (CNSM) (IEEE, 2017), pp. 1–9
8. M. Aazam, E.N. Huh, Fog computing micro datacenter based dynamic resource estimation and
pricing model for IoT, in 2015 IEEE 29th International Conference on Advanced Information
Networking and Applications (IEEE, 2015), pp. 687–694
9. A. Azari, G. Miao, Fundamental tradeoffs in resource provisioning for IoT services over cellular
networks, in 2017 IEEE International Conference on Communications (ICC) (IEEE, 2017),
pp. 1–7
10. W. Li, K. Liao, Q. He, Y. Xia, Performance-aware cost-effective resource provisioning for
future grid IoT-cloud system. J. Energy Eng. 145(5), 04019016 (2019)
11. S. Nastic, S. Sehic, D.H. Le, H.L. Truong, S. Dustdar, Provisioning software-defined IoT cloud
systems, in 2014 International Conference on Future Internet of Things and Cloud (IEEE,
2014), pp. 288–295
12. M. Sivaram, A.S. Mohammed, D. Yuvaraj, V. Porkodi, V. Manikandan, N. Yuvaraj, Advanced
expert system using particle swarm optimization based adaptive network based fuzzy inference
system to diagnose the physical constitution of human body, in International Conference on
Emerging Technologies in Computer Engineering (Springer, Singapore, 2019), pp. 349–362
13. M. Watfa, M. Moubarak, A. Kashani, Operating system designs in future wireless sensor
networks. J. Netw. 10(1), 1201–1214 (2010)
14. N. Yuvaraj, P. Vivekanandan, An efficient SVM based tumor classification with symmetry non-
negative matrix factorization using gene expression data, in 2013 International Conference on
Information Communication and Embedded Systems (ICICES) (IEEE, 2013), pp. 761–768
15. M. Sivaram, V. Porkodi, A.S. Mohammed, V. Manikandan, N. Yuvaraj, Retransmission
DBTMA protocol with fast retransmission strategy to improve the performance of MANETs.
IEEE Access 7, 85098–85109 (2019)
16. N.A. Pantazis, S.A. Nikolidakis, D.D. Vergados, Energy-efficient routing protocols in wireless
sensor networks: a survey. IEEE Commun. Surv. Tutorials 15(2), 551–591 (2012)
17. H. Malik, H. Pervaiz, M.M. Alam, Y. Le Moullec, A. Kuusik, M.A. Imran, Radio resource
management scheme in NB-IoT systems. IEEE Access 6, 15051–15064 (2018)
Less Polluted Flue Gases Obtained
with Green Technology During Precious
Metals Recovery from Unwanted
and Discarded Electrical and Electronics
Components
Rajendra Prasad Mahapatra, Satya Sai Srikant, Raghupatruni Bhima Rao,
and Bijayananda Mohanty
Abstract This paper deals with less polluted flue gas recovery with well-known
green technology heating process during precious and valuable metals recovery from
unwanted and discarded electrical and electronic components. This unwanted and
discarded electrical and electronic component is also said to be E-waste. During
recovery of valuable and precious metals like gold, silver, copper, etc. after e-waste
treatment with various methods, the researchers were unable to find the challenges
about the treatment of released toxic and poisonous flue gases as such gases severely
pollute the environment. This paper presents an innovative, clean as well as green
method for flue gas treatment while recovering the valuable and precious metals from
unwanted and discarded electrical and electronic components. A well known clean,
green as well as eco-friendly heating process, i.e. microwave heating was applied for
flue gas treatment. The results show that the treated flue gas not only minimizes toxic
effect but also helps environment to get less polluted. Also the obtained minimized
flue gases from the clean and green process were observed within the regulatory
range of Central Pollution Control Board.
Keywords E-waste ·Microwave heat treatment ·Leaching ·Flue gas
R. P. Mahapatra (B)·B. Mohanty
National Institute of Technology Mizoram, Aizawl 796012, India
e-mail: mahapatra.rp@gmail.com
R. P. Mahapatra ·S. S. Srikant
SRM Institute of Science and Technology, Modinagar 201204, India
R. B. Rao
CSIR-IMMT, Bhubaneswar 751013, India
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_63
715
716 R. P. Mahapatra et al.
1 Introduction
Recently, the usage of digital materials for both business and domestic purposes
has increased in an exponential way and hence problem of growing of scrapped
electronics products day by day.
In India, recycling and disposal of unwanted and discarded electrical and elec-
tronic component is mostly carried out with simple methods which severely pollute
the environment and causes health problems. Presently, scrapped electronics and elec-
trical materials are being processed with disassembling, upgrading and refining [1].
It has observed with many literatures that mechanical process separates the metallic
part from the scrapped components but at the same time, a huge amount of waste in
form of plastics and hazardous components were produced. Hence, it is impossible
to dispose the unwanted and discarded electrical and electronic components with
mechanical routes.
The halogenated flame retardants (HFRs) present in the scrapped electronics and
electrical components cause the poisonous and harmful dioxins formation. During the
individual metals recovery from E-waste with various leaching also causes harmful
dioxins and other toxic gases.
Microwave-assisted gasification systems were interfaced into a single system as
shown in Fig. 1, where E-waste materials were used for analyzing the toxic gas
production. However, attempts and goals were to extract the valuable metals from
electronic waste. Therefore, a process was developed, which has eliminated the
requirement of mechanical disassembling step. As microwave furnace generated the
temperature between 1000 and 1200 °C, the harmful and poisonous gases including
dioxin got decomposed. Due to this decomposition, the flue gas which produced
the limited sulphur dioxide, dioxins, carbon monoxide gas, various nitrogen oxides
components and others made eco-friendly process. This single-step process not only
produced valuable and useful metals in alloy form but also small quantity of waste
was generated. Copper has been recovered through microwave treated followed by
leaching and residue treated with aqua leaching has been enriched with valuable
components like gold and silver [2–6].
2 Experimental Procedures
2.1 Raw Materials
The transformers, capacitors, batteries and plastics were firstly dismantled (by
sharp cutting tools) from the collected unwanted and discarded electrical and elec-
tronics PCBs samples, i.e. monitors, mobile, printer cards, motherboard of TV and
computers, etc. These bared PCBs were crushed into a tiny particles having size
around 230–290 µm by the crusher. An amount of 400 g of SiC powder which
behaved as coupling agent [7–10] was combined with 10 kg of crushed tiny particles.
Less Polluted Flue Gases Obtained with Green Technology … 717
Fig. 1 Green and clean heating process arrangement of unwanted and discarded electrical and
electronic components with microwave for flue gas treatment (1) Argon gas (2). Microwave Furnace
(3) High temperature resisted Gas Pipe (4) Thermal oxidizer (5) Burner (6) Quench tower (7)
Scrubbers (8) Power Supply (9) Cooling Fan (10) Chimney (11), (13) and (17) scrubbing fluid
chamber, (12), (14) and (16) pumps to circulate and atomize scrubber solution
This mixture was then heated in the microwave furnace, as shown in Fig. 1. An inert
gas (here argon) as shown in Fig. 1was used to avoid oxidation reaction [7–9]inside
the furnace. The product after microwave heating (for duration 45 min) obtained
mostly consists of mixture of non-metal (ash mostly) and metal. The metallic portion
separated from the mixture (after cooling) was treated by acid leaching process using
Hydrochloric acid and then followed with cementation process [1].
2.2 Flue Gas Treatment
The investigation of flue gas has to be studied during microwave heating. The flue
gas was collected in secondary chamber as shown in Fig. 1by number 4. It consists of
thermal burner. This fluid solution is precipitated with electrostatic precipitator. The
gas solution was continuously scrubbed at location 7 as shown in Fig. 1. After scrub-
bing process, these solution has to be continuously stirred so that it can easily atom-
ized (as obtained in location 7). The entire process represented with a single flow-
sheet is shown in Fig. 2with Green and clean technology with microwave treatment
process.
718 R. P. Mahapatra et al.
Fig. 2 Flowsheet for flue gas cleaning treatment
3 Results and Discussions
3.1 Green Technology with Microwave Heat Treatment
During microwave heating (always treated to be green and clean process) with flue
gas arrangement at various time intervals, it has been analyzed that there was loss in
waste powdered sample. The metallic product with 3600 g and non-metallic product
with 900 g were produced for 10 kg of powdered sample after microwave treatment.
Less Polluted Flue Gases Obtained with Green Technology … 719
3.2 Treatment of Flue Gas
The typical metallic composition of the sample after microwave treatment is 12.99
% iron, 0.03 % gold, 9.79 % aluminium, 54.3 % copper, 1.13 % nickel, 0.28 % lead
and 0.07 % silver. With physical separation, the weight with 5500 gm containing
non-metallic part (i.e. 55 % of net loss) more of ash was easily removed. This loss
was in the form of dioxins, carbon monoxide, sulphur dioxide, furans, loss of ignition
(LOI), and it is known as flue gas. After scrubbing and filtering process, around 2.4 kg
of metallic part was obtained. The various oxide component of nitrogen and sulphur
were measured by Maritime 350 Testo gas analyzer, whereas carbon monoxide (CO)
was measured by Gas chromatography (GC-2979). The dioxins and furans compo-
nent was measured by SUPELCO. The hydrochloric acid (HCl) to be measured was
mixed with particulate filter, acidified water and then with alkaline water. With the
help of ion chromatography in other solution, the Cl−ions were measured. The flue
gas composition content was found to be in the range of NOx: 245–250 mg per
cubic metre, CO: 93–97 mg/m3,SO
2: 190–195 mg/m3, furans and dioxins: less than
0.15 ng/TEQ cubic metre, HCl: 80–85 mg/m3. After scrubbing and filtering process,
the total flue gas composition contents were analyzed as NOx: 240 mg/m3,CO:
50–55.5 mg/m3,SO
2: 18–20 mg/m3, Furans and dioxins: lesser than 0.12–0.15 ng,
HCl with acidic materials: 60–70 mg/m3. The assigned regulations set by Central
Pollution Control Board are NOx: 390–420 mg/m3,SO
2: 140–200 mg/m3, CO: 80–
110 mg/m3, Furans and dioxins: lesser than 0.15 ng/TEQ cubic metre, HCl and
other acidic components: 50–55 mg/m3. It is understood except the HCl components
factors (because of E-waste treatment), all other components of flue gas were found to
be under within allowable standard. Hence one can say that scrubbers, gas-cyclones,
absorbers, hydro-cyclones and electrostatic precipitators are very important control-
ling device to keep the environment clean. Moreover, the studies are still in progress
to make use of flue gases in the application for alternative source for LPG and CNG
and also for cooking gas.
4 Conclusion
The following conclusions were made for flue gas treatment during the recovery of
some valuable metal from scrapped electronic and electrical components:
•The hazardous unwanted and discarded electrical and electronic components (e-
waste sample) were treated in a microwave furnace for the useful and valuable
metal recovery.
•The microwaveheating is said t o beclean and green process treatment as it does not
pollute the environment. Also such innovative and greenery treatment generates
less waste with respect to the conventional process.
720 R. P. Mahapatra et al.
•The typical metallic compositions after microwave innovative heating process are
iron, copper, aluminium, lead, nickel and very few gold and silver components.
With physical separation, the non-metallic was removed easily.
•With continuous process of scrubbing and filtering at various levels, the contents
of flue gas were found to be well within the regulation range set by CPCB
(Government of India).
•The scrubbers, gas-cyclones, absorbers, hydro-cyclones and electrostatic precip-
itators are very important controlling device to keep the environment clean.
•Further studies are still in progress for flue gases treatment for cooking gas and
vehicle fuels applications.
•The entire innovative investigation shows system integrated with microwave
furnace and flue gas treatment can be concluded as perfect eco-friendly and green
process.
References
1. R.P. Mahapatra, S.S. Srikant, R.B. Rao, B. Mohanty, Recovery of basic valuable metals and
alloys from E-waste using microwave heating followed by leaching and cementation process.
Sadhana J. 44(10), 209–219 (2019)
2. R.P. Mahapatra, S.S. Srikant, R.B. Rao, B. Mohanty, Microwave heating and acid leaching
processes for recovery of gold and other precious metals from e-waste. Curr. Sci. 116(3),
463–468 (2019)
3. G. Chauhan, R.J. Prashant, K.K. Pant, K.D.P. Nigam, Novel technologies and conventional
processes for recovery of metals from waste electrical and electronic equipment: challenges
and opportunities—a review. J. Environ. Chem. Eng. 6, 1288–1304 (2018)
4. G.G. Wicks, D.E. Clark, R.L. Schulz, Microwave technology for waste management appli-
cations: treatment of discarded electronic circuitry (U). U.S. Department of Energy (DOE).
Contract No. DE-AC09-96SR1 8500 (1995)
5. S.S. Rath, P. Nayak, P.S. Mukherjee, G.R. Chaudhury, B.K. Mishra, Treatment of electronic
waste to recover metal values using thermal plasma coupled with acid leaching—a response
surface modeling approach. Waste Manage. 32, 575–583 (2012)
6. R. Khanna, R. Cayumil, P.S. Mukherjee, V. Sahajwalla, A novel recycling approach for trans-
forming waste printed circuit boards into a material resource. Procedia Environ. Sci. 21, 42–54
(2014)
7. S.S. Srikant, R.P. Mahapatra, T. Laxmi, Microwave energy for waste used engine oil. Inst. Eng.
J. 1, 255–261 (2015)
8. S.S. Srikant, P.S. Mukherjee, R.B. Rao, Microwave energy for waste management. Min. Proc.
Technol. 3, 948–953 (2013)
9. S.S. Srikant, P.S. Mukherjee, R.B. Rao, Prospects of microwave energy in material and mineral
processing. Turk. J. Eng. Sci. Technol. 2, 23–31 (2013)
10. C.A. Pickles, Microwaves in extractive metallurgy: part 1—a review of fundamental. Min. Eng.
1102–1111 (2009)
Secure Data Transmission in Mobile
Networks Using Modified S-ACK
Mechanism
P. Muthukrishnan and P. Muthu Kannan
Abstract Mobile ad hoc network is an important application in the current wifi
world. There is no need of any network infrastructure. Each node acts as a trans-
mitter that transmits the messages to the nearby nodes and receiver that receives
the messages. Each node also acts as a cluster head to receive information from the
virtually connected nodes. All grouped information is sent to base station. MANET
is self-governing networks. The dynamic nature of MANET, where the topology is
changing very rapidly and randomly. Wireless network medium used for message
dispersion is prone to attack, as it is accessible by all. It is a necessity to develop a
secure and efficient technique to identify cyber attackers to protect MANET from
attacks. The proposed technique ‘Modified S-ACK’ will provide secure and efficient
way of data transmission in Mobile Networks.
Keywords Wireless ·Adhoc network ·Nodes ·Malicious attackers ·MANET
1 Introduction
Mobile Ad hoc Network (MANET) comprises numerous mobile nodes that combine
to form a virtually connected network without any existing network infrastructure.
Each node in this network acts as a transmitter and receiver [1]. Each node receives
packets and transmits to the other nodes in this network. The structure of MANET
is shown in Fig. 1.
MANET is self-organizing, self-assembling wireless networks. Since the
mobililty is high, the network structure of ad hoc networks is dynamic. Each node
transmits and receives packets of data by random wireless channels [2]. If the distance
between two nodes for data transmission is more, multihop forward technique comes
P. Muthukrishnan (B)·P. Muthu Kannan
Saveetha School of Engineering, SIMATS, Chennai, India
e-mail: muthueng05@gmail.com
P. Muthu Kannan
e-mail: muthukannan@saveetha.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_64
721
722 P. Muthukrishnan and P. Muthu Kannan
Fig. 1 Structure of MANET
into picture. Since MANETS rely on wireless transmission, sharing of message must
be secured from vulnerabilities. Secured way of data transmission is needed to protect
the data integrity. An insecure ad hoc network poses threat to the entire network.
In MANET, there is no central system to identify and prevent anomalies. There
are many problems in MANET. One of the major problems is data security. How
secured is data transmission through wireless channels? Routing table sets the routing
protocols that lead nodes to transfer network topology information to build communi-
cation routes. This information is highly sensitive and may be a victim for malicious
network, who attacks the network and the applications running on it [3]. There are
two sources of threats to routing table. First is external attackers. An attacker could
partition a network and introduce traffic congestion in the routing protocols.
Second and most severe threat is from nodes that are compromised, which might
(i) misapply or share routing information to other nodes (ii) do vulnerable actions on
data that induce service failures [4]. Generally, ad hoc network attacks are classified
into passive attacks and active attacks. Active attacks are further categorized into
internal and external attacks. Internal attacks are from internal authorized nodes
inside the network [5]. It is difficult to identify those nodes that are attacking the
data. External attacks are done by external nodes outside the network. They do not
belong to the network and can be stopped by firewalls, decryption, and encryption
techniques.
2 Characteristics of MANET
MANET comprises many independent mobile nodes that communicate with each
other through radio waves. Mobile nodes that are within the range of other radio nodes
can communicate easily as the transmission between the nodes will be enabled. If
nodes are not within the range, then communication can be happened with the help
of intermediate nodes [6]. These networks can operate from anywhere and fully
Secure Data Transmission in Mobile Networks Using … 723
dispersed and do not need an infrastructure. The characteristics of MANET are
(i) Wireless communication, (ii) Nodes can be both a transmitter and a receiver,
(iii) There is no central system to control, (iv) Network topology is dynamic, (v)
Autonomous and self governing, as no infrastructure is needed, (vi) MANET can
work from anywhere, (vii) Limited security.
The dynamic network topology due to the mobility of nodes increases the chal-
lenge to design an ad hoc network [7]. Each radio terminal or a mobile node is
powered by rechargeable batteries (limited power source). There are three parts
of power consumption of radio terminal. One part of power consumption for data
processing inside the Radio Terminal. Second part is consumed to transmit its own
information to the target. Third part of power consumption is utilized when the
same Radio Terminal or the nodes are used as router. Energy consumption is the big
challenge in designing ad hoc networks [8].
Generally, mobile devices have less storage and low computational capabilities.
As their memory and computation is less, they are dependent on other resources for
data processing [9]. A reliable network structure must be enabled through efficient
and secure routing protocols.
3 Application in MANET
MANET has many applications. Some of the most common applications of ad hoc
networks are listed here. MANET is used on a day to day basis for the electronic
email, file transfer [10]. These days, emails are accessed from mobiles for faster
communication. Due to the competitive data rates offered by the Network Service
Providers, they are coming up with many new applications. Some of the applications
of MANET are (i) Military zone, where the network infrastructure is unavailable. Ad
hoc networks are self assembling, self governing. (ii) Disaster Recovery Management
(iii) Mining (iv) Robotics.
MANET wireless systems are essential in crisis times, natural disasters like flood,
earthquake, cyclone. During these times, restoring communication services is essen-
tial for people to lead their normal life. MANET does not need network infrastructure
and communication is established faster than wired network [11].
There is another application used in inter-vehicular communications. Here, the
secured communication happens between two vehicular nodes [12]. The vehi-
cles can communicate with each other which help to identify the real-time traffic
updates, weather conditions, route identification in city environment, and highway
environment. Such advanced network systems can be used for the future world.
724 P. Muthukrishnan and P. Muthu Kannan
Tabl e 1 Types of attacks
Active attacks Active attacks are done by the malicious nodes inside the routing protocol.
This modifies the data packets or sometimes creates the false stream. Active
attacks impacts the network performance
Passive attacks Passive attacks do not modify the data packets. Instead it will aim to get the
network structure, hierarchy, network information. These attacks are more
complex to identify, as they do not alter the data stream. Passive attacks
monitor the data transfer and collect the required network information
Wormhole attacks Wormhole attack means attackers monitor the network and record the
wireless information. They locate at different positions and create a tunnel.
This tunnel between two attackers is called wormhole. Now, the data stream
is forced to go through the tunnel
Denial of services This attacks the network or any system, so that data packets are lost and
does not reach the destination. DoS sends lot of void packets and creates a
network congestion. By this, the services are impacted and cannot be
accessed by the intended users
4 Issues in MANET
There are different issues in MANET. Some most common issues are (i) Data medium
(ii) Distribution of information is needed (iii) Synchronization is needed between the
sender and receiver that receives the message (iv) Some radio terminals are hidden
from a sender (v) Data transmission speed should be high, otherwise there will
be lag in data transfer (vi) Security. This is a critical issue. This paper revolves
around this issue and discussed in detail in the next paragraph. (vii) Real time data
synchronization is required for voice, video, images. (viii) Mobile storage is less.
It is necessary for resource reservation for an OS. (ix) Power control is required to
reduce the energy consumption at each node, routing table (x) Variation in sending
and receiving the data bits leads to bigger bandwidth usage [13].
The most common security attacks in MANET systems are tabulated [14], Table 1
shows different types of attacks.
Our proposed system states how security can be provided or improved in MANET
than the existing systems. In the proposed Modified S-ACK method, medium access
reducing handshake technique is imposed along with node ID.
5 Existing System
A number of secure routing schemes are introduced for detecting security vulnera-
bilities in MANET.
Secure Data Transmission in Mobile Networks Using … 725
Fig. 2 Schematic representation of TWOACK
5.1 Watchdog
It is responsible for detecting malicious node behavior in the network [15]. Watchdog
continuously monitors the behavior of the network and detects malicious nodes by
data transmission speed of the network.
5.2 TWOACK
TWOACK is to recover the drawback of watchdog. In this, if Node A communicates
with Nodes B, first Node A sends request signal to Node B. Upon receiving the
request signal from Node B, it sends two ACK signals to Node A. This helps to
identify the malicious node in the network. It also helps to avoid Data collision [16].
Figure 2represents schematic representation of TWOACK.
5.3 S-ACK
In S-ACK technique, it sends data packet involving group of n number of nodes.
Below example demonstrates to identify the malicious nodes in the network. Here,
S-ACK involves group of three nodes, namely node A, B, and C. Node A is the
source and it has to forward the packet to the destination node C. Node A sends Preq
1 to node B and node B sends Preq 1 signal to node C. Once Node C receives the
signal Preq 1, it responds to node B with an acknowledgement signal Pack1. Once
node B receives the acknowledgement signal Pack1, it forwards to node A. After
node A receives Pack1 signal, node A is ready to forward the data packet to node C.
In this network, if node A does not receive Pack1 signal, then other two nodes are
726 P. Muthukrishnan and P. Muthu Kannan
Fig. 3 Schematic
representation of S-ACK
considered to be malicious. In this way, the data packets are sent from the source to
the destination involving a group of three-three nodes. Figure 3shows the schematic
representation of S-ACK [17].
When comparing to Watchdog and TWOACK, this method shows more improve-
ment in finding malicious node in the network. False identification can be avoided
in S-ACK [18]. Time consumption is reduced as it does not wait for multiple
acknowledgements. Figure 4shows the flowchart of S-ACK.
6 Proposed System
In the proposed system, securing the data transmission in mobile networks using
S-ACK Mechanism. In this method, data packets are sent from source to destination
with the help of only two control signals Request To Send (RTS) and Clear To
Send (CTS). Data is forwarded in a separate channel. Below example illustrates the
operation of Modified S-ACK. Kindly refer Fig. 5for the proposed system.
Consider there are four nodes node A, B, C, and D. Node A is the source and
node D is the destination. Control signal RTS is initially sent from node A to node
B. If node B is free, it sends back CTS1 signal to the neighboring nodes (node A and
node C). Once node A receives CTS1 signal, it forwards the secure data to node B.
Secure Data Transmission in Mobile Networks Using … 727
Fig. 4 Flowchart of S-ACK
At the same time, node C also receives CTS1 control signal. Once node C receives,
it responds back with CTS2 signal to the neighboring nodes B and node D.
Every node has a unique identifier. Secured data is embedded with source Node
Id and destination Node Id in a frame format. Frame format has source Node Id,
Secure data, destination Node Id in this order. Once node B receives secure data, it
checks the destination Node Id of the received data. If the Node Id does not match, it
forwards to node C. The same process will continue till the data packets are received
by the destination node D. Once data is sent to node D, the destination Node Id
matches with the frame formatted data and receives the secure data.
Main advantages of our proposed approach are
1. Prevent data collision and detecting false misbehavioral nodes, consumes less
power.
2. Data is more secured compared to contemporary approaches.
3. Acknowledgment data packets in this proposed method are true.
4. Improves the packet delivery ratio.
728 P. Muthukrishnan and P. Muthu Kannan
Fig. 5 Schematic diagram
of modified S-ACK
7 Results and Discussion
One of the major security issues or attack in MANET is Security. Security issues
like Active attack, passive attack, Wormhole attack are overcome by some existing
security systems such as watchdog, TWOACK, S-ACK. TWOACK method is based
on acknowledgement packets. Hence, it is important to ensure that the acknowl-
edgment packets are true. But the disadvantage is that they are unable to identify
the misbehaving nodes with the presence of false report and bad acknowledgement
packets. Another drawback is the network congestion created by two acknowledge-
ment packets. Due to the limited power consumption of MANET, such network
congestion can impact the network performance. Figure 6shows the number of
attackers reduced in the proposed ratio and main the packet delivery ratio.
8 Conclusion
In the proposed modified S-ACK Mechanism, it is implemented framed data format
including Source Node Id, Secure Message, destination Node Id. Only the matched
Node Id receives the data. Data transmission speed is higher. It has advantages over
Secure Data Transmission in Mobile Networks Using … 729
Fig. 6 Packet delivery ratio versus number of attackers
the other existing methods like true nature of the nodes, less power consumption,
less receiver collision. All acknowledgment packets in the proposed system are true.
It demonstrates higher malicious-behavior-detection rates. This technique helps to
secure data in MANET by properly identifying the malicious nodes.
References
1. K. Al Agha, M.H. Berlin, T. Dang, A. Guitton, P. Minet, T. Val, J.B. Viollet, Wireless technology
for industrial wireless sensor networks 56(10):4266–4278 (2009)
2. E.M. Shakshuki, N. Kang, T.R. Sheltami TR, EAACK—a secure intrusion-detection system
for MANETs. IEEE Trans. Ind. Electron. 60(3) (2013)
3. G. Jayakumar, G. Gopinath, Ad hoc mobile wireless networks routing protocol—a review. J.
Comput. Sci. 3(8), 574–582 (2007)
4. D. Johnson, D. Maltz, Dynamic source routing in ad hoc wireless networks (Chap. 5), in Mobile
Computing (Kluwer, Norwell, 1996), pp. 153–181
5. N. Kang, E. Shakshuki, T. Sheltami, Detecting misbehavingnodes in MANETs, in Proceedings
of 12th International Conference on iiWAS, Paris, France, 8–10 Nov 2010, pp. 216–222
6. N. Kang, E. Shakshuki, T. Sheltami, Detecting forged acknowledgements in MANETs, in
Proceedings of IEEE 25th Int. Conf. AINA, Biopolis, Singapore, 22–25 Mar 2011, pp. 488–494
730 P. Muthukrishnan and P. Muthu Kannan
7. K. Liu, J. Deng, P.K. Varshney, K. Balakrishnan, An acknowledgment-based approach for the
detection of routing misbehavior in MANETs. IEEE Trans. Mobile Comput. 6(5), 536–550
(2007)
8. S. Marti, T.J.Giuli, K. Lai, M. Baker, Mitigating routing misbehavior in mobile ad hoc networks,
in Proceedings of 6th Annual International Conference on Mobile Computing Network,Boston,
MA, pp. 255–265 (2000)
9. N. Nasser, Y. Chen, Enhanced intrusion detection systems for discovering malicious nodes in
mobile ad hoc network, in Proceedings of IEEE International Conference on Communications,
Glasgow, Scotland, 24–28 June 2007, pp. 1154–1159
10. T. Sheltami, A. Al-Roubaiey, E. Shakshuki, A. Mahmoud, Video transmission enhancement in
presence of misbehaving nodes in MANETs. Int. J. Multimed. Syst. 15(5), 273–282 (2009)
11. T. Anantvalee, J. Wu, A survey on intrusion detection in mobile ad hoc networks, in
Wireless/Mobile Security (Springer, New York, 2008)
12. L. Zhou, Z.J. Haas, Network Security. IEEE, Nov 1999, pp. 1–12
13. D.B. Johnson, A. Perrig, Y.-C. Hu, SEAD: secure efficient distance vector routing for mobile
wireless ad hoc networks (Elsevier, 2003), pp. 175–192
14. J. Undercoffer, J. Parker, J. Pinkston, A. Joshi, On intrusion detection and response for mobile
ad hoc networks, in 2004 IEEE International Conference on Performance, Computing, and
Communications, USA, pp. 747–752 (2004)
15. A. Bertacchini, D. Brunelli, L. Larcher, L. Benini, D. Dondi, Modeling and optimization of
a solar energy harvester system for self-powered wireless sensor networks. IEEE Trans. Ind.
Electron. 55(7):2759–2766 (2008)
16. S. Vairachilai, A. Rubi, A survey on intrusion detection system in mobile adhoc networks. Int.
J. Comput. Sci. Mobile Comput. 2(12):389–393 (2013)
17. W. Liu, Y. Zhang, W. Lou, Y. Fang, Anonymous handshakes in mobile ad hoc networks, in
Military Communications Conference, pp. 1193–1199 (2004)
18. N. Kang, T.R. Sheltami, E.M. Shakshuki, EAACK—a secure intrusion-detection system for
MANETs. IEEE Trans. Ind. Electron. 60(3), 1089–1098 (2013)
An Anonymization Approach
for Dynamic Dataset with Multiple
Sensitive Attributes
V. Shyamala Susan
Abstract In recent days, the personal health information is collected in different
fields. When the data is shared for different reasons, it poses a major danger to
the field of health care. A number of anonymization methods are implemented to
maintain person’s privacy. The existing methods of anonymization support only
single sensitive and low-dimensional data. In our recent experiment, a method
of anonymization is expected in order to anonymize high-dimensional data with
multiple sensitive attributes. In line with the concept of k-anonymity and l-diversity,
it combines anatomization and improved slicing strategy. The experimental findings
show that it is limited to the static discharge of information only. In dynamic situa-
tions, the current technique may produce poor quality or high data loss. Hence, in this
proposed approach, an anonymization model is designed in such a way to anonymize
continuously growing dataset while assuring high utility.
Keywords Health care ·Privacy preservation ·Anonymization ·Anatomization ·
Improved slicing ·K-anonymity ·L-diversity
1 Introduction
Every organization publishes data that are collected from different users. When they
publish the data, the personal information of an individual may be disclosed. This
violates the privacy of the person and it needs to be protected from abuse [1]. To
maintain privacy of the personal information, anonymization procedures are intro-
duced. The dataset may be classified into different categories of attributes like iden-
tifiers, quasi-identifiers, sensitive attributes, and non-sensitive attributes [2]. The
k-anonymity generalization [3,4] and l-diversity bucketization [5] are the popular
privacy preservation approaches. The k-anonymity method in generalization [6,7]
disregards an enormous portion of information in the occasion of high-dimensional
data. But these approaches are suitable only for static data release and can handle
V. Shyamala Susan (B)
A.P.C. Mahalaxmi College for Women, Thoothukudi, India
e-mail: shyamalasusan@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_65
731
732 V. Shyamala Susan
single sensitive attribute. If all sensitive attributes are given a similar amount of
privacy, they may not provide predictable outcomes on the accessible informa-
tion. The real-world data contains Multiple Sensitive Attributes (MSA). To handle
that situation, our recent work proposed a model to anonymize MSA. It presents a
method of anonymization that integrates the assistances of anatomization and more
enriched slicing according to the rule of k-anonymity and l-diversity to handle high-
dimensional information with MSA. The algorithm of anatomization divides QI and
SA in the dataset and publishes unmodified QI and SA in two tables: QIT and ST.
The vertical partitioning stage in the enhanced slicing algorithm combines the corre-
lated SA in ST with the correlated QI attributes in QIT, thus significantly reducing
dimensionality through the use of the Modified Advanced Clustering Algorithm (m-
ACA). Similarly, Metaheuristic Firefly Algorithm with Minkowsi Distance Measure
(MFAMDM) approach ensures k-anonymity in QIT and l-diversity in ST in the hori-
zontal partitioning phase. The subsequent step anonymizes the dataset by linking the
two tables with the common group ID as quickly as the attributes are cut off.
The experimental findings indicate that with several SA the suggested system can
efficiently maintain data privacy. The anatomization approach significantly reduces
the loss of information and the slicing algorithm benefits preserve correlation and
utility, thus decreasing data dimensionality and loss of information. By minimizing
the time complexity, advanced clustering algorithms prove their effectiveness. More-
over, this research work tracks strictly the rule of k-anonymity, l-diversity, and thus
keeps away privacy.
Now, the next objective is to anonymize a continuously growing dataset with
MSA. The promising method is to anonymize the new dataset and then bring out
new records periodically. These published datasets can be used for analysis by the
researcher either autonomously or combine multiple datasets for better analysis. The
problem in the existing method is more time is spent in anonymization as many small
records are to be anonymized independently. When these small anonymized datasets
are merged, they may violate the privacy of the person. Also, many independent
anonymizations reduce data utility. Hence, a new method is to be designed in such a
way that an entire dataset needs to be anonymized whenever new records are added
to the existing dataset. In this paper, we apply an anonymization methodology for
continuously growing dataset with great data utility and minimum information loss.
2 Literature Review
In order to anonymize MSA, the following researches are used. The k-anonymity
and l-diversity model is introduced to safeguard the person’s privacy [8]. It used data
distortion to preserve privacy. A decomposition method is introduced to preserve
privacy [9]. But that model cannot be extended to dynamic data release. Similarly, a
model for dynamic data release is proposed [10], but that is not appropriate for high-
dimensional data. Liu et al. implemented a k-anonymity model for MSA [11].These
methods recognized characters with both type of attributes, but it is not appropriate
An Anonymization Approach for Dynamic Dataset … 733
for large dataset. An anonymization approach for multiple sensitive approaches is
introduced [12], but it is not tested with a real dataset. An anatomization with slicing
approach is introduced; but for Quasi-Identifiers (QI) and Sensitive Attributes (SA),
the slicing algorithm must be implemented individually [13]. Hassan et al. [14]
termed a model to reserve the confidentiality on multiple independent data publishing.
A method for MSA using (P,K) anonymization approach, but it is not limited to
low-dimensional data [1]. KC-slice methodology is suggested for dynamic data with
MSA set [15]. The privacy preservation approaches that are used in ecommerce
sites suffer from inference attacks [16]. Mehta and Rao [17,18] projected a novel
anonymization approach called Scalable k-Anonymization with minimum number
of iterations in MapReduce.
2.1 Proposed Work
In this chapter, we present our current work with a static dataset, which is appropriate
for MSA. Then, we extend our work to anonymize continuously growing dataset with
MSA.
2.2 The Anonymization Approach for Static Dataset
with MSA
Anatomization method is initially used to reduce information loss by directly
discharging QI and SA. It disassociates the correlation between the two attributes and
discharges double distinct tables, namely Quasi-Identifier Table (QIT) and Sensitive
Table (ST). Then, enhanced slicing approach is carried out in two phases. The first
phase groups the correlation between the attributes in QIT and ST and minimizes the
dimensionality. The improved slicing algorithm retains privacy over partitioning both
horizontally and vertically. Now, the SA is grouped by m-ACA in the vertical parti-
tioning phase and releases many ST each with the column group ID. Similarly, QIA
is grouped and released as QID. For the attributes being partitioned into columns,
the m-ACA is introduced, reducing the need to reassign the data point several times
in each iteration with optimum initial centroid detection. This can help to improve
the clustering velocity and minimize the complexity of the computation.
Algorithm:Modified Advanced Clustering Algorithm (m-ACA)
Input: Input dataset X=(x1,..xn), the count for clusters k
Output: Vertical partitioning groups
Procedure
Begin
1. Input dataset X=(x1,..xn)from the dataset
2. Estimate the close centroids and allocate to close cluster
734 V. Shyamala Susan
3. Identify least distance from cluster center
4. Calculate the distance between each dataset samples xito all the optimal centroids
cj
5. Repeat steps 1 to 5 until the number of data points extents a smaller amount of
distance value
6. End for;
7. Bring the two closest clusters together in a cluster
8. Recalculate the fresh cluster center for the merged cluster until it decreases the
amount of cluster to k
9. End for;
In the second phase, attributes are grouped into tuples by Metaheuristic Firefly
Algorithm (MFA) ensuring k-anonymity in Quasi-Identifier Table (QIT) and l-
diversity in the entire Sensitive Table (ST). In the last phase, with a shared group ID,
SA is shuffled in each cluster and connected to QI in QIT. This is to guarantee that an
intruder can identify a person’s sensitive value with a maximum of 1/l. This strategy
minimizes the loss of data and guarantees privacy through the slicing algorithm. The
advanced algorithm of chartering reduces time and complexity.
Algorithm : Tuple-partition
Input: Dataset , the parameter ℓ
Output:Sliced Bucket
Procedure
Begin
1. Q={T}, Sliced bucket =∅
2. While Q is not empty
{
2.1 Eliminate the first bucket from Q,Q=Q-{B}
2.2 Split the bucket into totally two different buckets using Firefly Algorithm(FA)
2.3 Check the tuple from Q
2.4 Fix the objective function f(X)for Q
2.5 Compute the light intensity function Iiby the objective function f(X)
2.6 Find attractiveness βby the minimum distance between the tuples
2.7 Using the intensity Iiand attractiveness βtuples are ranked
2.8 Highest ranking tuples form the bucket Q
2.9 if diversity-check(QIT,Q ∪ {B1,B2} ∪ SB,ℓ)
Q=Q U{B1,B2}
Else SB=SBU{B}
Return SB
}
3. End while
4. End
Thus, tuple partitioning guarantees -diversity in each bucket.
SLicing with Anatomization for Multiple Sensitive Attributes (SLAMSA)
The next step is to develop an anonymization group of SLicing with Anatomization
for Multiple Sensitive Attributes (SLAMSA). Here, the SA is associated with a
common group ID in each group with the QI characteristics in QIT. This guarantees
that an individual’s delicate value engaged in QIT can be extracted immediately from
an opponent with a maximum likelihood of 1/l. A greater lleads to greater privacy.
The method finishes when there is no longer any ungrouped SA or when it is not
possible to develop fresh group. If ungrouped documents are available, they will be
released as a single group. As this work follows both the principle of k-anonymity and
An Anonymization Approach for Dynamic Dataset … 735
l-diversity, it removes the overall complexity of privacy. Disclosures of membership
and identity are shielded by k-anonymity and exposés of attributes are eliminated by
l-diversity.
Algorithm : SLAMSA
Input: Dataset, k, l, Sensitive Attribute Table (SAT)
Output: Quasi-Identifier Table, horizontally partitioned Table;
Procedure
1. Begin
2. Divide the dataset into a set of sensitive attribute Tables and Quasi-Identifier Table w.r.t ST;
3. For each Table
{
3.1. For each tuple
Assign a numerical identifier starting from 1;
4. For i=st1to stm
4.1. Form st iw.r.t l, with no repetitive attribute values; (fitness)
4.2. End for
5. For each group of size < k
5.1. Select a group G of size < k randomly;
5.2. Find the distance between G and other groups;
5.3. Pick out the group G1 with the least distance;
5.4. Generalize G and G1;
5.5. end for;
6. Map generalized table and (Quasi-identifier Table)
7. End
2.3 Anonymization Approach for the Dynamic Dataset
Assume that an anonymized table Tis released for the static dataset. Now suppose
that new patient records R={r1,r2…rn}are to be inserted into the anonymized
table T. Let this anonymized table be T. To anonymize this table T, the required
conditions are that QI attributes in each group must be kanonymized and SA attributes
must be ldiverse. To ensure high data utility while anonymizing the dynamic dataset,
the following conditions are suggested.
Case 1: If the new records Rwhen added with Tsatisfy k-anonymity and l-
diversity and do not overlap with the existing anonymized table, then we can increase
the records to Tas a new anonymized group.
Case 2: If the new records in Rcannot be added directly into as an anonymized
group, then these records must be inserted into an existing anonymized table T.To
preserve privacy in the new anonymized table, each record is inserted by ensuring
k-anonymity and l-diversity.
Case 3: Sometimes addition or insertion of the new records into the anonymized
table Tmay not satisfy k-anonymity and l-diversity in each group. In such a case,
we split the anonymized table into two anonymized tables for better data quality.
736 V. Shyamala Susan
2.4 How to Enforce Anonymity in All the Above Three Cases
The add operation ensures anonymity by utilizing k-anonymity in QIT and l-diversity
in ST in each group. The insert operation may not produce anonymity in each
anonymized group when the new records are inserted in the anonymized group.
It may not satisfy k-anonymity in QIT and l-diversity in ST. This may cause a threat
to the dataset. In order to overcome such a situation, we insert the records into the
waiting list. The records will be inserted from the waiting list to the anonymized
group if it satisfies anonymity. If the records’ waiting time is more, then they are
identified to be published as a separate anonymized group. But these groups should
not overlap with the existing anonymized group. The process ends when there are
no ungrouped attributes in the waiting queue. The anonymized table Tcan be split
when it ensures k-anonymity and l-diversity in each group of the tables.
3 Performance Analysis
The experiment is evaluated through the dataset Cleveland Clinic Foundation and
Heart Disease and Health Institute of Cardiology, and the dataset is accessible in the
UCI machine learning repository. The dataset has 76 raw attributes, considering for
assessment only 12 attributes for evaluation. The age, gender, and social security
number are measured as QI, and further attribute that identifies the heart disease of
the patient is SA. The experiment is started with 557 records for both datasets and
compared with the existing KC and KCi methods. The KC slicing system utilizes the
same threshold value for all sensitive data that result in high data loss, while varying
threshold values are given to the SA in the KCi slicing system to decrease data loss.
3.1 Utility
The utilization of the proposed method is evaluated using reconstruction error. To
improve the utility, the correlated attributes must be arranged in a column which is
achieved by the slicing approach. The query.
Select count * from T
Where (sensitive items are present)
and (q1=val1)n(q2=val2)…n(qn=valn)
It is measured by obtaining the distance between actual probability distribution
function (Act) and estimated probability distribution function (Est) which gives KL
divergence
KLdivergenceActs,Ests=
∀cellC
Acts
Clog Acts
C
Ests
C
An Anonymization Approach for Dynamic Dataset … 737
Fig. 1 Reconstruction error
versus p(r=4)
0
0.05
0.1
0.15
0.2
0.25
4 6 8 10 12
KL-Divergence
p
KC (M=10) Kc(M=20)
KCi (M=10) KCi(M=20)
DynamicMSA (M=10) DynamicMSA(M=20)
Acts
C=Occurrences of sin C
Total Occurrences of sin T.
The quantity of elements in Gis “a,” and the quantity of the same QID tuples
is “b.” The query’s evaluated outcome converts a·b/a. For all tuples in Gwith the
same QID, either a or b=0, which means that a minimum reconstruction mistake
occurs. If Actsis same to Ests,KL
divergence =0 which means there is a minimum
reconstruction mistake.
It is necessary to evaluate the data quality from both static and dynamic datasets.
Change the values of P,M, and Rwhere Pis the degree of privacy between 4 and
20 for each experiment, Mis the number of SA ranging from 3 to 12, and Ris the
value of QID ranging from 2 to 4. Many queries are assessed by altering q1,q2…
qnand s1,s2…smrandomly. The average reconstitution error is measured, and then
the utility is identified.
The experimental effects are shown in the following figures.
Figure 1shows that the data quality is lower as compared to Kc and KCimethods.
But the quality is maintained regarding to data size. Increasing the degree of privacy
increases the reconstruction mistake and minimizes the utility. When the value of m
is increased, there is more value matching, which limits the quality which is depicted
in Fig. 2. Similarly, when Ris increased, the corresponding reconstruction error is
also increased due to high dimensionality but limits utility. This is depicted in Fig. 3.
The existing KC and KCitechniques show information loss as these tech-
niques make use of suppression in each group. But this method uses advantages
of anatomization that releases QID and SA directly that improves utility and limits
information loss.
4 Conclusion
In this proposed work, a module is designed to anonymize a dynamic dataset with
MSA. The correlation between the attributes identified using slicing algorithm and
738 V. Shyamala Susan
Fig. 2 Reconstruction error
versus m(r=4)
0
0.05
0.1
0.15
0.2
0.25
481012
KL-Divergence
m
KC(P=10) KC (P=20)
KCi (P=10) Kci(P=20)
DynamicMSA (P=10) DynamicMSA(P=20)
Fig. 3 Reconstruction error
versus r(m=10)
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
23456
KL-Divergence
r
KC (P=10) KC (P=20)
KCi (P=10) Kci(P=20)
DynamicMSA (P=10) DynamicMSA(P=20)
anatomization reduces the information loss. The Advanced Clustering algorithm
limits time and complexity, and the design ensures k-anonymity and l-diversity to
pervert privacy threats. This method can handle any number of SA in an effective
way.
References
1. A. Anjum, N. Ahmad, S.U. Malik, S. Zubair, B. Shahzad. An efficient approach for publishing
microdata for multiple sensitive attributes. J Supercomput. 1–29
2. S. Onashoga, B. Bamiro, A. Akinwale, J. Oguntuase, Kc-slice: a dynamic privacy-preserving
data publishing technique for multisensitive attributes. Inf. Secur. J. Global Perspect.
26(3):121–135
3. B. Gedik, L. Liu, Protecting location privacy with personalized k-anonymity: architecture and
algorithms. IEEE Trans. Mob. Comput. 7(1), 1–18 (2008)
4. K. El Emam, F.K. Dankar, Protecting privacy using k-anonymity. J. Am. Med. Inf. Assoc.
15(5):627–637 (2008)
An Anonymization Approach for Dynamic Dataset … 739
5. L. Ninja, L. Tiancheng, S. Venkatasubramanian, t-closeness: privacy beyond k-anonymity
and l-diversity. In: Proceedings of ICDE 2007, IEEE 23rd international conference on data
engineering, Istanbul, Turkey (2007)
6. C.C. Aggarwal, On k-anonymity and the curse of dimensionality, in Proceedings of the 31st
international conference on very large data bases, VLDB Endowment, 30 Aug 2005
7. D. Kifer, J. Gehrke, Injecting utility into anonymized data sets, in Proceedings of the ACM
SIGMOD, International Conference on Management of Data, 27 June 2006
8. T.S. Gal, Z. Chen, A. Gangopadhyay, A privacy protection model for patient data with multiple
sensitive attributes. IGI Global 2, 28 (2008)
9. Ye Y, Liu Y, Wang C, Lv D, Feng J (2009) Decomposition: privacy preservation for
multiple sensitive attributes, in International Conference on Database Systems for Advanced
Applications (Springer, Berlin), pp. 486–490
10. D. Das, D.K. Bhattacharyya, Decomposition +: improving l-diversity for multiple sensi-
tive attributes, in International Conference on Computer Science and Information Technology
(Springer, Berlin), pp. 403–412 (2012)
11. F. Liu, Y. Jia, W. Han, A new k-anonymity algorithm towards multiple sensitive attributes, in
IEEE 12th International Conference on Computer and Information Technology (CIT) (IEEE,
2012), pp. 768–772
12. P. Usha, R. Shriram, S. Sathishkumar, Multiple sensitive attributes based privacy preserving
data mining using k-anonymity. Int. J. Sci. Eng. Res. 5(4) (2014)
13. V.S. Susan, T. Christopher., Anatomisation with slicing: a new privacy preservation approach
for multiple sensitive attributes. SpringerPlus 5(1), 964
14. A.S.M. Hasan, Q. Jiang, H. Chen, S. Wang, A new approach to privacy-preserving multiple
independent data publishing. App. Sci. 8(783), 1–22 (2018)
15. S.A. Onashoga, B.A. Bamiro, A.F. Akinwale, J.A. Oguntuase, Privacy preserving data
publishing of multiple sensitive attributes: a taxonomic review. Presented at the International
Conference on Applied Information Technology. Federal University of Agriculture, Abeokuta,
Ogun State, Nigeria (2017)
16. D. Yang, Q. Bingqing, P. Cudre-Mauroux, Privacy-preserving social media data publishing
for personalized ranking-based recommendation. IEEE Trans. Knowl. Data Eng. (2018). ISSN
(Print):1041–4347, ISSN (Electronic):1558-2191
17. B.B. Mehta, U.P. Rao, Privacy preserving big data publishing: a scalable k-anonymization
approach using MapReduce. IET Softw. 11, 271–276 (2017). https://doi.org/10.1049/iet-sen.
2016.0264
18. B.B. Mehta, U.P. Rao, Toward scalable anonymization for privacy preserving big data
publishing. Recent Findings Intell. Comput. Tech. 708, 297–304 (2018). https://doi.org/10.
1007/978-981-10-8636-6.Proceedings of the 5th ICACNI 2017,vol.2
Newspaper Identification in Hindi
Subhabrata Banerjee
Abstract This paper reports the creation of a newspaper identification model in
Hindi. This work has two broad angles, namely, the detection of online content in
Indian languages, and second, the use of world-class data. Data is a primary problem
in Indic computing. This work tries to show with the use of proper data, which
may remove the data-driven challenge of Indic computing. The work used annotated
and unannotated gold standard Hindi data. Hand-annotated Hindi data for Named
Entity Recognition (NER) is used to carry out this task. The work shows a detailed
analysis of entity types present in each online newspaper. One generative and another
discriminative model, namely Naïve Bayes (NB) and Support Vector Machine (SVM)
based classifier, are applied here. In both the models, the success of around 90.0 in F-
measure is received. The work is done to help the ever-growing electronic document
user community. It will help a user to classify the Indic document types.
Keywords Indian language ·Hindi ·Gold standard data ·Newspaper
identification ·Classifier
1 Introduction
Document means the piece of text used for teaching or giving instructions. Document
is defined as “any concrete or symbolic indication, preserved or recorded, for recon-
structing or for proving a phenomenon, whether physical or mental” [1]. The history
of documents started in the days of Egyptian civilization. The document is divided as
secret, open, or private based on the parameter of ownership. There may be various
classifications of it based on content and website taxonomy. Documents come across
multiple storage formats, but document means nowadays the electronic file which
may contain information to be studied and stored in a retrievable form. With the
advent of this newer horizon of electronic use, the contents are also getting local-
ized. In our country, there are twenty-two Indian languages. In the days of efficient
S. Banerjee (B)
HCL Technologies, Noida, India
e-mail: subhabangalore@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_66
741
742 S. Banerjee
use of electronic devices, many of the languages have a sizeable amount of online
and offline data. Both governmental and non-governmental agencies are coming into
practice very fast. Days are near when we would need advanced writing assistants
like spell checker, grammar checker, or plagiarism analysis in these languages.
But to create such tools, we first need one sound content detection system. It
will give us a highlight of the peculiarities of such a document. We need to have
appropriate data to implement these systems. In our country, we did not have any
useful annotated general domain data, which was built by a group of experts. The
work is done in Hindi, and it is believed that the whole model right from data building
to content detection be replicated in other Indian languages soon. It will help us to
build a good base for the localization of software.
The problem is the identification of the source of online content. Here, the solution
is built upon online newspapers. In the days of substantial electronic material, it is
useful to make a tool for online detection of resources. The online content of every
language is increasing, and Hindi is no less. Almost all the major newspapers and
news channels have online content. Other than these, many government and business
houses have sizeable online content in Hindi. It is a boon for the readers and users to
have an increased number of content in their local language. But it is also an essential
problem for the cybercrime detection group to detect the source of online content.
The resolution here is to work on one online content detection system for Hindi
newspapers. An impressive result is reported.
The work is perhaps first of its kind in an Indian language. But there is a related
group of work being worked out in natural language processing. The associated works
are in the domain of author identification, sentiment analysis, and topic Modeling.
Various scholars investigate each of these fields, and I am trying to give reference
to only seminal contributions. References [2–4] gave pioneering work on author
identification. In the Indian language, [5] has worked on the author identification of
the Bengali language.
References [6–10] gave major works in sentiment analysis. References [11–13]
made a significant contribution to sentiment analysis in Indian languages. Refer-
ences [14–17] worked sizeably well in topic modeling. References [18,19] made an
excellent contribution to the topic modeling of Indian languages.
The problem here is tackled in two angles. There is one problem of classification
and the other issue of localization. The problem scope was in Hindi, so finding
appropriate data was severe. A gold standard data to build the solution, with hand label
entity tagging, is tried. Here, the resolution is given in three major segments. They
are the statement on the problem, the resolution part, and the results and discussion
part.
Newspaper Identification in Hindi 743
2Problem
The problem in its basic form is the identification of online newspapers or content.
The basis for the resolution is to exploit Indic data. The issue of this data is also
addressed here, briefly.
3 Resolution
In an age of surging electronic activity of document creation, it is almost necessary
to have text mining applications in each language. The recommendation here is to
detect the source of the content of Hindi newspapers, in the emerging electronic
behavior of the language.
In these days of good computing power, web resources, and openware, it is not
a very challenging work to build a statistical classifier and solve a document classi-
fication problem. But this is possible with a resource-rich language like English or
German. Finding proper resources for Indian languages is a daunting task.
To build this gap, I am working with various research teams in Kanpur and
Hyderabad.
This work has two segments of resolution. The first part discusses creating the
data, and the second part talks about building the classification issue.
3.1 Hindi
Hindi in Devanagari script is the lingua franca of India. It is also declared as one
of the primary official languages of India. The language originated mainly from the
Khariboli dialect spoken around the Delhi region. Hindi is a subgroup of Sauraseni
Prakrit. It has 33 consonants and around 11 vowels. The vocabulary consists of
Tatsama, Tadbhaba, Ardhtatsama, Desi, and Bideshi words. It is written left to right.
Chandrakanta by Devaki Nandan Khatri which was written in 1888 is considered
as a first important prose work while it is enriched by contributions of Munshi Prem-
chand, Swami Dayananda Saraswati, Bhartendu Harishchandra, Mahadevi Verma,
Suryakant Tripathi, Ramdhari Singh Dinkar, Haribansh Rai Bacchan, Kamleshwar,
Maithili Sharan Gupt, Nirmal Verma, Mohan Rakesh, Bhism Sahani, Mridula Garg,
Atal Behari Vajpayee, and many others.
It is an Indo-Aryan language. Hindi documents are written primarily using Hunte-
rian transliteration, International Alphabet of Sanskrit Transliteration (IAST), Indian
languages TRANSliteration (ITRANS), ISO 15919 Transliteration of Devanagari
and related Indic scripts into Latin characters (ISO 15919).
744 S. Banerjee
3.2 Gold Standard Data
To build gold standard data in Indian languages, few like-minded senior researchers
of Natural Language Processing (NLP) domain are trying to develop hand-annotated
gold standard data in various Indian languages.
Gold standard annotated corpora are necessary resources when building and evaluating NLP
systems. Manually labeled instances that are relevant to the specific NLP tasks must be
created. A useful gold standard should be rich in information and include a large variety
of documents and annotated instances that represent the diversity of document types and
instances at stake in a specific task. This is essential to (1) either train machine-learning-
based NLP systems, which need examples to learn from, or discover rules for rule-based
algorithms and (2) evaluate the performance of NLP systems [20].
Gold standard data generally refers to the best data given some parameters. These
parameters of data should be appropriately interpreted by an expert, which would be
valid for a considerable period. The data should have detailed guidelines, annotations
to be done by one expert or one editor, and data sets are collected and annotated using
standard methodologies. It may be like Brown corpus is following [21], or maxims
of annotation by Leech [22], or [23] guideline on annotation.
In Indic computing, many institutes and companies are working, but very few are
making their data public. A recent effort is coming from Technology Development
in Indian Languages (TDIL) and Jawaharlal Nehru University (JNU), but they are
pretty much in their infancy. We still do not have enough data for any researcher and
no proper fieldwork and analysis regarding the source of data and their taxonomies.
3.3 Our Solutions
An effort is being made to build gold standard data to be available in public domain
free for anyone, primarily in Hindi and Bangla with the Indian Institute of Tech-
nology (IIT)-Kanpur and International Institute of Information Technology (IIIT)-
Hyderabad. Seed data till now is collected by us after proper field study and is hand
annotated. The present results are based on this.
We have the aim of creating one million hand-labeled data of Parts of Speech
(POS) and NER in a few primary Indian languages. The data would be available
for free in the public domain. We have created, till now, around 0.2 million words
hand-annotated NER data in Hindi. Here I am using 0.1 million unbiased NER data
for our use retrieved from web pages of Dainik Bhaskar, Hindusthan, and British
Broadcasting Corporation (BBC) Hindi.
Figures 1and 2show samples of raw data and Hindi NER annotated data. A good
NER of Hindi is under testing with an F-measure of 90.0+. But it is beyond the scope
of this paper. We hope to publish its results soon.
After tagging the data, we stored the data in (x,y) format in Unicode Transfor-
mation Format (UTF)-8 encoded text files.
Newspaper Identification in Hindi 745
Fig. 1 Sample Hindi raw corpus
Fig. 2 Sample Hindi NER tagged corpus
An NB statistical classifier from Natural Language Toolkit (NLTK) and Support
Vector Machine Classifier (SVC) from scikit-learn is taken.
Naïve Bayes Classifier:
In machine learning, naïve Bayes classifiers are a family of simple “probabilistic
classifiers” based on applying Bayes’ theorem with strong (naïve) independence
assumptions between the features.
Naïve Bayes Classifier (NBC) is a probabilistic classifier. “The naïve Bayes clas-
sifier greatly simplifies learning by assuming that features are independent given
class. Although independence is generally a poor assumption, in practice, naïve
Bayes often competes well with more sophisticated classifiers.” [24]. In this work,
we have followed the NBC definition of [25].
Support Vector Machine:
“SVMs (Support Vector Machines) are a useful technique for data classification.”
[26]. It works out by separating hyperplane. Here, SVM, defined in [27], is used.
4 Results and Discussion
We have seen the presence of following distinct NER variation, as given in Table 1.
The results are impressive, as the papers could be detected nicely, though I
am expecting better results with the help of other classifiers. Some latest neural
network based models may also be tried. Though, the principle of Occam’s razor
and appropriate model selection acts as the best guideline.
746 S. Banerjee
Tabl e 1 Representation of NER tag variation across various Hindi online resources
Resources Person (PERS)
(%)
Location (LOC)
(%)
Date (DATE) (%) Organization
(ORG) (%)
BBC_HINDI 50 30 25 30
Hindusthan 40 60 30 10
Dainik Bhaskar 45 58 35 15
5 Conclusion
Here, one framework for the detection of contents is proposed in Hindi. It works
with two different algorithms and on two kinds of data, one preprocessed raw data
and the other a labeled data.
This kind of work with high precision gold standard data is a sizeable work of its
nature. The work is evaluated with F-measure. Our group has already finished work
on Hindi and Bangla NER, and as more enthusiasts are proposing to join our group,
with their own mother tongue. Extending the work in Telugu and Punjabi is under
active consideration. Perhaps a suite like NLTK in Indian languages is not a distant
dream.
References
1. S. Briet, Qu’est-ce que la documentation? (Documentaires Industrielles et Techniques, Paris,
1951)
2. G. Mikros, E. Argiri, Investigating topic influence in authorship attribution, in SIGIR 07 Work-
shop on Plagiarism Analysis, Authorship Identification, and Near-Duplicate Detection (PAN
07), ed. by B. Stein, M. Koppel, E. Stamatatos. CEUR-WS.org, July 2007
3. M. Kestemont, K. Luyckx, W. Daelemans, T. Crombez, Cross-genre authorship verification
using unmasking. English Stud. 93(3), 340–356 (2012)
4. M. Koppel, J. Schler, E. Bonchek-Dokow, Measuring differentiability: Unmasking pseudony-
mous authors. J. Mach. Learn. Res. 8:1261–1276 (2007)
5. S. Phani, S. Lahiri, A. Biswas, Authorship attribution in Bengali Language. Proc. ICON 2015,
277–282 (2015)
6. B. Liu, Sentiment analysis and subjectivity, in Handbook of Natural Language Processing, 2nd
edn, ed. by N. Indurkhya, F.J. Damerau (2010)
7. J. Karlgren, Affect, appeal, and sentiment as factors influencing interaction with multimedia
information, in Proceedings of Theseus/Image CLEF Workshop on Visual Information Retrieval
Evaluation, pp. 8–11 (2009)
8. P.J. Stone, D.C. Dunphy, M.S. Smith, The general inquirer: a computer approach to content
analysis (MIT Press, Cambridge, MA, 1966)
9. P. Turney, Thumbs up or thumbs down? Semantic orientation applied to unsupervised classifica-
tion of reviews, in Proceedings of the Association for Computational Linguistics, pp. 417–424
(2002)
10. B. Pang, L. Lillian, S. Vaithyanathan, Thumbs up? Sentiment classification using machine
learning techniques, in Proceedings of the Conference on Empirical Methods in Natural
Language Processing (EMNLP) (2002), pp. 79–86
Newspaper Identification in Hindi 747
11. M.S. Akhtar, A. Kumar, A. Ekbal, P. Bhattacharyya, A hybrid deep learning architecture for
sentiment analysis, in Proceedings of COLING 2016, the 26th International Conference on
Computational Linguistics: Technical Papers, pp. 482–493, Osaka, Japan, 11–17 Dec 2016
12. N. Bansal, A. Umair, A. Mukherjee, Sentiment analysis in Hindi. Retrieved from https://cse.
iitk.ac.in/users/cs671/2013/submissions/umair/project/report.pdf, IIT Kanpur Report (2013)
13. K. Starker, S. Chakraborty, A sentiment analysis system for Indian language Tweets,in Proceed-
ings of the Third International Conference on Mining Intelligence and Knowledge Exploration,
vol. 9468 (2015), pp. 694–702
14. D. Blei, Probabilistic topic models. Commun. ACM 55(4), 77–84 (2012)
15. D. Newman, S. Block, Probabilistic topic decomposition of an eighteenth-century newspaper.
J. Am. Soc. Inform. Sci. Technol. 57(5), 753–767 (2006)
16. M. Steyvers, T. Griffiths,” Probabilistic topic models.” In T. K. Landauer, D. S. McNamara,
S. Dennis, & W. Kintsch (Eds.), Handbook of Latent Semantic Analysis. Lawrence Erlbaum
Associates Publishers, Mahwah, NJ, US (2007), pp. 427–448
17. T. Hofmann, Probabilistic latent semantic indexing, in SIGIR ‘99 Proceedings of the 22nd
Annual International ACM SIGIR Conference on Research and Development in Information
Retrieval (1999), , pp. 50–57
18. D. Kanojia, A. Joshi, P. Bhattacharya, M.J. Carman, Using multilingual topic models for
improved alignment in English-Hindi MT. Proceedings of ICON 2015, 245–252 (2015)
19. S. Negi, Mining bilingual topic hierarchies from unaligned text, in Proceedings of the 5th
International Joint Conference on Natural Language Processing, Chiang Mai, Thailand, 8–13
Nov 2011, pp. 992–1000
20. L. Deleger, T. Lingren, Y. Ni, M. Kaiser, L. Stoutenborough, K. Marsolo, M. Kouril, K. Molnar,
I. Solti, Preparing an annotated gold standard corpus to share with extramural investigators for
de-identification research. J. Biomed. Inform. 50, 173–183 (2014)
21. B. Santorini, Part-of-speech tagging guidelines for the Penn Treebank project. Retrieved from
https://www.cis.uni-muenchen.de/~schmid/tools/TreeTagger/data/Penn-Treebank-Tagset.pdf
22. G. Leech, Corpus annotation schemes. Lit. Linguist. Comput. 8(4), 275–281 (1993)
23. E. Hovy, J. Lavid, Towards a ‘science’ of corpus annotation: a new methodological challenge
for corpus linguistics. Int. J. Transl. 22(1) (2010)
24. I. Rish, An empirical study of the naive Bayes classifier, in IJCAI 2001 Workshop on Empirical
Methods in Artificial Intelligence, vol. 3, No. 22 (2001)
25. M. Hall, A decision tree-based attribute weighting filter for Naive Bayes. Working Paper:
05/2006, Department of Computer Science, The University of Waikato, Hamilton, New Zealand
(2006)
26. H. Chih-Wei, C. Chih-Chung, L. Chih-Jen, A practical guide to support vector classification.
Technical Report, Department of Computer Science and Information Engineering, University
of National Taiwan, Taipei (2003), pp. 1–12
27. Y. Yang, X. Liu, A re-examination of text categorization methods, in Proceedings of the 22nd
Annual International ACM SIGIR Conference on Research and Development in Information
Retrieval, Berkeley, CA (1999), pp. 42–49
28. A.R. Balamurali, A. Joshi, P. Bhattacharya, Cross-Lingual sentiment analysis for Indian
languages using linked wordnets, Proceedings of COLING 2012: Posters, pp. 73–82, COLING
2012, Mumbai, Dec 2012
Firewall Scheduling and Routing Using
pfSense
M. Muthukumar, P. Senthilkumar, and M. Jawahar
Abstract Firewalls are used to safeguard essential networks from outdoor attacks to
guide network access based on the firewall access rules. The firewall system plays an
important aspect that protects from rule analyst and malignant attack which provides
security to all the internet users. The main purpose of this firewall system is to
manage the network access to or from a secured network. Some difficulty with the
process of firewall system is due to malfunction; it might be terrible to other fewer
secured systems on the internal network. The detecting malignant packets are very
significant in security issues. Therefore, we proposed a Similarity Index Algorithm
which is to detect the malignant packets in the firewall framework. The performance
of the proposed firewall system is tested using Netsim simulator environment in
terms of latency and malignant packet detection rate with respective to the number
of nodes or computers in network. Experimental studies were conducted in various
schemes using Similarity Index Algorithm. The result shows the enhanced perfor-
mance of packet delivery ratio. To achieve the above objective, three approaches
have been proposed for summarization. In the first approach, firewall access rule
routing and scheduling using pfSense scheme is employed. The pfSense is a soft-
ware tool that provides enthusiastic support to firewall system. The pfSense can be
enhanced through web system. This method adds two phases, namely rule fixture
and rule matching. The evaluation of the proposed approach is done with the help
of scheduling in the time interval. While comparing with existing approach, the
range can be calculated with 95% of latency. In the second approach, data accessing
can be processed by analyzing the real problem in packet confinement. This second
approach realizes the Deep Packet Confine (DPC) and Deep Packet Assessment
M. Muthukumar (B)
Department of Computer Science and Engineering, SRM Institute of Science and Technology
Delhi-NCR Campus, Ghaziabad, Uttar Pradesh, India
e-mail: mkumar7680@gmail.com
P. Senthilkumar ·M. Jawahar
Department of Computer Science and Engineering, NREC, Secunderabad, India
e-mail: psenthilmephd@gmail.com
M. Jawahar
e-mail: mjawahar@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_67
749
750 M. Muthukumar et al.
(DPA) to evaluate high traffic rate. The updated feature helps to assess disputation
analyses to determine security threat. The model explains packet framing and packet
filtrate. The framing model provides records to organize information and client data.
The filtrate investigates entry and exit by granting communication using specified
rules. The Deep packet assessment enables the client service, and summary is gener-
ated using diagnostic tools. In the third approach, experiment evaluation shows that
the proposed method has the capability of perceiving a movement percentage of
new attacks. It explains that the system detection can be developed by using Simi-
larity Index Algorithm. The Similarity Index Algorithm analyzes the inward packets
recognized using malignant packet detection and decides to precede packets through
gateway. The implementation in firewall contributes powerful security that can be
applied to all network traffic. The efficient summary is to assure data communication
mechanism in the networks. The result demonstrates that the gateway operation can
be turned on by investigating each packet. Finally, the result shows that the latency
and malignant packet detection rate of the proposed firewall architecture are 14.74 ms
and 87%, respectively.
Keywords Firewall system ·Firewall access rule ·pfSense
1 Introduction
In this modern world, Firewall plays an important role in network security that can
be a hardware or software or combination of both. It is used to protect the network
actions from the hackers and malicious nodes in network environment. In the year
1994, the idea of firewall and internet security has been introduced by Steven. M.
Bellovin. A firewall is a security system that detects and manages the succeeding
and departing packets based on the firewall access rules and regulation. The firewall
is placed within the boundary of each computer which is connected to internet by
the networks in the organization. Based on the predefined firewalls access rules, the
public network and private networks can be segregated. The security of the system is
dependent on the rules of the firewall configuration, or else undesired packet traffic
may pass or block the desired packets. The main function of the firewall is to control
the security policy and to protect the organization network from non legitimate traffic.
It also provides high flexible security to online computer users. Firewalls can be
achieved by testing all constrained and unconstrained network traffic according to
the predefined rules [1]. Figure 1shows the general firewall model communication
with LAN, WAN, and Enterprise or Organization Networks. The firewall is situated
at the junction point between the three networks such as LAN, WAN, and Enterprise
or organization network along with the internet connection.
Firewall Scheduling and Routing Using pfSense 751
Fig. 1 General firewall
model
1.1 Packet Filtrate
The packet filtrate is referred to as static packet filtering. Based on the IP address
of source and destination, the incoming and outgoing packets can be controlled and
monitored either to pass or halt the information.
The packet filtrate is a cost-effective security to attack against outside networks.
The general packet filtrate algorithm is as follows:
Input: List of packet header data, Access rule(r), Packet (P)
Output : Allow or disallow the packet (P)
if(Access rule(r)==accept)
then P ← accept
else if (Access rule(r) =disallow)
then P←Denied
else
r is not defined
process stop
end if
//repeat or iterate the process when access rule define
d
Stop
1.2 Firewall Access Rules
Firewalls can examine the firewall access rules. The access rules are network security
rules that can be set by the network authority to allow traffic to their respective web
hosting servers, FTP archives and daemon servers, thereby giving the computer
owners immense control over the traffic that flows in and out of their systems or
752 M. Muthukumar et al.
networks. In distributed firewalls, no two firewalls should have the same access
rules and regulation [2]. Normally, all the traffic in internet can be monitored and
controlled by giving the firewall installer a high level of security and protection over
the network.
1.3 Firewall Access Rule Design
The access rule design is a complete group of access rules which need to be designed.
This access rule explains which network traffic flows through the firewall, traffic that
enters or blocks the organization network (Al-Shaer et al. 2005).
Firewall Access Rule Steps
•The system engineers must be able to connect instantly with the firewall system.
•Firewall might not be able to connect instantly with any other network devices.
•Any other device should be able to connect directly with firewall system.
•The network traffic flow should be running instantly to the specified servers.
2 Firewall Access Rule Routing
The firewall routing is used to deliver the packet from source to destination and
sending it through one domain network environment to other domain environment.
Routing policy allows you to manage the routing in sequence between the routing
properties and the routing tables. These firewall systems support the following routing
areas:
•Stateful packet filter
•Network Address Translation (NAT)
•Filtrating based on Firewall access rule
•Packet Matching
•Packet Cleaning.
3 Software Tools––pfSense
The pfSense is a freely available firewall software based on allocation of operating
system. It is a physical arrangement of the computer system to make an enthusiastic
support of firewall system. It can be designed and enhanced through a web-based
system. The pfSense is generally set up as a boundary of firewall system, networking
devices like router, repeater, and wireless access point [3].
Firewall Scheduling and Routing Using pfSense 753
4 Firewall Access Rule System Implementation
The firewall access rule permits to keep the malignant users out and also expand
control over inherent risky users within your company. An access rule is to realize
the available information and services, present inherent for spoilage and whether
any security is already in place to inhibit misuse [4]. The policies are traffic rules,
regulations for the network which build up the internet. Network policy allows system
administrator to coordinate network elements to offer service to set of users. Each
system were permitted to connect with all other neighboring systems without any
limitation, then there would be no access rules for network. Network policy can be
stimulated into two different ways such as fixed and energetic. A fixed policy is a set
of action in a preplanned way according to a set of pre-built attributes. An energetic
policy is imposed in need, and it is based on mitigating condition.
Firewall Access Rule design
•The system engineers must be able to connect instantly with the firewall system.
•Firewall might not be able to connect instantly with any other network devices.
•Any other device should be able to connect directly with firewall system.
•The network traffic flow should be running instantly to the specified servers.
Figure 2represents the firewall access rule diagram. The firewall access rule can be
done by passing the packet, matching the same access rule, blocking the unspecified
access rule, and rejecting the unknown user. It can also identify the unauthorized and
authorized traffic in the networks.
Fig. 2 Firewall access rule diagram
754 M. Muthukumar et al.
Fig. 3 Firewall access rule scheduling diagram
5 Firewall Access Rule Scheduling Implementation
The firewall access rule scheduling can be planned to be active only at transparent
period of time. The scheduled access rule will proceed while they are not available,
when the planned stage is not agile.
Rules for Firewall scheduling
The twisted schedule resolves only when the firewall access rule operation will
be registered.
This access rule will not pertain at the edge of the schedule and will be served by
pfSense, as it is not adjacent. Figure 3represents the access rule scheduling.
6 Conclusion
Firewall is obvious to connect local corporate network to the Internet. It prevents
the corporate network from different threats and attacks. However, firewall tools can
be updated. The vulnerabilities, hazards, and current controls are analyzed along
with establishing the collateral policy. The goal of thesis is to determine access
rules for routing and scheduling using the pfSense scheme and also detecting the
malignant packets. During the initial stage of investigation, an exhaustive study on
firewall is made to achieve the framework. It regulates the factors of firewall access
rule authentication, authorization, and regulation to secure the threaten policy in
networks. The study on knowledge discovery techniques establishes firewall access
rules by extracting its network traffic based on the logging system.
A review on the Deep Packet Confine (DPC) to present complete network speed,
network packets consignment, and crossing a network with a high traffic flow rate
is done and investigated on the Deep Packet Assessment (DPA) for determining the
security threat and monitor the network traffic in real-time environment. Parametric
studies on the protocol diagnostic tools and methods to confine the process of network
logging traffic are successfully investigated. In the final phase, malignant packet
Firewall Scheduling and Routing Using pfSense 755
detection algorithm is proposed in firewall architecture which efficiently detects
the unauthorized packets from various network environments. The frameworks are
based on the computation of the energy index of the individual port in firewall
architecture. The performance of the proposed system is analyzed in terms of latency
and malignant packet detection rate. This simulation resolves 14.74 ms of latency
and 87% of malignant packet detection rate.
7 Limitations
•A firewall cannot guard from inner intrusion. It does anything to block internal
network intruders or intrusion attack from or within the network. In organization,
employee’s offense or inattention cannot be inhibited by firewall system.
•A firewall cannot prevent discrete employee or internet user with device from
twisting interior or exterior part of the network. This helps the user to completely
pass around the firewall system.
•A firewall system provides security if it is perfectly constructed and defined the
firewall access rule. A firewall administrator should design it to classify between
accept and denied network traffics.
•A firewall cannot bypass your password rules or contamination of passwords.
•A firewall is ineffectual across unspecialized exemption risk.
•Firewalls cannot measure against the hybrid attacks.
8 Future Work
It is examined that the future control for this research on the corporate network
security in the connection of firewall can be achieved by dealing with central research.
Using proposed Knowledge Discovering Technique with log-supported database
system, further inquisition can be drifted out on the impact of mesh packet filtrate.
The issues measured can be recognized. It is accessible to design new mechanism
that can authenticate acquaintance through proxy server. Finally, it can determine the
firewall manner by the whole research to prevent malicious attacks and to acquire
performance in network security.
References
1. E.S. Al-Shaer, H.H. Hamed, Modeling and management of firewall policies. IEEE Trans. Netw.
Serv. Manag. (2004)
2. M.J. Chapple, J. D’Arcy, A. Striegel, An analysis of firewall rulebase (mis)management
practices. ISSA J. 12–18 (2009)
756 M. Muthukumar et al.
3. K. Mamat, R.M. Saad, Home wireless network security using pfSense captive portal,
in Proceedings of 8th International Conference on IT in Asia 2013 (CITA’13)
(IEEE/SCOPUS/ISI). Accessed: 12 Apr 2016
4. M. Abedin, S. Nessa, L. Khan, E.S. Al-Shaer, M. Awad, Analysis of firewall policy rules using
data mining techniques. Int. J. Internet Protoc. Technol. (2010)
5. C. Chao, A flexible and feasible anomaly diagnosis system for internet firewall rules, in
Proceedings of the 13th Asia-Pacific Network Operations and Management Symposium (2011),
pp. 1–8
6. Z. Chen, W. Dong, H. Li, P. Zhang, X. Chen, J. Cao, Collaborative network security in multi-
tenant data center for cloud computing. Tsinghua Sci. Technol. 19, 1 (2014)
7. K.K.R. Choo, The cyber threat landscape: challenges and future research directions. Comput.
Secur. (2011)
8. T.H. Cormen, C.E. Leiserson, R.L. Rivest, C. Stein, Introduction to Algorithms (McGraw-Hill
Higher Education, 2009)
9. A. El-Atawy, K. Ibrahim, H. Hamed, E. Al-Shaer, Policy segmentation for intelligent firewall
testing, in Proceedings of the 1st IEEE ICNP Workshop on Secure Network Protocols (2005),
pp. 67–72
10. A.El-Atawy,T.Samak,A.Wali,E.S.Al-Shaer,F.Lin,C.Pham,S.Li,Anautomatedframework
for validating firewall policy enforcement, in IEEE International Workshop on Policies for
Distributed Systems and Networks (2007), pp. 151–160
11. M. Eslahi, M. Naseri, H. Hashim, N. Tahir, E. Saad, BYOD: current state and security chal-
lenges, in 2014 IEEE Symposium on IEEE Computer Applications and Industrial Electronics
(ISCAIE) (2014)
12. A. Gontarczyk, P. McMillan, C. Pavlovski, Cyber security zone modeling in practice, in
Proceedings of the 10th International Conference on Information Technology and Applications
(ICITA) (Sydney, Australia, 2015)
13. L. Guangcheng, Research and design of double firewall technology. Comput. CD Softw. Appl.
21, 70–81 (2012)
14. H. Guo, T. Tang, D. Wu, The research of private network secure interconnection scheme
in large-scaled enterprises, in Genetic and Evolutionary Computing (Springer International
Publishing, 2015)
15. H. Guochao, An improved strategy for intranet security based on two firewalls. Comput. Secur.
7, 36–38 (2012)
16. X. He, T. Chomsiri, P. Nanda, Z. Tan, Improving cloud network security using the tree-rule
firewall. Future Gen. Comput. Syst. 20, 1 (2014)
17. S. Jadhav, R. Agrawal, Fast and scalable method to resolve anomalies in firewall policies. Int.
J. Adv. Innov. Res. (IJAIR) (2013)
18. H. Jang, J. Jeong, H. Kim, J. Park, A survey on interfaces to network security functions in
network virtualization, in 2015 IEEE 29th International Conference on Advanced Information
Networking and Applications Workshops (WAI NA ) (IEEE, 2015)
19. J. Li, The research and application of multi-firewall technology in enterprise network security.
Int. J. Secur. Appl. 9(5), 153–162 (2015)
20. K. Kaur, D. Rao, Automation the process of unifying the change in the firewall performance.
Int. J. Comput. Sci. Netw. Secur. (IJCSNS) (2014)
21. T. Kaur, V. Malhotra, D. Singh, Comparison of network security tools-firewall, intrusion
detection system and Honeypot. Int. J. Enhanced Res. Sci. Technol. Eng. (2014)
22. S. Kikuchi, Y. Matsumoto (eds.), Performance modeling of concurrent live migration operations
in cloud computing systems using prism probabilistic model checker.Cloud Comput. (CLOUD)
(2011)
23. L. Kong, Research of building enterprise network security system based on cloud computing
technology. Biotechnol. Indian J. (2014)
24. S. Lar, X. Liao, A. Rehman, M. Qinglu, Proactive security mechanism and design for firewall.
J. Inf. Secur. 2(3), 122–130 (2011)
Firewall Scheduling and Routing Using pfSense 757
25. O. Osanaiye, H. Cai, K.-K.R. Choo, A. Dehghantanha, Z. Xu, M. Dlodlo, Ensemble-based
multi-filter feature selection method for DDoS detection in cloud computing. EURASIP J.
Wirel. Commun. Netw. (2016)
26. O. Osanaiye, K.K.R. Choo, M. Dlodlo, Distributed denial of service (DDoS) resilience in cloud:
review and conceptual cloud DDoS mitigation framework. J. Netw. Comput. Appl. (2016)
27. V. Prokhorenko, K.-K.R. Choo, H. Ashman, Web application protection techniques: a
taxonomy. J. Netw. Comput. Appl. (2016)
28. S. Dandamudi, T. Eltaeib, Firewalls implementation in computer networks and their role in
network security. J. Multidisc. Eng. Sci. Technol. (JMEST) 2(3) (2015)
29. K. Salah, R. Boutaba (eds.), Estimating service response time for elastic cloud applications, in
2012 IEEE 1st International Conference on Cloud Networking (CLOUDNET) (2012)
30. S. Shin, L. Xu, S. Hong, G. Gu, Enhancing network security through software defined
networking (SDN), in 25th International Conference on Computer Communication and
Networks (ICCCN) (IEEE, 2016)
31. A. Shukhman, P. Polezhaev, Y. Ushakov, L. Legashev, V. Tarasov, N. Bakhareva, Development
of network security tools for enterprise software-defined networks, in Proceedings of the 8th
International Conference on Security of Information and Networks (ACM, 2015)
32. Y. Smirnov, V. Halimon, Double layer gateway model for connection between production
network and enterprise network, in International Review of Automatic Control (IREACO)
(2015)
33. A.K. Vasu, A. Sudarsan, P. Ayyappan, A. Ganesh, V. Gokul, Improving firewall performance
by eliminating redundancies in access control lists. Int. J. Comput. Netw. (2014)
Undefeatable System Using Machine
Learning
Anand Sharma and Uma Meena
Abstract In machine learning, computer learns from its own experience, i.e., it
stores the tasks it performs and uses it as and when required. Machine learning is
widely used for gaming as it consists of many algorithms which help the computer
to self-learn and become an undefeatable system. Initially, the computer will have
no domain knowledge except the rules of the game but gradually it will learn the
game, and when all the combinations of the moves which can be played are learnt, it
will become the master player, i.e., undefeatable. Games like chess, Go, Shogi, i.e.,
Japanese chess use reinforcement learning which comes under the domain of machine
learning and maximizes the rewards [1–3]. The learning phase of the computer
includes both self-learning and learning by playing with human. The paper presents
a brief study of machine learning technologies which helps in making the system
undefeatable. That includes Integration of Machine Learning for Behavior Modeling,
Reinforcement Learning, and Goal-oriented Behavior.
Keywords Machine learning ·Game playing ·Reinforcement learning ·
Undefeatable system
1 Introduction
1.1 Puzzles
A puzzle is a game like crossword, number puzzles, word search puzzles, or logic
puzzles. They are generated to be a way of entertainment. Some are easy, some are
hard, and some require a little bit of head scratching. Puzzle games involve use of
A. Sharma
Department of Computer Science, SET-MUST, Lakshmangarh, India
e-mail: anand_glee@yahoo.co.in
U. Meena (B)
Department of Computer Science, SRMIST, Delhi-NCR Campus, Ghaziabad, India
e-mail: uma.b18@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_68
759
760 A. Sharma and U. Meena
your brain to solve challenges that become harder over time. It is a kind of brainteaser.
Envision a puzzle, where every one of the pieces appears the same, but then they all
have their little contrasts all over. Such little contrasts but then you cannot build the
puzzle until the point that you perceive the similitude. When you at last do, you are
prepared to begin building and piece by piece the picture begins to shape. Piece by
piece you match and you make. The pieces are the interminable potential outcomes
you have; the decision you make to consolidate one piece with another is the decision
you make which drives you to the finish of the puzzle.
1.2 Two-Player Perfect-Information Games
Two player games come under the category of multiplayer game which is played
by just two players. One of the players can be a human or computer. Few games
which come under this category are Chess Go, Tic Tac Toc, English Draughts, etc.
[4–6]. Here, the game yields three results, i.e., either it is win, lose, or draw. This
involves board game, dice game, or card game. These types of games involve timers
and scoring, i.e., the player who scores maximum in the given time bound wins the
game. These types of games include both indoor and outdoor games. Here, if we
have one of the players as the computer, then we can train it using Reinforcement
Learning, i.e., teaching the system tricks to win such that it becomes impossible for
the player to defeat the system, and we get only two cases as the result that is either
the system wins or it is a draw. Here, each player plays their move alternatively and
is aware of the previous moves, i.e., has perfect information about the game.
1.3 Imperfect-Information and Stochastic Games
In these types of games, the player plays according to the current state and actions
of the opponent player. It is a dynamic game with probabilistic transitions, i.e.,
not certain. In this, the process of playing according to the current state keeps on
repeating which may be finite or infinite times until the end of the game is reached.
These games are mostly used for analysis and modeling of discrete systems which
are being operated in an unknown environment. Zero-sum games come under this
category where the total gains are being added up and the losses are being separated,
and therefore the total sum becomes equal to zero. Zero-sum games are mostly
solved using the MinMax algorithm. The zero-sum property (if one wins, second
loses) means that any result of a zero-sum situation is Pareto optimal. All the games
where complete strategies are Pareto optimal are called a conflict game. Poker and
Gambling are the zero-sum games since the summation of the number of wins by
some players equals the number of losses of the others, and the information is also
incomplete as the opponent is not sure of what the player’s next move is. Markov
Undefeatable System Using Machine Learning 761
Fig. 1 Conceptual view of running artificial intelligence for games
chains are a way to statistically model random processes. They have been used in
many different domains from text generation to financial modeling.
2 Machine Learning in Games
2.1 Samuel’s Legacy
In 1947, Arthur L. Samuel, Professor of Electrical Engineering at the University
of Illinois and pioneer of Artificial Intelligence, brought the idea of developing a
checkers program. Checkers being a simpler game in comparison to chess seemed
to be a perfect province for representing the power of symbolic computing with a
quick programming project. He made the contribution in the tech world by making
the computers learn from their experience resulting in the process of making an
undefeatable system. The computer which he has trained to play the checkers game
has beaten America’s best checker player of that time. The steps which he used to train
the computer included storing of moves and searching for the saved moves which
is also called as rote learning. Then, he used reinforcement learning for evaluation
functions and finally he made a program that could teach itself by playing against a
copy of it. The program for checkers CHINOOK was the first program to conquest
a human world championship in any game. Later, chess and Go have been evolved
by using AI and Deep Learning (Fig. 1).
2.2 Book Learning
Human players do not always play with their mind instead sometimes they play with
their heart. So in order to tackle such kind of moves, the computer makes use of an
opening book which gives an easy way to cumulate their playing strength. Opening
762 A. Sharma and U. Meena
book is pre-computed for replying a set of positions and can be straightforwardly
programmed. It is also a very simple way to access the human knowledge through
machine which can be originated in game playing books. The opening books can
extend its libraries offline by identifying the lines that are frequently used/played and
compute the best move for those positions. Negamax search with added computer
evaluation is much more informative than straightforward frequency-based methods.
The alternate approach used is learning from the mistakes, i.e., it gets alert when the
same mistake is going to be repeated and hence prevents making the mistake again. If
the information is imperfect, then we use simulation learning instead of the traditional
one where the computer plays against itself and stores the information of the random
moves, and using the stored statistics evaluates the quality of the current move.
2.3 Learning Search Control
This method is used to reduce the time requisite for plan generation from experience.
The two approaches used in learning search control knowledge are explanation-based
learning and inductive learning. The inductive learning can learn from more than one
example at a single point of time and does not need the complete and tractable domain
theory. One of the examples for inductive learning is Grasshopper which works in
following five steps:
(i) Decision Extraction
(ii) Decision Clustering
(iii) Decision Characterization
(iv) Rule Generation
(v) Utility Optimization.
2.4 Explanation-Based Learning
The explanation-based learning takes an example and explains what it learns from
it. The relevant data is taken from the example by EBL and translated in a form
which can be used for problem solving. PRODIGY is a system which uses a single
architecture to integrate problem solving, planning, and learning. The control rules
in this can lead to the choice/rejection of the candidate and also leads to a partial
ordering of the candidates. The results from the problem-solving trace generated
by the central problem solver are used by the EBL module. It then constructs the
explanations that describe the domain and the architecture of the problem solver
using an axiomatized theory. Control rules are obtained from the results and then
added to the knowledge base which later on is used to guide the search process
effectively.
Undefeatable System Using Machine Learning 763
2.5 Learning Playing Strategies
It is a simple playing strategy and an interpretable procedure used to play game or
subgame. Learning such a procedure is altogether harder than to figure out how to
group positions as won or lost. Mesa effect occurs in the game of chess, i.e., problem
of choosing move from equally good moves. So in order to resolve this problem, a
simple strategy was used that checks the move which contains the shortest distance
to win.
3 Proposed Undefeatable System
This section will cover the three different approaches by which we can implement
undefeatable systems.
3.1 Integrated Machine Learning for Behavior Modeling
A decent player will learn the conduct of the adversary AI and start misusing it. The
misuse of these imperfections ought to be a part of the diversion; anyway the amuse-
ment ought to likewise continue testing skilled players. To utilize a machine learning
calculation, it is first important to decide the highlights [7]. We utilize supervised
learning since we are going about as an instructor to the learning algorithms. We
determine which cases ought to be thought of and which not. A supervised learning
algorithm will part our data set into two sections: a training set and a test set. The
training set serves to enable the machine learner to refresh its internal settings in
view of just these cases. The test set is utilized to assess how well the inward learned
capacity will perform on already imperceptible data.
3.1.1 The Problem and the Data Set
In this, the players look out to win the game, and the problem of the player is how to
do it. The data set is the set of tools or inputs to be applied, following the rules of the
game at the right time by analyzing the opponent’s behavior to win the game. The
best tool used at the best time leads to the survival of the fittest in the game. This use
depends on the right analysis and the learning algorithm which includes tricks and
tactics required for training of the computer.
764 A. Sharma and U. Meena
Fig. 2 Conceptual view of behavior modeling
3.1.2 Learning from the Data
The computer analyzes the moves of the opponent; in the initial stage, the perfor-
mance of the system might be low, and it may even lose to the opponent. But as the
number of iterations increases, the performance of the game also increases. After the
complete analysis of the whole possibilities of the opponent, the learning phase of the
game is completed. The most appropriate and optimized learning algorithm is used
to get the best outcome. For this, the comparison of different learning algorithms
should be made depending upon their execution time and their efficiency to give the
desired output.
3.1.3 Ensemble Methods: Boosting
It is a set of trained classifiers. An ensemble is more precise than single classifier. The
two popular methods used in this are boosting and bagging. Another term for boosting
is bias. The goal is to decrease the boosting, variance, and taming the stacking alias
ensemble in the supervised learning model [8]. The concept of boosting came from
the question by a researcher, that is, “Can a set of weak learners create a single
strong learner?” Initially, the hypothesis boosting problem was referred to as making
a strong learner out of a weak learner. Boosting signifies those algorithms which
attain hypothesis boosting hastily (Fig. 2).
3.2 Reinforcement Learning
Reinforcement learning comes under the domain of Machine Learning and is inspired
from human behavior. In this, the computer maximizes its performance by acting
correctly and attaining more points than penalties. It interacts with the environment
and learns. Here, the computer is not told what action it should take instead it should
decide on its own that which action will give the maximum performance. In rein-
forcement learning, it is not necessary to present the input and output pair and need
Undefeatable System Using Machine Learning 765
Fig. 3 Conceptual view of reinforcement learning for games
not correct the sub-optimal actions explicitly. It does not have direct input–output
mapping. A good example of reinforcement learning is chess where we have thou-
sands of moves, and it will be very difficult for a human to teach all the combinations
to the computer; therefore, the computer learns from its own experiences and builds
the knowledge base [9,10]. It uses Markov Decision process, i.e., repeating the
process of selecting the best action according to the environment. However, unlike
to the settings in supervised learning, the agent does not receive any information
regarding training from a domain expert. Instead, it may itself discover the different
actions and, while doing so, will get feedback from the environment—the so-called
reinforcement or reward—which can be utilized to rate the success of its own actions
[11] (Fig. 3).
3.3 Goal-Oriented Behavior
Goal-oriented behavior control rule consists of a series of behavior in order to ample
a task. Such type of control rules are delivered by reinforcement learning, planners,
or human programmers. In case of exceptional cases, it becomes difficult to build
a new behavior plan as done in all the working conditions. Therefore, planning
reinforcement learning is partially effective instead the computer may learn subgoals
and change their sequence in order to complete the task. Subgoals are the state which
computer necessarily goes through for the execution of task. This will save time, be
cost effective, and reduce search space. By using subgoals, the computer will be
able to explore more and accelerate learning if the environment is same or similar
by repeated use of subgoal. Subgoals also produce sequence of behavior for task
execution. Once the computer gets fully trained using the above techniques, it will
become an undefeatable system as it will know each and every trick of humans
(Fig. 4).
766 A. Sharma and U. Meena
Fig. 4 Conceptual view of goal-oriented behavior for goal attainment
4 Conclusion
Earlier, the computers were not that much powerful as humans, but with the advance-
ments in technology, they are giving a tough fight to humans. In the field of gaming,
whether it is puzzle or a video game, now the computers can beat the humans
with much more accuracy. In order to give computers the power to beat humans,
Machine Learning is used so that the result of the game is either the computer wins
or it is a draw. Machine learning is an application use of AI which provides the
computer an ability to learn automatically and develop from experience without
being programmed externally. It enables analysis of large amount of data at a faster
speed and gives accurate results. Therefore, we use it for making the system undefeat-
able by using the techniques like reinforcement learning and goal-oriented behavior
through which it will train itself.
References
1. Online chess games database, 365chess, 2017. https://www.365chess.com/
2. J. Schaeffer, The games computers (and people) play, in ed by M.V. Zelkowitz, Advances in
Computers, vol. 50, pp. 189–266 (2000)
3. M. Genesereth, N. Love, B. Pell, General game playing: overview of the AAAI competition.
AI Mag. 26(2), 62–72 2005)
4. G. Tesauro, Comparison training of chess evaluation functions, in ed by J. Fürnkranz, M. Kubat,
Machines that Learn to Play Games, Chapter 6, pp. 117–130 (2001)
5. O.E. David, N.S. Netanyahu, L. Wolf, Deepchess: End-to-end deep neural network for auto-
matic learning in chess, in International Conference on Artificial Neural Networks (Springer,
2016), pp. 88–96
6. D. Silver, J. Schrittwieser,K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker,
M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, D.
Hassabis, Mastering the game of go without human knowledge. Nature, 550, 354– 359 (2017)
7. B. Geisler, An Empirical Study of Machine Learning Algorithms Applied to Modeling Player
Behavior in a ‘First Person Shooter’ Video Game. M.S. Thesis, Department of Computer
Sciences, University of Wisconsin Press. Madison, 2002
Undefeatable System Using Machine Learning 767
8. Q. Le et al., Building High-level Features Using Large Scale Unsupervised Learning, in ICML
2012: 29th International Conference on Machine Learning (2012, June)
9. V. Mnih et al., Human-level control through deep reinforcement learning. Nature 518, 529–533
(2015)
10. M. Lai, Giraffe, Using deep reinforcement learning to play chess, Master’s thesis, Imperial
College London, 2015
11. S. Schiffel, M. Thielscher, Fluxplayer: a successful general game player, in Proceedings of the
AAAI Conference on Articial Intelligence (Vancouver, 2007), pp. 1191–1196
12. S. Schiffel, M. Thielscher, Fluxplayer: a successful general game player, in Proceedings of the
AAAI Conference on Artificial Intelligence (Vancouver, 2007), pp. 1191–1196
13. M. Bowling, Convergence and no-regret in multiagent learning, in Advances in Neural
Information Processing Systems 17 (NIPS) (2005), pp. 209–216
14. C. Claus, C. Boutilier, The dynamics of reinforcement learning in cooperative multiagent
systems, in Proceedings of the National Conference on Artificial Intelligence (Wiley, 1998),
pp. 746–752
15. S. Kapetanakis, D, Kudenko, Reinforcement learning of coordination in cooperative multi-
agent systems, in Proceedings of the National Conference on Artificial Intelligence (AAAI
Press, MIT Press, Menlo Park, Cambridge, 2002), pp. 326–331
Synchronization for Nonlinear
Time-Delay Chaotic Diabetes Mellitus
System via State Feedback Control
Strategy
Nalini Prasad Mohanty, Rajeeb Dey, Binoy Krishna Roy,
and Nimai Charan Patel
Abstract Recent studies disclose that the interaction of externally injected insulin
and blood glucose in some nonlinear model of diabetes mellitus results in chaotic
behavior. This work deals with the design of a state feedback synchronization control
technique for a nonlinear time-delay glucose–insulin regulatory system where the
β-cell kinetic stakes into account in the model. A state feedback control technique is
used for synchronization between master and slave system where the master system
is the healthy diabetes mellitus system and the slave system is the diseased diabetes
mellitus system. Here, Lipschitz condition lemma is used for approximating the
system nonlinearity. A Lyapunov–Krasovskii functional approach is adopted for
deriving the feedback control law which is responsible for the synchronization of
chaotic behavior present in the glucose–insulin regulatory system of diabetic patients.
Wirtinger inequality is utilized in the design that resulted in less conservative stability
conditions in terms of linear matrix inequality. The results of simulation for the
proposed synchronization technique confirm the efficacy toward healing the diabetes
mellitus.
Keywords Synchronization ·Nonlinear time-delay system ·State feedback
control ·Diabetes mellitus ·Wirtinger inequality
N. P. Mohanty (B)·R. Dey ·B. K. Roy
Electrical Engineering Department, National Institute of Technology Silchar, Silchar, Assam
788010, India
e-mail: nalini6111@gmail.com
R. Dey
e-mail: rajeeb.iitkgp@gmail.com
B. K. Roy
e-mail: bkr_nits@yahoo.co.in
N. C. Patel
Electrical Engineering Department, Government College of Engineering, Keonjhar, Odisha
758002, India
e-mail: ncpatel.iter@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_69
769
770 N. P. Mohanty et al.
1 Introduction
The technical name of diabetes is diabetes mellitus, which refers to the change in
the normal processes of human metabolism toward control mechanism disturbance
for blood sugar level. The presence of primary controlling element, i.e., insulin, is
either disregarded by body cells or not secreted. Generally, diabetes is categorized as
gestational diabetes (GD), type 1, and type 2. The GD type of diabetes is pregnancy-
based temporary condition, which affects approximately 2–4% of all pregnancies.
The immune system of patient interrupts insulin-releasing cells in Type 1 diabetes.
Usually 5–10% of patients are coming under this category. Whereas, the body of type
2 diabetic patient is unable to use insulin in proper way. This is the most common
category (around 90–95%) in the world [1].
A linear model for glucose tolerance test is suggested by Ackerman in the year
of 1964 [2]. In 1987, Bajaj introduced a basic nonlinear mathematical model using
β-cell kinetics for glucose–insulin feedback system [3]. A time-delay model which
incorporates two time-delays has been proposed by Sarika et al. [4]. The modified
model of Sarika et al. [4] is presented by Chuedoung et al. [5].
There are many time-delay chaotic systems (TDCS) are present in nature. There
are so many applications of TDCS in the field of engineering, physics, and biology
[6,7]. A system with time-delay can exhibit chaos [8,9]. The phenomena of chaos,
hyperchaos, multistability, etc., may be exhibited by systems because of the existence
of time-delays in the system dynamics.
Many synchronization techniques have been reported in the literature like state
feedback control [10], sliding mode control (SMC) for uncertain TDCS [11], and
synchronization and anti-synchronization of TDCS using SMC which are also
discussed in the literature [8,12].
This paper has five sections. Section 2contains the literature survey, background
and motivation, and system description. Section 3contains the design of state feed-
back controller (SFC) for synchronization of TDCS. The simulation result is analyzed
in Sect. 4. The concluding remark is given in Sect. 5.
2 System Description
The model defined by Chuedoung et al. [5] (Fig. 1) is represented as follows in the
form of delay differential equation.
˙x=r1yτ1zτ1−r2x+c1zτ1
˙y=R3N
z−r4xτ2+c2
˙z=r5y−ˆy(T−z)+Az −r6z(t−z)−r7z(1)
Synchronization for Nonlinear Time-Delay Chaotic Diabetes … 771
Fig. 1 Glucose–insulin system with two time-delays
where yτ1=y(t−τ1),zτ1=z(t−τ1),zτ2=z(t−τ2).
In Eq. (1), the insulin concentration, the glucose concentration, and the number of
β-cells are represented as x(t),y(t),z(t), respectively. The change in basal level and
glucose fasting level is represented as ˆy. Due to increase in blood glucose level, the
delay in insulin secretion is represented by τ1, which is based on clinical evidence
reported in [13]. Based on clinical evidence reported in [14], the delay in glucose
drop is due to increase in insulin level in human body, which is represented as τ2.
The total density and normal number of β-cells are represented as Tand N.
2.1 Bifurcation Diagram
The Bifurcation diagram is graphically represented by changing one of the system
parameter with respect to other parameters that are kept fixed. The bifurcation
diagram is illustrated in Fig. 2, with variation of the delay parameter τ2∈[0,7],
and keeping c1=0.1000,c2=0.8000,r1=0.4720,r2=0.2500,ˆy=1.4200,
R3=0.8200,r4=0.6000,T=1.5000,r5=0.3000, r6=0.3000,N=
1.2700,r7=0.2000,τ
1=0.0430 with initial conditions (ICs) (3.1,1.2,0.4).From
the bifurcation diagram, it is seen that for a range of τ2the system is stable, periodic,
and chaotic in nature. The system is observed stable with small delays; however, it
exhibits chaotic behavior with the increase in delays. The instantaneous reflection
of insulin as reported by Forrest et al. [15] was 14 out of the 20 monitored Jamaican
children with about 1 min response time. However, the delay for other children was
5–10 min.
772 N. P. Mohanty et al.
Fig. 2 Bifurcation of the system (1) with ICs (3.1,1.2,0.4)with respect to τ2
3 State Feedback Control Design for Synchronization
On the basis of (1), the healthy nonlinear time-delay chaotic diabetes mellitus system
(master system (MS)) can be represented in (2).
˙x=Ax +A1x(t−τ1)+A2x(t−τ2)+H¯λ(x,t)+Hϑ(x(t−τ1),t)+C
x=ψx,t∈[−max{τ12,τ
22},0]
(2)
where a1=R3N,a2=r5T,c3=r5ˆyT,A=r5ˆy+r6T−r7.
And the diseased nonlinear time-delay chaotic diabetes mellitus system (slave
system (SS)) is presented in (3).
⎧
⎨
⎩
˙y=Ay +A1x(t−τ1)+A2x(t−τ2)+H¯λ(y,t)
+Hϑ(y(t−τ1),t)+C+Bu(t)
x=ψx,t∈[−max{τ12,τ
22},0]
(3)
Let the error between the healthy and diseased uncertain system is represented as
e=x−y. The resulting synchronization error of MS and SS is represented in (4).
˙e(t)=Ae(t)+A1e(t−τ1)+A2e(t−τ2)+H¯λ(x,y,t)
+Hϑ(x(t−τ1),y(t−τ1),t)−Bu(t)(4)
For realizing the synchronization between the systems denoted in (2) and (3), the
control equation can be represented as
u=Ke (5)
Synchronization for Nonlinear Time-Delay Chaotic Diabetes … 773
where Kis represented as gain matrix. By merging Eqs. (4) and (5), the error system
becomes
˙e(t)=Ae(t)+A1e(t−τ1)+A2e(t−τ2)+H¯λ(x,y,t)
+Hϑ(x(t−τ1),y(t−τ1),t)−BKe(t)(6)
Assumption 1 Nonlinear terms ¯λand ϑsatisfy
||¯λ(x,t)−¯λ(y,t)|| ≤ ||L¯λ(x−y)||,
||ϑ(x(t−τ1), t)−ϑ(y(t−τ1), t)|| ≤ ||Lϑ(x(t−τ1)−y(t−τ1))|| (7)
where L¯λand Lϑwill represent the Lipschitz constant.
Lemma 1 For any positive definite matrix T =TTand any differentiable function
xin[r,s]→Rn, there exists
−
s
r
˙xT(ϕ)T˙x(ϕ )dϕ≤− 1
s−rT
0(r,s)T0(r,s)−3
s−rT
1(r,s)T1(r,s),
(8)
where
0(r,s)=x(s)−x(r)
1(r,s)=x(s)+x(r)−2
s−r
s
r
x(ϕ)dϕ.
Theorem 1 Consider the error system (6) satisfying Assumption 1. Suppose there
exist symmetric matrices P,Q1,Q2,Z1,Z2, and the scalars εkfor k =1,2such
that the succeeding LMI is satisfied.
⎡
⎣
τ
mTZ1τMTZ2
∗−Z10
∗∗ −Z2
⎤
⎦<0(9)
where
=eT
1P+TPe1+−T
1W11−T
2W22+ε1eT
1LT
¯λL¯λe1
−ε1eT
4e4+ε2eT
2LT
ϑLϑe2−ε2eT
5e5
=(A−BK)e1+A1e2+A2e3+He4+He5
1=col{e1−e2,e1+e2−2e6},
2=col{e1−e3,e1+e3−2e7},
W1=diag{Z1,3Z1},W2=diag{Z3,3Z3}.
774 N. P. Mohanty et al.
Proof An LKF candidate is constructed as
V(e,t)=V1(e,t)+V2(e,t)+V3(e,t)(10)
where
V1(e,t)=eT(t)Pe(t)
V2(e,t)=
t
t−τ1
eT(ϕ) Q1e(ϕ)dϕ+
t
t−τ2
eT(ϕ) Q2e(ϕ)dϕ
V3(e,t)=
0
−τ1
t
t+s
τ1˙eT(ϕ) Z1˙e(ϕ )dϕds+
0
−τ2
t
t+s
τ2˙eT(ϕ) Z2˙e(ϕ )dϕds
Hence, the time derivative of V(e,t)is expressed by
˙
V1(e,t)=2eTP[Ae(t)+A1e(t−τ1)+A2e(t−τ2)+H¯λ(x,y,t)
+Hϑ(x(t−τ1),y(t−τ1),t)]
=ζTeT
1P+TPe1ζ(11)
where =(A−BK)e1+A1e2+A2e3+He4+He5,
ζT=[eT(t),eT(t−τ1),eT(t−τ2),¯λT(x,y,t),ϑ
T(x(t−τ1),y(t−τ1),t),
1
τ1
t
t−τ1
eTdϕ, 1
τ2
t
t−τ2
eTdϕ]
˙
V2(e,t)≤eTQ1e−eT(t−τ1)Q1e(t−τ1)+eTQ2e−eT(t−τ2)Q2e(t−τ2)
=ζT{eT
1Q1e1−eT
2Q1e2+eT
1Q2e1−eT
2Q2e2}ζ
=ζTζ (12)
˙
V3(e,t)=ζTTτ2
1Z1+τ2
2Z2ζ −
t
t−τ1
τ1˙eT(ϕ) Z1˙e(ϕ )dϕ
−
t
t−τ2
τ2˙eT(ϕ) Z2˙e(ϕ )dϕ(13)
Considering Lemma 1,(13) becomes
−
t
t−τ1
τm˙eT(ϕ) Z1˙e(ϕ )dϕ−
t
t−τ2
d˙eT(ϕ) Z2˙e(ϕ )dϕ≤−ζTT
1W11ζ−ζTT
2W22ζ
(14)
Synchronization for Nonlinear Time-Delay Chaotic Diabetes … 775
where
1=col{e1−e2,e1+e2−2e6},
2=col{e1−e3,e1+e3−2e7},
W1=diag{Z1,3Z1,5Z1},W2=diag{Z3,3Z3,5Z3}
So combining (11)–(14), we get
˙
V(e,t)≤ζT{eT
1P+TPe1+−T
1W11−T
2W22
+Tτ2
1Z1+τ2
2Z2}ζ(15)
According to Assumption 1,wehave
ε1eTLT
Le−ε1¯λT(x,y,t)¯λ(x,y,t)=ζT(t){ε1eT
1LT
¯λL¯λe1−ε1eT
4e4}ζ(t)≥0,
(16)
ε2eT(t−τ1)LT
ϑLϑe(t−τ1)−ε2ϑT(x(t−τ1), y(t−τ1), t)ϑ(x(t−τ1),
y(t−τ1), t)=ζT(t){ε2eT
2LT
ϑLϑe2−ε2eT
5e5}ζ(t)≥0 (17)
Combining (15)–(17), we get
˙
V(e,t)≤ζT{eT
1P+TPe1+−T
1W11−T
2W22+Tτ2
1Z1+τ2
2Z2
+ε1eT
1LT
¯λL¯λe1−ε1eT
4e4+ε2eT
2LT
ϑLϑe2−ε2eT
5e5}ζ
After some mathematical simplification and Schure complement, the inequality
(9) can be produced. This proves Theorem 1.
Now an adequate condition is derived to obtain gain K.
Theorem 2 An adequate condition for the solution in Theorem 1is that there exist the
positive definite and symmetric matrices X ,Q1,Q2,Z1,Z2, matrix M of appropriate
dimensions, such that the following condition is satisfied.
⎡
⎢
⎢
⎣
eT
1XLT
¯λeT
2XLT
ϑ
∗−00
∗∗ −
ε1I0
∗∗ ∗ −
ε2I
⎤
⎥
⎥
⎦
<0 (18)
where
=eT
1X+XTe1+−T
1W11−T
2W22−ε1eT
4e4−ε2eT
5e5
=(AX −BM)e1+A1Xe2+A2Xe3+He4+He5
776 N. P. Mohanty et al.
=τ1T,τ
2T,=diagXZ−1
1X,X Z −1
2X,
1=col{e1−e2,e1+e2−2e6},
2=col{e1−e3,e1+e3−2e7},
W1=diagZ1,3Z1,W2=diagZ2,3Z2
Proof Based on the Schur complements and congruence transform we can obtain
(18).
4 Simulation Result
The effectiveness of the results is derived to establish the simulation for nonlinear
time-delay chaotic diabetes mellitus system as follows, considering the parameters
of systems (2) and (3):
A=⎡
⎣
−0.25 0 0
000
00.45 0.676
⎤
⎦,A1=⎡
⎣
000.1
00 0
00 0
⎤
⎦,A2=⎡
⎣
000
0.600
000
⎤
⎦
H=⎡
⎣
00 0
01.0414 0
000.3
⎤
⎦,H=⎡
⎣
0.4720 0 0
000
000
⎤
⎦,B=⎡
⎣
10
01
00
⎤
⎦
Attractor of the MS with ICs (3.1, 1.2, 0.4) and τ1=0.042,τ
2=0.07 is shown
in Fig. 3. Attractor of the SS with ICs (5.1, 1.8, 0.2) and τ1=3.63,τ2=3.36 is
shown in Fig. 4. We can obtain the controller gain matrix:
K=14.75 2.28 3.76
15.41.32 6.14
Figure 5shows the synchronized states of MS and SS. The synchronization
errors are given in Fig. 6. Figures 5and 6can conclude that all states of the SS
are synchronized with the MS after 6.2 min.
5 Conclusion
This piece of work deals with the problem of synchronization of the nonlinear time-
delay chaotic diabetes mellitus which has been addressed by utilizing a state feedback
Synchronization for Nonlinear Time-Delay Chaotic Diabetes … 777
Fig. 3 Attractor of the MS
with τ1=0.042,τ
2=0.07,
and ICs (3.1,1.2,0.4)
Fig. 4 Chaotic attractor of SS with τ1=3.63,τ
2=3.36, and ICs (5.1,1.8,0.2)
controller. Here, we are adding controlled input in two states of the system. The
simulation results reveal the successful achievement of the synchronization by the
proposed controller. The unhealthy uncertain time-delay chaotic diabetes mellitus
system follows the healthy uncertain time-delay chaotic diabetes mellitus system.
The future research topic is to design observer-based synchronization control for the
nonlinear time-delay chaotic diabetes mellitus system.
778 N. P. Mohanty et al.
Fig. 5 State responses of SFC-based MS and SS
Fig. 6 Error responses of
SFC-based MS and SS
References
1. Canadian Diabetes Association, Types of Diabetes, Jan 2017. http://www.diabetes.ca/about-
diabetes/types-of-diabetes
2. E. Ackerman, J.W. Rosevear, W.F. McGuckin, A mathematical model of the glucose-tolerance
test. Phys. Med. Biol. 9(2), 203 (1964)
3. J.S. Bajaj, G.S. Rao, J.S. Rao, R. Khardori, A mathematical model for insulin kinetics and
its application to protein-deficient (malnutrition-related) diabetes mellitus (PDDM). J. Theor.
Biol. 126(4), 491–503 (1987)
Synchronization for Nonlinear Time-Delay Chaotic Diabetes … 779
4. W. Sarika, Y. Lenbury, K. Kumnungkit, W. Kunphasuruang, Modelling glucose-insulin
feedback signal interchanges involving β-cells with delays. Differ. Equ. 1(2), 1 (2008)
5. M. Chuedoung, W. Sarika, Y. Lenbury, Dynamical analysis of a nonlinear model for glucose–
insulin system incorporating delays and β-cells compartment. Nonlinear Anal. Theory Methods
Appl. 71(12), e1048–e1058 (2009)
6. K. Ikeda, K. Matsumoto, High-dimensional chaotic behavior in systems with time-delayed
feedback. Phys. D 29(1–2), 223–235 (1987)
7. B. Li, Z. Zhao, R. Wang, G. Ding, Synchronization control design based on Wirtinger inequality
for uncertain coronary artery time-delay system with input saturation. IEEE Access 7, 76611–
76619 (2019)
8. N.P. Mohanty, R. Dey, B.K. Roy, Switching synchronisation of a 3-D multi-state-time-delay
chaotic system including externally added memristor with hidden attractors and multi-scroll
via sliding mode control. Eur. Phys. J. Spec. Top. 229, 1231–1244 (2020). https://doi.org/10.
1140/epjst/e2020-900195-4 (in press)
9. N.P. Mohanty, R. Dey, B.K. Roy, A new 3-D memristive time-delay chaotic system with
multi-scroll and hidden attractors. IFAC-PapersOnLine 51(1), 580–585 (2018)
10. T. Huang, C. Li, W. Yu, G. Chen, Synchronization of delayed chaotic systems with parameter
mismatches by using intermittent linear state feedback. Nonlinearity 22(3), 569 (2009)
11. L. Li-Xiang, P. Hai-Peng, G. Bao-Zhu, X. Jin-Ming, A new sliding mode control for a class of
uncertain time-delay chaotic systems. Chin. Phys. 10(8), 708 (2001)
12. N. Vasegh, F. Khellat, Projective synchronization of chaotic time-delayed systems via sliding
mode controller. Chaos Solitons Fractals 42(2), 1054–1061 (2009)
13. P. Palumbo, S. Panunzi, A. De Gaetano, Qualitative behavior of a family of delay-differential
models of the glucose-insulin system. Discrete Contin. Dyn. Syst. Ser. B 7(2), 399 (2007)
14. R. Prager, P. Wallace, J.M. Olefsky, In vivo kinetics of insulin action on peripheral glucose
disposal and hepatic glucose output in normal and obese subjects. J. Clin. Investig. 78(2),
472–481 (1986)
15. E. Forrest, P. Robinson, M. Hazel, Insulin, growth hormone and carbohydrate tolerance in
Jamaican children rehabilitated from severe malnutrition. MedCarib (1925)
Geo/G/1 System: Queues with Late
and Early Arrivals
Reena Grover, Himani Chaudhary, and Geetanjali Sharma
Abstract Consider various discrete-time systems where the time axis is divided to
put something into a particular space that is designed for it, known as spots. In view
of the services and evacuation could be taken as spot boundaries, where their periods
are integral multiple of spot span. In general, distributed service time is denoted by
Geo/G/1 and single-server system with discrete time has Bernoulli arrival process.
AGeo
x/G/1 system is the basis of all the extended models. Such particular points
are in fact identical with the distribution of the system states observed at an arbitrary
point on the continuous-time domain. This leads a random point which lies in the
middle of a spot on the continuous-time domain with probability 1.
Keywords Discrete time ·Early–late arrival ·Markov chain process ·Waiting
time ·Queue size
R. Grover (B)
Department of Mathematics, SRM Institute of Science and Technology
Delhi-NCR Campus, Ghaziabad, UP, India
e-mail: reenadr1980@gmail.com
H. Chaudhary ·G. Sharma
Department of Mathematics and Statistics, Banasthali Vidyapith,
Banasthali, Tonk, Rajasthan 304022, India
e-mail: himanichaudhary07@gmail.com
G. Sharma
e-mail: geetanjali.bu@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_70
781
782 R. Grover et al.
1 Introduction
1.1 Discrete-Time Systems
Let us call the entity of service a message, whose service time is always an integral
multiple of a spot period. Numbers of messages that arrive during successive slots
constitute a series of random variables which are independent and identical. When
given messages lie in probability λ(0<λ<l)and with probability 1 −λ, one can
examine a mean of 1/λ spots of interarrival time Iwhich is geometrically distributed.
Prob[I=lslots]=λ(1−λ)l−1
where l=1, 2….
Such an arrival process is called Bernoulli process. A Geox/G/1 system is the basis
of all the extended models.
In discrete-time queueing systems, one of the earliest investigations being made
by Meisling [1], Chu and Konheim [2], and Kobayashi and Konheim [3] provide
surveys of applications to computer communication networks. Bruneel and Kim [4]
analyze many general systems with discrete time ones as non-independent arrivals
and multiple servers.
1.2 Late Arrival Model
In late arrival model, the message arrives after a given spot just before border line of
the spot. So arrived messages look at departed messages which are about to go and
the departed message goes just after the message which has just reached. As message
isarrivedinthenth spot till the system is completed, the process and service will
start as (n+1)th spot starts. Hunter [5] specified this model as a delayed access with
the late arrival system. Gravey and Hebuterne [6] stated late arrival model as arrival
first policy.
Geo/G/1 System: Queues with Late and Early Arrivals 783
Therefore, the queue size distribution of first and second moments immediately
after a random spot period is
E[L]= λ2b2−λρ
2(1−ρ) +ρ(1)
E[L2]= λ3(b−3b2+2b)
3(1−p)+3λ2(b2−b)
2(1−p)+λ4(b2−b)2
2(1−ρ)2+ρ(2)
The discrete-time version of Little’s theorem is
E[L]=λE[T](3)
where Tis the time (measured in slots) for a message that stays for arrival to service
completion in the system, called response time of a message. Therefore,
E[T]= λb(2)−ρ
2(1−ρ) +b(4)
which is independent of the service discipline.
If we assume the “First Come First Served (FCFS)” discipline, after service
completion of messages left in system instantly is identical with messages which
784 R. Grover et al.
arrive in given duration during which the messages were in the system. Therefore,
if T(u) denotes the PGF for time Tthat a message stay in the system, then (Hunter
[5]).
(z)=T(1−λ+λz)(5)
The second moment of the distribution of Tis
E[T2]= λ(2b(3)−3b(2)+b)
6(1−ρ) +λ2(b(2)−b)2
2(1−ρ)2+(b(2)−b)
1−ρ+b(6)
E[W]= λb(2)−ρ
2(1−ρ) (7)
E[W2]= λ(2b(3)−3b(2)+b)
6(1−ρ) +λ(b(2)−b)2
2(1−ρ)2(8)
Prob[W=0]=W(0)=1−ρ
1−λ=Prob[L=0]
1−λ>Prob[L=0](9)
A loose argument to justify this inequality is that an arriving message does not
wait if it finds only one message in the system which is about to leave (Gravey [7]).
1.3 Early Arrival Model
In an early arrival model (Hunter [5]), the messages arrive early during a slot. If the
system is empty when a message arrives, its service is started immediately, including
the slot in which the message arrives when we calculate its waiting time. Gravey and
Hebuterne [6] call the early arrival model the departures first policy.
From the normalization condition, we get (cf. Hunter [5], Gravey and Hebuterne
[6])
(z)=(1−ρ)[A(z)−zA(z)]
(1−λ)[A(z)−z]
=(1−ρ)(1−z)B(1−λ+λz)
(1−λ+λz)[B(1−λ+λz)−z](10)
Comparing (z)with early arrival model with (z)with late arrival model, size
of the queue observed instantly afterwards the service completion given in first model
is quite less that is given for second model by the number whose PGF is given by
1−λ+λz, the number of arrivals (zero or one) in a slot. This is because “queue
size” is observed before the possible arrival point in the early arrival model.
Geo/G/1 System: Queues with Late and Early Arrivals 785
Since we count the slot in which a message arrives when we calculate the waiting
time in the early arrival model,
(z)=T(1−λ+λz)
1−λ+λz(11)
The queue size of an early arrival model is observed; it is also the queue size
observed at an arbitrary point during the same slot on the continuous-time domain.
2 Output Process
A sequence of those slot boundaries at which service is completed is called an
output process or a departure process. Consider the interdeparture time of a Geo/G/I
system, which is defined as “time interval” between two “successive service comple-
tions.” The correlation of two contiguous interdeparture times is denoted by τand
τ. Assuming that three successive service completions occur at the end of the n0th,
n0+τth, and n0+τ+τth slots, the joint probability is obtained.
(u,u)
∞
k=1
∞
=1
Prob[τ=k,τ=1]uku(12)
There are five cases regarding the instantly before and after the ends of the n0th
and n0+τth slots and the number of messages that arrive during τslots (Table 1).
2.1 Model
Consider Geo/G/1 system where service time is geometrically distributed as
b() =μ(1−μ)−1=1,2... (13)
where 1/mu is the mean service time.
Tabl e 1 Cases before and after the ends of the n0th and n0+τth slots
Case n0−0n0+0No. of arrivals during τslots n0+τ−0n0+τ+0
(a) 1 0 1 1 0
(b) 1 0 ≥2≥2≥1
(c) 2 1 0 1 0
(d) 2 1 ≥1≥2≥1
(e) ≥3≥2Any ≥2≥1
786 R. Grover et al.
(u,u)=λu
1−(1−λ)u·λu
1−(1−λ)u(14)
which implies that process of departure with rate λis a Bernoulli process.
2.2 Model
The queue size distribution {π∗
k;k=0,1,2...}of the “generating function π∗(z)”
observed either at the end of a slot during with empty system or immediately after a
service completion is
∗(z) π∗kzk=(1−ρ){B[(z)]−z(z)}
(1+λ−ρ){B[(z)]−z}(15)
The probability η(l)that there are l+1 slots between an arbitrary slot boundary
and the last Markov point on the condition that in the given time the system is
non-empty.
η() =1
b
∞
j=+1
b(j)=0,1,2... (16)
In this case, the arrivals during the first lslots contribute to the system state
observed at the starting of the last spot. Therefore, the probability Pk;k≥1, that
there are kmessages,
Pk=
k
j=1
πj
∞
=0
Prob[k−jmessage arrive in slot]η() k=1,2... (17)
Hence, the P(z) of size distribution is {Pk;k≥0}
P(z)P0+
∞
k=1
Pkzk=(1−ρ)(1−z)B[(z)]
B[(z)]−z(18)
This result is also given by Bruneel and Kim [4], Hunter [5], Konheim [3].
P(z)in(18) observed instantly after possible spot partition is generally different
from (z). Instantly, the service is completed in Geo*/G/1 system. However, they
are identical in the Geo/G/1 system.
Then,
(z)=ρ(z)1−(z)
λ(1−z)(19)
Geo/G/1 System: Queues with Late and Early Arrivals 787
holds for a Geox/G/1 system. Bruneel and Kim [4] state that the elementary
observation is true for enormous queuing models.
The event of “no service completion” includes not only a case in which the on-
going service is not completed but also a case in which the service is interrupted by
the server vacation.
E[zL(n+1)]={vE[zL(n)+(n+1)−1
L(n)≥1,service completion ]
+(1−v)E[zL(n)+(n+1)
L(n)≥1,no service completion ]} Prob[L(n)≥1]
+E[z(n+1)]Prob[L(n)=0](20)
However, from the theorem of the total probability,
E[zL(n)+(n+1)
L(n)≥1]=vE[zL(n)+(n+1)
L(n)≥1,service completion ]
+(1−v)E[zL(n)+(n−1)
L(n)≥1,no service completion]
(21)
3 Queue Size and Remaining Service Time
Following Dafermos and Neuts [8], Bruneel and Kim [4] and Hunter [5] recognize the
“joint distribution” of “queue size” and “remaining service time” instantly at every
spot boundary in the late arrival model. Let us consider the queue size L(n)instantly
afterward the nth slot, then the remaining time of service X(n)
+of the message will
be served at same instant, where n=−0, 1, 2….
(, )
() ()
LX
nn
+(,
() ()
LX
nn+
+
+11
nn+1n+2
Λ(n+1)
“Queue size” and “remaining service time” will be recognized instantly afterward
slot boundaries. Considering X(n)
+=0ifL(n)=0. Given L(n)=0ifX(n)
+=0 because
we observe the system after both arrivals, and a service completion possibly occur.
788 R. Grover et al.
The beginning of the first slot “queue size” and “remaining service time” be L(0) and
X(0)
+, respectively.
The sequence {L(n);n=0, 1, 2….} is not a “Markov chain.” However, the sequence
of pairs {L(n),X(n)
+;n=0, 1, 2….} forms a “Markov chain” having (0, 0) as state
space and all integer pairs (k,l) with “k≥1” and “l≥1.” The number of messages
which will arrive in nth slot is denoted by (n), and complete service time is denoted
by X.
L(n+1)=X(n+1)
+=0ifL(n)=X(n)
+=0 and (n+1)=0 (22a)
if L(n)=X(n)
+=1 and (n+1)=0 (22b)
L(n+1)=(n+1)and X(n+1)
+=Xif L(n)=X(n)
+=0or1,and (n+1)≥1
(23)
L(n+1)=L(n)+(n+1)and X(n+1)
+=X(n)
+−1,if L(n)=1 and X(n)
+>1
(24)
L(n+1)=L(n)+(n+1)−1 and X(n+1)
+=X,if L
(n)>1 and X(n)
+=1
(25)
A new service is therefore started at the beginning of the n+2nd slot.
In order to find the steady-state distribution for L(n),X(n)
+;n=0,1,2...r,
P0Lim
n→∞ Prob [L(n)=X(n)
+=0](26a)
k() Lim
n→∞ Prob [L(n)=K,X(n)
+=]k≥1,≥1 (26b)
and
(z,u)
∞
k=1
∞
=1
k()zku|z|≤1,|u|≤1 (26c)
From (22)to(25)
P0=P0λ(0)+1(1)λ(0)(27a)
k() =P0λ(k)b() +1(1)λ(k)b() +
k
j=1
j( +1)λ(k−j)
+
k+1
j=2
j(1)λ(k−j+1)b() k≥1,≥1 (27b)
Geo/G/1 System: Queues with Late and Early Arrivals 789
The normalization condition is
P0+
∞
k=1
∞
=1
k() =1 (27c)
Substituting (27b)into(26c),
(z,u)=[P0+1(1)]
∞
k=1
λ(k)zk
∞
=1
b()u+
∞
k=
∞
=1
zku
k
j=1
j( +1)λ(k−j)
+
∞
=1
b()u
∞
k=1
zk
k+1
j=2
j(i)λ(k−j+1)
By setting u=(z),
∞
j=1
j(1)zj=(1−ρ)z[1−(z)]B[(z)]
(z){B[(z)]−z}(28)
Substituting (28), then
(z,u)=(1−ρ)zu[1−(z)]{B(u)−B[(z)]}
[u−(z)]{B[(z)]−z}(29)
From (29), the PGF P(z) for the queue size instantly is denoted by
P(z)=P0+(z,1)=(1−ρ)(1−z)B[(z)]
B[(z)]−z(30a)
which agrees with (26); on the other hand, the PGF for the “remaining service time”
instantly afterward a random slot boundary in the busy span is indicated by
(1,u)
(1,1)=u[1−B(u)]
b(1−u)(30b)
4 Virtual Waiting Time and Unfinished Work
“Virtual waiting time” Wg(measured in slot) is in the n+1st slot so that a super
message could wait in analogous “Geo/G/1” system having “FCFS” discipline if it
will arrive in the n+1st slot. If L(n)=k≥1 and X(n)
+=≥2, it would be k−
1 complete service times plus l−1 slots. Finally, if L(n)=k≥2 and X(n)
+=1, it
would be exactly k−1 complete service times. Hence, the PGF W(u)forWis
790 R. Grover et al.
Wg(u)=P0+1(1)+
∞
k=1
∞
=2
k()[B(u)]k−1u−1+
∞
k=2
k(1)[B(u)]k−1
The “unfinished work U” instantly after the n+1st slot is the summand of
messages “service time” in the waiting line, and “remaining service time” of a
message will be served in that given time. Therefore, the PGF U(u)forUis given
by cf. Bruneel and Kim [4].
U(u)=Wg(u)[B(u)]= (1−p)(1−u)[B(u)]
[B(u)]−u(31a)
The mean of “unfinished work” is
E[U]=E[Wg]+ρ=λb(2)+λ(2)b(2)+ρ(1−2ρ)
2(1−ρ) (31b)
4.1 Relationship with the Continuous-Time System
Now let us assume the “continuous-time system” by considering limit →∞.
Firstly, the arrival and process will be changed. It is giventhat 1/ denoted number
of slots per “time unit;” the PGF denoted by c(z)for the number of messages that
will occur during a “time unit” is
c(z)=[(z)]1/ (32)
Let bcdenote the “mean” and b(2)
cdenote “second moment” of “service time” for
new “time units.”
It is obvious that band b(2)is related in this manner.
bc=b;b(2)
c=b22(33a)
ρ=λb=λcbc(33b)
is invariant, and by substitution we get
E[Wc]= λ2
cb(2)
c−λcρ+λ(2)
cbc
2λc(1−ρ) (34)
where E[Wc]=E[W]denotes “mean waiting time” expressed in new “time units.”
As →∞,theGeo
x/G/1 system approaches a continuous time Mx/01/1 system.
The PGF of the number of messages is represented by G(z) in a bunch,
Geo/G/1 System: Queues with Late and Early Arrivals 791
c(z)=eλ[G(z)−1](35)
if gand g(2) represents the “mean” and the “second factorial moment,” respectively,
of bunch size,
λc=λg;λ(2)
c=λ(2)
g+λ2g2(36)
Now, substituting (36)into(34),
E[Wc]= λg2b2
c+g(2)bc
2g(1−ρ) (37)
4.2 Early Arrival Model
For an early arrival model, let us first consider the Markov chain {Ln;n=1, 2….}
of the queue size Lninstantly [5].
L(n+1)=A−1+A
(n+1)Ln=0
Ln+A
n+1−1Ln≥1 (38a)
where denotes the number of messages that will arrive in a slot with the condition,
i.e., at least one message will arrive in that slot, and the number of messages is denoted
by A
n+1that will arrive between the time of service of the n+1st message given
that L=0. Number of messages A
n+1will arrive between the j−1 last slots if the
“service time” of the n+1st message is jslots. Therefore, PGF A(z)of A
n+1is
A(z)=
∞
j=1
b(j)[(z)]j−1=B[(z)]
(z)(38b)
Now,
(z)=(0)[A(z)−(z)A(z)]
A(z)−z
=λ(0)(0)[1−(z)]B[(z)]
[1−λ(0)](z){B[(z)]−z}(39a)
From the normalization condition ((1)=1),
(0)=[1−λ(0)][1−ρ]
λ(0)λ (39b)
792 R. Grover et al.
Hence,
(z)=(1−ρ)[1−(z)]B[(z)]
λ(z){B[(z)]−z}(40)
The system state in terms of “queue size” and “remaining service time” examined
at random point on the “continuous-time domain” is identical, and it is examined that
immediately after instantly a possible arrival point follows each “slot boundary” in
the early arrival model. Queue size instantly after a possible arrival point during the
nth slot is represented by L(n), and Message “remaining service” time which being
served at the same immediately is denoted by X(n). The PGF P(z) of “queue size”
observed at random point on “continuous-time domain” is given by (30a).
5 Conclusion
Hence, we reach the following conclusion about the comparison of the “late and
early arrival models.” Distribution of the “queue size” and the remaining in the stable
condition are similar in each model. The waiting time distribution of random message
is also the same in both models. The size of the queue instantly after the service
completion in the “early arrival model” is smaller in comparison to the “late arrival
model” by the number of messages included in a possible arrival. This is because the
observation point in the “early arrival model” is located before the possible “arrival
point.”
References
1. T. Meisling, Discrete-time queueing theory. Oper. Res. 6(1), 96–105 (1958)
2. W.W. Chu, A.G. Konheim, On the analysis and modeling of a class of computer communication
systems. IEEE Trans. Commun. 20(3, Part 2), 645–660 (1972)
3. H. Kobayashi, A.G. Konheim, Queueing models for computer communications system analysis.
IEEE Trans. Commun. 25(1), 2–29 (1977)
4. H. Bruneel, B.G. Kim, Discrete-Time Models for Communication Systems Including ATM,vol.
205 (Springer Science+Business Media, 1993)
5. J.J. Hunter, Mathematical Techniques of Applied Probability, Discrete Time Models: Techniques
and Applications, vol. 2 (Academic Press, New York, 1983)
6. A. Gravey, G. Hebuterne, Simultaneity in discrete-time single server queues with Bernoulli
inputs. Perform. Eval. 14(2), 123–131 (1992)
7. A. Gravey, On the Geo/D/1 and Geo/D/1/nqueues. Perform. Eval. 11(2), 117–125 (1990)
8. S.C. Dafermos, M.F. Neuts, A single server queue in discrete time. Caheirs du Centre de
Recherché Operationnelle 13, 23–40 (1971)
Intelligent Data Analysis with Classical
Machine Learning
Sanjeev Kumar Punia, Manoj Kumar, and Amit Sharma
Abstract In this advance era, artificial intelligence (AI) is the basic building block
for numerous innovation opportunities. The basic concept behind all artificial intel-
ligence applications is classical machine learning (CML) technique. The intelligent
data analysis (IDA) is used to analyze data in advanced reasoning systems. In this
paper, our motto is to analyze the data intelligently, i.e., intelligent data analysis
process using classical machine learning algorithms based on experimental data
set empirical hidden regularities. It is concluded that machine learning process is
the basic concept to implement any problem based on artificial intelligence, whereas
intelligent data analysis uses classical machine learning concept and helps in deciding
the next best available step during data analysis.
Keywords Classical machine learning (CML) ·Intelligent data analysis (IDA) ·
Artificial intelligence (AI) ·Internet of things (IoT)
1 Introduction
Presently, artificial intelligence is playing a big and active role at every step in
everyone’s life. Legg [1] explained that artificial intelligence is the requirement
of present and future era to enhance accuracy with efficiency through different
ways in every field. In parallel, artificial intelligence implements wide range of real
S. K. Punia (B)
JIMS Engineering Management Technical Campus, Greater Noida, Uttar Pradesh, India
e-mail: drsanjeevpunia@hotmail.com
M. Kumar
School of Computer Science, University of Petroleum and Energy Studies (UPES), Dehradun,
Uttrakhand, India
e-mail: wss.manojkumar@gmail.com
A. Sharma
SRMIST, NCR Campus, Modi Nagar, Ghaziabad, Uttar Pradesh, India
e-mail: Amit.krsharma123@gmail.com
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4_71
793
794 S. K. Punia et al.
applications outside research laboratories. The advanced artificial intelligence tech-
nique concept implemented in various areas like troop control weapons, internet of
things (IoT), and latest in the consulting of United States President election voting.
Bogost [2] mentioned that common understanding of artificial intelligence concept
is generally absent in many fields, and artificial intelligence is implemented using
machine learning, neural network learning, and deep machine learning. Generally,
deep machine learning is interrelated with intelligent data analysis and refers to the
explosive growth of artificial intelligence.
2 Notations in Classical Machine Learning and Intelligent
Data Analysis
Reference [3] explained that functional capabilities of artificial intelligence based
computer systems is a replica of human natural intelligence characteristics. The
cognitive reasoning is concerned with human intelligence as this also involves the
problem of learning, understanding, perceiving, analysis, and classification. Finn
et al. [4] explained that since origin artificial intelligence is used with all advanced
computer system applications to analyze, design, develop, implement, and test all
difficult problems. Any system is called intelligent if it is able to perform real
reasoning in data analysis field. The intelligent data analysis is the analysis of imple-
mented data using intelligent computer system framework. The classical machine
learning technique is an important tool in artificial intelligence that uses many
different mathematical models, algorithms, and methods for simulation. The classical
machine learning techniques perceive relations, learn analogies, generalize result,
and classify methods during difficult problem-solving process.
3 Bottleneck in Classical Machine Learning
Reference [5] demonstrates more functions and capabilities of classic intelligent data
analysis by comparing four different types of machine learning problems with some
machine learning tools limitations as follows.
A. Empirical Regularities
The different steps to solve traditional formulation machine learning problem are (i)
first, characterize the training use case description (precedents) sample into different
targets, (ii) second, forecast the new use case description (precedents) set into neces-
sary targets, (iii) third, predict interpolation–extrapolation into new object properties
targets, (iv) fourth, test quality construct hidden dependency into training sample, and
(v) finally, the result selection for predicting new sample into new object properties
targets. The solution of fixed training sample problem is to optimize the expansion
of new object as the results of different extrapolation samples are different. Hence,
Intelligent Data Analysis with Classical Machine Learning 795
the main realistic problem is the selection of best reliable extrapolation results. In
this, the major issue is the result detection of empirical regularities problems with
stable training set extension sample using classical machine learning.
B. Data Analysis Tools Combination
There are many classical machine learning tools to solve an interpolation–extrapo-
lation problem, but combining the different tools is a big problem. The comparison
of quality evaluation (e.g., bright and dry, green and small, etc.) is not a problem for
classical machine learning compared to intelligent data analysis.
C. Multi-model Training Sample Operation
In a research paper, reference [6] showed that target properties have to pay more atten-
tion for default acceptance in multi-model training sample operation during classical
machine learning process. The multi-model operation has multi causality problem
during proximity measurement in many machine learning samples. The emergence
problem predicts two different types of causation relationships (i) binary—the
different value attribute set causes different values effect and (ii) ternary—the specific
value attribute set with presence or absence of specific characters cause corresponding
effect. In these situations, initial sample (precedent) problem has to divide into two
sub-sample using special pre-processing tools. Later, these sub-samples are analyzed
using specific intelligent data analysis tools as per orientation in binary or ternary
causal relationship. These goal-oriented (focused) analyses of initial data sample
cannot be solved using traditional machine learning techniques.
D. Data Analysis Result Interpretation
Reference [7] explained that many problems are related with intermediate or final
interpretability results in classical machine learning techniques. These problems are
more difficult to implement, especially in neural networks and decision trees. The
constructive problem cannot be solved in classical machine learning using addiction
technique. Many complex problems have to process the initial training samples that
include poor sub-samples. Hence, choosing the right sample (precedent) from poor
sub-sample also plays an important role.
4 Method Automation in Intelligent Data Analysis
Reference [8] explained automation technique for intelligent data analysis using JSM
for scientific research support. Initially, main approach used in JSM automation
technique implementation is discussed. Later, the proposed solutions of all four
problems shown above in Sect. 3are compared.
A. Method Automation support in Intelligent Data Analysis
Reference [9] proposed JSM automatic method to formalize computer-oriented
inductive inference based schemes. The JSM automatic method use reasoning
796 S. K. Punia et al.
schemes to formalize many value-based logical reasoning. The JSM automatic
method uses abductive version of constructive explanation, synthesized using clas-
sical cognitive methods. The JSM automation method implements intelligent data
analysis framework using aided reasoning in (i) generation of empirical data using
extended precedents stable regularity and (ii) explanation of accumulative facts using
knowledge discovery implementation. Zabezhailo [10] proposed the data analysis
using partially defined relationship between object and properties. The analysis of
the data is based on their similar property assumption. The JSM automatic intelligent
data analysis framework is divided into three samples (precedents) as (i) £+—data set
with positive properties, (ii) £−—data set with negative properties, and (iii) £*—data
set with combined properties. Initially, separate data subset is prepared based on
the three classifications explained above. The object subset properties use inference
rules to implement JSM automatic intelligent data analysis framework. Polya [11]
explained that automatic intelligent data analysis JSM framework value is calculated
through object property as shown in Eq. 1
Ψ=Ψd∪Ψd(1)
where Ψd—Deductive inference rule set and Ψd—Plausible inference rule set.
Here, every rule of set Ψdis organized according to their distributive lattice pair.
The first lattice is analyzed for positive empirical dependency generation. The rule of
set Ψdis organized according to plausibility inference rule. The automatic intelligent
data analysis JSM framework uses classical machine learning methods to predict
the object properties of set Ψ. The different steps in prediction are (i) initially,
properties of set Ψis predicted based on sub-object Ψdand Ψdsets properties
to generate all dependencies, (ii) next, the prediction result of Ψset is added to
current Ψdand Ψdsets, and (iii) finally, JSM intelligent data analysis method is
repeated again and again to stabilize the Ψset value. Reference [12] also proposed
a constructive scheme (tentative test theory) to implement automatic intelligent data
analysis JSM framework. The tentative test theory is formulated through language,
and language is described using descriptive and argumentative function. The result of
methods is tested using falsification technique based on separate discovery regulari-
ties. Josephson [13] explained that value of automatic intelligent data analysis process
is evaluated based on four factors as (i) GCurrent—goal of automatic intelligent data
analysis process, (ii) TResult—tentative test theory result, (iii) VErrors—error correc-
tion approximate value, and (iv) GNext—next goal based on new automatic intelligent
data analysis process result. Finn [14] describes that JSM automatic intelligent data
analysis process uses the concept of inference rules sorted by their distributive lattices
pair relationship strength. The distributive lattices pair relationship strength is based
on JSM automatic intelligent data process reasoning generation quality. The JSM
automatic intelligent data analysis process is analog to human cognitive process
steps that include induction, abduction, inference, and forecast (IAIF) as shown in
Fig. 1.
Reference [15] explained that cognitive procedure scheme has significant
improvement than earlier proposed schemes. The authors generate a new formalized
Intelligent Data Analysis with Classical Machine Learning 797
Fig. 1 JSM automatic
intelligent data analysis
process
algorithm by modifying the existing empirical induction technique. The JSM auto-
matic intelligent data analysis process generates, analyzes, and corrects the empirical
knowledge base. The JSM automatic intelligent data analysis processes generate and
implement knowledge empirical discovery regularities hidden in expanded data sets.
B. Bottleneck Support in Intelligent Data Analysis
In this, we will discuss the solution of all four problems discussed above in Sect. 3
using JSM automatic intelligent data analysis process, namely (i) empirical regulari-
ties, (ii) data analysis tools combination, (iii) multi-model training sample operation,
and (iv) data analysis result interpretation.
Problem Solution 1: First, empirical dependencies extension is generated through
automatic intelligent data analysis process using classical machine learning. Then,
interpolation–extrapolation scheme is used to search stable dependencies in extended
data set. The searching of stable dependencies is based on empirical regularities.
This collects a huge open data set (precedents) collection based on JSM automatic
intelligent data analysis process fundamental characteristics.
Problem Solution 2: The different data analysis tools are combined and imple-
mented based on various JSM automatic intelligent data analysis process strategies.
For example, implementation of different plausible (p) and deductive (d) infer-
ence rules combination based on different distributive lattices. These combined auto-
matic intelligent data analysis process reasoning tools generate and evaluate forecast
dependency quality.
Problem Solution 3: The multi-model tools create many empirical dependencies
problem. The selection of adequate language for special pre-processing problems
uses automatic intelligent data analysis process based on classical machine learning.
The selected language analyzes the homogenous sub-arrays causality based on initial
empirical facts and implement for better result.
Problem Solution 4: The interpretability result is generated by causal depen-
dencies using automatic intelligent data analysis process based on classical machine
learning. The automatic intelligent data analysis process forecasts the generated result
798 S. K. Punia et al.
using abduction scheme. This controls the expandable element of initial training data
set (precedents) using causal dependencies.
5 Conclusion
The above discussion shows that intelligent data analysis process performs better
using classical machine learning method. The machine learning is the most popular
and developed method used almost everywhere in all artificial intelligence based
systems. However, in some areas, classical machine learning is not a best approach,
e.g., (i) the knowledge presentation in the special problem based on collection
theory and (ii) the reasoning formalization in cognitive procedure synthesis based
on induction, abduction, analogy and reasoning.
Presently, intelligent data analysis methods implement classical machine learning
which is used widely in artificial intelligence including infinite sub-areas. The clas-
sical machine learning has many implementations in different scientific and develop-
ment research area. In this paper, we conclude that intelligent data analysis process
became more efficient using classical machine learning process. This scenario
changes the use of intelligent data analysis process in non-trivial reasoning. The
advancement in machine learning provides an extended framework for better func-
tional data set analyzation. The traditional interpolation and extrapolation scheme
identifies new hidden empirical regularities in intelligent data analysis method with
new rules or laws using classical machine learning.
References
1. S. Legg, M. Hutter, A Collection of Definitions of Intelligence. Technical Report—IDSIA, 15
June 2007. https://arxiv.org/pdf/0706.3639.pdf
2. I. Bogost, Artificial intelligence has become meaningless. The Atlantic Monthly (2017). https://
www.theatlantic.com/technology/archive/2017/03/what-isartificialintelligence/518547/
3. S. Legg, M. Hutter, Universal intelligence: a definition of machine intelligence. Minds
Mach. 17(4):391–444 (2007). http://www.vetta.org/documents/legg-hutter-2007-universal-int
elligence.pdf
4. K. Finn Victor, I. Zabezhailo Michael, Intelligent data analysis and machine learning: are they
really equivalent concepts?, in Second Russia and Pacific Conference on Computer Technology
and Applications (RPC) (2017)
5. V.K. Finn, J.S. Mill, Inductive methods in artificial intelligence systems. Sci. Techn. Inf.
Process. Part I 38(6), 385–402 (2011), Sci. Techn. Inf. Process. Part II 39(5), 241–260 (2012)
6. O.P. Shesternikova, E.S. Pankratova, M.A. Agafonov, L.V. Vinokurova, V.K. Finn, Intelligent
system for chronic pancreatitis patient’s diabetes development prediction. Artif. Intell. Decis.
Support 4, 12–50 (2015). (in Russian)
7. A.Y. Volkova, Analyzing the data of different subject fields using the procedures of JSM method
for automatic hypothesis generation. Autom. Doc. Math. Linguist. 45(3), 127–139 (2011)
8. V.K. Finn, O.P. Shesternikova, On a new variant of generalized JSM method for automatic
scientific researches support. Artif. Intell. Decis. Support 1, 57–63 (2016). (in Russian)
Intelligent Data Analysis with Classical Machine Learning 799
9. M.A. Mikheenkova, A.Y. Volkova, Specification of the JSM intelligent system. Autom.
Documentation Math. Linguist. 47(4), 135–150 (2013)
10. M.I. Zabezhailo, V.K. Finn et al., Reasoning models for decision making: applications of JSM
method for intelligent control systems, in Architectures for Semiotic Modeling and Situation
Analysis in Large Complex Systems—Proceedings of the Workshop of 10th IEEE Symposium
on Intelligent Control (1995), pp. 27–29
11. G. Polya, Mathematics and Plausible Reasoning (Princeton University Press, Princeton, NJ,
1954)
12. V.K. Finn, Distributive lattices of inductive JSM procedures. Autom. Documentation Math.
Linguist. 48(6), 65–295 (2014)
13. J.R. Josephson, Smart inductive generalizations are abductions, in Abduction and Induction.
Essays on their Relation and Integration, ed. by P.A. Flach, A. Kakas (Kluwer Academic
Publishers, Dordrecht, Boston, London, 2010), pp. 31–44
14. V.K. Finn, The synthesis of cognitive procedures and the problem of induction. Autom
Documentation Math. Linguist. 43(3), 149–195 (2010)
15. V.K. Finn, M.A. Mikheenkova, Cognitive sociology from the knowledge discovery point of
view, in Proceedings of 2nd Russia and Pacific Conference on Computer Technology and
Applications (IEEE RPC) (2017)
Author Index
A
Agrawal, Anupam, 125
Ahmed, Md. Razu, 115,161
Ali, Md. Asraf, 115,161
Alphy, Anna, 593
Anita Christaline, J., 691
Ansari, Iffat Rehman, 13
Arokia Renjith, J., 179
Arshath Raja, R., 567
Arumugam, G., 559
Ayesha, Noor, 629
B
Babeetha, S., 1,277
Banerjee, Subhabrata, 741
Banu, Aisha, 617
Banu, W. Aisha, 657
Basu, Mainak, 83
Bhadoria, Vikas, 73
Bharathi Kannan, B., 575
Biswas, Ranjit, 13
Budihal., Suneeta V., 407
C
Chamundeeswari, G., 337,671
Chaudhari, Khushal, 97
Chaudhary, Himani, 781
Chaudhary, Pratibha, 645
Chauhan, Naman, 427
Chhettri, Devansh, 497
Chouhan, Vikas, 497
Chowdhary, Ramandeep Singh, 83
D
Daniel, A., 575
Dash, Suhransu Sekhar, 309
Dey, Rajeeb, 769
Dinesh Babu, A., 255
Divyavarshini, V. K., 671
Dubey, Praffulla Kumar, 435
Durgamahanthi, Vaishali, 691
Dwivedi, Bharti, 73
G
Ganesh Kumar, S., 369,395
Gangaprasad, Guntupalli, 23,519
Geetha, U., 607
Gobinath, D., 559
Gomathi, N., 707
Gomathy, C., 255
Govind, Nithyasri, 671
Grover, Reena, 781
Gupta, Nikhil Kumar, 531
Gupta, Saurabh, 435
H
Hanumantharaju, M. C., 447
Hariharan, V. L., 1,277
J
Jain, Ajit, 221
Jana, Gopal Chandra, 125
Janakiram, Kottnana, 23,519
Jawahar, M., 749
Jayan, Keerthi, 55
© Springer Nature Singapore Pte Ltd. 2021
S. S. Dash et al. (eds.), Intelligent Computing and Applications,
Advances in Intelligent Systems and Computing 1172,
https://doi.org/10.1007/978-981- 15-5566- 4
801
802 Author Index
K
Kala, A., 473
Kapoor, Sakshi, 267
Kar, Sanjeeb, 309
Karthick, S., 549,707
Khamari, Ramesh Ch., 97
Khatun, Mst. Arifa, 115,161
Kousik, N. V., 567,575
Kumar, Harmesh, 171
Kumar, Manoj, 793
Kumar, Mukul, 507
Kumar, Saroj, 171
L
Laxmisagar, H. S., 447
M
Magesh, Sandhya, 657
Mahapatra, Rajendra Prasad, 497,715
Mahapatra, R. P., 507
Majumdar, Pramathesh, 645
Malik, Medhavi, 245,427,585
Manoria, Manish, 419
Margarat, G. Simi, 541
Meena, Santosh, 35
Meena, Uma, 481,759
Menaga, D., 353
Mishra, Bharat, 419
Mittal, Ruchi, 235
Mohanty, Bijayananda, 715
Mohanty, Nalini Prasad, 769
Mukherjee, Saurabh, 143
Murugan, A., 559
Muruganantham, B., 55,383,507
Muthaiyah, Saravanan, 461
Muthu Kannan, P., 43,721
Muthukrishnan, P., 43,721
Muthukumar, M., 749
N
Naaz, Sameena, 13
Nayak, Smrutiranjan, 309
Naz, Insha, 13
Noor Alleema, N., 1,277
Noori, Sheak Rashed Haider, 115,161
P
Panda, Surya Narayan, 267
Pandey, Anjali, 245
Pandey, Ayushi, 645
Pandey, Sachi, 497
Patel, Nimai Charan, 769
Pattanayak, Himansu Sekhar, 203
Prasanna Bharathi, S., 285,297,337,671
Praveena Anjelin, D., 369
Premnath,S.P.,179
Punia, Sanjeev Kumar, 793
R
Rajakumari, P. Anitha, 549
Rajalakshmi, V., 473
Raja, R. Arshath, 549
Ramanjaneyulu, B. Seetha, 23,519
Rani, Anuj, 221
Rao, Raghupatruni Bhima, 715
Ravikumar, K., 541
Rayaguru, N. K., 323
Razeen, Faaez, 657
Rehman, N. Aaftab, 617
Revathi, S., 353
Roy, Binoy Krishna, 769
S
Sahayadhas, Arun, 115,161
Sandhya, 607
Sandhya, M., 617
Sangal,A.L.,203
Sangeetha, S., 1,277
Sankar, Sharmila, 607,617,657
Saraswat, Ritesh Kumar, 35
Sasikala, D., 473
Sehgal, Shankar, 171
Sekar, S., 323
Senthilkumar, P., 749
Shaardha, Chandra, 593
Shamili, P., 383,395
Shankar, Deepak, 617
Sharma, Amit, 793
Sharma, Anand, 759
Sharma, Arjun, 285,297,337
Sharma, Geetanjali, 781
Sharma, Girish Kumar, 481
Sharma, Himanshu, 497
Sharma, Karuna, 143
Sharma, Madhuri, 585
Sharma, Promila, 481
Sharma, Vinita, 35
Shirly Edward, A., 691
Shraddha, H., 1,277
Shukla, Shivam, 125
Shyamala Susan, V., 731
Author Index 803
Singh, Amit, 407
Singh, Archana, 645
Singh, Girijesh, 531
Singh, Juhi, 193
Singh, Sarbjeet, 171
Singh, Yagyiyaraj, 245
SivaSubramanian, S., 541
Srikant, Satya Sai, 715
Sriman, B., 383,395
Srivastava, Divyansh, 125
Suri, Harshit, 435
T
Taparia, Kriti, 507
Tiwari, Shubham, 73
Tyagi, Manoj, 419
V
Vasan, Vivek Ram, 285,297,337
Vasudevan, Amrita, 671
Verma, Harsh K., 203
Y
Yethi r a j , N . G . , 629
Yuva r a j , N . , 549,567,575
Z
Zaw, Thein Oak Kyaw, 461
