Content uploaded by Bhoj Raj Ghimire
Author content
All content in this area was uploaded by Bhoj Raj Ghimire on May 03, 2024
Content may be subject to copyright.
Content uploaded by Bhoj Raj Ghimire
Author content
All content in this area was uploaded by Bhoj Raj Ghimire on May 03, 2024
Content may be subject to copyright.
Content uploaded by Bhoj Raj Ghimire
Author content
All content in this area was uploaded by Bhoj Raj Ghimire on May 03, 2024
Content may be subject to copyright.
ISSN : 2635-4586
Date of Printing: January 07, 2022
Date of Publishing: January 12, 2022
Editor: Seongsoo Cho
Publication:
ICT-Advanced Engineering Society
Bima Build. #525, 20 Kwangwoon-ro, Nowon-gu, Seoul, Korea (01897)
info@ictaes.org; +82-2-940-8626 / 8637
Website: http://www.ictaes.org
Registration No.: 25100-2018-000027
© ICT-Advanced Engineering Society
8th Online International Conference on Advanced Engineering and
ICT-Convergence 2022 (ICAEIC-2022)
January 18, 2022
Organized by
ICT –Advanced Engineering Society,
Seoul, Korea (ICT-AES)
Bima Build. #525, 20 Kwangwoon-ro, Nowon-gu, Seoul, Korea (01897)
Email: info@ictaes.org, Tel.: +82-2-940-8626 / 8637
Organizing Committee (8th ICAEIC-2022) January 18, 2022
Honorary Chair:
Lochan Lal Amatya, President, SECEN, Nepal
Daeyoung Kim, Information Security Engineering, Jeju International University, Jeju, Korea
Ngo Quoc Viet, Professor, Ho Chi Minh City University of Education, Vietnam
Dae-Yeong Park, Professor, Busan National University, Korea
Dinesh Kumar Sharma, Professor / Advisor, Kathmandu Engineering College, Nepal
Hoon Kim, Professor, Incheon National University, Korea
General Chair:
Bhanu Shrestha, Professor, Kwangwoon University, Korea
Conference Chair:
Husni Teja Sukmana, Professor, Universitas Islam Negeri Syarif Hidayatullah, Indonesia
Conference Co-Chair:
Gi-Chul Yang, Professor, Mokpo National University, Korea
International Chair / Coordinator:
Yeonwoo Lee, Professor, Mokpo National University, Korea
International Co-chair:
Surendra Shrestha, Associate Professor, IoE, Tribhuban University, Nepal
Technical Chair:
Soonchul Kwon, Associate Professor, Kwangwoon University, Korea
Technical Co-Chair:
Untung Raharja, Professor, Raharja University, Indonesia
Dandy Pramana Hostiad, Professor, Technology and Business Institute STIKOM Bal
Min A Jeong, Professor, Mokpo National University, Korea
Yang-Ick Joo, Professor, Korea Maritime Ocean University, Busan Korea
Jeongjoon Kim, professor, Anyang University, Korea
Program Chair:
Youngman Kwon, Professor, Eulji University, Korea
Program Co-Chair:
Dadang Hermawan, Professor, Rectore, Technology and Business Institute STIKOM Bali
Evi Triandini, Professor, Technology and Business Institute STIKOM Bali
Jin-Mook Kim, Professor, Sunmoon University, Korea
Sang-Joon Lee, Professor, Chonnam National University, Korea
Kyeong Hur, Professor, Gyeongin National University of Education, Korea
Kwangchul Son Associate Professor, Kwangwoon University, Seoul, Korea
Publicity Chair:
Woo-Hyuk Kim, Professor, Incheon National University, Korea
Publicity Co-Chair:
Jinho Han, Associate Professor, Korean Bible University, Korea
Dong You Choi, Professor, Chosun University, Korea
Wooil Kim, Associate Professor, Incheon National University, Korea
Sungtek Kahng, Professor, Incheon National University, Korea
Youngmin Kim, Professor, Ajou University, Korea
Seung-Ho Kang, Professor, Donshin University, Korea
Naejoung kwak, Professor, Pai Chai University, Korea
Youn-Sik Hong, Professor, Incheon National University, Korea
Publishing Chair:
Seongsoo Cho, Professor, Kongju National University, Korea
Welcome Address
Prof. Dr. Husni Teja Sukmana
Conference Chair
Universitas Islam Negeri Syarif Hidayatullah, Indonesia
I am honored to be here as a Conference Chair of the 8th Online International Conference on Advanced
Engineering and ICT-Convergence 2022 (ICAEIC-2022) to welcome all distinguished guests and speakers
from various part of the world. Due to COVID-19 pandemic, we could not get together for the conference. So,
we are conducting virtual conference on zoom platform. The ICAEIC-2022 is dedicated to the research and
development of a broad range of Engineering and ICT-Convergence related topics, with focuses on theory,
simulation, design, realization, measurement, and applications. It is our sincere anticipation that the ideas and
technological solutions presented at this august gathering will contribute to the fields of Engineering and ICT
industry as the enabling force for positive development for the benefit of society.
We are conducting total 5 tracks including special session. For this conference, the organizing committee has
accepted 56 out of 82 papers excluding keynote papers for the presentations. Some papers have been rejected
to maintain the quality of the conference. In this conference, the papers from 6 countries will be presented. I
fully believe that these presentations and keynote speeches will be highly interesting and interactive.
I would like to express our sincere appreciation to all the committee members and many other helping hands
behind the scenes who have significantly contributed to set-up this conference. My special thanks go to
conference co-chair, international co-chair, all session chairs and the eminent persons who are attending here
in the virtual platform. If there is no problem of COVID-19, we can make face-to-face conference. Next 9th
international conference will be held in Jeju Island, Korea on July 13-15, 2022. I hope, we can meet together
over there. Thank you and have a good time in the virtual conference.
Thank you!
January 18, 2022.
Congratulatory Remarks
Prof. Dr. Bhanu Shrestha
Chairman of ICT-Advanced Engineering Society
Kwangwoon University, Korea
Ladies and Gentlemen,
First of all, I’d like to welcome all the distinguished guests and participants today for the 8th Online
International Conference on Advanced Engineering and ICT-Convergence (CAEIC-2022) online ICAEIC-
2022 conference. As you know that we have switched to face-to-face conference to virtual conference due to
pandemic of COVID-19. As we know that the technologies are rapidly increasing that makes the future society
to facilitate the people. That means, as a researcher, as a scientist, we are designing the future society for the
human welfare. Therefore, we need such conferences to change our society as well. At the same time, we are
living in the era of the 4th industrial revolution where information and communication technologies (ICTs)
with intelligence are the main components. Therefore, we have selected the emerging topic on artificial
intelligence, big-data, and convergence technologies for the conference. Also, this virtual platform will be
helpful to understand recently developed technologies and their impact on human life and society.
On behalf of the ICAEIC-2022 Organizing Committee, we thank keynote speakers, all the session chairs,
publication chairs, technical chairs, program chairs, international advisory board honorary chairs, and all
presenters. We would like to extend our special thanks to Conference Chair, Prof. Husni Teja Sukmana from
Syarif Hidayatullah State Islamic University, Indonesia, International Co-ordinator Prof. Yeonwoo Lee from
Mokpo International University, Seoul, Korea, International Co-chair, Prof. Dr. Surendra Shrestha from
Institute of Engineering, Pulchwok, Nepal, Publication chairs, Prof. Seongsoo Cho and Technical chair, Prof.
Sunchul Kwon for their invaluable guidance in organizing the conference. I hope, this conference will be a
success and fruitful and established a brotherhood relationship under Advanced Engineering and ICT-
Convergence.
Thank you very much.
January 18, 2022
Conference Program (8th ICAEIC-2022)
Tuesday, January 18, 2022, Korean Standard Time
A. Plenary Session (Time: 11:00 – 12:20)
MC: Prof. Dr. Surendra Shrestha, Institute of Engineering (IoT), Tribhuvan University, Nepal.
◾ Welcome Remarks (11:06-11:10): Prof. Dr. Husni Teja Sukmana, Conference Chair (11:11 -11:15)
Universitas Islam Negeri Syarif Hidayatullah, Indonesia.
◾ Congratulatory Remarks: Prof. Dr. Bhanu Shrestha, (Professor, Kwangwoon University, Korea, and
Chairman of ICT-AES, Korea)
B. Keynote Speech (Time: 11:15 –12:20)
◾ Keynote Session Chair: Prof. Dr. Yeonwoo Lee
Keynote Speaker ( 11:15-11:45 ): Emad Alsusa, Professor at School of Electrical and Electronic
Engineering, University of Manchester, Manchester, UK
Title: Overview of Research Issues on 5G, IoT, and D2D Communication Networks
Keynote Speaker ( 11:45-12:20 ): Prof. Dr.Bhoj Raj Ghimire, Prof. at Open University, Nepal
Title: Multidisciplinary Approach to COVID-19 Risk Communication
Break Time: 12:20 –12:30
B. Oral Session: 12:30 – 17:30 PM
Time : 12:30 –15:10
Track : A-1. AI
Session Chairs : Prof. Dr. Young Man Kwon | Prof. Dr. Evi Triandini
Track B: B-1. Big-data
Session Chairs : Prof. Dr. Sun Park | Prof. Dr. Ni Keut Dewi Ari Jayanti
Track : C-1. Big-data | RF
Session Chairs : Prof. Dr. Sun-Young Ihm | Prof. Dr. Gyanendra Pd. Joshi
Track : D. Cultural Conv.
Session Chairs : Prof. Dr. Dong Ho Kim | Prof. Lochan Lal Amatya
ix
AEICP Vol. 5, No. 1
Break Time Time : 15:10–15:20
Time : 15:20 –18:00
Track : A-2. AI
Session Chairs : Prof. Dr. Yeonwoo Lee | Prof. Dr. Subarna Shakya
Track : B-2. Big-data
Session Chairs : Prof. Dr. Hoon Kim | Prof. Dr. Nyoman Rudy
Track: C-2. Cultural Conv.
Session Chairs : Prof. Dr. Seong-Ho Kang | Prof. Dr. Luh Made Yulyantari
D. Closing Ceremony Time : 18:00 – 18:30
Certificate Distribution Announcement (Certificates + Best Paper Awards): One winner will be selected
from each session for the best paper award based on the evaluation of each session chair (by Prof. Dr.
Yeonwoo Lee).
Closing Remarks : Prof. Dr. Bhanu Shrestha, Chairman of ICT-AES, Korea.
Session Details
Thursday, January 18, 2022, Korean Standard Time
Keynote Speech
Time : 11:15 –12:00 Keynote Session Chair: Prof. Dr. Yeonwoo Lee
◾ Keynote Speaker ( 11:15-11:45 ): Emad Alsusa, Professor at School of Electrical and Electronic
Engineering, University of Manchester, Manchester, UK
Title: Overview of Research Issues on 5G, IoT, and D2D Communication Networks
◾ Keynote Speaker ( 11:45-12:20 ): Prof. Dr.Bhoj Raj Ghimire, Prof. at Open University, Nepal Title:
Multidisciplinary Approach to COVID-19 Risk Communication
Break Time : 12:00 –12:30
Parallel Oral Session Time : 12:30 – 15:10
Time : 12:30 –15:10 Track-A-1. AI (Artificial Intelligence)
Session Chairs : Prof. Young Man Kwon | Prof. Dr. Evi Triandini
ICAEIC-2022-138
Digital Twin Coaching for Openpose
(Qiuying Chen, Seong-Yeol An, Kyeong-Rak Lee, Sang-Joon Lee)
ICAEIC-2022-137
Reusable Fuzzy Extractor from Non-uniform Learning with Errors Problem
(Joo Woo, Jonghyun Kim, Jong Hwan Park)
ICAEIC-2022-117
Image Enhancement Malaria Parasite Detection System based on Giemsa-Stained Slide Images using
Homomorphic Filtering
(Edy Victor Haryanto S, Bob Subhan Riza, Juli Iriani, Noprita Elisabeth S, Universitas Potensi Utama)
ICAEIC-2022-123
CNN-based User Behavior Detection using Time-division Feature Data
(JinSuk Bang, NaeJoung Kwak)
ICAEIC-2022-165
Alphanumeric CAPTCHA Recognition Using CNN-BiLSTM Network
(Bal Krishna Nyaupane, Rupesh Kumar Sah, Bishnu Hari Paudel)
xi
AEICP Vol. 5, No. 1
ICAEIC-2022-121
Salient Object Detection using Parallel Symmetric Networks with Attention Mechanism
(Kyeongseok Jang, Donwoo Lee, Soonchul Kwon, Seunghyun Lee, Kwang Chul Son)
ICAEIC-2022-131
Partial Face Recognition with Deep Learning
(Jinho Han)
ICAEIC-2022-129
Energy Management System Model Applying a Nature-Inspired Optimization Algorithm to a Virtual Power
Plant
(Yeonwoo Lee)
Time : 12:30 –15:10 Track- B-1. Big-data
Session Chairs : Prof. Dr. Sun Park | Prof. Dr. Ni Keut Dewi Ari Jayanti
ICAEIC-2022-126
A Study on Mask Wearing Discrimination and Identity Identification
(NaeJoung Kwak, DongJu Kim)
ICAEIC-2022-114
Case Study Hotel Booking Demand: A Predictive Analytics Model
(Luh Made Yulyantari, Ni Ketut Dewi Ari Jayanti)
ICAEIC-2022-174
A study on the Method of Securing Documentary and Evidence Capability of the Suspect's Record
(Yoon Cheolhee, Hwang JeaHun, Cho MinJe, Lee Bong Gyou)
ICAEIC-2022-133
Multiple Imputation Study of Missing Marine Environmental Data
(Hyi-Thaek Ceong, Kyeong-Rak Lee, Sang-Joon Lee)
ICAEIC-2022-116
Verification of Big data-based Correlation Analysis for Autonomous Driving
(Dong-Jin Shin, Seung-Yeon Hwang, Mun-Yong Park, Dae-Sun Yum, Sung-Youn Cho)
ICAEIC-2022-175
A Study of Illegal Transaction Flow through Blockchain Forensics
(Yoon Cheolhee, Kyung Min Beak, Kyu Sic Oh, Jin-Mook Kim)
8th ICAEIC-2022
xii
ICAEIC-2022-170
Recycling of the Waste Paper Using the Cellulose Extracted from Pine Apple Leaves and Its Analysis
(Sabina Paudel, Bindra Shrestha)
ICAEIC-2022-176
A Study on the Web Service-Based Platform for Pregnant Women
(Jin-suk Bang, Jin-Mook Kim, Min A Jeong)
Time : 12:30 –15:10 Track- C-1. Big-data / RF
Session Chairs : Prof. Dr. Sun-Young Ihm| Prof. Dr. Gyanendra Pd. Joshi
ICAEIC-2022-135
A Deep Q-Network Based Intelligent Pumping System for the Rainwater Pumping Station
(Seung-Ho Kang)
ICAEIC-2022-145
Implementation of Public Mobility Tracing and Tracking System Using QR Code and Big Data to Support
Tourism Industry in Bali
(Evi Triandini, IB Suradarma, Sofwan Hanief, I Ketut Dedy Suryawan, I Gusti Rai Agung Sugiartha, I
Wayan Agus Hery Setiawan)
ICAEIC-2022-178
A Study on Exchange System by Cooling Dehumidification using Big-data Analysis
(Zhen-Huan WANG, Jin-Mook Kim, Young-Chul Kwon)
ICAEIC-2022-152
A Study on ICT Literacy Competency Model for Software Education
(Jin-Hee Ku)
ICAEIC-2022-161
IoT based Smart Home with Face Recognition Security
(Sunil Pokhrel, Anil K. Panta, Sachin Parajuli, Sudip Poudel, Rajan Adhikari, MK Guragai)
ICAEIC-2022-142
Design and Development of 3-stage Folding Electric Kickboard based on In-wheel Motor Method
(Duk-Keun An, Ik-Hyeon Kim, Si-Wook Sung, Dong-Cheol Kim, Sang-Hyun Lee)
xiii
AEICP Vol. 5, No. 1
ICAEIC-2022-182
A Study on Challenges of AI Ethics in Next Generation Wireless Networks
(Yeongchan Kim, Navin Ranjan, Sovit Bhandari, Hoon Kim)
ICAEIC-2022-157
Demonstration Project of Air Pollution Monitoring System in Urban Environment using LoRa : LoRa
Topology Improvement
(Ji-Seong Jeong, Jeong-Gi Lee, Chul-Seung Yang, Gi-won Ku)
Time : 12:30 –15:10 Track- D-1. Cultural Convergence
Session Chairs : Prof. Dong Ho Kim | Prof. Lochan Lal Amatya
ICAEIC-2022-134
Bibliometric Analysis of Blockchain and The Survey Development
(Hu Chenxi, Ren Chen,Sang-Joon Lee)
ICAEIC-2022-163
Enhanced PID Control of UVC 222nm Sterilization Lamp Combining Machine Learning based on
Independent Edge Computing Device
(Eon-Uck Kang, Duc-hwan Ahn)
ICAEIC-2022-139
Potential Scholarship Recipients with Simple Additive Weighting and Technique for Order of Preference by
Similarity to Ideal Solution
(Qurrotul Aini, Nurbojatmiko, Mega Ayu Silvianingsih)
ICAEIC-2022-180
Identify the CARLA Simulator Element of the RSS Model for Variable Angle Camera Application
(Min Joong Kim, Young Min Kim)
ICAEIC-2022-119
Building New Words Validation and Sentiment Analysis Model through AI Technique
(Dong Hyeon Kim, Da Bin Park, Seung Ri Park, Se Jong Oh, Ill Chul Doo)
ICAEIC-2022-169
Fabrication of Activated Carbon from Areca Nuts and Its Application for Methylene Blue Dye Removal
(Pramila Joshi, Sahira Joshi)
8th ICAEIC-2022
xiv
ICAEIC-2022-155
Pest Prediction and Diagnosis According to Deep Learning-based Environmental Change
(Man-Ting Li, Gan Liu, Hyun-Tae Kim, Sang-Bum Kim, Eun-Seo Song, Kyu-Ha Kim, SangHyun Lee)
ICAEIC-2022-156
A Study on the Anti-wrinkle Effect of Hydrogen Water on Mice’s Wrinkles Cause by UV and Its Toxicity
(Eun-Suk Lee, Ji-Ung Yang, In-sang Lee, Ki-jin Kwon, Dae-Gyeom Park)
Break Time Time: 15:10 –15:20
Time : 15:20 –18:00 Track-A-2. AI (Artificial Intelligence)
Session Chairs : Prof. Dr. Yeonwoo Lee | Prof. Subarna Shakya
ICAEIC-2022-128
A Novel Scheme of an Object-Preserving GAN Architecture for Data Augmentation
(Juwon Kweon, Jaekwang Oh, Soonchul Kwon)
ICAEIC-2022-166
Fake News Detection using Hybrid Neural Network
(Niroj Ghimire, Tanka Raj Pandey, Surendra Shrestha)
ICAEIC-2022-160
Efficient Semantic Segmentation by Using Down-Sampling and Subpixel Convolution
(Young-Man Kwon, Sung-Hoon Bae, Dong-Keun Chung, Myung-Jae Lim)
ICAEIC-2022-130
A Way of Implementing Brain Password System based on Graphical Password
(Gi-Chul Yang)
ICAEIC-2022-124
A Study on Determining Wearing a Hard Hat and Face Identification
(NaeJoung Kwak, KeunWoo Lee, DongJu Kim)
ICAEIC-2022-115
A Study on CNN-based Object Recognition Technology for Autonomous Vehicle
(Seung-Yeon Hwang, Dong-Jin Shin, Mun-Yong Park, Dae-Sun Yum, Sung-Youn Cho)
xv
AEICP Vol. 5, No. 1
ICAEIC-2022-150
Anti-Collision Control System for Unmanned Aerial Vehicle based on Reinforcement Learning
(Chung-Pyo Hong)
ICAEIC-2022-122
Super Resolution Method using Parallel Attention Mechanism Architecture
(Dongwoo Lee, Kyeongseok Jang, Chae-bong Sohn, Bhanu Shrestha, Seongsoo Cho, Kwang Chul Son)
Time : 15:20 –18:00 Track- B-2. Big Data
Session Chairs : Prof. Dr. Hoon Kim | Prof. Dr. Nyoman Rudy
ICAEIC-2022-140
Sentiment Analysis of Public Opinion on COVID-19 Vaccines on Twitter Social Media using Naive Bayes
Classifier Method
(Arini, Saepul Aripiyanto, Alvian Aristya, Iik Muhamad Malik Matin)
ICAEIC-2022-141
Design of BoP Big Data Processing System for BoP Data Analysis
(Sun Park, Byung-joo Chung, ByungRea Cha, JongWon Kim)
ICAEIC-2022-177
A Study of Estimate Information Service using DID (Distribute IDentification)
(Jeong-Kyung Moon, Jin-Mook Kim)
ICAEIC-2022-181
Suggestion of Maintenance Criteria for Electric Railway Facilities System based on Fuzzy TOPSIS Method
(Sunwoo Hwang, Joouk Kim, Youngmin Kim)
ICAEIC-2022-168
Synthesis and Characterization of the Phosphoric Acid Activated Carbon Prepared from Giant Cane
(Sahira Joshi, Shishir Baral)
ICAEIC-2022-146
Restaurant Skyline for Users with Physical Disabilities
(Chae-Eun Lim, Eun-Young Park, Jong Sup Lee, Sun-Young Ihm)
8th ICAEIC-2022
xvi
ICAEIC-2022-149
Creation of a Mesh by Applying Poisson Disk Sampling to PointCloud : Comparison of RTAB-Map and
VisualSFM
(Yujin Yang, Sohee Kim, Jungsuk Park, Dong Ho Kim)
ICAEIC-2022-173
Determination of Methylene Blue Number, Iodine Number, SEM and XRD to Characterize Activated
Carbon Prepared from Rudraksha (Elaeocarpus ganitrus) Bead
(Bishwas Pokharel, Rinita Rajbhandari (Joshi), Rajeshwar Man Shrestha)
Time : 15:20 –18:00 Track- C-2. Cultural Convergence
Session Chair : Prof. Dr. Seung-Ho Kang | Prof. Luh Made Yulyantari
ICAEIC-2022-171
Isolation of Cellulose from Sisal and Its Application for Removal of Arsenic from Aqueous Solution
(Manish Shrestha, Sanju Khatri and Bindra Shrestha)
ICAEIC-2022-179
Analysis of the Movement Path of Micro Dust in a Simulation-based Urban Environment for Specifying the
Installation Location of the Optical Particle Counter
(Jeong-Gi Lee, Chul-Seung Yang, Gi-won Ku, Ji-Seong Jeong)
ICAEIC-2022-143
Design and Development of Fan Boss to Reduce Noise and Vibration of Air Purifier Fan
(Hyeong-Sam Park, Doo-Sik Kim, Sang-Hyun Lee)
ICAEIC-2022-172
Characterization of Activated Carbon Prepared from Walnut (Jaglans regia) Shell
(Srijan Adhikari, Rinita Rajbhandari (Joshi), Rajeshwar Man Shrestha)
ICAEIC-2022-164
A Machine Learning Model for Unintentional Anomaly Detection and Control on TinyML Lite Platform
(Eon-Uck Kang, Woo-Sang Hwang)
ICAEIC-2022-154
A Study on the Influence of Psychological Factors of Parents on Children's Behavior
(Jung-Eun Lee, Chun-Ok Jang)
xvii
AEICP Vol. 5, No. 1
ICAEIC-2022-148
Implementation of Support Vector Machine (SVM) Method to Determine the Response of the Indonesian
People to the Administration of the Sinovac Vaccine
(Siti Ummi Masruroh, Obey Al Farobi, Khodijah Hulliyah, Husni Teja Sukmana, Yusuf Durachman, Nanda
Alivia Rizqy Vitalaya)
ICAEIC-2022-120
A Study on the Combination of Extractive Summary and AI-Based Abstractive Summary
(Woo Won Choi, Jeong Hyeon Kim, Min Kyu Park, Se Jong Oh, Ill Chul Doo)
C. Closing Ceremony Time : 18:00 – 18:30
• Certificate Distribution Announcement (Certificates + Best Paper Awards: One winner will be
selected from each session for the best paper award based on the evaluation of each session chair.
• Closing Remarks : Chairman of ICT-AES, Korea: Prof. Dr. Bhanu Shrestha
NOTE:
• The presentation time is 20 minutes (15 min-presentation and 5 min-Q&A) for each presenter except
Keynote speakers (25 minutes 5 min-Q&I).
• This is a tentative schedule for the virtual video conference. The time can be changed according to a
number of papers.
• All authors are requested to join in a virtual video conference before 15 minutes of your turn of
presentation.
Contact:
Bima Build. #525, 20 Kwangwoon-ro, Nowon-gu, Seoul, Korea (01897). Email: info@ictaes.org, Tel.: +82-
2-940-8626 / 8637 | Website: https://ictaes.org
Contents
1
Overview of Research Issues on 5G, IoT, and D2D Communication Networks
Emad Alsusa
1
2
Multidisciplinary Approach to COVID-19 Risk Communication
Bhoj Raj Ghimire, and Bhogendra Mishra
6
3
Digital Twin Coaching for Openpose
Qiuying Chen, Seong-Yeol An, Kyeong-Rak Lee, and Sang-Joon Lee
11
4
Reusable Fuzzy Extractor from Non-uniform Learning with Errors Problem
Joo Woo, Jonghyun Kim, and Jong Hwan Park
15
5
Image Enhancement Malaria Parasite Detection System based on Giemsa-Stained Slide
Images using Homomorphic Filtering
Edy Victor Haryanto S, Bob Subhan Riza, Juli Iriani, and Noprita Elisabeth S
20
6
CNN-based User Behavior Detection using Time-division Feature Data
JinSuk Bang, and NaeJoung Kwak
28
7
Alphanumeric CAPTCHA Recognition Using CNN-BiLSTM Network
Bal Krishna Nyaupane, Rupesh Kumar Sah, and Bishnu Hari Paudel
31
8
Salient Object Detection using Parallel Symmetric Networks with Attention
Mechanism
Kyeongseok Jang, Donwoo Lee, Soonchul Kwon, seunghyun Lee, and Kwang Chul Son
39
9
Partial Face Recognition with Deep Learning
Jinho Han
44
xix
AEICP Vol. 5, No. 1
10
Energy Management System Model Applying a Nature-Inspired Optimization
Algorithm to a Virtual Power Plant
Yeonwoo Lee
47
11
A Study on Mask Wearing Discrimination and Identity Identification
NaeJoung Kwak, and DongJu Kim
52
12
Case Study Hotel Booking Demand: A Predictive Analytics Model
Luh Made Yulyantari, and Ni Ketut Dewi Ari Jayanti
56
13
A study on the Method of Securing Documentary and Evidence Capability of the
Suspect's Record
Yoon Cheolhee, Hwang JeaHun, Cho MinJe, and Lee Bong Gyou
71
14
Multiple Imputation Study of Missing Marine Environmental Data
Hyi-Thaek Ceong, Kyeong-Rak Lee, and Sang-Joon Lee
76
15
Verification of Big data-based Correlation Analysis for Autonomous Driving
Dong-Jin Shin, Seung-Yeon Hwang, Mun-Yong Park, Dae-Sun Yum, and Sung-Youn Cho
80
16
A Study of Illegal Transaction Flow through Blockchain Forensics
Yoon Cheolhee, Kyung Min Beak, Kyu Sic Oh, and Jin-Mook Kim
84
17
Recycling of the Waste Paper Using the Cellulose Extracted from Pine Apple Leaves
and Its Analysis
Sabina Paudel, and Bindra Shrestha
89
18
A Study on the Web Service-Based Platform for Pregnant Women
Jin-suk Bang, Jin-Mook Kim, and Min A Jeong
96
19
A Deep Q-Network Based Intelligent Pumping System for the Rainwater Pumping
Station
Seung-Ho Kang
100
8th ICAEIC-2022
xx
20
Implementation of Public Mobility Tracing and Tracking System Using QR Code and
Big Data to Support Tourism Industry in Bali
Evi Triandini, IB Suradarma, Sofwan Hanief, I Ketut Dedy Suryawan,
I Gusti Rai Agung Sugiartha, and I Wayan Agus Hery Setiawan
105
21
A Study on Exchange System by Cooling Dehumidification using Big-data Analysis
Zhen-Huan WANG, Jin-Mook Kim, and Young-Chul Kwon
112
22
A Study on ICT Literacy Competency Model for Software Education
Jin-Hee Ku
117
23
IoT based Smart Home with Face Recognition Security
Sunil Pokhrel, Anil K. Panta, Sachin Parajuli, Sudip Poudel, Rajan Adhikari, and MK
Guragai
122
24
Design and Development of 3-stage Folding Electric Kickboard based on In-wheel
Motor Method
Duk-Keun An, Ik-Hyeon Kim, Si-Wook Sung, Dong-Cheol Kim, and Sang-Hyun Lee
127
25
A Study on Challenges of AI Ethics in Next Generation Wireless Networks
Yeongchan Kim, Navin Ranjan, Sovit Bhandari, and Hoon Kim
131
26
Demonstration Project of Air Pollution Monitoring System in Urban Environment
using LoRa : LoRa Topology Improvement
Ji-Seong Jeong, Jeong-Gi Lee, Chul-Seung Yang, and Gi-won Ku
140
27
Bibliometric Analysis of Blockchain and The Survey Development
Hu Chenxi, Ren Chen, and Sang-Joon Lee
145
28
Enhanced PID Control of UVC 222nm Sterilization Lamp Combining Machine
Learning based on Independent Edge Computing Device
Eon-Uck Kang, and Duc-hwan Ahn
150
xxi
AEICP Vol. 5, No. 1
29
Potential Scholarship Recipients with Simple Additive Weighting and Technique for
Order of Preference by Similarity to Ideal Solution
Qurrotul Aini, Nurbojatmiko, and Mega Ayu Silvianingsih
158
30
Identify the CARLA Simulator Element of the RSS Model for Variable Angle Camera
Application
Min Joong Kim, and Young Min Kim
168
31
Building New Words Validation and Sentiment Analysis Model through AI Technique
Dong Hyeon Kim, Da Bin Park, Seung Ri Park, Se Jong Oh, and Ill Chul Doo
173
32
Fabrication of Activated Carbon from Areca Nuts and Its Application for Methylene
Blue Dye Removal
Pramila Joshi, and Sahira Joshi
178
33
Pest Prediction and Diagnosis According to Deep Learning-based Environmental
Change
Man-Ting Li, Gan Liu, Hyun-Tae Kim, Sang-Bum Kim, Eun-Seo Song, Kyu-Ha Kim, and
SangHyun Lee
183
34
A Study on the Anti-wrinkle Effect of Hydrogen Water on Mice’s Wrinkles Cause by
UV and Its Toxicity
Eun-Suk Lee, Ji-Ung Yang, In-sang Lee, Ki-jin Kwon, and Dae-Gyeom Park
188
35
A Novel Scheme of an Object-Preserving GAN Architecture for Data Augmentation
Juwon Kweon, Jaekwang Oh, and Soonchul Kwon
192
36
Fake News Detection using Hybrid Neural Network
Niroj Ghimire, Tanka Raj Pandey, and Surendra Shrestha
197
37
Efficient Semantic Segmentation by Using Down-Sampling and Subpixel Convolution
Young-Man Kwon, Sung-Hoon Bae, Dong-Keun Chung, and Myung-Jae Lim
203
8th ICAEIC-2022
xxii
38
A Way of Implementing Brain Password System based on Graphical Password
Gi-Chul Yang
208
39
A Study on Determining Wearing a Hard Hat and Face Identification
NaeJoung Kwak, KeunWoo Lee, and DongJu Kim
212
40
A Study on CNN-based Object Recognition Technology for Autonomous Vehicle
Seung-Yeon Hwang, Dong-Jin Shin, Mun-Yong Park, Dae-Sun Yum, and Sung-Youn Cho
216
41
Anti-Collision Control System for Unmanned Aerial Vehicle based on Reinforcement
Learning
Chung-Pyo Hong
222
42
Super Resolution Method using Parallel Attention Mechanism Architecture
Dongwoo Lee, Kyeongseok Jang, Chae-bong Sohn, Bhanu Shrestha, Seongsoo Cho, and
Kwang Chul Son
226
43
Sentiment Analysis of Public Opinion on COVID-19 Vaccines on Twitter Social Media
using Naive Bayes Classifier Method
Arini, Saepul Aripiyanto, Alvian Aristya, and Iik Muhamad Malik Matin
230
44
Design of BoP Big Data Processing System for BoP Data Analysis
Sun Park, Byung-joo Chung, ByungRea Cha, and JongWon Kim
237
45
A Study of Estimate Information Service using DID (Distribute IDentification)
Jeong-Kyung Moon, and Jin-Mook Kim
241
46
Suggestion of Maintenance Criteria for Electric Railway Facilities System based on
Fuzzy TOPSIS Method
Sunwoo Hwang, Joouk Kim, and Youngmin Kim
245
47
Synthesis and Characterization of the Phosphoric Acid Activated Carbon Prepared
from Giant Cane
Sahira Joshi, and Shishir Baral
250
xxiii
AEICP Vol. 5, No. 1
48
Restaurant Skyline for Users with Physical Disabilities
Chae-Eun Lim, Eun-Young Park, Jong Sup Lee, and Sun-Young Ihm
255
49
Creation of a Mesh by Applying Poisson Disk Sampling to PointCloud : Comparison of
RTAB-Map and VisualSFM
Yujin Yang, Sohee Kim, Jungsuk Park, and Dong Ho Kim
258
50
Determination of Methylene Blue Number, Iodine Number, SEM and XRD to
Characterize Activated Carbon Prepared from Rudraksha (Elaeocarpus ganitrus)
Bead
Bishwas Pokharel, Rinita Rajbhandari (Joshi), and Rajeshwar Man Shrestha
265
51
Isolation of Cellulose from Sisal and Its Application for Removal of Arsenic from
Aqueous Solution
Manish Shrestha, Sanju Khatri and Bindra Shrestha
270
52
Analysis of the Movement Path of Micro Dust in a Simulation-based Urban
Environment for Specifying the Installation Location of the Optical Particle Counter
Jeong-Gi Lee, Chul-Seung Yang, Gi-won Ku, and Ji-Seong Jeong
275
53
Design and Development of Fan Boss to Reduce Noise and Vibration of Air Purifier
Fan
Hyeong-Sam Park, Doo-Sik Kim, and Sang-Hyun Lee
279
54
Characterization of Activated Carbon Prepared from Walnut (Jaglans regia) Shell
Srijan Adhikari, Rinita Rajbhandari (Joshi), and Rajeshwar Man Shrestha
284
55
A Machine Learning Model for Unintentional Anomaly Detection and Control on
TinyML Lite Platform
Eon-Uck Kang, and Woo-Sang Hwang
289
56
A Study on the Influence of Psychological Factors of Parents on Children's Behavior
Jung-Eun Lee1, and Chun-Ok Jang
294
8th ICAEIC-2022
xxiv
57
Efficient Semantic Segmentation by Using Down-Sampling and Subpixel Convolution
Siti Ummi Masruroh, Obey Al Farobi, Khodijah Hulliyah, Husni Teja Sukmana,
Yusuf Durachman, and Nanda Alivia Rizqy Vitalaya
298
58
A Study on the Combination of Extractive Summary and AI-Based Abstractive
Summary
Woo Won Choi, Jeong Hyeon Kim, Min Kyu Park, Se Jong Oh, and Ill Chul Doo
307
１
AEICP Vol. 5, No. 1
Keynote Speaker
Prof. Dr. Emad Alsusa
School of Electrical and Electronic Engineering, University of
Manchester, Manchester, U.K.
Brief Biography: Emad Alsusa (Senior Member, IEEE) received the Ph.D. degree in telecommunications
from the University of Bath, Bath, U.K., in 2000. In 2000, he became a Postdoctoral Researcher with
Edinburgh University, Edinburgh, U.K. In September 2003, he joined the Manchester University, Manchester,
U.K., where he is currently a Reader with the School of Electrical and Electronic Engineering.
His research interests include wireless networks and signal processing with particular focus on future ultra-
efficient radio resoure management, secret key exchange, and energy neutrality. He is also the U.K.
Representative of the International Union of Radio Science. He was the recipient of a number of awards,
including the Best Paper Award in IEEE PLC-2014 and the IEEE WCNC-2019. He was the Conference
General CO or Chair for the IEEE OnlineGreenCom in 2017 and Sustainability through ICT Summit in 2019
and the TPC Symposia-Chair of several IEEE conferences, including ICC 2015, VTC 2016, GISN 2016,
PIMRC 2017, and GLOBECOM 2018
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635-4586
©ICTAES 2018
２
8th ICAEIC-2022
Overview of Research Issues on 5G, IoT, and D2D Communication Networks
Abstract
The overview of novel concepts proposed by our previous reseach works on 5G, IoT, and D2D communication
networks are summarized in this paper. These novel ideas are a spectral efficient multiple access scheme, a
novel resource allocation scheme of D2D communication, and a physical layer security scheme applied to
MIMO FDD and TDD systems. Wih this overview and understanding of our selected research works, it is very
helpful to foresight the future research trends on 5G, IoT, and D2D communication networks.
Keywords: NOMA, SWIPT, 5G, IoT, D2D.
1. Introduction
The recent research issues on wireless and wired communication networks are mainly focusing on
enhancement techniques for optimizing complexity, power consumption and capacity of such networks. This
includes investigating and developing practical novel algorithms and network architectures, which is
categorized as following.
Wireless Networks
• Green communication networks
• Short-range high data rate Wireless LANs
• IoT and 5G systems
• Bandwidth efficient multiple access protocols
MAC Layer Techniques
• Dynamic & Adaptive Channel Allocation Techniques
• Scheduling and Radio Resource Management Algorithms
• Energy Efficient Protocols
• secure key exchange techniques
Interference cancellation, Receiver Algorithms and Architectures
• Linear and Non-linear Multiuser detection, Precoding and Pre-processing Techniques
• Spatial Modulation and Coding Techniques
• Interference Exploitation and Interference Alignment
Channel Modelling, Estimation and Equalization Techniques
• Maximum Delay and Dominant Paths Estimation
• Advanced Channel Equalization Techniques
• Channel Estimation Techniques under Harsh Conditions/Environments
３
AEICP Vol. 5, No. 1
Wired networks
• Smartgrid techniques
• Impulsive noise mitigation techniques
• Narrowband powerline system
This paper selects and reviews a couple of novel concepts proposed by our previous research works on 5G,
IoT, and D2D communication networks. This paper presents the overview of our selected research works on
spectral efficient multiple access scheme, a novel resource allocation scheme of D2D communication, and a
physical layer security scheme applied to MIMO FDD and TDD systems. Throughout the overview of our
works, the recent research trends and issues on 5G, IoT, and D2D communication network could be
foresighted.
2. Overview and Summary of Our Research Works on 5G, IoT, and D2D
2.1. Spectral Efficient Multiple Access Scheme: SWIP-NOMA
Orthogonal frequency division multiple access (OFDMA) has served as the multiple access scheme in 4G
owing to its significant performance against multipath fading as well as its higher SE compared to the
previously used multiple access schemes [1]. While an OFDMA-based system limits the achievable SE, non-
orthogonal multiple access (NOMA) can further improve the SE and hence it has received considerable
attention as a promising candidate for 5G [2-5]. The combination of simultaneous wireless information and
power transfer (SWIPT) and non-orthogonal multiple access (NOMA) is a potential solution to improve
spectral efficiency (SE) and energy efficiency (EE) of the upcoming fifth generation (5G) networks, especially
in order to support the functionality of the Internet of things (IoT) and the massive machine-type
communications (mMTC) scenarios.
Figure 1.
Illustration of a downlink of a
TS-based SWIPT-NOMA system.
We proposed a joint power allocation and time switching (TS) control for EE optimization in a TS-based
SWIPT NOMA system, satisfying the constraints on maximum transmit power budget, minimum data rate and
minimum harvested energy per-terminal. It is shown that numerical results validate the theoretical findings
４
8th ICAEIC-2022
and demonstrate that significant performance gain over orthogonal multiple access (OMA) scheme in terms of
EE can be achieved by the proposed algorithms in a SWIPT-NOMA system [6].
2.2. IoT, 5G and D2D communications
With the rise of the rapid development of the Internet-of-Things (IoT), 5G and device-to-device (D2D)
communications, these technologies are designed to support massive data access to meet the ever-increasing
demands, the complexity and heterogeneity of the network structure. For a more benign environment for D2D
communication, downlink/uplink decoupling (DUDe) has become an attractive approach [7-9].
We proposed an efficient joint cell-association, subchannel allocation, and power control scheme for
network sum-rate maximization in D2D-underlay heterogeneous networks in [10]. The proposed scheme
includes a minimum path-loss based DUDe to decide the uplink (UL) serving base-station, and a greedy
coloring scheme with modified Munkres algorithm to cluster the UEs, and allocate subchannels. Furthermore,
the difference of convex functions-based programming algorithm for power control is utilized to maximize the
network sum-rate while satisfying the UE transmit power and data rate constraints. The results presented show
that the proposed scheme can achieve superior performance to the coupled benchmarks. [10]
2.3. Physical Layer Security in MIMO FDD and TDD Systems
Along with the ongoing evolution of multiple antennas communication systems, new physical layer security
techniques are continuing to achieve higher levels of secrecy [11-14]. However, most physical layer
approaches concern time division duplex (TDD) channels, which are mainly relying on the use of the channel
reciprocity feature with a large computational burden.
We proposed a new physical layer security method, which is to utilize private random precoding for
exchanging the secret key bits in multiple-input multiple output (MIMO) systems in [15]. The principle of this
novel method is to exploit the precoding matrix index (PMI) in a manner that produces low correlation at the
adversary. A robust key exchange between the transmitter and the receiver is established by uniquely relating
the secret key bits to the channel precoding matrix using a private version of the universal codebooks.
Moreover, the proposed method is applicable in MIMO FDD and TDD systems. The results demonstrate that
the proposed secret key exchange using private random precoding can offer superior performance in terms of
the key agreement, secrecy level and computational load [15].
References
Andrews, J. G., Buzzi, S., Choi, W., Hanly, S. V., Lozano, A., Soong, A. C., & Zhang, J. C. (2014). What will
5G be?. IEEE Journal on selected areas in communications, 32(6), 1065-1082.
Liu, Y., Qin, Z., Elkashlan, M., Ding, Z., Nallanathan, A., & Hanzo, L. (2017). Non-orthogonal multiple access
for 5G and beyond. Proceedings of the IEEE, 105(12), 2347-2381.
Ding, Z., Lei, X., Karagiannidis, G. K., Schober, R., Yuan, J., & Bhargava, V. K. (2017). A survey on non-
orthogonal multiple access for 5G networks: Research challenges and future trends. IEEE Journal on Selected
Areas in Communications, 35(10), 2181-2195.
Ding, Z., Yang, Z., Fan, P., & Poor, H. V. (2014). On the performance of non-orthogonal multiple access in
5G systems with randomly deployed users. IEEE signal processing letters, 21(12), 1501-1505.
５
AEICP Vol. 5, No. 1
Cui, J., Liu, Y., Ding, Z., Fan, P., & Nallanathan, A. (2018). QoE-based resource allocation for multi-cell
NOMA networks. IEEE Transactions on Wireless Communications, 17(9), 6160-6176.
Tang, J., Luo, J., Liu, M., So, D. K., Alsusa, E., Chen, G., ... & Chambers, J. A. (2019). Energy efficiency
optimization for NOMA with SWIPT. IEEE Journal of Selected Topics in Signal Processing, 13(3), 452-466.
Smiljkovikj, K., Popovski, P., & Gavrilovska, L. (2015). Analysis of the decoupled access for downlink and
uplink in wireless heterogeneous networks. IEEE Wireless Communications Letters, 4(2), 173-176.
Tang, H., & Ding, Z. (2015). Mixed mode transmission and resource allocation for D2D communication. IEEE
Transactions on Wireless Communications, 15(1), 162-175
Mustafa, H. A., Shakir, M. Z., Sambo, Y. A., Qaraqe, K. A., Imran, M. A., & Serpedin, E. (2014, December).
Spectral efficiency improvements in HetNets by exploiting device-to-device communications. In 2014 IEEE
Globecom Workshops (GC Wkshps) (pp. 857-862). IEEE.
Shi, Y., Alsusa, E., & Baidas, M. W. (2021). Joint DL/UL Decoupled Cell-Association and Resource
Allocation in D2D-Underlay HetNets. IEEE Transactions on Vehicular Technology, 70(4), 3640-3651.
Zeng, K., Wu, D., Chan, A., & Mohapatra, P. (2010, March). Exploiting multiple-antenna diversity for shared
secret key generation in wireless networks. In 2010 Proceedings IEEE INFOCOM (pp. 1-9). IEEE.
Sayeed, A., & Perrig, A. (2008, March). Secure wireless communications: Secret keys through multipath.
In 2008 IEEE International Conference on Acoustics, Speech and Signal Processing (pp. 3013-3016). IEEE.
Quist, B. T., & Jensen, M. A. (2013). Bound on the key establishment rate for multi-antenna reciprocal
electromagnetic channels. IEEE Transactions on Antennas and Propagation, 62(3), 1378-1385.
Wu, C. Y., Lan, P. C., Yeh, P. C., Lee, C. H., & Cheng, C. M. (2013). Practical physical layer security schemes
for MIMO-OFDM systems using precoding matrix indices. IEEE journal on selected areas in
communications, 31(9), 1687-1700.
Taha, H., & Alsusa, E. (2016). Secret key exchange using private random precoding in MIMO FDD and TDD
systems. IEEE Transactions on Vehicular Technology, 66(6), 4823-4833.
６
8th ICAEIC-2022
Keynote Speaker
Prof. Dr. Bhoj Raj Ghimire
Nepal Open University
,
Nepal
Brief Biography: Graduate in computer application and doctorate in remote sensing and GIS, Bhoj Raj
Ghimire is an assistant professor and program coordinator at Nepal Open University, Lalitpur. He is a digital
governance expert and advisor to government bodies in Nepal. He is into ICT industry for more than a decade
and a half with national and international professional exposure. He blends cutting-edge technologies with
local knowledge, contextual business process and provides sustainable solution to complex business problems
via impactful and visible digital transformation. Among the professional forums, he is active in “Forum of
Digital Equality- FDE”( http://fde.org.np/) as vice president, a member of Multi Stake Group (MSG) for
Nepal Internet Governance Forum (NGIF), member of International Society for Photogrammetry and Remote
Sensing etc.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635-4586
©ICTAES 2018
７
AEICP Vol. 5, No. 1
Multidisciplinary Approach to COVID-19 Risk Communication
Bhoj Raj Ghimire1, and Bhogendra Mishra1,2
1Nepal Open University, Nepal
2Science Hub, Nepal
bghimire@nou.edu.np1, bmishra.rs@gmail.com1,2
Abstract
Hindu Kush Himalaya (HKH) region, also known as water tower for Asia, is a source of water for around two
billion people accross Asia. Deep understanding of water cycle is very essencial as it is very critical resource
in both the positive (for well being of human) as well as negative (disaster) aspects. Due to difficult terrain
and other regions, thre are very limited rainfall measurement stations in the regions. Luckily, there are many
gridded products which help researchers and academicians for the estimates. However, choosing the best one
is always crucial. We evaluated the performance of seven major prcipitation products using statistical
indicators in major river basins in the region for daily, decadal and monthly estimates. Daily estimates of the
products were poor in general. APHRODITE and TRMM were better while CHIRPS performed worst among
the compared products.
Keywords: HKH, gridded precipitation products, TRMM, APHRODITE.
1. Introduction
Majority of the river systems in Asia originates from the HKH. In Asia, the average precipitations are
projected to increase by 1-12% by the end of the century[1]. Thought the precipitation is projected to increase,
the uncertainity of the timing and intensity is also incrasing. The changed precipitation pattern is a big threat
to the sustainability of these reiver systems. Owin to sociaoeconomic and ecological importance of these river
systems, the effects of climate change are of extreme concern amon the planners and stakecholders in the
regions.
Accurate and detailed information on precipitation distrivution is essential to support hydrological study and
water accounting application. Ideally a dense distribution of the meterological staions would provide this
information. However, the distribution of such station in the region is very poor compared to other parts of the
world (Table 1) [2].
Table 1. GHCN precipitation station density
Region
GHCN Stations
Area per station (Sq km)
Africa
2083
14434
８
8th ICAEIC-2022
Asia
6152
4887
Australia
17424
414
Europe
5861
2254
North America
73964
328
South America
6439
2767
HKH
317
18652
GHCN – Global Historical Climatology Network; Source:
https://www1.ncdc.noaa.gov/pub/data/ghcn/daily/ghcnd-stations.txt
Gridded precipitation products developed from satellite imagery sources have well served these gaps and
stood as a strong alternative to the in-situ stations. These products come with varying spatial and temporal
resolution along with individual strengths and weakness. We have evaluated AFRODITE, TRMM, CHIRPS,
PERSIAN-CDR, CMORPH, and WFDEI-GPCC and compared their performances using statisticasl indicators
to understand their perofirmance in Ganga and Mekong River Basin.
2. Materials and Methods
Aphrodite is Asia based product whereas CHIRPS, CMORPH, PERSIANN-CDR and TRMM are near
global and WFDEI CRU and WFDEI CPCC are global products with spatial medium spatial resolution of 0.25
and 0.5 dgre. Some characteristics of the products is highlighted in Table 2.
Table 2. Gridded precipitation product summary
Products
Temporal
coverage
Spatial coverage
Spatial resolution
Temporal
resolution
Primary
sources
APHRODITE V1801_R1
1998 - 2015
Asia
0.25o
Daily
Rain-gauge
CHIRPS v2.0
1983 – present
near global
0.05o
daily
Satellite
CMORPH
2000 – present
Near Global
0.25o
daily
Satellite
PERSIANN-CDR v1 rev1
1983 – present
Near global
0.25o
daily
Satellite
WFDEI CRU
1979 - 2018
Global
0.5o
daily
Rain-gauge
WFDEI GPCC
1901-2016
Global
0. 5o
daily
Rain-gauge
TRMM 3B42 v7
1998 – present
near global
0.25o
3-hourly
Satellite
Two stations one each on higher and low altitude were selected for analysis for Ganga and Mekong River
basins. Daily, dekadal and monthly estimates were investigated for Kathmandu (1337m) and Dehradun
(682m), and Diqing (3320) and Qamdo (3307m) to compare the ground data for 2001-2015 on point-to-pixel
basis[3]. Continuous statistics (correlation coefficient, mean absolute error, bias RMSE) were used to assess
their performance in estimating and reproducing rainfall amounts while categorical statistics (probability of
detection and false alarm ratio) were used to evaluate the rain detection cababilities.
９
AEICP Vol. 5, No. 1
3. Result
In case of Ganga River basin (Figure 1), GPCC, and TRMM were foud best in Kathmandu whereas TRMM
performed best for Hehradun. CHIRPS was worst for botyh Kathmanduy and Dehradun stations. Aphrodite
was best for probability of detection in both the stations. TRMM, CRU and FPCC provided the last false alarm
ratio.
Figure 1.
Performance in Ganga River Basin.
For Mekong basin, APHRODITE and TRMM were found best for Diqing and Qamdo station respectively
wherereas CHIRPS was the worst for botht the stations. Again, APHRODITE and TRMM were better for
probability of detection and PERSIANN CDR and GPCC stood better for fasle alarm ratio respectively.
Figure 2 summariges the result
１０
8th ICAEIC-2022
Figure 2.
Performance in Mekong basin.
4. Conclusion
Since the false alarm ratio of daily estimates are very high, it is suggested not to use in the HKH region,
while decadal and monthly estimates of some products can be well considered. In general APHRODITE and
TRMM performed best in all the inidcators for all the stations in the region and CHIRPS performed the worst.
Very low number of stations in the region must have contirubuted to the worse performance of the CHIRPS
data. The researchers and professionals are suggested to evaluate the products for their study area and objective
of the study before using these gridded products in their study.
References
[1] R. K. Pachauri and A. Reisinger, Climate Change 2007 Synthesis Report. 2007.
[2] NOAA, “GHCND Station.” [Online]. Available: https://www1.ncdc.noaa.gov/pub/data/
ghcn/daily/ghcnd-stations.txt. [Accessed: 01-Dec-2021].
[3] M. Dembélé and S. J. Zwart, “Evaluation and comparison of satellite-based rainfall products
in Burkina Faso, West Africa,” http://dx.doi.org/10.1080/01431161.2016.1207258, vol. 37, no.
17, pp. 3995–4014, Sep. 2016, doi: 10.1080/01431161.2016.1207258.
１１
AEICP Vol. 5, No. 1
Digital Twin Coaching for Openpose
Qiuying Chen1, Seong-Yeol An1, Kyeong-Rak Lee2, and Sang-Joon Lee1
1Interdisciplinary Program of Digital Future Convergence Service, Chonnam National University
2BK4 Education and Research Team of Subtraction Platform, Chonnam National University
chenqiuying11@gmail.com1, syan97@daum.net1, kryi0807@jnu.ac.kr2, s-lee@jnu.ac.kr1
Abstract
With the increasing richness of online fitness resources, autonomous fitness has become a new sporting trend.
However, due to the lack of action guidance and correction from professional fitness coaches, autonomous
fitness usually cannot guarantee the fitness effect and is prone to sports injuries. Therefore, real-time
monitoring of the accuracy of fitness actions is required. Using Digital Twin Coaching, we can fill the gap left
by this issue as it is tailored to work remotely on any modern device. We use OpenPose to establish the
HumanFIT AI dataset for fitness posture evaluation and feedback. Also it is to lay the foundation for
developing AI models and services related to human behavior recognition fields such as home workout, fitness
and daily life in smart cities.
Keywords: digital twin coaching, onpepose, home workout, fitness.
1. Introduction
As defined by El Saddik, Digital Twins (DT) are “digital replications of living as well as nonliving entities
that enable data to be seamlessly transmitted between the physical and virtual worlds” [1]. Digital twins support
digital transformation by copying and simulating new business models and decision support systems [2].
Current companies working with Digital Twins include General Electric, Philips, Sisco Systems, Microsoft,
Oracle, and IBM [3]. We can see that the major players in the technology field have not lagged behind this
trend, and we will see this trend more and more in the future.
Another field in which DT technology can benefit is virtual coaching, more specifically, sports or fitness
coaching. One of the most promising applications for humans is the DT for health and well-being [4]. Barricelli
et al. propose a system called SmartFit, a collection of DTs for health and fitness tracking, fromnutrition and
sleep hours, to physical training [5]. They use machine learning algorithms such as support vector machines
and K-nearest neighbors to advise trainees. Virtual fitness coach came into being. In traditional gyms, the
business of personal fitness trainers includes recommendation of exercise items before fitness, guidance, and
correction of exercises during fitness, and follow-up and supervision after fitness. Among them, action
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１２
8th ICAEIC-2022
guidance and correction are a necessary part, which can ensure the effect of sports and effectively reduce sports
injuries.
The continuous update of human body key point detection technology makes it possible to realize low-cost,
high-precision fitness monitoring functions. The key point detection of the human body is to predict the
important joint points of the human body, and then estimate its action posture, or reedit the action posture.
Methods based on the detection of key points of the human body usually have the advantages of low cost of
use, convenient installation, and diverse scenarios, but there are relatively few studies on the combination of
real-time monitoring of fitness movements. In this paper, we use OpenPose, the champion of the COCO key
point competition in 2016 [6]. To establish the HumanFIT AI dataset for fitness posture evaluation and
feedback to lay the foundation for developing AI models and services related to human behavior recognition
fields such as home training, fitness center exercise, rehabilitation treatment, and daily life in smart cities.
2. Materials and methods
2.1. Data collection
The data of this study were the Korea Intelligent Information Society Agency AI hub data (The latest data
is updated to October 2020). It is composed of fitness posture image AI data. The dataset consists of 1 Json
file by extracting 1-3 video images per second from video information of 200,000 Clips (15 seconds per case),
processing a total of 3 million video images with 24 keypoints.
2.2. Data organization
A detailed human action class for each exercise should be constructed including not only the type of exercise,
but also the point information on the correct posture to perform the exercise accurately. In addition, in line
with the characteristics of movement, which should be able to analyze various body shapes in detail, from the
previously widely used 17 keypoints, neck, palm, spine1/2, and instep were added and annotated as 24
keypoints as shown in figure 1.
Figure 1.
Human keypoint set.
１３
AEICP Vol. 5, No. 1
Through detailed information on correct and incorrect postures, service users can receive real-time feedback
on how their current postures and movements are wrong. The data set to be built uses image information shot
with an RGB camera so that users can receive feedback through easy-to-use devices such as smartphones and
external RGB cameras.
3. Data processing
3.1. Inspection procedure
When processing data, video images extracted from 5 cameras are simultaneously compared and analyzed
by 5multiview, and if the motions do not match, the original data is excluded from processing, transferred to
the editing stage, and re-edited from the captured image information. The processing data is checked whether
the 24keypoint labeling standards are met, and easy corrections are made by the inspector.
3.2. Sampling inspection by secondary random sampling
1% or more of the total processed video images are randomly extracted and inspected in the same way as
the processed data inspection of the first total inspection. And it is verified through AI model and utilization
service.
4. Result & Discussion
The key point for bone prediction was detected through the model. As shown in Figure 2, it was confirmed
that the accuracy was very low. Re-test by resizing the resolution for accuracy.
Figure 2.
Image test result.
Re-test by resizing the resolution for accuracy. As a result of retesting at 640 x 600 with a lower resolution,
the accuracy is improved. Expected to require pre-processing of sample images as shown in Figure 3.
１４
8th ICAEIC-2022
Figure 3.
Image re-test result.
5. Conclusion
The openpose model trained on the HumanFit dataset proposed in this paper is determined from the input
video, and on this basis, the accuracy is adjusted and determined. On this basis, this article wants to strengthen
the diffusion of digital twin in academic and reallife fields. In the next research, we will use deep learning to
determine which poses are the most accurate motion poses. If the movement posture is incorrect, then it is
necessary to detect which keypoint is the most important factor that affects the movement's mistakes, and the
most error-prone posture. In this way, we can provide more advantageous suggestions for home workout
companies. Then the quality of people's life will be further improved, and a complete combination of digital
twin and coaching will be realized.
References
[1] El Saddik A. (Apr. 2018), ‘Digital Twins: The Convergence of Multimedia Technologies’, IEEE
Multimed., vol. 25, no. 2, pp. 87–92, DOI: 10.1109/MMUL.2018.023121167.
[2] VanDerHorn E. and Mahadevan S. (Jun. 2021), "Digital twin: Generalization, characterization and
implementation", Decis. Support Syst., vol. 145, p. 113524, DOI: 10.1016/j.dss.2021.113524.
[3] ‘Top 10 Digital Twin Vendors for 2019 | EM360’. https://em360tech.com/business_agility/tech-news/top-
10-digital-twin-vendors (accessed Nov. 27, 2020).
[4] Wei S, Ramakrishna V, Kanade T, et al (2016). Convolutional Pose Machines[C]. Computer Vision and
Pattern Recognition (CVPR). Las Vegas, USA: IEEE Press, 4724-4732.
[5] Barricelli B. R., Casiraghi E., Gliozzo J. , Petrini A. , and Valtolina S. (2020), ‘Human Digital Twin for
Fitness Management’, IEEE Access, vol. 8, pp. 26637–26664, DOI: 10.1109/ACCESS.2020.2971576.
[6] CAO Z，SIMONT，WEI S. (2017)etal. Realtimemultivperson 2D pose estimation using part affinity
fields[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
Piscataway:IEEE，2017:1302-1310. DOI:10.1109/ CVPR. 2017.143.
１５
AEICP Vol. 5, No. 1
Reusable Fuzzy Extractor from Non-uniform Learning with Errors Problem
Joo Woo1, Jonghyun Kim2, and Jong Hwan Park3
1,2Graduate School of Information Security, Korea University, Seoul, South Korea
3Department of Computer Science, Sangmyung University, Seoul, South Korea
woojoo0121@korea.ac.kr1, yoswuk@korea.ac.kr2, jhpark@smu.ac.kr3
Abstract
A fuzzy extractor (FE) can extract an almost uniform random string from a noisy source with sufficient entropy.
FE generates a helper data and a random string from biometric data and reproduce the random string from
similar biometric data using the helper data. Fuller et al. (Asiacrypt 2013) proposed a computational FE.
However, their construction only can tolerate sub-linear fraction of errors and has inefficient decoding
algorithm. In this paper, we propose a new efficient computational FE. In consequence, our scheme is reusable
and tolerates linear errors. We provide the formal security proof for the proposed FE based on the Non-
uniform Learning with Errors problem.
Keywords: Fuzzy extractor, Reusability, Biometric authentication, Non-uniform Learning with Errors.
1. Introduction
Authentication requires a secret drawn from some source with sufficient entropy. Biometric information,
such as fingerprints, iris patterns and facial features, can be a promising candidate due to its uniqueness and
usability. However, utilizing biometric data has two obstacles; First, once biometric information is leaked to
an adversary, it is not easy to change because of its immutability. Second, whenever one generates biometric
data, small errors occur naturally depending on the various conditions and environments.
Since the concept of a fuzzy extractor (FE) was first proposed by Dodis et al.[1], it has been regarded as one
of candidate solutions for key management utilizing biometric data. Fuller et al.[2] proposed a construction of
computational FE based on the Learning with Errors (LWE) assumption. However, their construction can only
tolerate sub-linear errors and it provides no guarantee of reusability. Reusablity is the security model
formalized by Boyen [3], which is the security notion in the case that several pairs of extracted string and
related helper data issued from correlated biometric data are revealed to an adversary. Later, Apon et al.[4]
modified Fuller’s construction[2] to satisfy the reusability. Nevertheless, it fails to overcome the limitions of
Fuller’s construction: sub-linear error tolerance and inefficient decoding algorithm. Canetti et al.[5]
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１６
8th ICAEIC-2022
constructed a reusable FE based on a powerful tool “digital locker”. However, their construction can only
tolerate sub-linear fraction of errors as well.
In this paper, we propose a new computational FE. In consequence, our scheme is reusable, and can tolerate
linear errors using new decoding algorithm. To cope with the error resulted from the difference between the
biometric data used for registration and the biometric data used for authentication, we encode the extracted
key with two cryptographic primitives: Error Correction Code (ECC) and EMBLEM encoding method[6].
These points lead to improvement of efficiency for the reproduce algorithm and error tolerance for the
biometric data. We provide the formal security proof for the proposed FE based on the Non-uniform LWE
problem[7]. In addition, we provide parameter sets satisfying the security requirements.
2. Preliminaries
2.1. Error Correction Code (ECC)
The central idea of ECC is that the sender encodes the message with redundancy bits to allow the receiver
to detect a limited number of errors that may occur anywhere in the message, and to correct these errors without
re-transmission. In this paper, we adopt (,,) BCH code, where is the block length, is
the message length and is the hamming weight of error.
2.2. EMBLEM Encoding method [6]
We adopt the simple encoding method to tolerate errors. Using the EMBLEM encoding, decoding (during
decryption) is done by simply extracting the most significant bit from each coefficient, in contrast to the Regev
encoding, which requires a rounding operation per each bit.
2.3. Non-uniform Learning with Errors [7]
For integers =()2 and a noise distribution over , and a distribution over , the non-
uniform learning with errors problem ,,,-NLWE is to distinguish between the two distributions
(,= +) and (,),
Where =(),×,
,, and
.
The advantage of an adversary in solving decisional NLWE problem is defined as follows:
()= | [(,)= 1][(,)= 1]|.
3. Definitions
3.1. Reusable Computational FE
A FE consists of two PPT algorithms (Gen, Rep). Gen is the generation algorithm which takes as input
１７
AEICP Vol. 5, No. 1
and outputs (,) with a public helper string and an extracted string {0,1} and Rec is the
reproduction algorithm which takes the public helper string and and outputs an extracted string
{0,1}. The reusability is the security of a reissued pair (,) Gen(), given some pairs (,)
Gen() for correlated and . We follow the security model in [8].
4. Computational FE with Reusability
4.1. Construction
In this paper, we present a computational reusable FE based on NLWE problem. The details of our construction
are as follows.
Gen(): To generate a helper data and an extracted string , it proceeds as follows:
1. Sample × and uniformly where [
,
1].
2. Sample {0,1}
3. Set = (,++ Encode()) where Encode(.)= EMBLEM.Encode(ECC.Encode(. ))
4. Output (,)
Rep(,): To reproduce the string from the helper data = (,) and , it proceeds as follows:
1. Compute =
2. Compute = Decode() where Decode(.)= ECC.Decode(EMBLEM.Decode(. ))
3. Output
Figure 1. Construction of proposed FE scheme.
4.2. Correctness
If dis (,), then we can set |()+|21 where =1with
overwhelming probability through proper parameter setting. If it is the case, then by the correctness of
１８
8th ICAEIC-2022
EMBLEM encoding method, EMBLEM.Decode() = .(). Now, it is easy to get with
ECC.Decode. If |()+|21, then EMBLEM.Decode() = .()
where the hamming weight of is less than with proper parameter setting. Now, by the correctness of
ECC, we can get ECC.Decode(EMBLEM.Decode()) = .
4.3. Security
Theorem 1. If NLWE problem is hard, then our construction is a computational FE with reusability.
Proof. We show the reusability for our construction by defining a sequence of games and proving the adjacent
games indistinguishable. For each for = 0, … ,4, is defined as the event that an adversary
correctly guesses.
Game : It is the original game ,
(0) in which =,+++()(w)
and =,++() Gen(). Clearly, we have []=[,
(0)1] .
Game : It is the same as , except that Gen(+) and Gen(), now is changed to
=,+() and =,+() where ,[/,/]
×,,
, and ,
{0,1}. Therefore, |[][]|().
Game : It is the same as , except that the coin = 0 is replaced to = 1, that is, the challenger
gives (,) to the adversary. In this game, the helper data and all for = 1, … , have no
information about and , are uniformly random. Therefore, []=[].
Game : It is the same as , except that =,+() and =,+() now is
changed to =,+++() and =,++(). That is, it is the
original game ,
(1). Therefore, we have []=[,
(1)1] . Also, we get |[]
[]|().
As a result, we get [,
(1) 1[,
(0) 12(). □
4.4. Parameter Setting
We provide two parameter sets. For 128-bit security with 20% error tolerance rates, we set =256,=
128,=511, and =130 when BCH(511,130,55) code is used. For 256-bit security with 10% tolerance
rates, we set =256,=256,=1023, and =258 when BCH(1023,258,106) code is used. We
estimate the concrete hardness of NLWE based on the reduction between NLWE and original LWE and use
the LWE-estimator Albrecht et al. presented[9].
5. Conclusions and Future Work
In this paper, we propose a new computational FE improved upon across all aspects of efficiency and
１９
AEICP Vol. 5, No. 1
tolerance error rates. In consequence, our scheme is reusable, and can tolerate linear errors using new decoding
algorithm using ECC and multi-bits encoding method. These points lead to improving the efficiency for the
reproduce algorithm and supporting linear fraction of errors for the biometric data. We provide the formal
security proof for the proposed FE based on the Non-uniform LWE problem.
In future work we plan to integrate the robustness to our FE scheme, which is the security notion enable to
verify the integrity of the helper data. Moreover, we will provide optimized parameter sets and
implementations for our scheme through a rigorous analysis of the relation between false acceptance rate and
false rejection rate. Lastly, we are planning to study some deep learning-based extraction technique to extract
features from raw biometric data such as face, fingerprints, and iris.
Acknowledgement
This work was supported by Institute of Information & communications Technology Planning & Evaluation
(IITP) grant funded by the Korea government (MSIT) (No.2021-0-00518, Blockchain privacy preserving
techniques based on data encryption).
References
[1] Dodis, Y., Reyzin, L., & Smith, A. (2004). Fuzzy extractors: How to generate strong keys from biometrics
and other noisy data. In Proceedings of the Advances in Cryptology (EUROCRYPT 2004), pp. 523–540.
[2] Boyen, X. (2004). Reusable Cryptographic Fuzzy Extractors. In Proceedings of the 11th ACM Conference
on Computer and Communications Security, pp. 82–91.
[3] Fuller, B., Meng, X., & Reyzin, L. (2013). Computational Fuzzy Extractors. In Proceedings of the
advances in Cryptology (ASIACRYPT 2013), pp. 174–193.
[4] Apon, D., Cho, C., Eldefrawy, K., & Katz, J. (2017). Efficient, Reusable Fuzzy Extractors from LWE. In
Proceedings of the International Conference on Cyber Security Cryptography and Machine Learning, pp.
1–18.
[5] Canetti, R., Fuller, B., Paneth, O., Reyzin, L., & Smith, A. (2016). Reusable fuzzy extractors for low-
entropy distributions. In Proceedings of the Advances in Cryptology (EUROCRYPT 2016), pp. 117–146.
[6] Seo, M., Kim, S., Lee, D. H., & Park, J. H. (2020). EMBLEM:(R) LWE-based key encapsulation with a
new multi-bit encoding method. International Journal of Information Security, pp. 383-399.
[7] Boneh, D., Lewi, K., Montgomery, H., & Raghunathan, A. (2013). Key homomorphic PRFs and their
applications. In Annual Cryptology Conference. pp. 410-428.
[8] Wen, Y., & Liu, S. (2018, July). Reusable fuzzy extractor from LWE. In Australasian Conference on
Information Security and Privacy, pp. 13-27.
[9] Albrecht, M. R., Player, R., & Scott, S. (2015). On the concrete hardness of learning with errors. Journal
of Mathematical Cryptology, 9(3), 169-203.
２０
8th ICAEIC-2022
Image Enhancement Malaria Parasite Detection System based on Giemsa-
Stained Slide Images using Homomorphic Filtering
Edy Victor Haryanto S1, Bob Subhan Riza2, Juli Iriani3, and Noprita Elisabeth S4
1,2,3,4G Universitas Potensi Utama, Indonesia
edyvictor@gmail.com1, bob.potensi@gmail.com2, juli@potensi-utama.ac.id3, novryelisa@gmail.com4
Abstract
Malaria is an infectious disease caused by a parasite called plasmodium and this disease can be transmitted
to humans through a female mosquito called Anopheles which is still developing, especially in the tropics, in
research aimed at helping to improve the images taken from patients infected with malaria through a
microscope, the image taken is BMP format then the image is converted to grayscale and then dicroped the
existing object to facilitate the parasite in the image and the method used is Homomorphic Filtering, after
which binarization is carried out with the Otsu method, the result is the malaria parasite object in the picture
looks better.
Keywords: otsu, homomorphic filtering, parasite malaria, grayscale.
1. Introduction
Speed high enough spread of malaria parasites. The risk of death of the patient will be high if the diagnosis
is received late. On the other hand there are many obstacles. Morphological changes in malaria parasites, as
well as the variety of strains (strains) make it difficult to identify the malaria parasite. Another problem is the
low quality of the microscope at some medical centers that affect the process of diagnosis, but for gold
microscope use method is still experiencing weakness because the process is long and many misdiagnosys
made [1].
Developments in the field of detection of malaria have undergone rapid development [7] has developed a
malaria detection by using decision support system, but still experiencing problems because of not applying
the edge detection and filter. Nasir suggests in research of malaria, where Nasir using a segmentation approach
clustering K-Mean, but in this study is still the basic steps of detection of malaria parasites in the blood image
[5]. The importance of research on malaria and the development of databases on malaria was also proposed by
Ghosh [3], where the database is used as an enrichment type and form of the malaria parasite. And used as a
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２１
AEICP Vol. 5, No. 1
comparison in the detection of malaria. Plasmodium parasite detection in the image of red blood cells may
help diagnose malaria quickly and accurately, especially in areas that do not have medical expertise.
Broadly speaking there are three types of image enhancement methods, namely (a) image enhancement with
binarization and thresholding methods, (b) image enhancement with hybridization method between
binarization / thresholding and other methods, and (c) image enhancement without thresholding method.
2. Literature Review
2.1 Malaria Parasite
Malaria is an infectious disease caused by the parasite Plasmodium which can be characterized by fever,
hepatosplenomegaly and anemia. Plasmodium live and breed in human red blood cells. The disease is naturally
transmitted through the bite of a female Anopheles mosquito.
Plasmodium species in humans is [6] :
1. Plasmodium falciparum (P. falciparum).
2. Plasmodium vivax (P. vivax)
3. Plasmodium ovale (P. ovale)
4. Plasmodium malariae (P. malariae)
5. Plasmodium knowlesi (P. knowlesi)
Figure 2. Malaria Parasite.
2.2. Homomorphic Filtering
In image processing, homomorphic filtering is one method that can use for to compensate the effects of
uneven illumination on the image and enhance the appearance of simultaneous image compression varying
intensity and contrast enhancement [2] [4]. According to this model, an image has the following equation:
(,)=(,)(,) (1)
２２
8th ICAEIC-2022
where(,) is an image which is the result of multiplication (product) which is a component of (,) the
illumination with which a reflectance component. To separate two independent components and facilitate their
separate, algorithm transform on Eq. (1) has been taken, Thus Spake
(,) = ln (,)
= ln (,)+ ln (,) (2)
Then, the Fourier transform of Eq. (2) is calculated:
ℑ
{(,)} =
ℑ
{ln (,)}
=
ℑ
{ln (,)}+
ℑ
{ln (,)} (3)
Or : (,) = (,)+ (,) (4)
where (,) dan (,) are the Fourier transformation of ln (,)danln (,). After being moved in
the frequency domain, then the image is processed by using appropriate filters in order to be able to achieve
the original purpose of which is to attenuate low-frequency and high-frequency strengthens resulting in image
enhancement and image sharpening by formula
(,) = H(,)(,)
=(,)(,)+(,) (,) (5)
where (,) is the Fourier transform of the image that has been processed. So as to obtain results that
actually needs to be returned to the spatial domain by the formula:
(,) =
ℑ
-1{(,)}
=
ℑ
-1{(,)(,)}+
ℑ
-1{(,)(,)} (6)
By defining (,) =
ℑ
-1{(,)(,)} (7)
and (,) =
ℑ
-1{(,)(,)} (8)
Eq. (6) can be expressed as follows : (,)=(,) + (,) (9)
The last step is eliminating the logarithmic operation is done early in the process by performing an
exponential operation in order to obtain the desired enhanced image is denoted by g (x, y), namely
g(,)=(,)
=(,)(,)
= (,)(,) (10)
２３
AEICP Vol. 5, No. 1
where (,)=(,) and (,)=(,) is lighting and the reflection component of the output of
each - each image. The H(,) normally used in this procedure is the Butterworth high passfilter defined as
(,)=( )
(/(,))+ (11)
where is the cut off distance measured from the origin, (,) is distance from the origin of centered
Fourier transform, and n is the order of the Butterworth filter.
3. Methodology
3.1. Data Acquisition
The original image here is an image of the blood taken from patients with suspected malaria, the image is
Plasmodium falciparum (P. falciparum). The original image obtained from the original data from hospitals that
have been collected will be used in the analysis.
Figure 2. Original Image
3.2. Algorithm
The results from the acquisition of the data stored on the computer is cut into small sections. Each - each
section is converted to grayscale image first. Then the result of the grayscale image filtering using
Homomorphic Filtering. Results filtration each - each section later in binarization with binarization Otsu [8]
and the results are put back together into a complete image. Next binarization results were evaluated by
questionnaire to some correspondents to assess the results of image enhancement.
２４
8th ICAEIC-2022
Original
Image Cropping
Image Grayscale
Image
Binarization
Otsu Enhanced
Image
Homomorphic
Filtering
Figure 3. Algorithm Based Process
4. Result And Discussion
4.1 Result Homomorphic Filtering
Homomorphic applied in this study using a Butterworth filter that is adopted from [8]. In this paper, the
reflectance component of images are Considered for further processes. Therefore, homomorphic filtering with
a Butterworth high pass kernel is a convenient way to Achieve this Butterworth filter component formula is as
follows : (,)=( )1
1 + (/(,))+
In Figure 4 is an original image that has not been modified. Figure 5 is a grayscale image of the image,
Figure 6 is the result of the image after applied Homomorphic Filtering.
Figure 4. Original image
２５
AEICP Vol. 5, No. 1
Figure 5. Grayscale image
Figure 6. Homomorphic Filtering
4.2 Otsu Binarization
After the homomorphic filtering and adaptive median filtering, then it is Followed by binarization process
using Otsu method. Results from the process Otsu binarization method shown in Figure 7. The approach taken
by the Otsu method is by performing discriminant analysis is to determine a variable that can distinguish
between two or more groups that arise naturally. Discriminant analysis will maximize these variables in order
to divide the foreground and background objects.
Figure 7. Otsu Binarization
２６
8th ICAEIC-2022
4.3 Discussion
From the results obtained can be seen that the homomorphic filtering has been able to perform image
enhancement in the detection of malaria. On the results of image enhancement, all the malaria parasite picture
looks fine, but there are some images of blood cells are somewhat opaque. this is caused by the Otsu
binarization result from the input image is categorized as part of the background and rated the white pixel so
that the image somewhat blurred in some parts. In addition, the image has a lot of noise, some parts of the
picture are categorized including foreground and rated the pixel black, cause there is still some residual noise.
Although the results ideally every part of the background is converted to black and every backgound changed
to white. Otsu binarization error can be resolved by performing the filtering process using homomorphic
filtering before Otsu binarization process.
5. Conclusion
Based on these results it can be concluded homomorphic filtering with binarization Otsu good to improve
the malaria parasite. Steps taken to repair the image of the malaria parasite is using homomorphic filtering
filtration process and then binarization is done by using the Otsu binarization method. Otsu method is able to
provide very satisfactory results. This method succeeded in producing images of malaria parasites in the area
with malaria image is always expressed with black and white outwardly, so that the geometric shape of the
image of malaria in the binary image becomes very clear. Binarization process image of malaria by setting the
threshold value T at a certain value on the results of image enhancement using homomorphic filtering malaria
parasite shows malaria image detection can be improved better.
Reference
[1] Arco, J. E., Górriz, J. M., Ramírez, J., Álvarez, I., & Puntonet, C. G. (2015). Digital image analysis for
automatic enumeration of malaria parasites using morphological operations. Expert Systems with
Applications, 42(6), 3041-3047.
[2] Bock, R., Meier, J., Nyúl, L. G., Hornegger, J., & Michelson, G. (2010). Glaucoma risk index: automated
glaucoma detection from color fundus images. Medical image analysis, 14(3), 471-481.
[3] Ghosh, S., & Ghosh, A. (2013, December). Content based retrival of malaria positive images from a
clinical database. In Image Information Processing (ICIIP), 2013 IEEE Second International Conference
on (pp. 313-318). IEEE.
[4] Gonzalez, R. C., & Woods, R. E. (2009). Digital image processing 3rd edition.
[5] Nasir Abdul, A. S., Mashor, M. Y., & Mohamed, Z. (2012, December). Segmentation based approach for
detection of malaria parasites using moving k-means clustering. In Biomedical Engineering and Sciences
(IECBES), 2012 IEEE EMBS Conference on (pp. 653-658). IEEE.
２７
AEICP Vol. 5, No. 1
[6] Rakshit, P., & Bhowmik, K. (2013, December). Detection of presence of parasites in human RBC in case
of diagnosing malaria using image processing. In Image Information Processing (ICIIP), 2013 IEEE
Second International Conference on (pp. 329-334). IEEE.
[7] Prasad, K., Winter, J., Bhat, U. M., Acharya, R. V., & Prabhu, G. K. (2012). Image analysis approach for
development of a decision support system for detection of malaria parasites in thin blood smear images.
Journal of digital imaging, 25(4), 542-549.
[8] Shahamat, H., & Pouyan, A. A. (2014). Face recognition under large illumination variations using
homomorphic filtering in spatial domain. Journal of Visual Communication and Image Representation,
25(5), 970-977.
[9] S. E. V. Haryanto, M. Y. Mashor, A. S. A. Nasir and H. Jaafar, "A fast and accurate detection of Schizont
plasmodium falciparum using channel color space segmentation method," 2017 5th International
Conference on Cyber and IT Service Management (CITSM), 2017, pp. 1-4, doi:
10.1109/CITSM.2017.8089290.
[10] S. E. V. Haryanto, M. Y. Mashor, A. S. A. Nasir and H. Jaafar, "Malaria parasite detection with histogram
color space method in Giemsa-stained blood cell images," 2017 5th International Conference on Cyber
and IT Service Management (CITSM), 2017, pp. 1-4, doi: 10.1109/CITSM.2017.8089291.
[11] I. Muhimmah, N. Harniawati and N. Lusiyana, "Characteristics determination of infected erithrocytes by
plasmodium falciparum as a diagnostic of malaria, based on microscopic images," 2017 International
Conference on Advances in Computing, Communications and Informatics (ICACCI), 2017, pp. 1844-
1847, doi: 10.1109/ICACCI.2017.8126113.
[12] H. A. Nugroho et al., "Segmentation of Plasmodium using Saturation Channel of HSV Color Space," 2019
International Conference on Information and Communications Technology (ICOIACT), 2019, pp. 91-94,
doi: 10.1109/ICOIACT46704.2019.8938471.
２８
8th ICAEIC-2022
CNN-based User Behavior Detection using Time-division Feature Data
JinSuk Bang1, and NaeJoung Kwak2
1Department of AI Convergence, HoSeo University, Korea
2Division of AI Software Engineering (Information Security), Pai Chai University, Daejeon, Korea
bluegony @hodeo.edu1, knj0125@pcu.ac.kr2
Abstract
In this paper, we propose a method to detect user behavior by applying the information of the 3-axis
accelerometer and 3-axis gyro sensor embedded in the smartphone to a convolutional neural network. Human
behavior has different starting and ending times including the duration of signal data constituting the motion
according to the size and range of the motion. We proposed a CNN-based behavior detection model that detects
the behavior of a smartphone user by extracting features divided according to time sections by inputting sensor
data.
Keywords: Deeplearning, CNNI, Time-series data classification, Human activity recognition, Sensor.
1. Introduction
Human Activity Recognition (HAR) refers to recognizing actions by collecting and analyzing data related
to human motion using various sensors. Since the spread of smartphones, interest in wearable devices has
increased and diversified, it has become possible to easily collect a large amount of data closely related to life
and simple to apply to real life. Therefore, studies are being conducted to analyze and recognize behavior
patterns to understand user behavior [1].
Recently, as the performance of deep learning has improved due to the development of hardware and
algorithms, interest in behavior recognition methods using deep learning is increasing. Deep learning is a
machine learning algorithm that uses deep neural networks, and unlike other machine learning, it automatically
extracts features, so generalized model learning is possible. Therefore, a lot of research is being conducted in
the field of user behavior recognition based on smartphone sensors [2].
In order to analyze and understand human behavior patterns, it is necessary to distinguish between static
and dynamic motions, which are physical and most basic motions [3]. In this experiment, according to the data
set, the user's behavior range was defined as three motions of standing, walking, and running, and six motions
of walking, standing, sitting, lying, climbing stairs, and going down stairs.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２９
AEICP Vol. 5, No. 1
In this paper, we propose a method for detecting human behavior using data collected through the smart
phone's built-in sensor. For behavior recognition, the deep learning algorithm of the convolutional neural
network is applied to extract various time-division features based on the time axis of the input data, and the
behavior is detected through the feature fusion convolutional neural network.
1.
The proposed method
Figure 1 shows the overall structure of the proposed system. The proposed method consists of a time division
layer that divides input data, a convolutional neural network that learns about the divided features, and a feature
fusion classifier that classifies motions with multi-layered features.
Figure 3.
Architecture of the proposed method.
The time division layer learns data divided by time based on input data. The time division information
extracted through the time division layer has information divided based on time in the signal data as well as
the entire data area, and it helps to accommodate even the actions performed in a short time.
The convolutional neural network is the part that learns the characteristics of time-divided data. The
structure of the convolutional neural network used in the proposed model consists of two types of modules.
The first module consists of three convolutional layers and one downsampling layer, and the second module
consists of two convolutional layers and one downsampling layer. Also, all convolutional layers are composed
of BN, ReLU, and Conv(1X3). Figure 2 shows the structure of the proposed convolutional neural network.
The feature fusion classifier classifies motion by fusing some feature maps of the convolutional neural
network and features extracted from the FC layer. The proposed classifier accommodates both long and
shorttime intervals by fusing multi-layered features.
Figure 2.
The structure of the proposed convolutional neural network.
Figure 3 compares the precision and recall results according to the application of the time division layer in
the proposed model. (a) is the result of not using the time division layer, and (b) is the result of the neural
network to which the time division layer is applied. In the case of the AGBS data set, it can be seen that both
the precision and recall are high regardless of whether the time-division layer is used or not, because the
composition of the motions, such as walking, running, and standing, is characterized by dynamic and static
３０
8th ICAEIC-2022
motions. On the other hand, in the case of the UHD data set, in the case of the neural network that does not
use the time division layer, it can be seen that the precision and recall of the sitting and standing motions are
relatively low compared to other motions.
Figure 3.
Comparison of precision and recall results according to application of time division layer
3. Conclusion
In this paper, a system that can detect the user's behavior using the sensor embedded in the smartphone is
proposed. When a pattern of motion data such as human behavior is applied to the proposed method, it can
show high performance because different features are extracted according to time. In the future, research on
how to reflect other features of smartphones is needed.
References
[1] D. Roggen, G. Troester, P. Lukowicz, L. Ferscha, J. Millan, R. Chavarriaga. (2013). Opportunistic human
activity and context recognition. IEEE Computer, 46(2), 36-45.
R. A. voicu, C. Dobre, L. Bajenaru, and R. I. Ciobanu. (2019). Human physical activity recognition using
smartphone sensors. Sensors, 19(3), 458.
S. M. Lee, H. Y. Joe, and S. M. Yun. (2016). Machine learning analysis for human behavior recognition based
on 3-axis acceleration sensor. Information and communication, 33(11), 65-70.
(a)
(b)
３１
AEICP Vol. 5, No. 1
Alphanumeric CAPTCHA Recognition Using CNN-BiLSTM Network
Bal Krishna Nyaupane1, Rupesh Kumar Sah2, and Bishnu Hari Paudel3
1,2,3Department of Electronics and Computer Engineering, Institute of Engineering, Tribhuvan University
bkn@wrc.edu.np1, rupesh@wrc.edu.np2, bishnuhari@wrc.edu.np3
Abstract
The Completely Automated Public Turing Test to Tell Computers and Humans Apart (CAPTCHA) is a critical
human-machine distinguishing technique for websites to prevent malicious software attacks. Websites can
improve their security and prevent malicious internet attacks by using CAPTCHA verification to determine
whether the end-user is a human operator, or attacking programs, or other computerized agents that attempt
to mimic human understanding. The most prevalent type is text-based CAPTCHA, which is designed to be
easily recognized by a human operator but difficult to recognize by computerized agents or robots. In this
research, a combination of Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term
Memory (BiLSTM) model has proposed. The CNN-BiLSTM network model is constructed to recognize the 5-
characters alphanumeric CAPTCHA. The CNN-BiLSTM model first builds a CNN network to extract features
from the input CAPTCHA images. Then, these feature sequences are input to the two-layer BiLSTM networks
to extract detail structural features of the CAPTCHA images, and finally, the output sequence is generated as
the 5-alphanumeric characters. The proposed model can recognize alphanumerical CAPTCHAs. To train our
proposed model, we have used a Kaggle dataset of 3,000 CAPTCHA images. After training and testing, we
achieved a better recognition and build a more robust model for noise in alphanumeric CAPTCHA recognition.
The problem of over-fitting is minimized through kernel regularization and dropout methods.
Keywords: CAPTCHA, CNN-BILSTM, alphanumerical, extract features, human operator.
1. Introduction
CAPTCHA is a computer test that can tell the difference between humans and robots. As a result,
CAPTCHA could be used to protect web services, websites, login credentials, and even semiautonomous
vehicles and driver assistance systems from various cyber security threats, attacks, and penetrations [1]. With
the rapid growth of the Internet sector, there is an increasing number of network security challenges.
CAPTCHA technology is used in a variety of network security and information security applications [2].
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
３２
8th ICAEIC-2022
Cyber-attacks frequently result in circumstances where computer programs take the place of humans and
attempt to automate services such as sending a large number of unwanted emails, accessing databases, or
influencing online pools or polls. One of the most common types of cyber-attacks is the distributed denial-of-
service (DDoS) assault, in which the target service is overloaded with unexpected traffic either to find the
target credentials or to temporarily halt the system. In the advancement of cyber security systems, using a
CAPTCHA system is one of the old yet extremely successful options. As a result, the attacking machines can
be identified, and odd traffic can be blocked or ignored to avoid damage [1].
This research introduces the CNN-BiLSTM model, which solves the challenge of complicated alphanumeric
CAPTCHA identification while also addressing issues with typical neural network training such as model
complexity and output layer parameter redundancy. The proposed method's efficiency is demonstrated by the
test results on a test dataset. The remaining information is divided into three sections: Section 2 introduces
related work, while Section 3 focuses on materials and technique. The experimental data and analyses are
described in Section 4. The conclusion is then delivered.
2. Related Work
Methods for recognizing alphanumeric-based CAPTCHA have also been conducted previously using a
different approach. Researchers have developed many deep learning approaches for CAPTCHA images
recognition in recent years.
In 2019, J. Wang et al. [3] implanted 4-layered called Dense Convolutional Network (DenseNet) with cross-
layer connection for CAPTCHA recognition. They trained the model on three sets of datasets of Chinese or
English CAPTCHA with four or five characters. The maximum obtained accuracy rate of CAPTCHA with the
background noise is above 99.9%. In 2019, Y. Shu et al. [4] build the model for 4-characters CAPTCHA
recognition using an end-to-end deep CNN-RNN network model. The CNN model was constructed with
Residual Convolution Block to avoid the gradient vanishing problems. The CNN extracts the features of input
images of CAPTCHA, and then further deep internal features are extracted by the two-layer GRU network and
recognize the CAPTCHA. The model was trained on four different datasets of size: 60000 CAPTCHA images
containing 50,000 training datasets and 10,000 test datasets. In 2021, O. Bostik1 et al. [5] introduced a novel
technique for training the CAPTCHA system that did not require the construction of a manual dataset. The
researched used brute-force attacks combined with transfer learning to assess the CAPTCHA system's
vulnerabilities. They have used a 15-layers convolutional neural network and fine-tuning of transfer learning
to break the 5-digit text-based CAPTCHA system. The maximum obtained accuracy was 80% when the 6 fine-
tuning steps with 50000 attacks per step.
3. Materials and Methodology
In this section, image acquisition, transfer learning, pre-trained VGG models, proposed methodology, fine-
tuning and hyperparameter used for pre-trained models and performance metrics for evaluation are briefly
explained to fulfill the objective of this research.
３３
AEICP Vol. 5, No. 1
3.1. Dataset
Two different Kaggle datasets has been used in this research. The Dataset-1 has 2,000 five-characters and
Dataset-2 has 1,070 CAPTCHA images. The main objective of this research is to recognize alphanumeric
characters of CAPTCHA images. As mentioned in Table 1, the training dataset of Dataset-1 includes 1600
CAPTCHA images; the validation dataset includes 300 CAPTCHA images, and the testing dataset includes
100 CAPTCHA images. Similarly, the training dataset of Dataset-2 includes 800 CAPTCHA images; the
validation dataset includes 200 CAPTCHA images, and the testing dataset includes 70 CAPTCHA images.
There are thirty-six possible combinations of alphanumeric characters in CAPTCHA images. One
CAPTCHA image has only five random alphanumeric characters. OneHotEncoding method has used for
CAPTCHA images labeling before input to the CNN network.
Table 1. Dataset Descriptions.
Dataset-1
Dataset-2
Training Dataset
1600
800
Validation Dataset
300
200
Test Dataset
100
70
Figure 1. Alphanumeric CAPTCHA images of Dataset -1
Figure 2. Alphanumeric CAPTCHA images of Dataset -2
3.2. Block diagram of Proposed Model
Figure 3. The block diagram of purposed model
３４
8th ICAEIC-2022
Figure 3 shows all the components and model used during this research. As shown in figure 1 and 2, two
different datasets have used for model training and testing. The CAPTCHA images size is 180 × 50. The
purposed model has two different architectures: CNN and BiLSTM. The final layer has five fuly connected
layers one for each alphanumeric characters of CAPTCHA image. Thr ReLU activation function has used for
the inner layer of CNN for the feature extraction and the softmax activation function has used at the output
layers for character recognition.
3.3. Convolution Neural Network
A Convolutional Neural Network (CNN) is a multilayer Deep Learning network in which the output of one
layer becomes the input of the next layer. An input, one to multiple hidden layers, and an output are the most
common components. CNN can take an image as input, assign learnable weights to distinct objects in the
image, and distinguish between them. The purpose of the convolution process is to extract high-level features
from the input image. The first convolution layer is mainly responsible for capturing Low-Level information
such as edges, color, gradient direction, and so on. With the addition of layers, the architecture adjusts to the
high-level characteristics as well, giving us a network that understands the images in the same way that we do
[6]. Unlike other machine learning classification algorithms, CNN requires a significantly low level of
preprocessing. By using relevant filters, a CNN is able to successfully capture the spatial and temporal
relationships from the input image.
After the success of AlexNet's in 2012, CNNs have gone through a number of upgrades. Convolutional layer,
pooling layer, activation function, loss function, regularization, and optimization are six significant
advancements on CNNs [6]. The major role of the pooling layer is to reduce the spatial size of features. It also
helps keep the model's training process smooth by extracting rotational and positional invariant dominating
features. Maximum pooling and average pooling are the two types of pooling. Max Pooling returns the
maximum value from the portion of the image covered by the kernel. On the other hand, average pooling
returns the average of all the values from the kernel's section of the image.
3.4. Bidirectional Long Short-Term Memory
Hochreiter and Schmidhuber proposed Long Short-Term Memory Neural Networks (LSTM) in 1997. They
are an RNN extension that can solve the problem of gradient vanishing using its memory, which allows it to
read, write, and delete data through three gates: the first allows or blocks updates (Input Gate); the second
disables a neuron if it is not important based on the weights learned by the algorithm, which determines its
importance (Forget Gate); and the third is a control gate of the neuron state in the output (Output Gate) [ [7].
Long Short-Term Memory (LSTM) and Bi-directional Recurrent Networks (BiRNN) are combined in Bi-
LSTM. Recurrent Neural Network (RNN) is a specific development of Artificial Neural Networks (ANN) that
processes sequences and time-series data. RNN offers the advantage of encoding input dependencies. Bi-
LSTM is created when Bi-RNN and LSTM are joined. So Bi-LSTM excels due to the advantages of LSTM in
the form of storage in cell memory and Bi-RNN with access information from the context before and after [8].
It allows the Bi-LSTM to benefit from the LSTM's feedback for the next layer.
３５
AEICP Vol. 5, No. 1
4. Experimental Result and Analysis
To conduct this research,8th generation Intel i7 processors with four cores and 8GB of RAM have been
used to run code through a Jupiter notebook of the Anaconda environment. The proposed model is trained on
two different datasets one after another to complete this research.
The main objective of this research is to tune the hyperparameters to improve the accuracy of CAPTCHA
image recognition using preprocessed image as an input. In addition to this, train the CAPTCHA images by
using CNN-BiLSTM model. In each model, different combinations of hyperparameters have been tuned to
obtain the best results. The models had been tested on the test dataset after the completion of the training phase.
Result in Terms of Training Accuracy and Test Loss
(a)
(b) (c)
Figure 6. (a)The model training accuracy plot of individual characters in CAPTCHA images of Dataset- 2; (b)
The avarage model training accuracy vs. validation accuracy plot of CAPTCHA images of Dataset-
2; (c) The model training loss vs. validation loss plot of CAPTCHA images of Dataset-2.
３６
8th ICAEIC-2022
Initially, the proposed model was trained on Dataset 2. The training accuracy of individual characters,
average model training accuracy, and validation accuracy, and model training loss vs. validation loss plot of
CAPTCHA images of Dataset 2 is shown in Figure 6. Even though we had implemented the L1 and L2 kernel
regularization with the Dropout function in the dense layer, there are still overfitting problems. To achieve a
better result than the state-of-the-art methods, we had built more than ten different architectures of CNN-
BiLSTM networks by changing different hyperparameters such as learning rate, number of convolution layer,
number of neurons in convolution layer, dropout rage, BiLSTM layers, and BiLSTM neurons. Among them,
the better one result has been presented in figure 6.
(a)
(b) (c)
Figure 7. (a)The model training accuracy plot of individual characters in CAPTCHA images of Dataset- 1; (b)
The avarage model training accuracy vs. validation accuracy plot of CAPTCHA images of Dataset-
1; (c) The model training loss vs. validation loss plot of CAPTCHA images of Dataset-1.
The proposed model was also trained on Dataset 1. The training accuracy of individual characters, average
model training accuracy, and validation accuracy, and model training loss vs. validation loss plot of
CAPTCHA images of Dataset 1 is shown in Figure 7. with additional number of images in Dataset 1, L1 and
３７
AEICP Vol. 5, No. 1
L2 kernel regularization, and dropout layer in the dense layer significantly reduce the overfitting problems. To
achieve a better result, we built different architectures of CNN-BiLSTM networks by changing different
hyperparameters such as learning rate, number of convolution layer, number of neurons in convolution layer,
dropout rage, BiLSTM layers, and BiLSTM neurons. Among them, the better one result has been presented in
figure 7.
5. Conclusion and Future Enhancement
In this research, we have implemented a CNN-BiLSTM model to recognize CAPTCHA images for two
different datasets. The onHotEncoding technique has been used for CAPTCHA image labeling for the training
process. The labeled images are input images for the CNN architectures. The CNN structure is used to extract
the basic features from the CAPTCHA images, and then BiLSTM architecture is used to extract the deep
features of the alphanumeric characters from the images. During Training, we have tuned the number of
hyperparameters such as learning rate, number of convolution/BiLSTM layer, number of neurons in the layer,
dropout rage, and L1 & L2 regularization parameters to achieve a better result than the state-of-the-art methods.
The model training and validation accuracy for Dataset-1 is higher than the Dataset-2 keeping the same
hyperparameters. The average validation accuracy of the model for Dataset-1 is around 98%, and the average
validation accuracy for Dataset-2 is around 81%.
In the future, other encoding techniques and robust image preprocessing steps can be applied to enhance the
features of CAPTCHA images. Also, the performance of the proposed method can be improved by adding a
large number of training images for the training phases. In addition to this, more hyperparameters can be tuned
after the feature extraction process.
References
[1] Z. Noury and M. Rezaei, "Deep-CAPTCHA: a deep learning based CAPTCHA solver for vulnerability
assessment," arXiv:2006.08296v2, 2020.
[2] Y. Hu, L. Chen and J. Cheng, "A CAPTCHA recognition technology based on deep learning," in 13th
IEEE Conference on Industrial Electronics and Applications (ICIEA), 2018.
[3] J. Wang, J. Qin, X. Xiang, Y. Tan and N. Pan, "CAPTCHA recognition based on deep convolutional
neural network," Mathematical Biosciences and Engineering, vol. 16, no. 5, pp. 5851--5861, 2019.
[4] Y. Shu and Y. Xu, "End-to-End Captcha Recognition Using Deep CNN-RNN Network," in IEEE 3rd
Advanced Information Management,Communicates,Electronic and Automation Control Conference
(IMCEC 2019), 2019.
[5] O. Bostik, K. Horak, L. Kratochvila, T. Zemcik and S. Bilik, "Semi-supervised deep learning approach to
break common CAPTCHAs," Neural Computing and Applications, Springer, pp. 1-11, 2021.
[6] J. Gua, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liub, X. Wang, L. Wang, G. Wang, J. Cai
and T. Chen, "Recent Advances in Convolutional Neural Networks," Pattern Recognition, vol. 77, no.
0031-3203, pp. 354-377, 2018.
[7] S. Hochreiter and J. Schmidhube, "LONG SHORT-TERM MEMORY," Neural Computation, vol. 9, no.
３８
8th ICAEIC-2022
8, pp. 1735-1780, 1997.
[8] I. N. Yulita, M. I. Fananya and A. M. Arymuthy, "Bi-directional long short-term memory using quantized
data of deep belief networks for sleep stage classification," Procedia computer science, vol. 116, pp. 530-
-538}, 2017.
３９
AEICP Vol. 5, No. 1
Salient Object Detection using Parallel Symmetric Networks with Attention
Mechanism
Kyeongseok Jang1, Donwoo Lee2, Soonchul Kwon3, seunghyun Lee4, and Kwang Chul Son*
1,2Department of Plasma Bio Display, Kwangwoon University, 20, Gwangun-ro, Nowon-gu, seoul, Korea
3Department of Smart Convergence, Kwangwoon University, 20, Gwangun-ro, Nowon-gu, Seoul, Korea
4Ingenium College of Liberal Arts, Kwangwoon University, 20, Gwangun-ro, Nowon-gu, Seoul, Korea
*Department of Information Contents, Kwangwoon University, 20, Gwangun-ro, Nowon-gu, seoul, Korea
ksjang1234@kw.ac.kr1, led0121@kw.ac.kr2, ksc0226@kw.ac.kr3, shlee@kw.ac.kr4, kcson@kw.ac.kr*
Abstract
In this paper, we propose a network that does not use pre-trained weights and backbone networks. Channel
attention and spatial attention were used to improve the feature loss between learning, and training was
performed by configuring the network in parallel. The proposed model showed improved results than the UCF
and RFCN models.
Keywords: Deep Learning, Salient Object Detection, Channel Attention, Spatial Attention.
1. Introduction
Humans use light reflected from objects through their eyes to identify objects and extract important
information to make visual decisions. It extracts characteristic information such as the color and size of objects
from all scenes that are input and uses them. In various fields of computer vision, various researches such as
object detection and recognition are being conducted using various characteristic information [1-2]. Salient
Object Detection (SOD) is a model that extracts objects from the background by utilizing various features
without using classes, unlike general object detection models [3]. In order to increase the detection rate of
objects such as Fully Convolutional Network (FCN) and U-Net, various methods such as modifying the model
structure and using Backbone Network were used [4-5]. In this paper, the asymmetric parallel structure SOD
model is proposed based on a basic U-net structure that does not use the backbone network and weights
obtained through pre-learning.
2. Attention Mechanism
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
４０
8th ICAEIC-2022
Attention Mechanism was proposed to perform context attention in the local area in a Recurrent Neural
Network (RNN)-based model [6-7]. In models such as Convolutional Neural Network (CNN)-based Image
Classification and Detection, image features are emphasized, and parts other than features are relatively
suppressed [8]. Attention Mechanism in Vision field is largely divided into Channel Attention and Spatial
Attention, and each Attention utilizes Average Pooling and Max Pooling. Figure 1 and Figure 2 show Channel
Attention and Spatial Attention of Attention Mechanism.
Figure 1.
Channel Attnetion Module.
Channel Attention performs Max Pooling and Average Pooling on input features, shares Multi-Layer
Perceptron, and performs Attention using the relationship between channels of features.
Figure 2.
Spatial Attention Module.
Spatial Attention performs Max Pooling and Average Pooling like Channel Attention but performs Attention
using spatial information by performing Concatenate and Convolution on the features created through Pooling.
3. Proposed Method
This paper proposes a backbone network and a network that does not use pre-trained weights. The SOD
model using the existing Backbone Network has the disadvantage that the structure of the Backbone Network
is included in the structure of the model, making structural modification difficult. In addition, if the weights
learned in advance are used, the size of the data input to the model training is fixed. Based on the most basic
U-net structure, the proposed model consists of each layer of the encoder and decoder stages in parallel. Figure
3 shows the structure of the proposed network.
４１
AEICP Vol. 5, No. 1
Figure 3.
Structure of Proposed Method.
Attention Mechanism was utilized to improve the loss of features occurring in the encoder process of
feature extraction and compression of input images, and Multi-Scale Convolution was used together to
construct an encoder block. In the decoder process of extending and restoring the extracted features to the size
of the input image, the lost features were improved through the feature map and concatenation of the encoder
stage, and an improved saliency map was generated.
4. Result
DUTS-Train Dataset (10,553 chapters) was used for training the proposed model, and DUTS-TEST (5,019
chapters) and DUTS-OMRON (5,168 chapters), ECSSD (1,000 chapters), HKU-IS (4,447 chapters) were used
for validation. was used. The optimizer used in training is the Adam Optimizer and Mean Squared Error (MSE)
is used as the Loss Function. Figure 3 shows the saliency map output after learning through the proposed
network and compared with the DUTS-OMRON Dataset.
Figure 3.
Comparison of Saliency Map of other models with the method proposed by DUTS-OMRON Dataset. (a)
Original Image (b) Groundtruth (c) Proposed Method (d) UCF (e) RFCN
４２
8th ICAEIC-2022
Figure 4.
Comparison of Saliency Map of other models with the method proposed by HKU-IS Dataset. (a) Original Image
(b) Groundtruth (c) Proposed Method (d) UCF (e) RFCN
Figures 3 and 4 compare the saliency map output through training of the proposed network with the results
of UCF and RFCN, which are other salient object detection models. The proposed model improves the
characteristics lost during learning by using the attention mechanism. The saliency map of the proposed model
showed improved feature loss than the saliency map of UCF and RFCN.
5. Conclusion
In this paper, we propose an asymmetrical parallel network without using pre-trained weights and backbone
networks, keeping the basic U-net structure. In the existing SOD model, the size of the input image was fixed
using weights learned in advance, and it was difficult to modify the structure of the network using the Backbone
Network. Since the proposed model does not use pre-trained weights, the size of the input image can be
adjusted, and since the backbone network is not used, the network structure is constructed in a parallel structure
while maintaining symmetry. In addition, by using the attention mechanism to improve the loss of features,
the results were improved compared to the UCF and RFCN models.
References
[1] Su, C., Xu, Z., Pathak, J., & Wang, F. (2020). Deep learning in mental health outcome research: a scoping
review. Translational Psychiatry, 10(1), 1-26.
[2] Ferrag, M. A., Maglaras, L., Moschoyiannis, S., & Janicke, H. (2020). Deep learning for cyber security
intrusion detection: Approaches, datasets, and comparative study. Journal of Information Security and
Applications, 50, 102419.
[3] Borji, A., Cheng, M. M., Jiang, H., & Li, J. (2015). Salient object detection: A benchmark. IEEE
transactions on image processing, 24(12), 5706-5722.
[4] LONG, Jonathan; SHELHAMER, Evan; DARRELL, Trevor. Fully convolutional networks for semantic
segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.
p. 3431-3440.
４３
AEICP Vol. 5, No. 1
[5] Ronneberger, O., Fischer, P., & Brox, T. (2015, October). U-net: Convolutional networks for biomedical
image segmentation. In International Conference on Medical image computing and computer-assisted
intervention (pp. 234-241). Springer, Cham.
[6] ZAREMBA, Wojciech; SUTSKEVER, Ilya; VINYALS, Oriol. Recurrent neural network regularization.
arXiv preprint arXiv:1409.2329, 2014.
[7] VASWANI, Ashish, et al. Attention is all you need. In: Advances in neural information processing
systems. 2017. p. 5998-6008.
[8] Park, J., Woo, S., Lee, J. Y., & Kweon, I. S. (2018). Bam: Bottleneck attention module. arXiv preprint
arXiv:1807.06514.
４４
8th ICAEIC-2022
Partial Face Recognition with Deep Learning
Jinho Han
Department of Liberal Studies (Computer), Korean Bible University, Seoul, South Korea
hjinob@bible.ac.kr
Abstract
Partial face recognition (PFR) technologies are very important in biometric authentication of the real world.
Before deep learning PFR methods had been developed based on feature vector representation. After emerging
deep learning, object detections and PFR have used the deep learning method. This paper explains how to
recognize the identities of partial faces with deep learning convolution neural networks (CNN). The training
dataset consists of holistic faces and three partial feces which are face patches of eyes, nose, and mouth. Four
groups of training datasets are trained respectively. When the test datasets are tested, the accuracy of the
partial face is the max value of four accuracy values of four groups. In the experiment, the proposed method
shows a valuable result.
Keywords: partial face recogniton; holistic face; deep learning; CNN; face recognition.
1. Introduction
Partial Face Recognition (PFR) is a solution to identify partial facial images. Wang & Deng [2] explained
increasing requirements of PFR because of the partial facial images acquired from modern digital devices like
CCTV cameras, mobile phones, and robots. Before deep learning PFR methods have been developed based on
feature vector representation [2-3]. Liao et al. [2] proposed a partial alignment-free face representation method
based on Multi-Keypoint Descriptors (MKD). Chen & Gao [3] introduced Stringface that was string-based
face recognition integrating the organization of intermediate-level features into a high-level structure. After
emerging deep learning, object detections and PFR have used the deep learning method [4-8]. Wang et al. [4]
suggested a face attention network (FAN) that could detect the partial faces occluded by mask, sunglasses. He
et al. [5] combined a fully convolutional network with sparse representation classification and proposed
dynamic feature matching (DFM) for PFR. Yang et al. [6] used CNN of two-channel architecture and the
shared layer for partial face verification. Elmahmudi & Ugail [7] proposed a CNN-based feature extraction
method for partial facial recognition. Hörmann et al. [8] proposed a novel approach with attentional pooling
of a ResNet's intermediate feature maps. But their methods extracted face features before adopting CNN and
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
４５
AEICP Vol. 5, No. 1
were not efficient. This paper explains how to verify partial facial images using CNN only and introduces a
novel efficient method. In the experiment, the proposed method showed a meaningful result.
2. Proposed method
The proposed PFR method uses four types of training datasets for CNN: holistic face, eyes, nose, and mouth.
Fig.1 shows the way of making training datasets from holistic faces. The PFR system has five types of CNNs.
CNNclassification is trained with all four types of datasets to know the test image holistic face, eyes, nose, or mouth.
When arbitrary images of partial faces are tested, they are tested what type they are among four types by
CNNcalssification. And then they are tested one of the other four CNNs. When it is the partial face with the eyes,
it is tested by CNNeyes and gets acceyes. If the test image cannot be verified its type of four, it is tested by all
four CNNs. Finally, the max accuracy is selected as the accuracy of the test image among four types of
accuracies. Fig. 2 shows the testing process of the method.
Figure 1. Making training datasets from a holistic face.
The PFR method always outputs four kinds of accuracies after its validation process: accholistic, acceyes, accnose,
and accmouth. Among four accuracies the biggest value becomes the final accuracy. The accuracy is calculated
by the equation as follows:
Figure 2. An example of a validation process of testing partial face of eyes.
４６
8th ICAEIC-2022
Accuracy = Max{acc,acc,acc,acc} (1)
If the tested partial face with eyes has 91% accuracy of its identity, it may be thought that accholistic= 0
acceyes= 0.91 accnose= 0, and accmouth= 0 after good working by CNNcalssification. And the max function outputs
0.91.
3. Conclusion
This paper proposed a simple but efficient PFR system. It used only CNN architecture for training partial
face images without a feature extraction process from input datasets before training process. In the experiments,
reasonable results were made. And the future work will be the study of training and testing the method with
more partial face data.
Acknowledgment
This work was supported by the National Research Foundation of Korea (NRF) grant in 2019 (NRF-
2019R1G1A1004773).
References
[1] Wang, M. & Deng, W. (2021). Deep Face Recognition: A Survey. Neurocomputing Vol. 429, 14 March
2021, (pp. 215-244)
Liao, S., Jain, A. K., & Li, S. Z. (2013). Partial face recognition: Alignment-free approach. IEEE Transactions
on Pattern Analysis and Machine Intelligence (TPAMI), 35(5), (pp.1193–1205).
Chen, W. & Gao, Y. (2010). Recognizing Partially Occluded Faces from a Single Sample Per Class Using
String-Based Matching. Computer Vision - ECCV, (pp. 496--509)
Wang, J., Yuan, Y., & Yu, G. (2017). Face Attention Network: An Effective Face Detector for the Occluded
Faces. arXiv preprint arXiv:1711.07246, 2017.
He, L., Li, H., Zhang, Q., & Sun, Z. (2018). Dynamic Feature Learning for Partial Face Recognition.
IEEE/CVF Conference on Computer Vision and Pattern Recognition 2018, (pp. 7054-7063), doi:
10.1109/CVPR.2018.00737.
Yang, L., Ma, J., Lian, J., Zhang, Y., & Liu, H. (2018). Deep representation for partially occluded face
verification. J Image Video Proc. 2018, 143 (2018). https://doi.org/10.1186/s13640-018-0379-2
[2] Elmahmudi, A., & Ugail, H. (2019). Deep face recognition using imperfect facial data. Future Gener.
Comput. Syst. 2019, 99, (pp. 213–225).
Hörmann, S., Zhang, Z., Knoche, M., Teepe, T. & Rigoll G. (2021). Attention-based Partial Face Recognition,
https://arxiv.org/abs/2106.06415v2.
４７
AEICP Vol. 5, No. 1
Energy Management System Model Applying a Nature-Inspired Optimization
Algorithm to a Virtual Power Plant
Yeonwoo Lee
1Department of Electronics & Information Communication Engineering, Mokpo National University,
Chonnam, Republic of Korea
ylee@mokpo.ac.kr
Abstract
In the context of the renewable energy resources (RES), an efficient energy management system (EMS) is
regarded as a key technology in the microgrid (MG) integrated virtual power plant (VPP). With an optimal
scheduling algorithm. The EMS is able to provide the optimal control in operation and management of the
usage of RES such as photovoltaic (PV) and small-size wind-turbine (WT) energies. In this paper, the efficient
EMS application model is considered with applying a metaheuristic nature-inspired algorithms such as and
particle swarm optimization (PSO), as to minimize the objective function in the context of power fluctuation
and the cost for power curtailment in a microgrid integrated VPP network. Thus, this paper presents the
simulation model on the effect of metaheuristic algorithms in the MG integrated VPP network, which shows
the better energy management scheme than that of the conventional management strategy.
Keywords: Energy Management System, Microgrid Network, Virtual Power Plant, Renewable Energy
Sources, Nature-Inspired Algorithm
1. Introduction
Recently, the concept of microgrid network (MG) and its applications have become important issues. MG
is well known as a promising way to guarantee a reliable energy supply and demand of users such as prosumers
and consumers. This can be realized by aggregating conventional generators, renewable energy systems
(RESs), and energy storage systems (ESS) along with different loads, to form a self-sufficient and flexible
system. One of most generally considered RESs is a photovoltaic (PV) prosumer who utilizes renewable
energy of solar power, and contributes on distribution of energy resources. A smart microgrid is defined as an
energy production sources such as distributed energy resources (DERs), energy storage (battery) facilities,
energy flow/distribution management such as microgrid operator (MGO). In general, the MGO is regarded as
a traditional energy management of the microgrid, aiming at minimizing the cost of the distributed renewable
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
４８
8th ICAEIC-2022
energy resources (DERs) as well as energy pricing cost. An energy management system (EMS) has an
important role in a MG in order to operate MG’s components and DERs efficiently. The major goals of
sustainability in the EMS are economic operation, reliability, and environmental impact, which is usually
economically related to minimizing the operating costs of the MG or maximizing its income. However, due to
new regulations of the network operators, a new objective related to the minimization of power peaks and
fluctuations in the power profile exchanged with the utility network has taken great interest in recent years.
Thus, an optimization problem to minimize the operating costs of the MG or to maximize its income is the
major open issue to be discussed. For this operating of optimization procedure, we should take into account
the adjustment parameters to control the performance of the EMS, as well as considering DERs, the overall
income. Of course, it is guaranteed only if the power peaks and fluctuation in the power profile is minimized
or stable with certain utility. Therefore, the parameter adjustment becomes an optimization problem, which
can be formulated in terms of a cost or utility function to be solved by the metaheuristic nature-inspired
algorithms.
Thus, an optimization problem of energy management of VPP to minimize the operating costs of the MG
or to maximize its profit revenue is the major open issue to be discussed. For this operating of optimization
procedure, we should take into accout the adjustment parameters to control the performance of the EMS, as
well as considering DERs, the overall income. Of course, it is guaranteed only if the power peaks and
fluctuation in the power profile is minimized or stable with certain utility. Therefore, the parameter adjustment
becomes an optimization problem, which can be formulated in terms of a cost or utility function to be solved
by the metaheuristic nature-inspired algorithms.
2. The Proposed Energy Management System
2.1 Virtual Power Plant
The proposed microgrid integrated VPP model with the EMS applying nature-inspired optimization
algorithms is illustrated in Fig. 1, where multiple microgrids include one or more prosumers (or DERs)
producing/consuming energy resources by PV, wind-turbine (WT) and battery storage (BS). As shown in Fig.
1, VPPs are expected to play a central role in future UESs by providing improved automation and control
capabilities in smart grids through the aggregation and virtualization of dispersed DERs, including power
generators, storage units, and active loads. The virtual power plant (VPP) is an innovative technology of the
power system, and it can effectively integrate, aggregate, and manage both conventional and renewable energy
power plants to achieve rational power allocation with limited and changeable resource availabilities [1]. VPPs
refers to heterogeneous power plants, which usually include distributed renewable energy power plants
(DERs), traditional fossil–fuel-fired power plants such as conventional heat-power plant (CHP), energy storage
facilities, and dispatchable loads as shown in Fig. 1. Through the coordination of the VPPs in the Aggregator,
the impact of fluctuations in renewable energy generation can be abated. Recently many research works focus
on key system parameters with uncertainity such as renewable energy availability, load demands, and energy
prices. These uncertainties might further intensify the complexity of the decision-making process. Therefore,
efficient mathematical programming techniques for planning electric power systems with consideration of
uncertainties and complexities are desired.
４９
AEICP Vol. 5, No. 1
Figure 1.
The EMS applying a nature-inspired optimization algorithm in a MG-itegrated VPP network with multiple
DERs.
2.2. Nature-Inspired Optimization Algorithm
Optimization techniques have played an important role in helping decision makers manage planning
problems in an effective and efficient way. Meta-heuristic optimization algorithms based on nature-inspired
concepts are dealing with selecting the best alternative solution given objective function. Some examples of
these nature-inspired algorithms that can be found in the literature are genetic algorithms (GAs), differential
evolution (DE), ant colony optimization, particle swarm optimization (PSO), firefly algorithm, Cuckoo Search
Algorithm (CSA), among others. In 1995, Particle swarm optimization (PSO) was developed by Kennedy and
Eberhart in 1995 based on swarm behavior in nature, such as fish and bird schooling [2]. The PSO is an
optimization algorithm inspired by swarm intelligence of fish and birds and even by human behavior. The
multiple agents, called particles, swarm around the search space, starting from some initial random guess. The
swarm communicates the current best and shares the global best so as to focus on the quality solutions. Since
then, PSO has generated much wider interests and forms an exciting, ever-expanding research subject, called
swarm intelligence. PSO has been applied to almost every area in optimization, computational intelligence,
and design applications. It is known that PSO is better than traditional search algorithms and even better than
genetic algorithms for many types of problems [2, 3].
2.3. Application of PSO to Optimize the Objective Function in VPP
Many research on trying to solve problems of energy resource optimization in VPPs to either minimize the
operating cost or maximize profit. The most common techniques that obtain the best solutions are as Particle
swarm optimization algorithm (PSO) [4, 5] and genetic algorithms (GAs).
The PSO algorithm is inspired by birds’ social behavior in flight. Each particle is characterized by a position
５０
8th ICAEIC-2022
and velocity vector that directs its movement in the search space. This movement is guided by the optimal
particles in the current moment. This algorithm is simple to implement, as there are few parameters that need
adjustment. In addition, it produces an efficient search for the optimal solution with a shorter calcu- lation time
and less memory usage. On the other hand, genetic algorithms simulate biological evolution and natural
selection based on learning, adaptation, and evolution. In each iteration, these algorithms consider a series of
starting solutions. The main advantages are their flexibility and mode of operation, as they simultaneously
determine several solutions instead of determining them sequentially, as done in common mathematical
techniques. In other words, they explore the solution space quickly and intelligently. In addition, they more
strongly avoid local optimal solu- tions even with highly complex problems. However, obtaining appro- priate
solutions demands that special attention be paid to the selection of the algorithm parameters, such as the
population size and mutation rate. For example, if the population size is very small, the algorithm does not
adequately explore the entire solution space, which may result in a local optimum. The basic steps of the
Particle Swarm Optimization (PSO) can be summarized as the pseudo code shown in Algorithm [3].
Algorithm 1. Particle Swarm Optimization
Objective function f(x), x = (x
1
,..., x
d
)T
Initialize locations xi and velocity vi of n particles.
Find g from min{f(x1),..., f(xn)} (at t = 0)
while (criterion)
for
loop over all n particles and all d dimensions
Generate new velocity vit+1 using following Eq. (1)
vit+1 = vit + αε1 [g − xit] + βε2 [ x1(t) − xit] (1)
where ε1 and ε1 are two random vectors, and each entry takes the
value between 0 and 1. The parameters α and β are the learning
parameters or acceleration constants, which can typically be taken
as, say, α ≈ β ≈ 2.
Calculate new locations xit+1 = xit + vit+1
Evaluate objective functions at new locations xit+1
F ind the current best for each particle xi
end for
Find the current global best g
Update t = t + 1 (pseudo time or iteration counter)
end while
Output the final results xi and g
2.4. Objective Function in VPP
Many research on trying to solve problems of energy resource optimization in VPPs to either minimize the
cost of VPP revenue and minimize the power fluctuation.
The objective function is re-written as:
５１
AEICP Vol. 5, No. 1
min
(,) min
[(,)+()]
. (1)
The total amount of surplus energy () can be defined as
()=(),, ()
() (2)
and the DER constraints and enegy cost are respectively,
,,
() ,, ()() ,,
() (3)
()= ,, ()
()+()()()()
(4)
3. Discussion and Further Study
Such approaches based on metaheuristic optimization algorithms are is applied to the EMS of a residential
grid-connected electro-thermal microgrid, improving its performance in terms of smoothing, i.e., minimizing
power peaks and fluctuations, the power profile exchanged with the utility network. The finding of
cost/objective function to be optimize in terms of power profile, economic cost, energy sharing, etc, is left to
be studied in further our research work.
References
[1] Hatziargyriou, N., Asano, H., Iravani, R., & Marnay, C. (2007). Microgrids. IEEE power and energy
magazine, 5(4), 78-94.
KennedyJ, Eberhart R. Particle swarm optimization. In Proceedings of the IEEE International Conference on
Neural Networks, Piscataway, NJ, USA; 1995. p. 1942–48.
Yang, Xin-She. Nature-inspired optimization algorithms. Academic Press, 2020
Qiu J, Meng K, Zheng Y, Dong ZY. Optimal scheduling of distributed energy resources as a virtual power
plant in a transactive energy framework. IET Gener, Transm Distrib 2017;11:3417–27.
adayeghparast S, SoltaniNejad Farsangi A, Shayanfar H. Day-ahead stochastic multi-objective
economic/emission operational scheduling of a large scale virtual power plant. Energy 2019; 172: 630–46.
Hropko D, Ivanecký J, Turˇcek J. Optimal dispatch of renewable energy sources included in virtual power
plant using accelerated particle swarm optimization. In: Proc 9th int conf ELEKTRO 2012. Rajeck teplice;
2012. p. 196–200.
５２
8th ICAEIC-2022
A Study on Mask Wearing Discrimination and Identity Identification
NaeJoung Kwak1, and DongJu Kim2
1Division of AI Software Engineering (Information Security), Pai Chai University, Daejeon, Korea
2Department of the research division, POSTECH Institute of Artificial Intelligence
knj0125@pcu.ac.kr1, kkb0320@postech.ac.kr2
Abstract
As the COVID-19 virus has become a daily routine, wearing a mask has become compulsory, so it is important
to wear a mask in real time. In addition, self-authentication using real-time camera input is widely used as an
important security method in the entering system. However, it is necessary to temporarily remove the mask
due to wearing a mask to self-authentication using camera, which increases the risk of virus spread. . In this
study, we performed to determine whether a mask is worn or not using Yolov5s and Facenet, and to identify
who is wearing a mask for security. It was confirmed that the proposed model discriminates whether or not to
wear a mask well and also performs identity identifictioon while wearing a mask.
Keywords: YOLO, object detection, real-time object, COVID-19, face detection, Identity Identification.
1. Introduction
Due to COVID-19 which is prevalent around the world, people have been avoiding face-to-face contact.
Due to high contagiousness of COVID-19 virus, wearing a mask is mandatory, and inspection of wearing a
mask is increasing. Under such circumstances, wearing masks is inspected in public places and indoors and
outdoors. However, since most people identify non-masked people with their own eyes, man-power is required,
and sometimes there may be parts that cannot be checked. Therefore, an automatic detection algorithm of
wearing mask is required.
In addition, a system that checks the identity of the visitors is required for security at the time of entry.
Most automated systems for identity identification use face recognition techniques, but with the existing face
recognition system, if you wear a mask, you may have to take off the mask when checking your identity. This
needs improvement because it not only makes the authentication process cumbersome for users, but also
increases the risk of virus spread.
In this paper, we implement a system that detects whether a mask is worn in real time using YOLOv5[1]
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
５３
AEICP Vol. 5, No. 1
among deep learning techniques for real-time object detection. In addition, we implement a system that can
confirm the identity of the person using facenet [2] among the face recognition models while wearing a mask.
2. The proposed method
In this study, we implement a system that determines whether a mask is worn or not and checks the identity
of the person who enters. In order to determine whether a mask is worn or not, the model was implemented by
transfer learning using the s-model, which is the fastest among YOLOv5. The identity identification model
was derived through transfer learning after learning the Facenet using Masked VGG2 among the datasets
provided in [3]. Figure 1 shows the system structure of the proposed method.
Figure 4.
System structure of the proposed method.
The experiment was conducted in the following environment on Ubuntu 16.04.7 LTS.
Table 1. Test environments.
OS
GPU
CPU
RAM
CUDA
cuDNN
software
Deep Learnig model
Ubuntu 18.04.5 LTS
Tesla V100-SXM2
16 core Intel(R) Xeon(R)
Gold 5120 CPU @ 2.20GHz
177GB
10.1
7.6.0
python/pytorch
YOLOv5 S model/Facenet
To determine wearing a mask, we labeled the Kaggle dataset[4] and data collected from the web
mask/no_mask using a labeling tool [5]. As the dataset for face recognition classification, Masked VGG2
５４
8th ICAEIC-2022
among the datasets provided in [3] was used as Facenet training data, and 9 classes of same persons taken with
a smartphone were used as data for transfer learning.
Figure 2 shows the mask detection result using the proposed model. Figure 3 shows the results of the wearing
or not wearing mask and the identy identification using the proposed model and test images. The results show
that the mask is worn or not and the identity identification is well done.
Figure 5.
The result of mask detection.
(a) mask detection (b) identity identification
Figure 3.
The mask wearing discrimination andeidentity identification using
test images.
3. Conclusion
In this paper, we implemented and verified a deep learning model that can determine whether mask is worn
or not using transfer learning of the Yolov5s model and automatically recognize the identity of a person using
Facenet transfer learning. For the data set, data collected from the web and the dataset provided by Kaggle
were used for mask detection. For identification, Masked VGG2 as face mask dataset were used as facenet
training data and data taken with a smart phone was used as data for transfer learning. The proposed model
determine whether to wear a mask and performed identity identification well.
Acknowledgement
This work was supported by Institute of Information & communications Technology Planning & Evaluation
(IITP) grant funded by the Korea government(MSIT) (No.2020-0-02029-002)
References
[1] ultralytics/yolov5. https://github.com/ultralytics/yolov5
５５
AEICP Vol. 5, No. 1
F. Schroff et al.(2015, June), FaceNet: A Unified Embedding for Face Recognition and Clustering.
Procedding of Conf. on Comput. Vision Pattern Recogn. (pp. 815-823). IEEE.
Masked dataset: https://github.com/SamYuen101234/Masked_Face_Recognition
Kaggle Dataset: https://www.kaggle.com/aditya276/face-mask-dataset-yolo-format
Labeling Tools (labelmg). https://github.com/tzutalin/labelImg.
５６
8th ICAEIC-2022
Case Study Hotel Booking Demand: A Predictive Analytics Model
Luh Made Yulyantari 1, and Ni Ketut Dewi Ari Jayanti 2
1,2Information System, Institute of Technology and Business STIKOM Bali
yulyantari@stikom-bali.ac.id1, daj@stikom-bali.ac.id2
Abstract
In the hospitality industry, harnessing the power of data helps decision makers solve specific challenging tasks
including improving occupancy forecasting, setting competitive room rates, selecting the most profitable
distribution channels, and identifying and targeting guests who are most profitable. This study uses a data set
about booking information for city hotels and resort hotels, and includes information such as time of booking,
hotel categories, and others. The data stretches from 2015 to 2017. All hotel data requires anonymity, so all
personally identifiable information has been removed from the data. These data sets are ideal for using
Exploratory Data Analysis (EDA) or prediction model development. This study focuses on exploratory data
analysis to support analyzing and discovering the properties of the data which can be useful in selecting the
appropriate statistical model in predictive analysis, namely the Naïve Bayes algorithm and the Generalized
Linear Model. The results obtained are hotel category forecasting data as well as the trend of changing interest
of potential visitors based on two hotel categories, namely resort hotels and city hotels. Based on the results
of the evaluation of the resulting prediction model, it can be seen from the two classification algorithms,
namely Naïve Bayes and Generalized Linear Model, that each algorithm has provided results with a high
degree of accuracy and a small difference in the values of the two algorithms for all types of evaluation. This
can be used as a basis for using the tested model on other data.
Keywords: hotel; predictive analysis; Naïve Bayes; Generalized Linear Model.
1. Introduction
The hotel business generates a lot of data literally all the time. New data occurs when tourists book
accommodation online, front office managers check on guests, or other events. In simple terms, it can be said
that new data occurs when an event occurs and is related to the hotel.
Valuable facts and figures seem endless, but the utilization of the data is not optimal. If not managed
properly, most of the information is lost or not used, resulting in no profit. This study will discuss the data
management approach used in the hospitality industry to increase revenue and improve customer experience.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
５７
AEICP Vol. 5, No. 1
Data management is the policy and practice of treating data as a valuable resource. Its goal is to transform
information into meaningful insights, enabling cost and operating optimization, as well as cutting costs while
increasing the resulting profit. In the hospitality industry, harnessing the power of data helps decision makers
to complete specific tasks including:
a) improve occupancy forecasting,
b) set competitive room rates,
c) choose the most profitable distribution channel,
d) optimize procurement operations,
e) increase guest loyalty, and
f) identify and target the most profitable guests.
The main challenge for hotel data management is the diversity of information availability. It can be extracted
from many websites, metasearch platforms, social media, internal documents, reports, and systems. There are
several pillar datasets to consider from the start. Therefore, the research data is taken from the Hotel Booking
Collection, Datasets, written by Nuno Antonio, Ana Almeida, and Luis Nunes for Short Data, Volume 22,
February 2019 [1]. This data set contains reservation information for city hotels and resort hotels, includes
information such as when the reservation was made, length of stay, number of adults, children and / or babies,
and the number of available parking spaces.
These data sets are ideal for conducting Exploratory Data Analysis (EDA) or prediction model development
[2]. This study will focus on exploratory data analysis to support analyzing and discovering the properties of
the data which can be useful in selecting the appropriate statistical model in predictive analysis for the hotel
industry, city hotels and resort hotels. Results enable hoteliers to accurately predict net demand and make
better forecasts, improve cancellation policies, establish better overbooking strategies, and use firmer pricing.
2. Literature Review
Many other researchers have conducted research in the field of predictive analytics. The purpose of research
in this field is primarily to improve organizational performance through forecasting [3]. Data sets contained in
an organization may also only need an illustration of how existing data can be easily used for predictive
analytics, it is possible that the data is already there, but data users are not aware of its suitability [4]. Many
business organizations require decision making based on data sets using predictive analytics [5]. One of them
is research on forecasting the number of canceled hotel bookings to reduce uncertainty and increase revenue
[6]. This research continues to be a model, then the final result is the creation of automatic machine learning
for a decision support system for forecasting hotel booking cancellation [7]. Several references related to the
creation of predictive analytics in the hotel sector, none have focused on distinguishing the types of bookings
based on differences in hotel types, especially in the form of city hotels and resort hotels.
3. Methodology
CRoss Industrial Standard Process for Data Mining (CRISP-DM) is a methodology applied in the
implementation of this research. CRISP-DM is one of the most widely used process models in predictive
５８
8th ICAEIC-2022
analytics projects [8]. CRISP-DM provides a six-step process: business understanding, data comprehension,
data preparation, modeling, evaluation, and deployment. The following sections offer a brief description of
each of these steps. CRISP-DM breaks down the data mining process into six main phases:
3.1. Business Understanding
In the first stage of the CRISP-DM process is to understand what to achieve from a business perspective,
one of which is the main purpose of the hospitality business, namely:
a) Improving occupancy forecasting,
b) set competitive room rates,
c) select the most profitable distribution channels,
d) optimizing procurement operations,
e) increase guest loyalty, and
f) identify and target the most profitable guests.
3.2. Data understanding
The second phase of the CRISP-DM process requires obtaining the data listed in the project data source.
This initial collection includes loading of data if this is necessary for data understanding. In this study, the data
source used is about hotel bookings for two types of hotels, namely city hotels and resort hotels. An explanation
of the details of the contents of the data has been explained in the previous section. At this stage, the data
quality verification process will be used in the research. The process of cleaning the data is carried out at the
next stage.
3.3. Data preparation
The data preparation stage is the research stage to decide which data to use for analysis. Criteria that may
be used to make these decisions include data relevance to data mining purposes, data quality, as well as
technical constraints such as data volume limits or data types. Data selection includes the selection of attributes
(columns) as well as the selection of records (rows) in tables. There are several main stages, namely:
a) Data cleaning
b) Build the necessary data
c) Integration of hospitality data for all columns
After the three stages are carried out, it is continued with the creation of Data vizualitation, which is a
representation of data in the form of graphs to facilitate the process of reading the movement of data.
3.4. Modeling
Modelling will be made in the form of predictive analysis results, namely using time series forecasting
methods. One method that can be used is exponential smoothing method. The selection of this method with
the reason, that in the method is a procedure that continuously improves forecasting with an average
(smoothing) the past value of a time-runtut data by decreasing (exponential). This is thought to reduce the rate
of forecasting errors.
５９
AEICP Vol. 5, No. 1
3.5. Evaluation
In the evaluation stage will be evaluated the results in the previous stage. The method that will be used in
this stage is the process of measuring the accuracy of the results of predictive analysis at the modeling stage.
3.6. Deployment
At this stage it will depend on the results of the evaluation stage to be able to define the strategy in the next
development. If the evaluation results show high data accuracy, then the modeling data can be used
immediately, but if the evaluation results show a low level of accuracy, then it is necessary to define the next
development strategy.
The sequence of phases is not strict and moves back and forth between different phases as is always required.
The arrows in the process diagram show the most important and frequent dependencies between phases. The
outer circle in the diagram symbolizes the cyclic nature of data mining itself. The data mining process continues
after the solution has been used. Lessons learned during the process can trigger new business questions that
are often more focused, and subsequent data mining processes will benefit from previous experience. Depiction
of CRISP-DM model can be seen in Figure 1.
Figure 1. CRISP-DM Model.
4. Result and Discussion
4.1. Business Understanding
Hotels can be categorized based on their location. In this study, the hotel category will be seen from two
different types of locations. City hotels, namely hotels that are in the middle of an urban area, are usually also
intended for people who want to stay temporarily or stay for a relatively short period of time, city hotels are
６０
8th ICAEIC-2022
also often referred to as transit hotels because they are often visited by businesspeople. In addition, there are
Resort Hotels, namely hotels that are in mountainous areas as well as on the beaches and so on. This resort
hotel will be aimed at people who want to stay or rest on holidays or for those who want to take a vacation [9].
The predictive analysis that will be carried out aims to provide an overview to hotel managers regarding
guest visit information, so that later it can be used as a basis for making decisions related to increasing company
income. There are several data information that will be generated, namely:
a. Booking times are seen from the week category (weekdays or weekends) in two hotel categories,
namely resort hotels and city hotels.
b. The number of visits seen from the week category (weekdays or weekends) at hotels in each country.
c. The month that has the highest visits to resort hotels and city hotels.
4.2. Data Understanding
As previously mentioned, this study uses a data set on booking information for city hotels and resort hotels,
and includes information such as time of booking, length of stay, number of adults, children and / or infants,
and the number of available parking spaces. and others. The data stretches from 2015 to 2017. All hotel data
requires anonymity, so all personally identifiable information has been removed from the data. This data comes
from the Hotel Booking Collection article, Datasets, and is downloaded via Kaggle.com [1]. Details of the
types of data contained in it are: hotel, is_canceled, lead_time, arrival_date_year, arrival_date_month,
stay_in_weekend, stay_in_week, adults, children, babbies, meal, country, market_segment, is_repeated_guest,
distribution_channel, previous_room_cancelation, reserved_room_type, booking_type, agent_changes_type,
booking_type, agent_changes_type, assigned_room, booking , company, days_in_waiting_list, customer_type,
adr, requires_car_parking_space, total_of_special_request, reservation_status, reservation_status_date. The
total data contained in this hotel booking data set is 119,391 data.
The phase of data understanding, or data preprocessing is the stage for checking the data to be used. Based
on the analysis objectives that have been set in the previous stage, Table 1 is a description of the attributes
used. Figure 2, Figure 3, Figure 4, Figure 5, Figure 6, Figure 7 is a data description for each attribute for valid
data, mismatched data, missing data, unique data, and the most common data.
Table 1. Atribute hotel
Name
Type
Description
hotel
String
Hotel (H1 = Resort Hotel or H2 = City Hotel)
arrival_date_year
Numeric
Year of arrival date
arrival_date_month
Numeric
Month date of arrival
stay_in_weekend
Numeric
Number of weekend nights (Saturday or Sunday) occupied or booked by guests to
stay at the hotel
stay_in_week
Numeric
Number of working nights (Monday to Friday) the guest stayed or booked to stay
at the hotel
country
String
Country of origin. Categories are represented in ISO 3155–3: 2013 format
６１
AEICP Vol. 5, No. 1
Figure 2. Atribute Hotel.
Figure 3. Atribute Arrival_date_year.
Figure 4. Atribute Arrival_date_month.
Figure 5. Atribute Stays_in_weekend_nights.
６２
8th ICAEIC-2022
Figure 6. Atribute Stays_in_week_nights.
Figure 7. Atribute Country.
4.3. Data Preparation
Data validation was carried out on 119,391 data in the data source. Validation is carried out based on the
completeness and validity of the contents of each record. The data validation stage did not find valid data, so
all preliminary data were used in this study, with data spanning from 2015 to 2017, while the attributes involved
in modeling using predictive analysis were six attributes. Figure 8 shows a summary of the data from all the
attributes that will be used in the modeling stage.
Figure 8. Data Atribute.
６３
AEICP Vol. 5, No. 1
This step takes advantage of previously undertaken data exploration and quality verification to create a final
data set for use in predictive model development. It begins with deletion from the original observation data set
(rows) based on prior considerations. Figure 9, Figure 10, and Figure 11 illustrate a set of records based on the
hotel, stay_in_weekend, and stay_in_week attributes.
Figure 9. Hotel Attribute Monthly Data Overview.
Figure10. Monthly Data Overview The stays_in_week_nights attribute.
Figure 11. Monthly Data Overview The stays_in_weekend_nights attribute.
６４
8th ICAEIC-2022
4.4. Modeling and Evaluation
The modeling used is a predictive analysis. Different algorithms give different results, new models are
developed using different classification algorithms and then choose the one that presents a better performance
indicator. The classification algorithm used is Naïve Bayes and Generalized Linear Model. In the two
forecasting algorithms, the total data processed was 119,391, while the processed testing data was 47756.
Figure 12, there are some initial data changes in the forecasting results of the Naïve Bayes algorithm for
hotel type attributes.
Figure 12. Forecasting Results of the Naïve Bayes Algorithm.
The data from the forecasting results are then processed to see the trend of the data generated from the Naïve
Bayes forecasting algorithm. Table 2 and Figure 13 show the types of data that do not change and the types of
data that change between real data and forecast data.
Table 2. Trend of naïve bayes forecasting data
Type of Data
Total
Percentage
Permanent
17132
36%
Change
30624
64%
Total
47756
100%
６５
AEICP Vol. 5, No. 1
Figure 13. Graph of Naïve Bayes Forecasting Data.
The trend of forecasting is greater for changes in hotel types, either from the choice of resort hotel to city
hotel, or from city hotel to resort hotel. The trend of data change is shown in Table 3 and Figure 14.
Table 3. The tendency of data to change in naïve bayes forecasting results
Type of Change
Total
Percentage
Resort Hotel to City Hotel
819
3%
City Hotel to resort hotel
29805
97%
Total
30624
100%
Figure 14. Graph of Changing Data Types in Naïve Bayes Forecasting Results.
4.4.1. Generalized Linier Model
The results of forecasting the Generalized Linear Model algorithm are shown in Figure 15, there are some
initial data changes in the attributes of the hotel type.
６６
8th ICAEIC-2022
Figure 15. Forecasting Results of Generalized Linear Model Algorithms.
The data from the forecasting results are then processed to see the trend of the data generated from the
Generalized Linear Model forecasting algorithm. Table 4 and Figure 16 show the types of data that do not
change and the types of data that change between real data and forecast data.
Table 4. The trend of generalized linear model forecasting results.
Type of Data
Total
Percentage
Permanent
20824
44%
Change
26932
56%
Total
47756
100%
Figure 16. Graph of Generalized Linear Model Forecasting Data.
６７
AEICP Vol. 5, No. 1
The trend of forecasting is greater for changes in hotel types, either from the choice of resort hotel to city
hotel, or from city hotel to resort hotel. The trend of data change is shown in Table 5 and Figure 17.
Table 5. The trend of changing data in generalized linear model forecasting.
Type of change
Total
Percentage
Resort Hotel to
City Hotel
2636
10%
City Hotel to
resort hotel
24296
90%
Total
26932
100%
Figure 17. Graph of Changed Data Types of Generalized Linear Model Forecasting Results.
Evaluation is conducted for both classification algorithms for several components, namely accuracy (Figure
18), classification error (Figure 19), Area Under Curve (Figure 20), precision (Figure 21), and Receiver
Operating Curve (Figure 22).
a) Accuracy
Figure 18. Accuracy.
６８
8th ICAEIC-2022
b) Classification Error
Figure 19. Classification Error.
c) AUC (Area Under Curve)
Figure 20. Area Under Curve (AUC).
d) Precision
Figure 21. Precision.
６９
AEICP Vol. 5, No. 1
e) Ratio of Reciever Operating Curve (ROC)
Figure 22. Ratio of Reciever Operating Curve (ROC).
Based on the results of the evaluation of the resulting prediction model, it can be seen from the two
classification algorithms, namely Naïve Bayes and Generalized Linear Model, that each algorithm has
provided results with a high degree of accuracy and a small difference in the values of the two algorithms for
all types of evaluation. This can be used as a basis for using the tested model on other data.
Analyze hotel data from booking data seen from check-in and cancellation data, to see forecast data for
cancellations that may occur. In addition, testing other classification algorithm models such as ARIMA.
5. Conclusion
There are five stages carried out in research according to the CRISP-DM method providing a six-step
process: business understanding, data understanding, data preparation, modeling, evaluation, and deployment.
There are six data attributes used in the study, namely hotel, arrival_date_year, arrival_date_month,
stay_in_weekend, and stay_in_week, and country. The results obtained are hotel category forecasting data and
the trend of changing interest of potential visitors based on two hotel categories, namely resort hotels and city
hotels. The evaluation of the prediction model generated from the two classification algorithms, namely Naïve
Bayes and the Generalized Linear Model, has given results with a high degree of accuracy and a small
difference in the values of the two algorithms for all types of evaluation. This can be used as a basis for using
the tested model on other data.
Acknowledgement
We would like to thank Institute of Technology and Business STIKOM Bali (https://stikom-bali.ac.id/) for
their funding in this research.
７０
8th ICAEIC-2022
References
[1] J. Mostipak, "Dataset: Hotel Booking Demand," 25 Maret 2020. [Online]. Available:
https://www.kaggle.com/jessemostipak/hotel-booking-demand.
[2] R. Kurniawan, Cara Mudah Belajar Statistik: Analisis Data & Eksplorasi, Jakarta: Prenadamedia, 2019.
[3] S. Akter, S. F. Wamba, A. Gunasekaran, R. Dubey and J. S. Childe, "How to Improve Firm Performance
Using Big Data Analytics Capability and Business Strategy Alignment," International Journal Production
Economics, vol. 182, pp. 113-131, 2016.
[4] T. Schoenherr and C. S. Pero, "Data Science, Predictive Analytics, and Big Data in Supply Chain
Management: Current State and Future Potential," Journal of Business Logistics, pp. 1-13, 2015.
[5] C. Carlberg, Predictive Analytics: Microsoft Excel, Indianapolis, Indiana, USA: Wiley, 2012.
[6] N. Antonio, A. d. Almeida and L. Nunes, "Predicting hotel booking cancellations to decrease uncertainty
and increase revenue," Tourism & Management Studies, vol. 2, no. 13, pp. 25-39, 2017.
[7] N. Antonio, A. d. Almeida and L. Nunes, "An Automated Machine Learning Based Decision Support
System to Predict Hotel Booking Cancellations," Data Science Journal, vol. 18, no. 32, pp. 1-20, 2019.
[8] D. Abbott, Applied predictive analytics: Principles and techniques for the professional data analyst,
Indianapolis, IN, USA: Wiley, 2014.
[9] M. P. d. E. K. R. Indonesia, Standar Usaha Hotel, Jakarta: Peraturan Menteri Pariwisata dan Ekonomi
Kreatif Republik Indonesia, 2013.
７１
AEICP Vol. 5, No. 1
A study on the Method of Securing Documentary and Evidence Capability of the
Suspect's Record
Yoon Cheolhee1, Hwang JeaHun2, Cho MinJe3, and Lee Bong Gyou4
1, 2, 3 Laboratory of Autonomous Vehicle and Block-chain, Korean national police University, Korea
4 Laboratory of Block-chain in Graduate School of Information, Yonsei University, Korea
bertter@police.ac.kr 1, sharproo@police.go.kr 2 epiworld@police.go.kr 3 bglee@yonsei.ac.kr 4
Abstract
Under the current law, it is time to change the application of biometric authentication, which cannot be stolen
or transferred from the existing accredited certificate method, as a way to replace the suspect's newspaper
report from paper documents to electronic documents. Signs and seals that are important in court can be
improved with biometric authentication (eg, fingerprints, faces) that cannot be stolen or transferred. In
addition, leakage of personal information and infringement of information self-determination can be prevented
by applying a blockchain-based DID for integrity verification. This thesis is concerned with the infringement
of personal information and the right to self-determination of personal information due to fake fingerprints of
the suspect's newspaper record, and the concerns and problems that a third party can use it for authentication
by viewing and copying the fingerprints imprinted on the suspect's newspaper record during the trial process.
A method that can be improved through recognition and application of blockchain-based DID was proposed.
Keywords: Electronic documents, blockchain, DID, digital forensics, biometrics, non-repudiation.
1. Introduction
The Korea Information System of Criminal Justice (KICS) digitalizes various criminal procedure documents
between investigation, prosecution, trial and enforcement agencies based on the 'Act on the Promotion of
Electronic Criminal Justice Procedures' to provide “paperless investigations” and “paperless investigations”.
It is a national-based computer network developed for the purpose. The suspect's interrogation report
(hereinafter referred to as "interrogate report") prepared in criminal justice procedures is a representative
electronic document distributed in the criminal justice information system. The part where the interrogate
report is currently being used as an electronic document is being used in the criminal justice process for simple
traffic cases. In the case of drunk driving and unlicensed driving, which are simple traffic cases, the interrogate
report agreed to by the suspect is written as an electronic document is finished [1]. In addition, in the general
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
７２
8th ICAEIC-2022
case interrogate report, the police officer writes the suspect's statement as it is, and if the statement is the same,
it is printed out as a paper document and signed at each store with a fingerprint. This is a procedure for securing
evidence capacity by maintaining continuity and integrity of the document of evasion, which is a paper
document.
The Criminal Justice Information System was developed to implement the ‘paperless e-criminal procedure’
and to provide advanced criminal justice information services in line with the information age of the 21st
century [1]. However, the interrogate report is still printed as a paper document, and it is signed with the
suspect's fingerprint. In this way, the prepared interrogate report contains various problems such as the
procedural burden of being served directly from the police station to the prosecution and the leakage, loss, and
damage of personal information stored in the documents.
Although security tokens are used to solve the problem of leaking public certificates, there is a problem that
these security tokens can also be stolen or transferred. Therefore, accredited certificates can be stolen and
transferred, are weak in security, and do not fit the global authentication method, so it needs to be replaced
with another authentication method. Therefore, biometric information is not required to be stored and
memorized, there is no risk of loss, and cannot be stolen or transferred. Recently, financial institutions and
public institutions are replacing digital document signature methods from public certificates to biometric
authentication methods.
2. Proposed System
Paragraph 3 of Article 23「Name, Seal, etc.」of Police Agency Ordinance No. 858 of the Criminal
Investigation Regulations stipulates that “Investigation documents shall be signed by each store”. The purpose
of this regulation is to secure the continuity and integrity of all investigation documents used in criminal justice
procedures. The purpose of securing the continuity of investigation documents is to prevent forgery or
falsification of documents, such as inserting or deleting documents in prepared documents.
Most of the investigation documents prepared by the criminal justice system are fingerprints. However, there
is a problem in that the fingerprints used for human beings can be used to unlock crime-related smartphones
by taking a picture and creating a fake fingerprint with silicon.
Biometric authentication is attracting attention as an additional or alternative authentication method to
existing authentication methods such as passwords and public certificates in that storage and memorization are
unnecessary, there is no risk of loss, and it is impossible to steal or transfer. Therefore, the problem of
fingerprints used for accredited certificates and seals can be solved by using an accredited digital signature
with a biometric authentication method that cannot be stolen or transferred. After filling out the interrogate
report, which is a representative electronic document, and checking the contents, it is possible to secure the
documentary by using the fingerprints of the suspect, investigator, and participating police officer to
authenticate electronically. And because the authentication method of FIDO, a biometric authentication
standard, uses the same PKI method as the public certificate, identity verification, integrity, confidentiality,
and non-repudiation are possible.
FIDO [Figure 1] U2F supports two-step online authentication method. The FIDO U2F authentication
method can enhance security by adding a separate authentication method to the existing encryption
７３
AEICP Vol. 5, No. 1
infrastructure. As before, the user logs in using a user name and password, and authenticates using a second
factor device such as a FIDO security key (eg USB dongle key, smart card, etc.). Through a strong second
authentication factor, it is possible to simplify complex passwords that were difficult to use and remember to
otp 4-digit PIN numbers without affecting security [12].
Figure 6. FIDO methods of authentication.
The authentication data generated in this way can be used on a blockchain-based DID platform. The
components for self-sovereign identity management consist of DID and DID documents used as identifiers
and authentication means, VC (verifiable credentials) used as storage IDs, and VP (verifiable presentations)
used as IDs for submission.
Figure 7. Document on DID distribution.
For the distribution of the suspect's evidence, the main participants are the issuer who issues the VC, the
holder who processes the VC after it is issued and submits it to the verification agency, the verifier that receives
the VP from the user and verifies the authenticity of the VP, and the DID and There is Identifier Registry, a
distributed repository that stores ID-related information. The DID (Decentralized Identifier) infrastructure for
self-sovereign identity management is implemented on a DID-compatible blockchain, distributed ledger, or
７４
8th ICAEIC-2022
distributed network data platform. Investigated documents are stored in a database on a blockchain platform
for the purpose of trial, and the collected data transforms evidence information into DID using public keys,
authentication protocols, and service endpoint datasets required for encrypted interactions between entities. is
managed by
After that, distribution for analysis and investigation, digital forensic activity applied to investigation, and
data transmission for trial are possible. As shown in [Figure 9], in the case of DID, data integrity and
authenticity determination can be handled by the institution itself through the public key shared by institutions
connected to the distributed network. In this process, the integrity of the authentication process is guaranteed
through the consensus algorithm it can be done.
3. Conclusion
In order to replace paper documents used in criminal justice procedures with electronic documents,
equivalence, integrity, security, and confidentiality conditions are required. First, equivalence means that paper
documents and electronic documents must be the same. And integrity is maintained when digital signatures
are certified digital signatures. And the accredited digital signature does not specify only the accredited
certificate. Therefore, since the biometric authentication method is the same as the public authentication
method, it can be introduced as much as possible. In addition, if electronic documents are digitally signed with
biometric authentication, they are encrypted, and security and confidentiality can be maintained by using
blockchain-based DID.
As in the “Nth Room” incident, the identity of public officials was stolen and transferred to social service
personnel, resulting in the problem of illegal personal information search and exposure that you have to. In
particular, biometric authentication is an authentication method that cannot be stolen or transferred. In
particular, it is necessary to actively introduce a sensitive system such as personal information, and at the same
time, it is time to apply DID, a self-sovereign identity based on blockchain.
Acknowledgement
This work was supported by Institute of Information & communications Technology Planning &
Evaluation(IITP) grant funded by the Korea government(MSIT) (No.2020-0-00901, Information tracking
technology related with cyber crime activity including illegal virtual asset transactions).
References
[1] S. H.Park, S.J.Lee, “A Study on the Documentability of Electronic Documents in Simple Traffic Case
Criminal Justice Procedures”, Police Science Research, vol 19, 1, pp. 34-56, Mar 2019.
J.D. Kim, B,S. Moon “Biometric authentication emerges as the wearable market grows”, LG Business Insight,
2015
K.Y. Moon, “iometrics Technology Status and Prospect”, TTA Jounal No 98, 38-47,
７５
AEICP Vol. 5, No. 1
S&T Market Report, “Biometric technology and market trends”, Commerciatizations Prornotion Agency for
R&D Coutcomes, 2016, 2
Sang-su Jang, “A Study on the Effect Fintech on the Information Scurity Industry”, Internet&Security Focus,
4-32, 2015.
７６
8th ICAEIC-2022
Multiple Imputation Study of Missing Marine Environmental Data
Hyi-Thaek Ceong1, Kyeong-Rak Lee2, and Sang-Joon Lee3
1School of Culture Contents, Chonnam National University, Korea
2BK4 Education & Research Group of Subtraction Platform, Chonnam National University, Korea
3Interdisciplinary Program of Digital Future Convergence Service, Chonnam National University, Korea
htceong@jnu.ac.kr1, kryi@nate.com2, s-lee@jnu.ac.kr3
Abstract
Using real marine environmental data, step-by-step work from data pre-processing to multiple imputation (MI)
is described, and further, we iteratively extracted the pooled estimate for each missing imputation method and
compared it with the estimate of the regression model in the original data set through dot plot and multiple
imputation evaluation criteria.
Keywords: missing values, multiple imputation, multivariate imputation by chained equation (MICE),
preprocessing, imputed complete dataset, marine environmental data.
1. Introduction
In most statistical packages, the default setting for missing values is listwise deletion. Even in cases such as
boxplot, plot, and lm (regression analysis function) in R, if there are missing values, the case containing the
missing values is completely removed and the analysis is performed. In this case, it is known that it not only
causes problems with the mean and descriptive statistics, but also affects the results of regression analysis that
analyzes the relationship between variables. Therefore, data analysis must be performed after sufficient
understanding of missing values and missing imputation methods to reduce problems with bias or statistical
power and increase the validity and reliability of research results [1-4].
This paper uses real field data to identify missing value patterns from the pre-processing stage, and perform
regression analysis on the original data set. Next, after generating m = 20 imputed complete datasets by
multiple imputation, regression analysis is performed on each data sets. By combining each result, a pooled
estimate is derived. This process was repeated 100 times. Here, six missing imputation methods (pmm,
norm.predict, norm.nob, norm.boot, cart, rf) were tested. 100 pooled estimates and 95% confidence intervals
were derived for each missing imputation method and we compared them with the regression coefficients of
the original data set with listwise deletion. A dot plot shows the expected value of the pooled estimate for each
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
７７
AEICP Vol. 5, No. 1
missing imputation method. Multiple imputation evaluation criteria were calculated and their meaning was
examined in relation to the dot plot.
2. Materials and methods
2.1. Data collection
The data of this study were the physical environment and water quality environment data from 2015 to 2019
of the Korea Marine Environment Corporation. This data was investigated in spring and summer according to
the Marine Ecosystem Basic Survey Standard Manual [5,6] of the Korea Marine Environment Corporation
(https://www.koem.or.kr). This data consists of water temperature, salinity, depth, chlorophyll-a concentration,
DO (Dissolved Oxygen), SPM (Suspended Particle Materials), Transparency, pH, PON, POC, DSi, DIP, DIN,
NO2, NO3, NH4. Observation data in the surface and bottom layers are mixed.
2.2. multiple imputation(MI) using MICE package in R
For comparison of imputation methods, four indices (RB, PB, CR, AW) of the multiple imputation
evaluation criteria presented in Flexible imputation of missing data (second edition) [9] were used. Q means
the estimate (regression coefficient) of the original dataset regression model, and means the pooled estimate
in Table 1.
Table 1. Multiple imputation evaluation criteria.
Criteria
Definition
Raw bias(RB)
RB = E()
Percent bias (PB)
PB =100 × |(())
|
Coverage rate (CR)
the proportion of confidence intervals that contain the true value
Average width (AW)
the average width of the confidence interval
3. Result & Discussion
3.1. Data preprocessing & missing data pattern
We merged the data with a full (outer) join that extracts all the data that exists in both tables using the
physical and water quality environment data of the five years (2015-2019) data as the key year, season, station,
and depth. A total of 1,583 data were obtained. As a result of checking the missing pattern, it was found that
the Chla missing rate was more than 30%. This causes a problem of lowering the replacement accuracy due to
a high missing rate, so when the depth value was filtered to a depth of 1 m, a total of 729 data was obtained.
７８
8th ICAEIC-2022
3.2. Comparison of 6 imputation methods
By performing the experiment, 100 pooled estimates and 95% confidence intervals were derived for each
missing imputation method. And by considering the regression coefficient of the original data set with listwise
deletion as the true value estimate, Multiple imputation evaluation criteria were calculated (Table 2).
Table 2. Multiple imputation evaluation criteria by 6 imputation methods.
(a) Transparency
(b) Salinity
RB
PB
CR
AW
RB
PB
CR
AW
pmm
-0.043
53.265
1
0.154
pmm
0.138
45.168
0
0.162
norm.predict
-0.039
49.019
1
0.149
norm.predict
0.182
59.473
0
0.129
norm.nob
-0.020
25.239
1
0.149
norm.nob
0.178
58.331
0
0.134
norm.boot
-0.017
21.329
1
0.164
norm.boot
0.176
57.5
0
0.151
cart
-0.088
110.13
0
0.135
cart
0.161
52.63
0
0.152
rf
-0.052
65.325
1
0.139
rf
0.151
49.532
0
0.160
(c) DO
(d) SPM
RB
PB
CR
AW
RB
PB
CR
AW
pmm
-0.014
7.299
1
0.132
pmm
0.075
39.644
0.01
0.135
norm.predict
-0.026
13.846
1
0.129
norm.predict
0.077
40.94
0
0.133
norm.nob
-0.022
12.08
1
0.13
norm.nob
0.085
44.804
0
0.134
norm.boot
-0.021
11.506
1
0.131
norm.boot
0.085
44.865
0
0.136
cart
-0.016
8.682
1
0.13
cart
0.084
44.641
0
0.134
rf
-0.010
5.43
1
0.131
rf
0.090
47.386
0
0.134
4. Conclusion
A series of processes were performed starting from the pre-processing stage of the real marine environmental
data, confirming the missing data patterns, and performing correlation and regression analysis on the original
dataset, and followed by multiple imputation to extract pooled estimates. Of course, there is a limitation in that
the estimate of the regression model of the original dataset with listwise deletion was regarded as the estimate
of the true value and compared with the estimate of the multiple imputation model. However, it is meaningful
that we explored a way to do MI(multiple imputation) in the observation data using real field data.
Acknowledgement
This research was a part of the project titled "Research center for fishery resource management based on the
information and communication technology" (2021, grant number 20180384), funded by the Ministry of
Oceans and Fisheries, Korea.
７９
AEICP Vol. 5, No. 1
References
[1] Alruhaymi, A. and Kim, C. (2021) Why Can Multiple Imputations and How (MICE) Algorithm Work ?.
Open Journal of Statistics, 11, 759-777. doi: 10.4236/ojs.2021.115045.
Kilkon Ko, Hyunwoo Tak and Bora Lee. (2014). Impact of Missing Values on Survey Research and Relevancy
of Multiple Imputation Techniques. Korean Journal of Policy Analysis and Evaluation, 24(3), 49-75. doi :
10.23036/kapae.2014.24.3.003
Kilkon Ko and Hyunwoo Tak. (2016). The Treatment of Missing Values using the Integrated Multiple
Imputation and Callback Method. Korean Journal of Public Administration, 54, 1-4.
Jaehyun Kim. (2020). A Study on the Multiple Imputation of Missing Values: Focus on Fine Dust Data. The
Society of Convergence Knowledge Transactions, 8(4), 149-156. doi:10.22716/sckt.2020.8.4.044
Korea Marine Environment Management Corporation (KOEM). 2021. National Marine Ecosystem Monitoring
Guide in Korea. http://koem.or.kr/
Marine Environment Information Portal (MEIS). 2021. National Marine Ecosystem Comprehensive Survey
Documents List. http://meis.go.kr/
８０
8th ICAEIC-2022
Verification of Big data-based Correlation Analysis for Autonomous Driving
Dong-Jin Shin1, Seung-Yeon Hwang2, Mun-Yong Park3, Dae-Sun Yum4, Sung-Youn Cho*
1,2,3,4,*Novacos Co., Ltd., Anyang, South Korea
{djshin1, syhwang2, pmy3, daesun12144, scho*}@novacos.co.kr
Abstract
Over the past decade, Korea has seen an average of 44,000 fires per year, about 3,600 fires per month and
122 fires per day, injuring about 1,856 people every year and killing about 325 people. Fire safety education
is being conducted to prevent fire damage, but opinions are being raised that fire safety education is not going
well. Therefore, this paper intends to suggest improvement directions by identifying which factors cause fires
through multiple regression and correlation analysis in the Seoul area using R, a big data analysis tool.
Keywords: Fire Safety Education, BigData, R Prmgramming, Multiple Linear Regression, Correlation
Analysis.
1. Introduction
Recently, about 44,000 fires have occurred annually in Korea over the past 10 years, with about 3,600 fires
a month and 122 fires a day, injuring about 1,856 people every year and killing about 325 people. Also,
property damage is increasing due to the fire outbreak, and the damage caused by the fire exceeds 400 billion
won on average each year. To prevent such fire damage, Korea provides fire safety education to improve
citizens' ability to cope with autonomous disasters. According to current regulations on fire safety management
in Korea, public institutions must conduct fire drills at least twice a year and train with fire stations at least
once a year. However, there is a survey that shows that fire education and training are not going well. Despite
fire drills at least twice a year and firefighting training conducted jointly with firefighters at least once a year,
a survey conducted at two schools showed that 89.6% of them said that fire drills were not being conducted
properly [1]. Overseas, fire prevention training regulations are stricter than in Korea. In Canada, the training
regulations were tightened by week due to the awareness of fire, and the training was conducted once a year
without notice. U.S. medical institutions, which have strengthened disaster prevention requirements and made
evacuation drills a routine, have recorded zero deaths from fires since 2014 [2].
Therefore, in order to strengthen fire prevention education in Korea, we will investigate how fire safety
education affects fire prevention in Seoul with data on gas explosion factors, electrical factors, mechanical
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
８１
AEICP Vol. 5, No. 1
factors, chemical factors, and the number of residents who received fire safety education [3-4].
2. Analysis Results Using R
This chapter introduces data and analysis methods used for efficient fire safety training analysis. The data
on the number of residents who received fire prevention education per year in Seoul and detailed damage data
on each detailed factor of fire per year are used.
For multiple regression analysis, the dependent variable is the number of residents, and the number of
detailed damage data is used as the independent variable to analyze what factors cause the most fires by year.
Multiple regression analysis was performed using the built-in lm() function of R programming [5]. Figure 1
shows the output from multiple regression analysis.
Figure 1.
Results of multiple regression analysis of the number of residents and fire factors.
In Figure 1, Pr(>|t|) is a P=value is called a significance probability. If it is less than 0.05, it can be said to
be a significant variable in the model of multiple regression analysis [6]. In other words, gas and care data are
important variables in multiple regression analysis. The gas data is a fire factor caused by gas explosion, and
the care data is a fire factor caused by carelessness in detail. It is classified close proximity to combustibles,
burning paddy fields, cigarette butts, embers flame flower garden neglect, playing with fire, washing laundry,
garbage incineration, Weld Cutting Polishing, Oil handling, cooking food, fireworks, Other (negligence)
matters.
The result of correlation analysis using the corrplot package to determine which fire factor is applied as the
largest factor in the care data is shown in Figure 2 [7]. As a result of checking the correlation of care data in
Figure 2, the correlation between 0 and +1 is a positive correlation, between -1 and 0 is a negative correlation,
and it can be said that there is no relationship as it is closer to 0. The major findings were cptc (close proximity
８２
8th ICAEIC-2022
to combustibles), cb (cigarette butts), effgn (embers flame flower garden neglect), wl (washing laundry), gi
(garbage, incineration), and cf (cooking food).
Figure 2.
Correlation Analysis Results for Care Data.
3. Conclusion
In this paper, multiple regression analysis and correlation analysis were performed through R programming,
a big data analysis tool, using the number of residents who received fire prevention education and various fire
factor data. The main factors according to the fire prevention education were gas explosion and negligence,
and among the negligence factors, factors such as cptc, cb, effgn, wl, gi, and cf were confirmed to cause fire.
Therefore, it is thought that fire prevention can be actively reduced if fire prevention education is conducted
８３
AEICP Vol. 5, No. 1
on factors measured through the results of this analysis rather than education and prevention of factors in
various fields for fire prevention.
Acknowledgement
This work was supported by Korea Institute of Police Technology (KIPoT) grant funded by the Korea
government (KNPA) (No.202100200201, Development of Advanced Technology for Perception Improvement
on Traffic Entities and Risk Mitigation on Adverse Driving Conditions for Lv.4 Connected Autonomous
Driving)
References
[1] Fire Safety Education, is it working really well?, https://www.joongang.co.kr/article/22014787#home
[2] Advanced countries have reduced fire deaths through sprinklers and training, https://newstapa.org/
article/Q2FDE
[3] Fire prevention education resident data, https://www.data.go.kr/data/15046372/fileData.do
[4] Fire Factor Data, https://nfds.go.kr/stat/general.do
[5] Imai, K., Keele, L., Tingley, D., & Yamamoto, T. (2010). Causal mediation analysis using R. In Advances
in social science research using R (pp. 129-154). Springer, New York, NY.
[6] Ji won P., Byounghee L., Suhyun Lee., & Sangwoo Kim. (2020). Correlation of motor function, balance,
and cognition in patients with stroke, 27(1), 56-65.
[7] An Introduction to corrplot Package, https://cran.r-project.org/web/packages/corrplot/vignettes/corrplot-
intro.html
８４
8th ICAEIC-2022
A Study of Illegal Transaction Flow through Blockchain Forensics
Yoon Cheolhee1, Kyung Min Beak2, Kyu Sic Oh3, and Jin-Mook Kim4
1Korean Nation Police University, Korea
2Korean Nation Agency, Korea
3Korean Seoul Police Agency, Korea
4Division of IT Education, Sunmoon University, Korea
bertter@police.ac.kr 1, bugzero@police.go.kr2, djghost@police.go.kr3, calf0425@sunmoon.ac.kr4
Abstract
Cybercrimes, such as the concealment and laundering of illegally acquired funds using the dark web or virtual
assets, have become more advanced, and the scale and range of the damage in-flicted are drastically
increasing; thus, the development of technology to combat them, is urgently needed. Although economic losses
incurred nationwide have been continually increasing due to virtual assets, such as unauthorized financing,
damage from ransom ware, hacking into ex-changes, and the sale of illegal goods such as narcotics, there are
currently insufficient measures for analysis; for example, to track virtual assets or illegal transactions. In the
case of virtual assets, which are connected to the dark web in cyberspace, there are technical difficulties in
identifying and apprehending criminals who engage in such illegal dealings, while additional crimes, dis-
tributing drugs or pornography and concealing criminal funds concurrently take place. In this study, we
consider how to automate the tracking of virtual assets and implement such a system by using blockchain
forensics.
Keywords: Blockchain forensic, digital forensics, unauthorized financing forensics, virtual asset tracking.
1. Introduction
The domestic and international environment for cybercrime is evolving into new forms, involving
cybercrime information, fake news, aiding and abetting suicide, online drug deals, hacking, online fraud, and
smishing. These activities take place the Internet and social media, with transnational accessibility and
tremendous reach. Criminals are exploiting illegal virtual assets to undertake criminal activities, while keeping
in mind the characteristics of the distribution of diverse criminal data. Meth-ods or systems for the automated
tracking of illegal virtual assets have been consid-ered as a solution for such problems. If the on-site
investigator can track the transfer of virtual currencies and identify the actual time of deposit at the virtual
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
８５
AEICP Vol. 5, No. 1
asset exchange, it is possible to freeze the illegal virtual assets and confiscate them. Virtual assets or
cryptocurrency transactions are becoming more active through the use of blockchain technology, but crimes
involving virtual assets, such as hacking or illegal funds being circulated as virtual assets, are also on the rise,
despite the high level of security of blockchain. Due to regulations in most countries, virtual assets currently
cannot re-place cash and are, rather, used as a means for the storage of value, rather than as payment. Therefore,
if virtual assets were to be used to commit crimes, criminals could deposit virtual assets in the virtual asset
transaction exchange and then cash them through dealings within the transaction exchange [1].
Generally, virtual asset transactions are recorded in each node of the blockchain, such that they can be
checked by anyone, but a centralized transaction exchange rec-ords transactions among clients within the
transaction exchange only in its internal database and not in the blockchain and, so, it cannot be accessed from
outside the ex-change. Consequently, if the virtual assets suspected of being the subject of a crime are deposited
into the virtual asset transaction exchange and then exchanged into other virtual assets, or if they are cashed
and then withdrawn, this poses a problem, as it causes a significant delay in the investigation due to insufficient
cooperation from the transaction exchange in question. The goal of blockchain forensics is to provide meth-
ods and a system for the automated tracking of illegal virtual assets that enable the freezing or confiscation of
virtual assets when the flow of illegal virtual assets is tracked and the time of deposit at the virtual asset
exchange is identified [2].
2. Blockchain Forensics Transaction Process Analysis
Virtual assets use blockchain when creating, distributing, and storing transmitted and received data; one
data unit in a blockchain is called a transaction, and a group consisting of an unspecified number of transactions
is called a block [6].
Figure 8. Process of Block Tracking.
８６
8th ICAEIC-2022
As each block has its own sequential block number, which is stored after part of the preceding block data
and part of the following block data are combined and hash is processed, all of the following block data need
to be changed in order to change a part of the past block data; therefore, It is widely accepted that manipulating
past block data within blockchain technology is virtually impossible. A hash is added to the in-ternal data for
each transaction, generating a unique value, which differentiates it from other transactions; this is called a
transaction ID (TXID). In addition, the data distribu-tion process is illustrated in Figure 1, as explained in the
following.
3. Using Blockchain Forensics for Automated Analysis of Illegal transactions
Blockchain database forensics analysis identifies and stores the address within the virtual asset
exchange and then performs an analysis on all transaction data in the blockchain, including addresses
outside the virtual asset exchange. An internal address extraction analysis to identify addresses within
the virtual asset exchange, an analysis to determine an attempt to deposit or the completed deposit of
illegal virtual assets, and an analysis to determine the time of an attempt to deposit or a completed
deposit of illegal virtual assets to an address within the virtual asset exchange are applied. When such
analyses are conducted, it is possible to identify whether the transfer of a virtual asset took place from
an address that was established in advance outside the virtual asset exchange, to an address within
the virtual asset exchange that was issued in advance solely for depositing, or the central management
address of the virtual asset exchange where the virtual asset is deposited. If the circulation of cash
and virtual assets is assessed, as in Figure 6 (if the number of transactions is abnormal), the address
is flagged as suspicious and a system to track the address in question is set up.
Figure 9. Process of tracking virtual asset transaction.
８７
AEICP Vol. 5, No. 1
4. Conclusion
The process of consistent amounts of exploited virtual assets being sent regularly to unidentified multiple
wallets in a short amount of time demonstrates that the mix-ing and swapping process may be automated. As
exploited virtual assets undergo mixing and swapping, there are multiple suspects, and it appears that the
exploited virtual assets are being laundered by a virtual asset transfer program, or by using a mixing and
swapping transaction exchange. It is possible to see that the exploited vir-tual assets are divided and stored, in
order to prepare for problems occurring in the wallet to which virtual assets are transferred, by using the
personal wallets and virtual asset exchange wallets of multiple unspecified persons. In addition, virtual asset
transaction exchanges taking place abroad are used, due to the fees involved in cashing out exploited virtual
assets, the possibility of converting assets to cash, and the possi-bility of withdrawal after converting to cash.
The automated system for tracking illegal virtual assets uses the clustering meth-od to match and process
the address information and user information of the virtual assets. As one individual can create and own
multiple addresses, and all transactions pertaining to these addresses are carried out within the blockchain, it
is possible to an-alyze the user’s payment pattern by analyzing the transaction details, even though ob-taining
user information only with the address is not possible. Ultimately, it is possible to conduct user matching
through use of the blockchain forensics method and, as the recipient of a particular virtual asset (e.g., bitcoin)
can be the owner but the sender cannot have several owners, if there are multiple sender addresses, it is assumed
that they have the same owner and the blockchain forensics method can be applied. By us-ing this blockchain
forensics clustering technique, the address and transaction history information can be classified and stored in
a database, while the stored clustering data can be referenced, in order to visualize the transaction history per
user and display it. Furthermore, information on the most concerning areas of cyberspace can be shared, illegal
transactions can be prevented, and management expenses can be reduced by utilizing all the network users
participating in data transactions. If attempts at illegal transactions involving virtual assets are blocked and an
AI-based judgment algorithm is applied, as detailed in this paper, then an advanced method of regular patrol
process research based on the monitoring of virtual assets could be made possible. In addition, swift monitoring
and responses, due to a concentrated investigation capacity and high work efficiency, as well as the automation
of the work process, can be expected.
Acknowledgement
This work was supported by Institute of Information & communications Technology Planning &
Evaluation(IITP) grant funded by the Korea government(MSIT) (No.2020-0-00901, Information tracking
technology related with cyber crime activity including illegal virtual asset transactions)
References
[1] Mougayar, W. The Business Blockchain: Promise, Practice, and Application of the Next Internet
Technology; John Wiley & Sons: Hoboken, NJ, USA, 2016.
Kim, J. A Study on the Factors Affecting the Intention of Acceptance of Block Chain Technology. Ph.D.
Thesis, Soongsil University, 2016.
８８
8th ICAEIC-2022
Kim, T. Block Chain Concept and Case Analysis by Field; Ji-Jong 487; 2017.
No, H. A terminal agent for smart signage linked to the Internet. In Proceedings of the Korean Information
Science Society Conference, 2016; pp. 360–361.
Park, S. Block chain paradigm and fintech security. J. Korean Inst. Commun. Sci. Inf. Commun. 2017, 34, 23–
28.
Lone, A.H.; Mir, R.N. Forensic-chain: Ethereum blockchain based digital forensics chain of custody. Sci. Pract.
Cyber Secur. J. (SPCSJ) 2018, 1, 21–27.
Kim, S.; Chang, S.; Lee, S. Consumer Trend Platform Development for Combination Analysis of Structured
and Unstructured Big Data. J. Digit. Converg. 2017, 15, 133–143.
Park, H.; Song, M. A study on the research trends in library & information science in korea using topic
modeling. J. Korean Soc. Inf. Manag. 2013, 30, 7–32.
Oh, J.S. Identifying research opportunities in the convergence of transportation and ICT using text mining
techniques. J. Transp. Res. 2015, 22, 93–110.
Na, S.T.; Kim, J.H.; Jung, M.H.; Ahn, J.E. Trend analysis using topic modeling for simulation studies. J. Korea
Soc. Simul. 2016, 25, 107–116.
Bae, J.; Han, N.; Song, M. Twitter issue tracking system by topic modeling techniques. J. Intell. Inf. Syst. 2014,
20, 109–122.
Park, S. Virtual Asset Fraudulent Transaction Cybercrime Activity Information Tracking Technology; Annual
Report Dec. 2020.
Wang, X.; McCallum, A. Topics over time: A non-Markov continuous-time model of topical trends. In
Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
Philadelphia, PA, USA, 20–23 August 2006; pp. 424–433.
８９
AEICP Vol. 5, No. 1
Recycling of the Waste Paper Using the Cellulose Extracted from Pine Apple
Leaves and Its Analysis
Sabina Paudel1, and Bindra Shrestha2
1,2Department of Chemistry, Tri-Chandra Multiple Campus, Tribhuvan University, Nepal
paudelsabina17@gmail.com1, binraghu@yahoo.com2
Abstract
This work is mainly focused on the feasibility of extraction of cellulose from an agricultural waste material,
pineapple leaf through simple and effective procedure. This process of extraction of cellulose includes acid
hydrolysis and bleaching process. The final product was characterized by FTIR and XRD analysis. Cellulose
is most abundant linear polymer which consists D-anhydroglucose joined together by beta-1, 4-glycosidic
linkage. It is a polysaccharide polymer. The typical peaks of cellulose are usually observed at 2ß around
15degree and 22.6 degree. This extracted cellulose was used in paper recycling process. FTIR analysis of
recycled paper showed that adding cellulose helps to improve the quality of recycled paper. The study of those
different properties of recycled paper and comparing with the properties of sample paper it has been concluded
that paper can be recycled again. It can be used further for making different other products like tissues, cards,
envelops, paper bags, paper of lower grades, egg crates, boxes, etc. Since deinking and bleaching did not
affect the chemical composition of pulp, papers formed in three different sets are quite similar in physical
properties.
Keywords: cellulose, FTIR, XRD, extraction, pineapple waste, deinking, recycle.
1. Introduction
Modern civilization cannot be imagined without the use of paper. Paper is well known as the past and the
present event recorder. The primary source of raw materials for the production of paper is plants [1]. It is
known that paper production has various effects on the environment. Likely there are technologies which can
moderate the negative impacts on the environment and causes the positive economic effect. One of the
processes is the recycling [2].
Recycling is one of the best methods of disposal simply because it goes a long way to preserve the
environment. The best part of the recycling is that it has both economic and environmental benefits. Recycling
can prevent the waste of potentially useful materials which reduce energy, air pollution and water pollution
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
９０
8th ICAEIC-2022
[3]. Recycling process involves a series of steps to produce new products. Recycling helps to extend the life
and usefulness of something that has already served its initial purpose by producing something that is useable
It helps to develop the tangible measures to express the environmental benefits associated with the recycling
of various materials. Recycling almost always has environmental impacts of its own. Paper recycling is one of
the most well established methods applied to waste materials in today's world [4]. Paper recycling is the process
of converting waste papers into the useable papers and paper products. Recycling process is carried out mainly
in seven processes: sorting, Re-pulping, screening, cleaning, Deinking, Refining and paper making [5].
As we are talking about paper, the main source of paper pulp is wood. 90% of paper pulp is made of wood.
Wood is the hard, fibrous substance found beneath bark in the stems and branches of trees and shrubs. It is
plentiful and replaceable. Since a new tree can be grown where one has been cut, wood is considered as
renewable natural resources. Almost 90% of virgin fibers come from woody plants that are being used in paper
industry. However, in more and more region of world, recycled fiber is becoming major source of papermaking
fibers. Paper production accounts for about 35% of felled trees and represents 1.2% of the world’s total
economic output. Recycling one ton of newsprint saves about one ton of wood while recycling one ton of
printing or copier paper saves more than two tons of wood. So, paper recycling has been the area of interest
nowadays. Addition of Cellulose during the recycling process can increase the quality, durability and the
strength of the paper.
Cellulose is most common organic compound which is considered as the structural component in the herbal
cells and tissues. The term cellulose was first used by Anselme Payen in 1838. Cellulose is the linear polymer
consisting of D-anhydroglucose units joined together by ß-1,4-glucosidic linkages [6].
In this work, cellulose was extracted from the waste pineapple leaf which consists of cellulose,
holocellulose, hemicelluloses and lignin along with some gum and resin. Pineapple leaves are of high
mechanical value and very effective for weakness by providing energy and improve blood circulation.
Cellulose can be extracted from the plants woods as well as leaf using some of the chemical and mechanical
methods. Cellulose is focused as research interest owing to its environmental friendly benefit. Due to its
nontoxicity, biological and biodegradation, thermal and mechanical properties, renewable and easy
modification it has received great importance for the research. Cellulose can produce considerable derivatives
by the means of chemical and physical modifications. Cellulose derivative is microcrystalline cellulose (MCC)
which is generally generated by the acid or enzymatic hydrolysis of cellulose with high cellulose content [4].
Normally pineapple leaf fiber contained higher cellulose content than wood fiber. Here the pineapple leaves
were used in the extraction of cellulose which is further used in the enhancement of the quality of recycled
paper. Pineapple is produced 16-19 million tons annually in the world. It is one of the most popular typical
tropic fruits. Pineapple is mainly consumed as fresh fruit and juice. The disposal of pineapple peel and leaves
in the large scale causes environmental issue. Pineapple peel and leaves is principally composed of cellulose,
hemicelluloses, lignin and pectin. Recently it is reported that pineapple peel as well as leaves is used for
isolating cellulose and preparing cellulose based hydrogels [5].
2. Materials and Methods
2.1. Extraction of Cellulose from Pineapple leaves
９１
AEICP Vol. 5, No. 1
Pineapple leaf was used for the extraction of cellulose. The leaves were washed properly, air dried and
grinded. About 100 gm of powdered leaf was washed twice with distilled water and let them dried in air. The
dried material was mixed with 1000 mL nitric acid (1.0 M) for 24 hrs and washed again with distilled water.
It was then dried in an air oven at 50-60 oC. The dried material was treated with 1000 ml sodium hydroxide
(1M) for 24 hrs. Then the mixture was filtered and the dark brown filtrate was collected in a plastic container.
The filtrate was treated with sulphuric acid (5.0 M) with continuous stirring until the constant pH around 5-6
was obtained. After some time cellulose was separated. The mixture was filtered and the residue was washed
thoroughly with distilled water. Thus obtained cellulose was dried and 26 gm of dried cellulose was collected.
2.2. Recycling of waste paper
About 60 g of used writing paper was soaked in distilled water for 2 days. The soaked paper was re-pulped.
The pulp was divided into three parts. One part of pulp was kept separately to make the re-pulped paper and
two parts of pulp was mixed with 100g Ariel detergent which is used as deinking agent and left for 24 hours.
Then the deinked pulp was washed under running tap water for about 10 times to remove the ink particles
separated out from the pulp. The deinked pulp was mixed with 7g of bleaching powder to bleach the pulp and
left for 24 hours. After being bleached, the pulp was again washed about 5 times to remove the bleaching
powder. The bleached pulp was divided into three parts. One part of the pulp was used to make the bleached
hand sheet paper.
The remaining pulp was then mixed with cellulose extracted from pineapple leaves to make the mixture of
1:2 and 1:4 ratios. Two different hand sheets were made from these mixtures.
2.3. Characterization of Cellulose and papers
The extracted cellulose is characterized by XRD (Bruker D2 phaser) and FTIR (Model: IRPrestige-21). The
recycled paper, deinked paper and cellulose mixed recycled paper are also characterized by their FTIR spectra.
Similarly some other basic properties of papers likebasic weight (grammage), bulk, moisture content and ink
absorption of papers are also studied.
• Basic Weight (Grammage) =
. Its unit is gm/m2.
• Bulk =
. Its unit is m3/gm.
• Moisture content = . .
. × 100%
• Ink absorption : For ink absorption, first of all a small drop of ink was placed in the sample paper with
the help of a capillary tube. Then, the time required to spread and dry the ink drop was noted by the help
of a stopwatch. The diameter of the ink drop was measured.
• Water absorption : Water absorption was tested by weighing the amount of water absorbed by paper in
different time intervals. The weight of paper was first measured then, it was kept in a beaker containing
150 mL water. Then the water absorbed by the paper after 5mins was measured by weighing it. Before
weighing, the excess water present on the paper was removed by letting to drip by suspending it in the
air in glass rod and keeping it in between the folds of dry filter paper. After weighing, it was again
soaked in water and weighed after 5mins. The process was repeated till 30mins.
９２
8th ICAEIC-2022
2. Results
Figure 1. XRD Spectra of Cellulose. Figure 10. FTIR pattern of Cellulose.
Figure 3. FTIR pattern of Cellulose and recycled paper. Figure 4. FTIR pattern of cellulose mixed and
deinked paper.
Some basic physica; properties of papers
• Radius of sample and recycled paper = 0.095 m
• Area of sample and recycled paper = 0.028 m2
Papers
Weight
at 27
0
C
Weight
at 80
0
C
Thickness
(mm)
Grammage
(gm/m
2
)
Bulk
(m
3
/gm)
Moisture
content (%)
Ink absorption
(mm)
Sample
2.21
1.96
0.107
78.92
2.99×10-6
11.31
1.5
Repulped
2.08
1.83
0.188
74.28
5.26×10-6
12.02
3.8
Deinked
2.11
1.77
0.115
75.35
4.34×10-6
16.11
3.3
Recycled (1:2)
2.01
1.72
0.128
71.78
3.58×10-6
14.42
3.0
Recycled (1:4)
2.02
1.70
0.164
72.14
4.59×10-6
15.84
3.1
９３
AEICP Vol. 5, No. 1
Water absorption
Time
(min)
Wt. of Sample
Paper (gm)
Wt. of Repulped
Paper (gm)
Wt. of Deinked
Paper (gm)
WT of Recycled
(1:2) Paper (gm)
Wt. of Recycled
(1:4) Paper (gm)
5
2.21
2.08
2.11
2.01
2.02
10
4.52
4.60
4.57
4.84
4.82
15
4.58
4.71
4.61
4.98
4.97
20
4.60
4.72
4.65
5.0
4.98
25
4.59
4.72
4.62
5.0
4.96
30
4.56
4.70
4.62
4.98
4.94
4. Conclusions
Cellulose has a crystalline structure, and differs from hemicelluloses and lignin, which are amorphous in
nature. The crystalline structure of cellulose is due to hydrogen bonding and Van der Waals force between
adjacent molecules [7].
The XRD pattern for the cellulose prepared from pine apple leaves is shown in figure 3.1. The typical
peaks of cellulose I are usually observed at 2Ɵ values of around 150 and 22.60 [8-9]. We found that cellulose
has typical peak at 2Ɵ values of 190, consequently indicating that cellulose is type I. The peak obtained is
sharper and the sharper diffraction peak was correlated with higher degree of crystallinity in cellulose structure
[10]. The higher crystalline structure hindered an enzymatic hydrolysis of cellulose [11]. In other word enzymes
like cellulose can hydrolysis amorphous cellulose faster than crystalline cellulose. So, less crystalline structure
can be used in various ways inn food and biomass industries because they can be easily hydrolyzed by enzymes.
The FTIR pattern of cellulose prepared from pine apple leaves is shown in figure 2. FTIR spectroscopy was
used to confirm that the lignin and hemicelluloses have been removed during cellulose isolation step through
analysis of its functional groups. The peak intensity band at 3356 cm−1 is attributed to O–H stretching vibration.
The bands at 2893 and 1357 cm−1 are characteristics of C–H stretching and –CH2 bending, respectively. The
peaks at 1635 and 972 cm−1 are attributed to the H–O–H stretching vibration of absorbed water in carbohydrate
and the C1–H deformation vibrations of cellulose, respectively. The peak at 1730 cm−1 represents the ester
linkage of carboxylic group of ferulic and p-coumaric acids in hemicelluloses. It shows that hemicellulose
has not been completely removed during the chemical process [12]. Peak intensity at 1514 cm−1 is attributed to
the C=C stretching vibration in the aromatic ring of lignin. However, the cellulose did not show the C=C
stretching at that region. It indicates that lignin was well removed by chemical process [13–14].
The peak at 972 cm−1 in is connected with glycosidic –C1–H deformation, a ring vibration, and –O–H
bending. These characters imply the β-glycosidic linkages between the anhydro glucose units in cellulose. The
rise of intensity peak at 1057 cm−1 confirms that the cellulose content increased due to the alkaline treatment
[14]. Consequently, based on the result, it was clearly confirmed that the extraction of cellulose from pineapple
leaves was successfully accomplished.
The FTIR patterns of different recycled papers are shown in the figure 3 and 4. All of them have similar
pattern to that of cellulose. This indicates that most portion of paper comprises cellulose or cellulose derivatives.
The transmittance value of waste paper is relatively high while that of deinked paper is low. On addition of
９４
8th ICAEIC-2022
cellulose transmittance of paper has increased which is near about the original sample paper. This shows that
addition of cellulose during paper recycling process helps to maintain the quality of the paper.
The test carried out in sample paper and recycled paper showed recycled papers had low grammage and
bulk content value compared to sample paper. The water absorption was high in recycled papers as compared
to sample papers. Ink absorption of sample paper was very low and took much time to dry while ink absorption
of recycled papers were high consuming less time to dry. The study of those different properties of recycled
paper and comparing with the properties of sample paper it has been concluded that paper can be recycled
again. It can be used further for making different other products like tissues, cards, envelops, paper bags, paper
of lower grades, egg crates, boxes, etc. Since deinking and bleaching did not affect the chemical composition
of pulp, papers formed in three different sets are quite similar in physical properties.
Acknowledgments
The authors would like to acknowledge Department of Chemistry, Tri-Chandra Multiple Campus,
Tribhuvan University, Nepal for providing lab facilities to do this research work. We are thankful to
Department of Plant Resources for FTIR and Nepal Academy of Science and Technology (NAST) for XRD of
the samples papers.
References
[1] Singh, N., & Singh, C.B. (2015). A case study approach to sensitize the hoteliers. Research gate, 1-13.
[2] Cabalova, I., Kacik, F., Geffert, A., & Kanikova, D. (2001). Effect of paper recycling and its
environmental impact. Environmental management in practice, 17-350.
[3] Misman, M., Alwi, S.R.W., & Manan, Z.A. (2008). State of the art for paper recycling. International
conference on science and technology (ICSTIE).
[4] Pivnenka, K., Erikssona, E., & Astrupa, T. F. (2015). Waste paper for recycling overview and
identification of potentially critical substance. Research gate, 1-27.
[5] Amasuomo, E., & Baird, J. (2016). The concept of waste and waste management. Journal of management
and sustainability, 88-96.
[6] Eichhom, S.J., Baillie, C. A., Zafeiropouios. N., & Mawaikambo, L. Y. (2001). Current international
research into cellulosic fibers and composites. Journal of material science, 2107-2131.
[7] Zhang, YHP., Lynd, LR. (2004). Toward an aggregated understanding of enzymatic hydrolysis of
cellulose: noncomplexed cellulase systems. Biotechnol Bioeng, 88, 797–824.
[8] Klemm, D., Heublein, B., Fink, HP., Bohn, A. (2005). Cellulose: fascinating biopolymer and. sustainable
raw material. Angew Chem Int Ed, 44, 3358–3393.
[9] Mahadeva, SK., Yun, S., Kim, J. (2011). Flexible humidity and temperature sensor based on cellulose-
polypyrrole nanocomposite. Sens Actuators A, 165, 194–199.
[10] Alemdar, A., Sain, M. (2008). Isolation and characterization of nanofibers from agricultural residues–
wheat straw and soy hulls. Bioresource Technol, 99, 1664–1671.
[11] Yoshida, M., Liu, Y., Uchida, S., Kawarada, K., Ukagami, Y., Ichinose, H., Kaneko, S., Fukuda, K. (2008).
９５
AEICP Vol. 5, No. 1
Effects of cellulose crystallinity, hemicellulose, and lignin on the enzymatic hydrolysis of Miscanthus
sinensis to monosaccharides. Biosci Biotechnol Biochem, 7, 805–810.
[12] Chen, W., Yu, H., Liu, Y. (2011). Preparation of millimeter-long cellulose I nanofibers with diameters of
30–80 nm from bamboo fibers. Carbohydrate Polymers, 86, 453–461.
[13] Rosa, S. M. L., Rehman, N., Miranda, V S. M., Nachtigall, B., Bica, C.I.D. (2012). Chlorine-free
extraction of cellulose from rice husk and whisker isolation. Carbohydrate Polymers, 87, 1131–1138.
[14] Nguyen, H.D., Thuy Mai, T.T., Nguyen, N.B., Dang, T.D., Phung Le, M.L., Dang, T.T. (2013). A novel
method for preparing microfibrillated cellulose from bamboo fibers. Advances in Natural Sciences:
Nanoscience and Nanotechnology, 4, Article ID 015016.
９６
8th ICAEIC-2022
A Study on the Web Service-Based Platform for Pregnant Women
Jin-suk Bang1, Jin-Mook Kim2, Min A Jeong*
1Department of AI Convergence, Hoseo University, Korea
2Division of IT Education, Sunmoon University, Korea
*Department of Computer Engineering, Mokpo National University, Korea
bluegony@hoseo.edu1, calf0425@sunmoon.ac.kr2, majung@mnu.ac.kr3
Abstract
Pregnant women have a wide variety of suspected diseases due to symptoms, making it difficult to diagnose
the disease on their own, and even if they successfully diagnose the disease, they often delay treatment because
they do not know the exact treatment plan. This situation raises the anxiety of health-conscious mothers and
threatens their health, so a platform is needed for pregnant women to get accurate information on diseases.
Therefore, this paper aims to develop a customized big data-based web platform for pregnant women to
provide customized services that can present disease-specific diseases to pregnant women using web platforms
that are compatible with various IT devices and optimized for search engines, and to guide them to disease-
specific treatments and specialized hospitals to help them manage the disease.
Keywords: Big-data, Pregnant woman’s disease, Web service, Framework.
1. Introduction
The Due to the COVID-19 outbreak, even going to the hospital has become a world of fear. In these days,
when many parts of the world are changing to non-face-to-face due to concerns about contagion, providing
personalized medical services through the web could provide a better solution to many patients.
In particular, the elderly, including pregnant women, are reluctant to go to the hospital because they are
more likely to fall into a more dangerous situation when exposed to diseases. In addition, it is difficult to obtain
information about diseases or specialized hospitals. In this case, we will provide you with information about
your current disease status or a suitable specialized hospital.
If personalized medical services are provided, pregnant women will be able to comfortably predict their
physical condition in advance and check the current condition, thereby reducing concerns about their health.
Nowadays, smart devices are widely available and easy access to the Internet, so if there is a web service
platform for disease management in pregnant women, anyone can easily access it. Therefore, in this paper,
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
９７
AEICP Vol. 5, No. 1
based on medical big data, the design of a system model that searches for pregnant women's current disease,
emergency treatment, and nearby specialized hospitals is presented. In Chapter 2, big data-related research, in
Chapter 3, a web service platform using big data customized for pregnant women is designed, and in Chapter
4, the construction method of each component is described, and in Chapter 5, conclusions and future research
direction was presented.
2. Big-data for Medical sevice
Big-data is a large amount of data with a short generation rate and variety. In order to utilize various data
collected from various environments, a preprocessing process is required. The preprocessing process refers to
a process of consistently refining data by removing unnecessary data. The big data preprocessing process is
shown in Figure 1.
Figure 11. Pre-process of Big-data system.
Figure 1 shows the big data preprocessing process, and the composition is largely composed of Data
Transformation, Data Cleaning, Data Set, Data Reduction, and Data Discretization. . Data transformation is a
transformation through an inference engine to correct inconsistent data. Data cleaning is the process of filling
in data gaps, removing noise, and resolving inconsistent data. A data set is the combining of data from multiple
sources to form a coherent data. Data reduction is to reduce redundant or unnecessarily listed data while
ensuring that unique characteristics are not compromised.
Medical big data is a collection of data that can be searched, viewed, statistical, aggregated, and statistically
analyzed by recording the health, medical care, and nursing records of one citizen on a daily and constant basis
in the form of electronic data. It is classified into life science, clinical epidemiology, and health prevention,
and the types of big data in clinical epidemiology to be dealt with include medical treatment fee specification
data, patient registration data, government statistics, and other data.
９８
8th ICAEIC-2022
3. Proposed system
Figure 2 shows the overall configuration of our proposed system. The disease self-diagnosis service,
symptom search service, hospital search service, and disease-related exercise provision service are
organized into large categories using a disease management platform customized for pregnant women.
Figure 12. Structure of our proposed system.
The proposed system has four components. Each component is described. First, when a disease is
searched for in the disease self-diagnosis service, the search results for disease name, disease
information, treatment and precautions are displayed. In treatment, specialized hospitals and drug
information are searched more specifically. In the notes section, you search for food information
about beneficial foods and foods to avoid to help with disease treatment.
Second, in the symptom search service, the disease list is searched and the result is found using the
same DB in conjunction with the disease self-diagnosis service.
Third, the hospital search site searches for specialized hospitals in the vicinity and finds
information.
Finally, the fourth component, disease-related exercise provision service, provides a service for
exercise prescription if there is a part that can be prevented or treated with exercise, even if it is not a
major disease.
4. Conclusion
There have been many existing studies on personal medical information customization, but studies that
provide a platform for web service-based medical information for pregnant women, a specific class, are
insufficient. The web platform proposed in this study has the advantage of being able to check the diseases of
pregnant women at an early stage through disease and symptom search, and supports a system that can match
with professional medical staff through hospital search. In particular, it will be possible to relieve anxiety by
providing information about taking medications during pregnancy.
９９
AEICP Vol. 5, No. 1
In the future, a specific system will be built with a blueprint, and it can be used anytime, anywhere through
a mobile device. In addition, we will study technologies that can be fused with various IoT home appliances.
In addition, we will specifically look for ways to apply various algorithms to AI-based technology in this study.
References
[1] Yoon-Ae Ahn, Han-Jin Cho, Hospital System Model for Personalized Medical Service, Journal of the
Korea Convergence Society, Vol. 8. No. 12, pp. 77-84, 2017.
JI-Soo Kang, Kyungyong Chung, Heterogeneous Lifelog Mining Model in Health Big-data Platform, Journal
of the Korea Convergence Society, Vol. 9. No. 10, pp. 75-80, 2018.
Kim, SungHyun·Hwang, HyunSeok, Developing a Personalized Disease and Hospital Information Application
Using Medical Big Data, dmxrue Journal of Information Technology, Vol.15, No.2 pp.7-16, December 2016.
J. Ekanayake, S. Pallickara, and G. Fox, “MapReduce for Data Intensive Scientific Analyses,” the 4th IEEE
International Conference on eScience, pp. 277-284, 2008.
J. Kim and C. Jeong, “A Study on Phon Call Big Data Analytics,” Journal of Information Technology and
Architecture, Vol. 10, No. 3, pp. 387-397, 2013.
Y. Lim and E. Choi, “The CPS with the Hadoop ecosystems,” the 7th International Conference on Application
of Information and Communication Technologies, pp. 1-4, 2013.
J. Dean and S. Ghemawat, “MapReduce: Simplified Data Processing on Large Clusters,” Communications of
the ACM, Vol. 51, Iss. 1, pp. 107-113, 2008.
Adomavicius, G. and Tuzhilin, A., “Toward the next generation of recommender systems: A survey of the
state-of-the-art and possible extensions”, IEEE Transactions on Knowledge and Data Engineering, Vol.17,
No.6, pp.734-749, 2005.
Straus, S.E., et al., Evidence-based medicine: how to practice and teach EBM., 2005.
Groves, P., et al., 2013. “The ‘big data’ revolution in healthcare”, McKinsey Quarterly, Vol.2.
blueB, https://www.blueb.co.kr
Watson Health, https://www.ibm.com/watson/health/?lnk=mpr_buw
１００
8th ICAEIC-2022
A Deep Q-Network Based Intelligent Pumping System for the Rainwater
Pumping Station
Seung-Ho Kang
Department of Information Security, Dongshin University, Republic of Korea
drminor@dsu.ac.kr
Abstract
Recent climate change due to global warming is causing unpredictable changes in rainfall. Therefore, it is
difficult to adequately cope with the sudden change of reservoir due to global warming with a simple pumping
policy. In this paper, we propose a Deep Q-Network based rainwater pump operation method that can select
the appropriate number of operating pumps to maintain the proper water level using the information such as
rainfall, water volume, and water level of the reservoir.
Keywords: Climate Change, Rainwater pumping system, Reinforcement Learning, Deep Q-Network, SWMM
1. Introduction
The rainwater pumping station is equipped with several pumps and operates the pumps according to the
appropriate rules to drain the rainwater flowing into the reservoir. In general, rainwater pumping stations use
a simple rule-based pumping policy that increases or decreases the number of operating pumps based on the
water level of the reservoir. However, the rule-based policy is difficult to adequately cope with the sudden
change of the reservoir due to the sudden increase in rainfall.
Various studies have been conducted for the optimal operation of rainwater pump systems [1-4]. Most of
these are studies on pump operation automation or optimal operation using simple rules.
Reinforcement learning-based machine learning methods are suitable to determine the optimal policy for
pump operation in real time. The ultimate goal of the deep reinforcement learning-based rainwater pump
system is to keep the water level of the reservoir as low as possible by operating an appropriate number of
pumps during rain. Another target to consider together here is to keep the change in the number of running
pumps as small as possible to minimize power loss and pump wear. Although various algorithms exist in deep
reinforcement learning, in this study, we use the Deep Q-Network(DQN) model to solve the problem.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１０１
AEICP Vol. 5, No. 1
2. Simulation environment
Gasan rainwater pumping station located in Geumcheon-gu, Korea was used as a simulation environment.
Gasan Pump Station consists of three 100 cmm (cubic meter per minute) units and two 170 cmm units. Instead
of all pump operation types, only six combinations, [0,0,0,0,0], [1,0,0,0,0], [1,1,0,0,0], [1,1,1,0,0], [1,1,1,1,0],
[1,1,1,1,1], that are actually operated at the Gasan pumping station were used. Storm Water Management
Model(SWMM)[5] was used to simulate the rainwater pumping station.
3. Deep Q-Network based pump operation system
The problem to be solved by the rainwater pump system based on reinforcement learning can be defined as
a multi-objective optimization problem that must satisfy various objectives under predefined constraints[6-8].
The objectives aim to minimize the maximum water level in the reservoir and minimize the number of pump
changes. The reason to minimize the frequency of pump change is necessary to prevent power consumption
and pump wear due to frequent change.
A Deep Q-Network(DQN) was used to solve the pump combination selection problem. The neural network
that is responsible for decision-making according to the environment is a four-layer fully connected neural
network composed of an input/output layer and two hidden layers. The number of nodes in each layer is 17,
15, 20, and 6, respectively. The number of nodes in the output layer represents 6 possible pump combinations,
and each node outputs a Q-value for the corresponding action.
The elements of the input vector observed from the environment include rainfall, reservoir inflow, reservoir
water volume, and reservoir water level. Among them, the amount of rainfall is obtained by using simulated
rainfall data, the amount of inflow into the reservoir is a value obtained by simulating SWMM in units of 1
minute, and the volume and level of the reservoir are obtained by reflecting the operation results of the
operating pump combination. In addition, a recursive input element is needed to make the model temporal. As
these input factors, two recent consecutive pump combinations and the previous water volume of the reservoir
are used. The Gradient Descent algorithm is used to minimize the squared error between the Q-value estimated
by the DQN and the target Q-value for a state-action pair in the training step. Replay buffer and epsilon-greedy
policy are used to improve learning efficiency and expand search space.
The reward function was designed so that the model using reinforcement learning can make decisions every
minute about the pump combination that minimizes the highest water level in the reservoir while refraining
from changing the pump as much as possible. First, the reward for lowering the water level of the reservoir is
the difference between the water volume of the previous reservoir (pre_vol) and the volume of the reservoir
(cur_vol) when the currently selected pump combination is operated as shown in equation (1).
volume_reward = pre_vol
–
cur_vol (1)
In order to minimize the change in the number of pumps in operation, the previous two pump combinations
in a row and the current pump combination were used so that the smaller the number of pump changes, the
higher the reward.
１０２
8th ICAEIC-2022
pre_change_reward = change_count(ppre_action, pre_action) (2)
cur_change_reward = change_count(pre_action, cur_action) (3)
change_reward = 10 - (pre_change_reward + cur_change_reward) (4)
Here, ppre_action denotes the pump combination before last, pre_action denotes the previous pump
combination, and cur_action denotes the current pump combination. The change_count() function counts the
number of changes between two pump combinations. The constant 10 in equation (4) is the sum of the
maximum possible number of pump changes in pre_change_reward and cur_change_reward, respectively.
By weighted summation of the two rewards defined above, the final reward is defined as follows.
reward = w1 * volume_reward + w2 * change_reward (5)
4. Experiments and performance anaysis
4.1 Rainfall Simulation
Huff's dimensionless accumulation curve method was used as a method for time distribution of probability
precipitation. Huff's method has the characteristic of using the past good observational data, so its reliability
is recognized and it is the most widely used.
The rainfall sample was generated by dividing the rainfall period into 10 periods of 30, 60, 120, 180, 240,
720, 1080, 1440, 2880, and 4320 (minutes) for a 10-year cycle. Since 4 samples were generated for each
rainfall period, a total of 40 samples were used as training data for the model. And the same number of samples
were used as test data.
4.2 Experiment
Figure 1 shows the simulation results of a pump operating system based on DQN for one 60-minute rainfall
sample for a 10-year cycle. The experiment was conducted for up to 1 hour after the end of the rain. And the
amount of rainfall, accumulated rainfall, inflow into the reservoir, and outflow through the pump can be seen
in Figure 1(a). Figure 1(b) shows the water level in the reservoir over time when the pump combination is
selected and operated. Figure 1(c) shows the pump combination selected every minute according to the Q-
Value, the output result of DQN, and the cumulative number of pump changes according to the change of the
pump combination. Figure 1(d) shows the amount of outflow through the pump at the reservoir and the amount
of water stored in the reservoir.
Table 1 shows the average value obtained by simulating the pump operation against the test data of all 40
10-year rainfall samples using the proposed DQN model.
１０３
AEICP Vol. 5, No. 1
Figure 1.
Case of rainwater pumping operation for simulated 10-year cycle 60-minute rainfall.
Table 1. 10-year cycle rainfall test results.
Average maximum water level Average pump change frequency
Propose model 8.41 26.75
First of all, the average highest water level of the reservoir is 8.41m. The height of the reservoir is 9m, but
there is no case of overflow. It can be said that it was operated stably. However, the average number of pump
changes is very high, about 27 times. This is because the operation of the pumping according to the proposed
model attempts to pump out all the water in the reservoir. In particular, frequent changes of the operating pump
combination occur in the latter half of the period, when the rainfall stops and the water level in the reservoir is
lowered. This is because the system must satisfy the condition that the pump capacity of the pump combination
must not exceed the residual water quantity of the reservoir while the system tries to pump out all the running
water in the reservoir.
5. Conclusion
In this paper, a model was designed and implemented using the Deep Q-Learning algorithm to automate the
pump operation of the rainwater pumping station. In order to evaluate the performance of the deep
reinforcement learning-based pumping system, the environment and 10-year rainfall data of the actual Gasan
１０４
8th ICAEIC-2022
rainwater pumping station were prepared and simulated using SWMM.
The system was learned to find an optimal operation method by designing an appropriate reward function
to minimize the water level of the reservoir and the changes of operating pump combination. It was confirmed
that the learned system was operated stably without exceeding the height of the reservoir. However, the system
was designed to remove all of the running water in the reservoir, showing a large number of pump changes.
This paper is expected to contribute greatly to inducing new research in this reinforcement learning-based
pump automation system where it is impossible to obtain labeled learning data. In the future, it is necessary to
properly define the problem so that it can more reflect the actual environment. In addition, research on the
improved learning algorithm and model selection to reduce the number of pump changes is needed.
Acknowledgement
This work was supported by the National Research Foundation of Korea under Grant NRF-
2020R1I1A3071599.
References
[1] Yun, S.U. & Lee, J.T. (1995). A study on the development of the operation models for storm water
pumps in detention pond. J. KWRA, 28(6), 203-215.
Sim, J.H., Joo, W.C., & Lee, W.H. (1992). An adaptive control of inland pumping station using self-tuning
of fuzzy control technique. KWRA Hydro Research Paper, 291-299.
Feng, X., Qiu, B., Yang, X., & Pei, B. (2011). Optimal methods and its application of large pumping station
operation. J. Drainage and Irrigation Machinery Engineering, 2(10).
Zhuan, X., & Xiaouha, X. (2013). Optimal operation scheduling of a pumping station with multiple pumps.
Applied Energy, 104, 250-257.
https://www.epa.gov/water-research/storm-water-management-model-swmm
Baumeister, T., Brunton, S.L., & Kutz, J.N. (2018). Deep learning and model predictive control for self-
tuning mode-locked lasers. J. Optical Society of America B, 35(3), 617-626.
Lee, X.Y., et al. (2019). A Case Study of Deep Reinforcement Learning for Engineering Design: Application
to Microfluidic Devices for Flow Sculpting, J. Mechanical Design, 141(11), 111401(10 pages).
Richard, S.S. & Andrew, G.B. (2018) Reinforcement Learning: An Introduction second edition. The MIT
Press.
１０５
AEICP Vol. 5, No. 1
Implementation of Public Mobility Tracing and Tracking System Using QR
Code and Big Data to Support Tourism Industry in Bali
Evi Triandini1, IB Suradarma 2, Sofwan Hanief3, I Ketut Dedy Suryawan4, I Gusti Rai Agung Sugiartha5,
And I Wayan Agus Hery Setiawan6
123456Institute Tecnology and Business STIKOM Bali
evi@stikom-bali.ac.id
Abstract
The COVID-19 pandemic has made the economy of Bali, which is so far dependent on tourism sector, slump.
It even experienced a contraction up to 12.28% in the third quarter – 2020. However, Bali is designated as a
COVID -19 free tourist destination during these conditions. The designation provides encouragement for Bali
to increase efforts in tackling COVID-19. In Taiwan, Big Data Analytics has been implemented successfully
to help identify COVID-19 cases and generate real-time alerts through analysis of clinical visits, travel history,
and clinical symptoms. Regarding the designation of Bali as a COVID-19-free tourist destination and the use
of mobile technology and big data analytics to tackle COVID-19, it is necessary to conduct research related
to mobile-based public mobility and big data. The problem that has been identified in this study is the
unavailability of public mobility data that can be used to track people exposed to COVID-19, especially in
Denpasar City. The purpose of this study is to produce information tracing on the possibility of people being
exposed to COVID-19 based on mobility data from positive COVID-19 patients and notifications to their
smartphones to check with Health Facilities. This study conducts in Denpasar City in 146 public places such
as academic institution, government office, and shopping mall as the object research. The obtained data are
being processed and analyzed using Long-Term Short Memory (LTSM) method in Python. The results from
this study are top 10 places with the highest mobility out of 146 places studied were dominated by academic
institutions especially junior high schools in Denpasar City with the highest amount of mobility of more than
8000 people.
Keywords: COVID-19 pandemic, tourism sector, Big Data, clinical visits, mobile technology.
1. Introduction
The COVID-19 pandemic has made the economy of Bali, which is so far dependent on tourism sector,
slump. It even experienced a contraction up to 12.28% in the third quarter – 2020. Businessmen in the sectors
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１０６
8th ICAEIC-2022
of accommodation and food-beverage; transportation and warehouses; as well as services are the most affected.
The average turnover decline of the businessmen in those sectors was 90.93%[1]. However, Bali is designated
as a COVID -19 free tourist destination during these conditions. The designation provides encouragement for
Bali to increase efforts in tackling COVID-19. With regards to efforts to achieve the target as a COVID-19
free tourist destination, the mobility of the Balinese people has begun to be relaxed, which means that there is
a reduction in the limitations of people mobility so that productivity gradually increases significantly[2]. It is
feared that the reduction of people’s mobility restrictions will lead to an increase in the spread of COVID-19.
This is in line with the research conducted by [3]. The results of the study indicate that the large number of
residents in an area is not the main factor influencing the spread of this pandemic, but rather from the way of
interaction among individuals in the society. The number of residents has a negative correlation with the spread
of COVID-19.
Various innovations and technological applications have been developed to fight the coronavirus
pandemic[4]. The pandemic has implications for the design, development, and use of information systems and
technology[5]. There is an urgent need to understand what role can be played by the researchers of information
systems and technology during this pandemic [6]. Researchers and practitioners in the field of information
systems and technology can help analyze COVID-19 pandemic data and engage in research on relevant topics,
such as facilitating jobs while doing social distancing, contactless trading (e-commerce), facial recognition
while wearing masks or in other crises, COVID-19 applications in terms of privacy, crowdsourcing, donating
data, and others.
Several new technology applications such as COVID-19 contact tracing application and mobile-based
chatbot have been developed recently to fight this pandemic. The application of the technologies can help
reduce the impact of the coronavirus pandemic on society and organization. The effective and innovative use
of new technologies can help identify the spread of the coronavirus in society, monitor the condition of infected
patients, improve the treatment of patients infected with COVID-19, and help develop medical treatments and
vaccines [7]. Big Data Analytics can be used to identify people in need of quarantine based on their travel
history, predict the COVID-19 curve, accelerate development of antiviral drugs and vaccines, and advance the
understanding of the spread of COVID-19 across space and time [6].
Based on this description, namely regarding the designation of Bali as a COVID-19-free tourist destination
and the use of mobile technology and big data analytics to tackle COVID-19, it is necessary to conduct research
related to mobile-based public mobility and big data. The problem that has been identified in this study is the
unavailability of public mobility data that can be used to track people exposed to COVID-19, especially in
Denpasar City. The purpose of this study is to produce information tracing on the possibility of people being
exposed to COVID-19 based on mobility data from positive COVID-19 patients and notifications to their
smartphones to check with Health Facilities. The information tracing availability is expected to support the
readiness and achievement of Bali as a COVID-free tourist destination. Tracing capabilities that are much
more valid and measurable (based on data), are expected to attract the exposed residents for testing as quickly
as possible. In the end, it makes the active case curve more sloping and ready to support the opening of the
Bali tourism industry.
１０７
AEICP Vol. 5, No. 1
2. State of the Art
Technologies that have huge potentials such as Internet of Things (IoT), Artificial Intelligence (AI), Big
Data, Cloud computing, blockchain, and 5G can help in the Infection Prevention and Control of the COVID-
19 pandemic [4]. Mobile applications via smartphones and video conferencing tools also can be used to track
individual mobilities, remind people not to visit COVID-19 hotspots, and help doctors diagnose patients
through video services [6]. Big Data Analytics can be used to identify people in need of quarantine based on
their travel history, predict the COVID-19 curve, accelerate the development of antivirus drugs and vaccines,
and advance understanding of the spread of COVID-19 across time and space. In Taiwan, Big Data Analytics
has been implemented successfully to help identify COVID-19 cases and generate real-time alerts through
analysis of clinical visits, travel history, and clinical symptoms [7], [8], [9]. Researchers [4] assert that mining
data from various Big Data sources will be useful to find information and knowledge, which can be turned into
policy for appropriate actions that benefit health agencies in disease control and prevention. Other researchers
also support using Big Data Analytics to help governments track the spread of disease and monitor population
movements.
Understanding the various human mobility during the COVID-19 pandemic and how it responds to different
directions is incredibly valuable but remains challenging due to the lack of baseline data and the complexity
of the various confounding effects [10]. Restricting individual movement has shown an inspiring effect in
containing the spread of the virus in many countries. The mobility information generated by this study can
further help agencies identify community places that attract more travelers from a higher risk of infection.
Lastly, attention should also be focused on underserved and vulnerable populations, particularly low-income
groups, to help them overcome the challenges of adhering to new stay-at-home orders and social distancing
norms [4], [10], [11].
3. Method
This study is collaborative research involving two partners, the Department of Communication, Information
and Statistics of Denpasar City and PT. Bamboomedia Cipta Persada. This research involves five Penta helix
components, namely academics, government bureaus, industry, non-profit institutions, and society. This study
will be conducted in Denpasar City in 146 locations as the object research. The chosen object research location
is a location that represents a place for people to gather, for example, schools, government and private
institutions, markets, supermarkets, modern shopping centers (malls), hotels, hospitals, places of worship,
banks, tourists’ attractions, and bus terminals. The researchers have installed a QR code at the designated
location to get the data number of people who came and check-in from September 2021 to October 2021. The
obtained data will be processed and analyzed using Long-Term Short Memory (LTSM) method in Python. The
results of data analysis will get the top ten places in Denpasar City with the highest mobility along with the
highest crowd time.
１０８
8th ICAEIC-2022
4. Results
4.1. Data Collection
The data used in this study is data from the Denpasar community who check-in at locations or points that
have been installed with QR Codes. The locations with QR Codes are 146 locations. The number of QR Code
installations achieved in the study exceeded the targets, the targets of this research are 100 locations, but in
this research QR Codes have been installed in 146 locations. Locations for the QR Code installation include
schools, colleges, urban village offices, village offices, banjar offices, Denpasar Mayor's office, Denpasar
Kominfo office, shops, restaurants, markets, and mass media offices. The data used in this research are the
ones recorded in the system from September to October 2021. The data number of people who check-in at
point locations in two months is 39,877.
4.2. Data Processing and Analysis
The results of data processing and analysis carried out using the Long-Term Short Memory (LTSM) method
from 39,877 people obtained from 146 places in Denpasar City will obtain top ten places with the highest
mobility and the highest crowd hours to help the government trace and track peoples’ mobility in Denpasar
City. The data processing with the Long-Term Short Memory (LTSM) method is using Python programming
language. The Python libraries that are used to process and analyze the data are 'pandas', 'matplotlib', and
'seaborn'.
Top 10 Highest Mobility Places in Denpasar City
Data visualization from the results of the data analysis is shown in the form of a horizontal bar plot by
utilizing the 'matplotlib' library by replacing the 'dataframe' index with the 'Place' column, then deleting the
'place' column to prevent data redundancy using the codes in Figure 1.
Figure 1.
Horizontal Bar Plot Data Visualization.
After compiling the program code, the data visualization of top 10 highest mobility places in Denpasar City
will be displayed as shown in Figure 2. Surprisingly, among of public places that become the object research
from this study, academic institutions in Denpasar City, especially junior high schools dominate the top 10
highest mobility places in Denpasar City.
１０９
AEICP Vol. 5, No. 1
Figure 2.
Top 10 Highest Mobility Places in Denpasar City.
Highest Crowd Time
The processed data will be extracted to see the highest crowd time (dates and hours) at the research object
places using the 'dt_day' function for 'Date' and 'dt_hour' for 'Hour' in Python. The information extraction
results are visualized in the form of Distribution Plot or Distplot to see the data distributions which can display
the highest crowd time using the code program as shown in Figure 3.
Figure 3.
Distribution Plot Data Visualization.
１１０
8th ICAEIC-2022
After running the code program in Figure 3, the crowd dates and hours in public places is displayed as
shown in Figure 4. The results from the data analysis are shown that the highest mobility hours in Denpasar
City is around 3 PM and followed by 7 AM. One of the factors that indicate the high people mobility around
that hours is because the majority of people in Denpasar City start their activities at 7 AM and come home
from workplaces or school at 3 PM.
Figure 4.
Distribution Plot Data Visualization.
5. Conclusion
This research has resulted in 146 points installed with QR Codes for tracing and tracking the public mobility
in Denpasar City. There are 39,977 people’s mobility who check in at the point locations. Based on the data
stored in the system and after the simulation carried out using the user data of COVID-19 suspect, then the
system can generate information for tracing people who have been in contact with the suspect. Notification
can be sent to the people. Based on the results of the analysis, top ten places with the highest mobility out of
146 places studied were dominated by academic institutions especially junior high schools in Denpasar City
with the highest amount of mobility of more than 8000 people. Offline learning which began to be held in
Denpasar City is one of the factors that causes schools to become places with the highest mobility in Denpasar
City.
References
[1] National Agency for Disaster Management (BNPB), Panduan Teknis Ideathon: Bali Kembali. Denpasar,
Bali, Indonesia: National Agency for Disaster Management (BNPB), 2020.
I. Fadil, “Dampak Pandemi, Ekonomi Bali Tumbuh Minus Selama 9 Bulan (Pandemic Impact, Bali’s Economy
Grows Minus for 9 Months),” Denpasar, 2020.
R. A. Ghiffari, “DAMPAK POPULASI DAN MOBILITAS PERKOTAAN TERHADAP PENYEBARAN
PANDEMI COVID-19 DI JAKARTA,” Tunas Geografi, vol. 9, no. 1, p. 81, Jul. 2020, doi:
10.24114/tgeo.v9i1.18622.
１１１
AEICP Vol. 5, No. 1
H. J. Smidt and O. Jokonya, “The challenge of privacy and security when using technology to track people in
times of COVID-19 pandemic,” Procedia Computer Science, vol. 181, pp. 1018–1026, 2021, doi:
10.1016/j.procs.2021.01.281.
M. K. Sein, “The serendipitous impact of COVID-19 pandemic: A rare opportunity for research and practice,”
International Journal of Information Management, vol. 55, p. 102164, Dec. 2020, doi:
10.1016/j.ijinfomgt.2020.102164.
W. He, Z. (Justin) Zhang, and W. Li, “Information technology solutions, challenges, and suggestions for
tackling the COVID-19 pandemic,” International Journal of Information Management, vol. 57, p. 102287,
Apr. 2021, doi: 10.1016/j.ijinfomgt.2020.102287.
S. Johnstone, “A Viral Warning for Change. The Wuhan Coronavirus Versus the Red Cross: Better Solutions
Via Blockchain and Artificial Intelligence,” SSRN Electronic Journal, 2020, doi: 10.2139/ssrn.3530756.
C. J. Wang, C. Y. Ng, and R. H. Brook, “Response to COVID-19 in Taiwan,” JAMA, vol. 323, no. 14, p. 1341,
Apr. 2020, doi: 10.1001/jama.2020.3151.
S. Wang et al., “A fully automatic deep learning system for COVID-19 diagnostic and prognostic analysis.,”
The European respiratory journal, vol. 56, no. 2, 2020, doi: 10.1183/13993003.00775-2020.
S. Hu, C. Xiong, M. Yang, H. Younes, W. Luo, and L. Zhang, “A big-data driven approach to analyzing and
modeling human mobility trend under non-pharmaceutical interventions during COVID-19 pandemic,”
Transportation Research Part C: Emerging Technologies, vol. 124, p. 102955, Mar. 2021, doi:
10.1016/j.trc.2020.102955.
W.-L. Shang, J. Chen, H. Bi, Y. Sui, Y. Chen, and H. Yu, “Impacts of COVID-19 pandemic on user behaviors
and environmental benefits of bike sharing: A big-data analysis,” Applied Energy, vol. 285, p. 116429,
Mar. 2021, doi: 10.1016/j.apenergy.2020.116429.
１１２
8th ICAEIC-2022
A Study on Exchange System by Cooling Dehumidification using Big-data
Analysis
Zhen-Huan WANG1, Jin-Mook Kim2, and Young-Chul Kwon3
1Department of Mechanical Engineering, Binzhou University, China
2Division of IT Education, Sunmoon University, Korea
3Department of Mechanical Engineering, Sunmoon University, Korea
huan_2020@hotmail.com1, calf0425@sunmoon.ac.kr2, yckweon1@sunmoon.ac.kr3
Abstract
In the present study, effects of reducing white smoke at a WSR (white smoke reduction) heat exchange system
were studied in the winter season. For this purpose, the heat transfer processes on SA (supply air) and EA
(exhaust air) were investigated by computer simulation using Solidworks. The flow simulation was used to
analyze the heat flow on the heat exchange system under uniform conditions. In order to evaluate the
performance of the WSR heat exchange system, W(water)/SA recovered capacities and the temperature /
absolute humidity reduction rate were calculated. Also, the mixing process of SA and the EA is presented in
the psychrometric chart to confirm the possibility of reducing white smoke.
Keywords: Cooling Dehumidificatio, Heat exchange system, Mixing zone, Psychrometric chart, White Smoke.
1. Introduction
The white smoke phenomenon is mainly observed in high-energy facilities that discharge high-temperature,
high-humidity and humid air to the outside. White smoke is white air emitted to the outside through a stack,
etc., and appears in the process of changing into minute droplets when hot water vapor contained in exhaust
mixes and diffuses with cold outside air. It occurs mainly in winter when the temperature difference between
exhaust and outside air is large, and white smoke contains environmental pollutants, fine dust, and dust.
Therefore, in order to reduce environmental pollution including air quality, it is necessary to develop white
smoke reduction technology.
Wang et al. [1] investigated the heat transfer and flow characteristics of a wave heat exchanger according to
changes in inlet conditions through computational analysis to investigate the heat flow characteristics of the
WSR system. Cho et al. [2] researched a method for reducing white smoke experimentally by manufacturing
a pilot WSR system. Takata et al. [3] reported that the air mixing ratio of the dry and wet parts in the white
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１１３
AEICP Vol. 5, No. 1
smoke reduction system affects the generation of white smoke through fluid simulations and actual
experiments.
In this study, the outdoor air conditions in the winter season, in which the white smoke phenomenon is easily
observed, were selected. The white smoke reduction effect of the WSR heat exchange system and the amount
of heat recovered from the SA and W were investigated numerically using Solidworks. For the analysis, the
WSR heat exchange system applied with air-to-air and air-to-water wave heat exchangers was modeled, and
heat transfer characteristics on the supply and exhaust side were predicted.
2. Analysis method and conditions
In order to understand the heat transfer characteristics by the heat flow process of the WSR heat exchange
system, it was assumed that the inlet velocities of the air side and the water side were uniform. For heat flow
analysis, 3D modeling was performed using Solidworks 2016, and air side heat flow analysis of the heat
exchange system was performed under uniform conditions using Solidworks flow simulation. The working
fluid state was assumed to be incompressible and steady state, and a model was used to analyze the
turbulent flow. As boundary conditions, uniform flow velocity was applied to the inlet, atmospheric pressure
was applied to the outlet, and adhesion conditions were applied to the surface. For the convergence of the
analysis, the critical error range was set to 10-4. About 7.8 million regular hexagonal grids were created using
Solidworks flow simulation, and the grid dependence was considered to derive more accurate computational
analysis results.
Figure 1. Schematic of WSR heat exchange system.
１１４
8th ICAEIC-2022
Fig. 1 shows the schematic diagram of the WSR heat exchange system (H: 3.5m, L: 2.2m, W: 1.1m), This
system supplies cold SA and W to cool the hot and humid EA flowing into ①, and discharges the exhaust to
a low temperature state to reduce the generation of white smoke. The arrows indicate the flow directions of W
(water), SA (supply air), and EA (exhaust air). The number indicates the position of the data measurement
section. HXs 1∼5 are cross-flow wave heat exchangers of the same size (HXs 1, 5; air-air, HXs 2, 3, 4; air-
water). In order to investigate the white smoke reduction characteristics of the WSR heat exchange system, an
incompressible steady state was assumed for flow field analysis including heat transfer. The conditions of EA
and W were fixed, and the temperature and absolute humidity of SA, W, and EA in the measurement sections
①-⑫ were obtained while changing SA velocity. As the analysis conditions for the WSR heat exchange
system, the inlet temperature, velocity, and relative humidity of EA were fixed at 71 °C, 3.5 m/s, and 60%,
respectively, and the inlet temperature and velocity of W were fixed at 6 °C and 0.15 m/s, respectively. The
SA velocity (3, 5, 7 m/s) was changed, and the temperature and relative humidity were 0°C and 40%. From
these results, the amount of heat recovered from SA and W, and the reduction rate of temperature and absolute
humidity of EA were calculated.
3. Analysis results
Fig. 2 shows the temperature distribution according to the change in the velocity of SA in the inlet/outlet
section and pipe flow section (A∼H) of the air-air heat exchangers HX5 and HX1 connected to the supply
system of the WSR system. The 0℃ supply air flowing into ⑧ passes through HX5 and HX1 and absorbs heat
from the high-temperature exhaust. Therefore, the temperature of the supply air changes as it goes downstream.
As the velocity of SA increased, the turbulence intensity between C-D on the right side of SA system increased
to about 8.7~14.5% due to the active convective flow in the pipe. And the SA outlet temperature of section H
decreased from about 28 to 15℃. The SA and EA introduced from at ⑫ and ⑥ exchanges supply and exhaust
heat by strong turbulent flow in the mixing zone. The strength of turbulent flow in the mixing zone was very
high, about 27.3 to 60.8% in the SA velocity range. The temperature and absolute humidity of EA of ① are
greatly reduced at the outlet ⑦ due to the mixed flow in the mixing zone. The decrease in temperature and
absolute humidity was 41℃ and 0.1216 kg/kg.
Figure 2. Temperature distribution of HXs according to SA velocity.
(a) 3m/s
(b) 5m/s
(c) 7m/s
１１５
AEICP Vol. 5, No. 1
Fig. 3 shows the temperature distribution of EA to the front-view (right) and side-view (left) of EA and W
pipes of the WSR heat exchange system. The faster the SA velocity, the more active the heat flow in the mixing
zone. And then the outlet temperature of EA decreased. As the SA velocity increases, the heat flow process
between SA and EA systems becomes more active, so that forced convection heat transfer increases. The
temperature at the outlet ⑦ of the mixing zone decreased significantly as the SA velocity increased, and then
the temperature decrease rate increased significantly to about 42.5, 48.5, and 54.8%. Also, the amount of heat
recovered from W was about 4.69, 5.41 and 6.27 kW, which increased by about 15.4% (at 5 m/s) and 33.7%
(at 7 m/s) based on the velocity of 3 m/s.
Figure 3. Temperature distribution of EA and W according to SA velocity.
4. Conclusions
In order to investigate the heat flow characteristics of the supply/exhaust piping of the WSR heat exchange
system according to the SA velocity, a computational analysis method was conducted under winter conditions.
The following conclusions were obtained by investigating the temperature and absolute humidity of the WSR
heat exchange system. The analysis results of the SA/EA unit and the mixing zone suggested that the outlet
temperature and absolute humidity at the mixing zone could be significantly lowered due to the forced
convection heat flow process as the SA velocity increased. The reduction in temperature and absolute humidity
in the mixing zone was 40-44% and 43-49% of the total EA system, respectively. This shows the temperature
and absolute humidity reduction effect is clear. And the amount of heat recovered from EA by W was about
35% of that of EA heat as the SA velocity increased.
References
[1] Wang, Z. H., Byun, S. J., Cha, J. M. and Kwon, Y. C. (2017), A study of computational analysis on the
heat transfer and flow characteristics of wave heat exchanger. Korea J. Society of Mechanical Technology,
19(4), 452-457.
Cho, G. H., Lee, S. J., Kwon, Y. C. and Kim, D. H (2017). Performance test of pilot white smoke reduction
system. Korea Academia-Industrial Cooperation Society Spring Conference, 772-774.
(a) 3m/s
(b) 5m/s
(c) 7m/s
１１６
8th ICAEIC-2022
Takata, K., Michioka, T., Kurose, R. (2016). Prediction of a visible plume from a dry and wet combined
cooling tower and its mechanism of abatement. Atmosphere, 7(4), 59.
Dassault Systems Solidworks Co. (2016). Solidworks Simulation Manual.
１１７
AEICP Vol. 5, No. 1
A Study on ICT Literacy Competency Model for Software Education
Jin-Hee Ku
Division of Software Liberal Arts, Mokwon University, Korea
jhku2020@gmail.com
Abstract
With the evolution of digital technology, convergence and connectivity are being emphasized not only in daily
life but also in the overall industrial field, and innovative educational methods suitable for the environment of
the 4th industrial revolution are continuously being explored at universities. Due to the big change in the
educational environment called digital transformation, software education has grown in importance, and a
method that can objectively measure ICT literacy competency is required. Recently, interest in ICT literacy as
well as creativity, problem-solving ability, and communication ability as a necessary ability to cultivate core
competencies of college students is increasing. In particular, the ICT literacy competency model also requires
evolution as ICT literacy affects students' teaching and learning as a whole, and ICT in the classroom
continues to evolve over time and as new technologies emerge. The purpose of this study is to develop an ICT
literacy competency model to strengthen ICT literacy competency by objectively diagnosing ICT literacy levels
and to establish effective software curriculum strategies.
Keywords: ICT Literacy, Competency Model, Software Education, Digital Transformation.
1. Introduction
ICT literacy includes computer literacy, information literacy, and knowledge literacy as the core
competencies that creative talents must possess in the knowledge-based society of the 21st century. It can be
defined as including the ability to select and evaluate information, reproduce and share knowledge, beyond the
use of information from computers or the Internet. In addition, the competency model refers to the
systematization of competencies including knowledge, skills, attitudes, and strategies required to perform a
job or role. In this context, competency modeling that includes both ICT literacy and ICT-using literacy is
required. Figure 1 shows the relationship between ICT literacy, that is, the ability to demonstrate ICT
technology itself and literacy using ICT. In other words, it includes ICT-based literacy as well as ICT literacy,
and it means thinking critically and creatively about information and communication technology as a citizen
of a global community while using ICT safely, responsibly and ethically.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１１８
8th ICAEIC-2022
Figure 1.
Literacy Domain using ICT.
Recently, as software education is mandatory in universities, interest in SW competency diagnosis is
increasing, but in most cases, it is a study on a tool to diagnose software coding competency. There are few
studies on the status of college students' ICT literacy and the level of difference between students and what
needs to be diagnosed to determine the level. The purpose of this study is to develop an ICT literacy
competency model to strengthen ICT literacy competency by objectively diagnosing ICT literacy levels and
to establish effective software curriculum strategies.
2. ICT literacy framework
ICT literacy is a concept that encompasses both cognitive and non-cognitive attributes necessary to solve
general information-related problems encountered while living in the information society. In other words, ICT
literacy includes not only knowledge related to the use of information processing processes, digital
technologies, communication tools, networks, etc., but also skills, abilities, and attitudes related to ICT such
as ethical and legal perceptions about information use. Table 1 shows the definitions of ICT literacy in previous
studies.
Table 1. Definition of ICT literacy.
No
Definitions of ICT literacy in previous studies
1
Ability to collect, produce, process, preserve, transmit and utilize information using software technology necessary for
operation of information devices and information management
2
Ability to solve problems by having a sound information ethics awareness, using ICT, recognizing what information is
needed, accessing, finding, processing, and effectively utilizing information
3
The ability to access, manage, integrate, evaluate, and produce information using digital technologies, communication
tools, and networks to function in a knowledge-based society.
4
ICT literacy consists of ICT proficiency, cognitive proficiency, and technical proficiency. Evaluated as a competency factor
Competence is an enduring disposition that enables an individual to predict an individual's behavior in a
variety of situations, including motivation, traits, self-concept, skills, and knowledge. The ultimate goal of
students' ICT literacy is to develop an integrated ability to function productively in a knowledge-based society.
１１９
AEICP Vol. 5, No. 1
Therefore, it would be desirable to integrate and enhance cognitive ability, which is a basic skill in daily life,
and technical proficiency, a component of digital literacy. Figure 2 defines the skills (proficiency) that ICT
literacy wants to develop, that is, cognitive and technical skills, and ICT skills.
Figure 2.
Components of the ICT literacy Competency.
3. ICT literacy competency model
A competency model for measuring ICT literacy was developed through literature study and expert review.
Through previous studies on the concept and competency of ICT literacy, ICT literacy competency elements
and concepts were classified, and large domains of competency were extracted. In addition, domestic and
foreign prior studies related to ICT literacy competency were analyzed to establish the 'content element' of
each competency diagnosis area. The developed ICT literacy competency model was evaluated for validity on
a 5-point Likert scale by an expert review, and feedback such as corrections and additional supplements was
reflected. Based on this, the final ICT literacy competency model was developed.
3.1. ICT Literacy competency elements and assessment domain
ICT literacy competency elements for developing items by domain were set as follows
○ Knowledge means understanding the basic functions and usage of ICT.
○ Skills refers to the ability to use technology to access, retrieve, store, manage, integrate, evaluate, create,
share, and participate in knowledge and information.
○ Attitude refers to the ability to evaluate information in a critical and reflective attitude.
Table 2. Large domain and the concept.
Assessment domain
Elements
Concepts
Accessing Information
Knowledge, Skills
Ability to identify information needs to determine how information is retrieved
and retrieved
Creating Information
Knowledge, Skills
Ability to generate information and knowledge through the modification,
integration, application, design, and authoring of information
Managing Information
Knowledge, Skills
Ability to organize, classify, and store information for retrieval and reuse
Sharing Inforamtion
Skills, Attitude
Ability to share knowledge and create, exchange, and collaborate on
information suitable for a variety of readers, situations, and media
Ethic Information
Knowledge, Atttitude
Ability to use ICT appropriately by recognizing social, legal and ethical issues
１２０
8th ICAEIC-2022
Table 2 shows the large domains of the ICT literacy competency model, competency elements for each
domain, and their concepts. Also, after extracting the middle domains for each large domain, the small domain
and its measurement items are developed.
3.2. Design for competency level types
Table 3 shows examples of competency level types in the ICT competency model framework. It will be
used to analyze the results of data measured through the ICT literacy diagnosis system built on the basis of the
ICT literacy competency model.
Table 3. ICT literacy competency level types
Competency and
performance level
Basic stage
Development stage
Maturity stage
overall performance level
Interest in ICT, but lack
of understanding and
poor use of ICT in all
domains
Recognizes the value and need
of ICT and uses ICT in daily
life, but lacks technological
application capability
Recognizes the value and need of ICT,
and is proficient in accessing, creating,
managing, and sharing knowledge and
information using ICT
Overall competency level
poor-fair
fair
good
Competency score
(out of 100 points)
0 to 40 points
40 to 80 points
80 to 100 points
Additional terms
n/a
Information access, creation,
and management
competencies, more than 40%
All major competencies, more than
40%
* =
(W: weight of Large domain, L: Large domain)
4. Conclusion
In this study, in order to measure ICT literacy competency, we developed a framework for performing ICT
literacy diagnostic tools that evaluate whether students can actually create products, and proposed to measure
competency in terms of knowledge, attitude, and skill. Therefore, compared to the existing measurement tools
to measure the competencies of survey participants based on their self-awareness, the developed performance-
type ICT literacy diagnostic tool is designed to objectively diagnose the student's performance level and
determine the actual proficiency in handling ICT.
References
[1] Australian Curriculum and Reporting Authority(ACARA) (2011). National Assessment Program-ICT
Literacy Years 6 & 10 Report 2011. ACARA, Sydney.
[2] Educational Testing Service(ETS) (2002). Digital Transformation - A Framework for ICT Literacy. A
report of International Information and Communication Literacy Panel. Educational Testing Service, USA.
１２１
AEICP Vol. 5, No. 1
[3] Kim, H. J. (2016). Exploring College Students’ Perceptions and Educational Experiences of Digital
Literacy. Journal of Learner-Centered Curriculum and Instruction, 16(8), 937-958.
[4] Jin, S. H., (2006). A Comparative Study on ICT Literacy Education Between South and North Korea.
Unification Stratrgy, 6(2), 299-331.
[5] Rha, I. J., Lee, J. H. (2009). Korean In-Service Teachers' ICT Literacy: The main components and the
structure of general cognitive and technical capabilities. The Journal of Educational Information and
Media, 15(4), 21-45.
[6] OECD (2015). OECD skills outlook 2015: Youth, skills and employability. OECD publishing, Paris.
[7] OECD (2019). PISA 2021 ICT Framework. OECD Publishing, Paris.
[8] Pernia, E., (2008). Strategy Framework for Promoting ICT Literacy. UNESCO.
[9] UNESCO-UNICEF (2013). Making education a priority in the post-2015 development agenda. Available
at http://www.unicef.org/education/files/Making_Education_a_Priority_in_the_Post-2015_Develop-
ment_ Agenda.pdf
１２２
8th ICAEIC-2022
IoT based Smart Home with Face Recognition Security
Sunil Pokhrel1, Anil K. Panta2, Sachin Parajuli3, Sudip Poudel4, Rajan Adhikari5, and MK Guragai6
1,2,3,4,5,6Department of electronics and computer engineering, Purwanchal campus, Institute of
Engineering Tribhuwan University, Dharan, Nepal
sunil073bex@ioepc.edu.np1, anil073bex@ioepc.edu.np2, sachin072bex@ioepc.edu.np3,
sudippoudel987@gmail.com4, rajan073bex@ioepc.edu.np5, mkguragai@gmail.com6
Abstract
This project features the controlling of home appliances from a centralized system along with face recognition
security. In today’s world, home security is the major concern. Various traditional method s that are being
followed till date are easy to break which eventually leads to burglary. To protect the home, we need to install
a costly security system. To overcome this problem, in this paper, we are proposing the system with the help
of face recognition to develop the smart door lock and unlock system. It also gives us the facility to monitor
our home remotely and take appropriate actions if anything goes wrong. Proposed system has the feature of
alerting the admin even when he/she is inaccessible to the internet which provides more security. Along with
this, cost-effective and secured home automation is also featured in this paper.
Keywords: centralized system, face recognition security, home security, protect.
1. Introduction
In this digital world, all kinds of things around us need to be automatic and safe, reducing human
effort. More and more electronic circuits are appearing every day to make today's life easier and
simpler. At present context, energy crisis is a big problem that everyone is facing. Therefore, it is
necessary to save energy for establishing an efficient and convenient system. This paper details the
design and development of IoT based security surveillance systems in buildings with Wi-Fi network
connectivity. Upon detecting the face, the controller enables the camera for capturing the event, alerts
the user by placing the live video of that event on the webpage along with smart home with automation
features installed on it.
IoT is an evolving technology that allows hardware devices to be controlled via the internet. Here
we propose to use IoT to control devices to automate modern homes via the internet. This system
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１２３
AEICP Vol. 5, No. 1
uses three loads that are used for lighting the house and a fan. Our user-friendly interface allows users
to easily control these devices via the internet. In this system, we use a central computer as the
processor. A camera is used to recognize faces. This system uses two loads to be demonstrated in our
project which are light and buzzer. It has notification automation that sends the alert message in the
owner's device like mobile (but we have used desktop to send notification for demonstration). It has
a user-friendly web application that allows the admin to monitor the house and control the home
appliances through the internet. When someone breaks into the house then the camera monitors the
house compares the face of the person with the sample images of the database. If the compared face
matches with the sample dataset then nothing happens, but when it doesn't match with the sample
dataset then it sends the notification to the owner. When the owner receives the notification he goes
to the login panel of the web application and enters his username and password. If he is the authentic
owner, then he gets to access the real-time live streaming of the camera footage. He can view the full
real-time video streaming and get a clear idea about who is on his property. He can alert the person
on his property by controlling the home appliances using IoT i.e., turning on light, turning on a buzzer,
etc. If the owner doesn't see the situation in his favor he could call the police and report the case where
the video footage can prove the evidence. This automation prototype is intended to demonstrate the
good functionality and scalability of the system and does not include all the desired features of the
system. There are countless features and devices that could have been integrated into the system, but
only one user interface was designed to stay within the scope of the project. The future of home
automation systems is to make homes even smarter. Standardization will create smart homes that can
control appliances, lighting, environment, energy consumption, security, and extend connectivity to
other networks.
2. Methodology
Proposed methodology consists of Node MCU, LED, Webcam, Server, Processor, and Operating
System. The whole system here is divided into two main sections i.e. home automation and security
system integrated there in the system. For home automation Node MCU is an integral part. Loads of
the house are all connected in the Node MCU centrally and for ac system realization relay modules
can also be integrated into the system. Here all the devices of the house can be controlled via the web
application. Buttons are provided to each electrical appliance individually and the user can control all
appliances via those buttons on the web page. On the other hand, for the security system face
recognition system is integrated. Here in this system CPU plays the integral part which is fed with
the face recognition program. The webcam integrated with the processor gets the human face detected
by the camera in front of the door.
１２４
8th ICAEIC-2022
Figure 13. Block Diagram of the proposed system.
If the detected face is unknown, then the owner of the house gets notified by the system via the
notification system. As the owner gets notified then the owner of the house login to the web interface
system and sees the activity around through the live streaming in the web browser. From the given
system owner can turn on the alert system integrated into the system and also take the necessary
required action.
As we have controlled the home appliances via the web application where all the devices in the
home appliances can be controlled via the buttons switch in the web page. Here the server connects
to the hotspot created by the node MCU and the data are sent and received via port 80.
At the same time for security system face recognition system is implemented in the system. The
web camera fitted in the system is in constant surveillance of the home and as soon as some unknown
person comes in front of the system it detects and sends a notification to the owner of the house. The
owner can log in to the system and monitor along with that he can turn on the buzzer or take necessary
action. In this way, home automation and security system using IoT can be implemented.
The error may occur while the user tries to log into the system using the wrong credentials. The
system may sometimes show errors while recognizing the face.
１２５
AEICP Vol. 5, No. 1
Figure 2. Flowchart for Face Recognition System. Figure 3. Flowchart for Home Automation System
Figure 14. Interfacing of node MCU with loads.
１２６
8th ICAEIC-2022
3. Discussion and Conclusion
The technology used in our paper can be used for many organizations, colleges, malls, etc. by
making slight modifications in the program and system. By using more datasets and appropriate
algorithms the system can be made more efficient and faster. With greater efficiency, we can control
home appliances and a security system can be implemented in the home. Time consumption and
electrical wastage are reduced. The operating software required is free and easily available and the
system can be implemented at a cheaper cost. This system can be used in various applications for
future scope.
References
[1] Dwi Ana Ratna and Wati Dika Abadianto, "Design of Face Detection and Recognition System for Smart
Home Security Application", 2017 2nd ICITISEE.
[2] Jiakailin Wang, Jinjin Zheng, Shiwu Zhang, Jijun He, Xiao Liang, and Sui Feng, "A Face Recognition
System Based on Local Binary Patterns and Support Vector Machine for Home Security Service
Robot", 2016 9th International Symposium on Computational Intelligence and Design.
[3] Sandesh Kulkarni, Minakshee Bagul, Akansha Dukare, and Archana Gaikwad, "Face Recognition System
Using IoT", IJARCET, vol. 6, no. 11, November 2017, ISSN 2278-1323.
[4] Ravi Kishore Kodali, Vishal Jain, Suvadeep Bose, and Lakshmi Boppana, IoT Based Smart Security and
Home Automation System, IEEE, 2016.
[5] Anwar Shaik and D. Kishore, "IoT based Home security system with alert and door access control using
Smart Phone", IJERT, December 2016.
[6] Rhythm Haji, Arjun Trivedi, Hitarth Mehta, and A.B. Upadhyay, "Implementation of Web-Surveillance
using Raspberry Pi", International Journal of Engineering Research & Technology (IJERT), vol. 3, no.
10, October 2014.
１２７
AEICP Vol. 5, No. 1
Design and Development of 3-stage Folding Electric Kickboard based on In-
wheel Motor Method
Duk-Keun An1, Ik-Hyeon Kim2, Si-Wook Sung3, Dong-Cheol Kim4, Sang-Hyun Lee5
1,2Institute of Technology SECONDWHITE Co., Ltd, Korea
3,4Institute of Technology CSE Co., Ltd, Korea
5Department of Computer Engineering, Honam University, Korea
dk@secondwhite.com1, gill@secondwhite.com2, swsung@csetp.com3, ironkd@csetp.com4,
leesang64@honam.ac.kr5
Abstract
Recently, awareness of environmental pollution is increasing, and new means of transportation are attracting
attention as exhaust gas regulations and regulations on vehicles operating in urban areas are being
strengthened or spread. Therefore, to reduce air pollution sources, e-mobility in the form of electric kickboards
that are powered by electricity is emerging as an alternative. In this paper, the design of a light hand-type
electric kickboard is adopted as a three-fold folding method so that women or the elderly can easily carry it.
In order to overcome the shortcomings of this, we intend to develop a low-noise, low-vibration and high-
efficiency drive system by applying an in-wheel type motor.
Keywords: In-Wheel Motor, e-mobility, Outer Rotor, Electric kick board.
1. Introduction
Currently, personal mobility is spreading as a new paradigm of transportation among Korean youth.
Recently, awareness of environmental pollution is increasing due to fine dust and global warming, and as a
result, movements to strengthen exhaust gas regulations and vehicle regulations in the city center are spreading,
so new means of transportation are attracting attention. Therefore, the use of e-mobility powered by electricity
to reduce air pollution sources is widely used among young people.
Electric kickboards use an electric power source as power, and personal mobility can be used by one or two
people. As a means of transportation, it can be used in a variety of ways, such as short-distance, commuting,
and leisure activities depending on the purpose, and the number of users is continuously increasing [1].
Accordingly, it has been reported that the occurrence and risk of unexpected safety accidents occur due to
the uncertain durability life and damage to major parts due to irregular road surface driving, and transmission
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１２８
8th ICAEIC-2022
of vibration according to the road surface condition during driving has a negative effect on the human body
[2].
Therefore, in this paper, the design of a hand-type electric kickboard that can be easily moved by women or
the elderly is adopted as a folding method that is a three-stage folding type. To overcome this, we are going to
develop a low noise, low vibration and high efficiency drive system by applying an in-wheel type motor.
2. Design of the proposed model
In this study, electric kickboard design for women or the elderly is a combination type that divides the high-
strength aluminum body into three modules and folds and unfolds, and a high-efficiency in-wheel method to
reduce vibration and noise of the outer rotor type motor was designed as shown in Figure 1.
Figure 1.
Concept Design.
Figure 2 presents a new design of the in whel outer rotor method to solve the noise of the motor shaft and
the vibration problem when driving on a curved road.
Existing motors have problems with shaft motor noise and vibration.
As an alternative to the problem, the use of an in-wheel outer rotor
type motor.
Figure 2.
Adopted In Wheel Outer Rotor method.
It is a system that independently drives the In Wheel Outer Rotor (IWOR) by installing a drive motor inside
the wheel mounted on the wheel of the electric scooter. Since all wheels are individually controlled, stability
is good when cornering and there is no energy wastage in the process of power transmission, so fuel economy
It has the advantage of having a large improvement effect.
１２９
AEICP Vol. 5, No. 1
By using the IWOR, it is possible to eliminate powertrain components such as the engine and to control the
driving force of each wheel, and because it independently controls the driving force of each wheel in a manner
similar to torque vectoring, environmental performance, safety, and convenience as shown in Fig. 3, it is
possible to improve the vehicle speed, reduce the body weight, solve the problem of power loss, which is a
disadvantage of electric vehicles, and secure additional interior space.
Figure 3.
Design of Outer Rotor Motor ASSY.
Figure 4 is the configuration diagram of the controller of the electric kickboard.
Figure 4.
Configuration diagram of controller of electric kickboard.
3. Results
The front wheel hub motor fitted to the front wheel of the electric kickboard studied in this paper should be
located between the forks, and it was designed to be narrower than the rear wheel hub motor. The width of the
motor manufactured to fit the gap of the fork was designed to be 100mm.
In addition, the rear hub motor for the rear wheel is wider than the front wheel because the freewheel or
sprocket is installed on the wheel where the gear is installed, and the width of the rear wheel hub motor is
designed to be about 137~145mm. A 10mm space was secured for the steel frame transformation allowance.
Figure 5 shows the concept product of the proposed study. It is a 3-stage folding electric kickboard that can
obtain synergistic effects in two aspects of design and performance.
１３０
8th ICAEIC-2022
Figure 5.
The final concept product of the proposed 3-stage folding electric kickboard.
4. Conclusion
Currently, most of the Personal Mobility products introduced and operated in Korea are purchased from
low-priced Chinese products and used. The electric kickboard has the advantage that it can be used in a variety
of ways, such as short-distance, commuting, and leisure activities, depending on the purpose as a means of
transportation. However, despite these advantages, the electric kickboard has an uncertain durability life and
damage to major parts due to irregular road surface driving, contains the occurrence and risk of unexpected
safety accidents, and the transmission of vibration according to the road surface condition during driving It has
been reported to have a negative effect on the human body.
Therefore, in this paper, the design of a hand-type electric kickboard that can be easily carried by women or
the elderly was adopted and developed in a folding method that is a three-stage folding type. In order to
overcome this problem, a low noise, low vibration and high efficiency driving system was developed and
applied by applying an in-wheel type motor.
Acknowledgement
This work was supported by the Technology Innovation Program (20015242, Development of x-Ev(Electric
Vehicle)Built-in Type Smart Electric Kickboard design product for Last mile mobility market activation)
funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea)“.
References
[1] Choi, K. K., & Cho, J. U. (2020). A Study on Structural Analysis at Front Collision According to the
Shape of Electric Kick Board, J. Korean Soc. Mech.Technol. Vol. 22, No. 3, pp.457-462.
[2] Kim H. Shin, G. (2020), Changes in standing andwalking performance after riding electric kick
scooter,Proceedings of 2020 Fall Conference of ESK.
１３１
AEICP Vol. 5, No. 1
A Study on Challenges of AI Ethics in Next Generation Wireless Networks
Yeongchan Kim1,2, Navin Ranjan2, Sovit Bhandari2, Hoon Kim1,2*
1Department of Electronics Engineering, Incheon National University, Incheon, South Korea
2IoT & Big Data Research Center, Incheon National University, Incheon, South Korea
kyc0288@inu.ac.kr, ranjannavin07@gmail.com, sovit198@gmail.com, hoon@inu.ac.kr
Abstract
It is expected the next-generation wireless networks (6G) will provide services with more advanced AI
(artificial intelligence) functions in various mobile environments. With the recent spread and advancement of
AI, the issues related to AI ethics are of great concern. Therefore, design or implementation of 6G AI
technology must incorporate AI with ethics. This paper surveys a volume of research on design principles and
implementation guidelines for resolving ethical issues in AI applications and also presents reviews on some
technical challenges required for 6G technology
Keywords: 6G, AI, Ethics.
1. Introduction
The next-generation wireless network (6G) is expected to play a full-fledged role as an
infrastructure/platform to provide innovative services in the era of the 4th industrial revolution based on AI
and big data along with meeting the rapidly growing demand for wireless data, and the functions of AI such
as super-giant AI AI technology in the mobile environment, such as advancement, autonomous driving, and
interactive services, is expected to expand the application of daily services.
Accordingly, research is being conducted for network evolution to support and utilize AI functions in 6G,
and the issue of safe and reliable application of AI technology is emerging. In particular, in consideration of
the more advanced and generalized level of AI in 6G, it is expected that the requirements regarding the
corresponding ethical level will arise. The ethical issue of AI has recently become a global issue at home and
abroad, including international organizations and major countries, and in particular, the level of AI ethics will
increase in importance as one of the key considerations in the implementation and utilization of AI functions
around 2030, when 6G will be commercialized. Accordingly, from the early to mid 2020, the 6G pre-research
stage, it is expected that the competition for prior research to develop ethical AI and considerations on AI
ethics will begin in earnest. Therefore, securing technological competitiveness through AI ethics issues,
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１３２
8th ICAEIC-2022
technical requirements analysis, and prior research on key element technologies is very important in 6G and
AI convergence technologies and industries.
With this need, research to establish requirements for AI ethics in the 6G environment based on global AI
ethics principles and guidelines is starting. In other words, the basic principles of global AI ethics have been
proposed for the past three to four years, and detailed standards and design guidelines have been prepared for
them. AI ethical standards and RD requirements that follow them are being announced.
Accordingly, it is necessary to establish AI ethics requirements based on this in 6G, which is standardizing
related things while aiming for a vision such as super-giant AI. In the major 6G use cases, it is necessary to
prepare a learning data set to achieve major AI ethical standards and levels such as fairness, accountability,
and privacy protection, and to conduct prior research on learning model improvement. In addition, there is a
need to conduct research to derive and establish requirements from a network perspective that supports AI
ethics and function in the 6G environment.
This paper introduces the current status of AI evolution, AI ethics principles, and design guidelines to
consider the issues of ethical AI implementation. In addition, AI ethical issues in 6G are derived and major
research issues are presented to solve them. In particular, it includes the latest research trends of 6G AI and
network infra structure and research opportunities of 6G AI related issues.
2. AI Evolution to Strong AI
Weak AI and strong AI are concepts best used in “Chinese Room Argument” proposed by Professor John
Searle in 1980. In general, it is described as strong AI that implements the human mind through complex
information processing, and weak AI that simply imitates a part of human ability or aims for such a task.
Futurist Ray Kurzweil predicted the emergence of strong AI in 2045 and moved it forward to 2030, which is
also an expected time point for 6G to appear.
Figure
1. Quarterly news mentions (of AI and ‘ethics’) 2014-Q3 2018.
Today, weak artificial intelligence is actively used in medical, business, education, and services. Unlike the
strong AI that can have a self like a human, there is a limit in that it can think only in a limited part of human
cognitive ability. Some argue that weak AI is not “true AI” because of its inability to think on its own.
１３３
AEICP Vol. 5, No. 1
The above-mentioned Chinese room argument is an argument to oppose the Turing test, which determines
whether artificial intelligence is intelligent on this problem. It is not necessary to understand itself. In other
words, even if an object claimed to be an artificial intelligence can speak and act like a human on the surface,
it cannot be regarded as having intelligence because it is only the result of thorough calculation and information
processing. Moreover, this is not the end, there is a more fundamental problem. People generally think of free
will human consciousness as something distinct from an algorithmically responsive program, but in fact, the
existence of free will has never been proven.
Simply put, as mentioned above, if you are trying to differentiate between ‘a thing that is truly conscious
like a human’ and ‘something that only appears to be conscious from the outside like a program’, first of all,
what is ‘a thing that is truly conscious like a human’? It should be defined from the beginning, but there is still
no conclusion that most scholars agree on. However, at the moment, most engineers focus on solving practical
problems in front of them rather than philosophical problems, so it can be said that it is not a very important
issue right now. Right now, strong artificial intelligence research is being conducted with the goal of directly
studying how the brain works or simulating it rather than making it based on theory. However, it is clear that
either way it will be difficult to see results in the near future. After all, ‘what is human intellect?’ Unless there
is an answer to the question, the road to true artificial intelligence seems long.
3. AI Ethics Principles and Design Guidelines
There have been lots of works on AI Ethics principles and design guidelines as shown in Table 1. AI has
already provided beneficial tools that are used every day by people around the world. Its continued
development, guided by the following principles,
Table 1. Study on AI Ethics.
Institution
Principles or Guidelines
Aslomar
Aslomar AI principles (’17)
IEEE
Ethically aligned systems (’17)
OECD
Recommendation of the Council on AI (’17)
EU
Ethics guidelines for trustworthy AI (’19)
UNESCO
Preliminary Study on the Ethics of AI (’19)
UNESCO
Recommendation on the Ethics of AI (’19)
IEEE
New standard to address ethical concerns during system design (’21)
１３４
8th ICAEIC-2022
Figure
2. Framework for Trustworthy AI (EU).
Figure
3. AI Ethics Readiness Framework (IEEE).
１３５
AEICP Vol. 5, No. 1
Will offer amazing opportunities to help and empower people in the decades and centuries ahead. The content
of the recommendation suggested by OECD, in which the recommendation aims to realize the advantages AI
brings to society and reduce the risks it entails. It ensures that digital transformations promote human rights
and contribute to the achievement of the Sustainable Development Goals, addressing issues around
transparency, accountability and privacy, with action-oriented policy chapters on data governance, education,
culture, labour, healthcare and the economy.
• Protecting data
The Recommendation calls for action beyond what tech firms and governments are doing to guarantee
individuals more protection by ensuring transparency, agency and control over their personal data. It states
that individuals should all be able to access or even erase records of their personal data. It also includes actions
to improve data protection and an individual’s knowledge of, and right to control, their own data. It also
increases the ability of regulatory bodies around the world to enforce this.
• Banning social scoring and mass surveillance
The Recommendation explicitly bans the use of AI systems for social scoring and mass surveillance. These
types of technologies are very invasive, they infringe on human rights and fundamental freedoms, and they are
used in a broad way. The Recommendation stresses that when developing regulatory frameworks, Member
States should consider that ultimate responsibility and accountability must always lie with humans and that AI
technologies should not be given legal personality themselves.
• Helping to monitor and evaluate
The Recommendation also sets the ground for tools that will assist in its implementation. Ethical Impact
Assessment is intended to help countries and companies developing and deploying AI systems to assess the
impact of those systems on individuals, on society and on the environment. Readiness Assessment
Methodology helps Member States to assess how ready they are in terms of legal and technical infrastructure.
This tool will assist in enhancing the institutional capacity of countries and recommend appropriate measures
to be taken in order to ensure that ethics are implemented in practice. In addition, the Recommendation
encourages Member States to consider adding the role of an independent AI Ethics Officer or some other
mechanism to oversee auditing and continuous monitoring efforts.
• Protecting the environment
The Recommendation emphasises that AI actors should favour data, energy and resource-efficient AI
methods that will help ensure that AI becomes a more prominent tool in the fight against climate change and
on tackling environmental issues. The Recommendation asks governments to assess the direct and indirect
environmental impact throughout the AI system life cycle. This includes its carbon footprint, energy
consumption and the environmental impact of raw material extraction for supporting the manufacturing of AI
technologies. It also aims at reducing the environmental impact of AI systems and data infrastructures. It
incentivizes governments to invest in green tech, and if there are disproportionate negative impact of AI
systems on the environment, the Recommendation instruct that they should not be used.
１３６
8th ICAEIC-2022
Figure
4. Aslomar AI Principles.
Emerging technologies such as AI have proven their immense capacity to deliver for good. However, its
negative impacts that are exacerbating an already divided and unequal world, should be controlled. AI
developments should abide by the rule of law, avoiding harm, and ensuring that when harm happens,
accountability and redressal mechanisms are at hand for those affected.
4. 6G AI Ethical Issues and Challenges
4.1. 6G Vision on AI
6G wireless aims at bridging the “physical world” and the “cyber world”; it is about a new paradigm shift:
from connected people and things (information world) to connected intelligence (intelligent world). 6G
wireless is the technology to deliver AI to everyone, anywhere and at any time.
The first shift in paradigm is about going from an AI intelligence enhanced network, which is the 5G system
today and its future releases, to an AI native communication platform. ITU-T Focus Group(FG)-ML5G
proposes a logical ML pipeline, i.e., a set of logical entities (each with specific functionalities) that can be
combined to form an analytics function. Each functionality in the ML pipeline is defined as a ML Pipeline
node. In addition to supporting the concept of ML pipeline by design, 6G Wireless is expected to incorporate
outer semantic channels, and all intelligence will be connected following a defence-in-depth strategy –
augmented by a zero- trust model–through digital twinning, using B5G/6G wireless, and machine reasoning
will meet ML at the edge.
4.2. 6G AI Ethics Issues
A new IEEE Standard regarding AI ethics has been active. A set of processes by which organizations can
include consideration of ethical values throughout the stages of concept exploration and development is
established. Around the year, 2030 in which 6G begins to be commercialized, this kind of standard on AI ethics
１３７
AEICP Vol. 5, No. 1
is also reflected to 6G.
Management and engineering in transparent communication with selected stakeholders for ethical values
elicitation and prioritization is supported by this standard, involving traceability of ethical values through an
operational concept, value propositions, and value dispositions in the system design. Processes that provide
for traceability of ethical values in the concept of operations, ethical requirements, and ethical riskbased design
are described in the standard. All sizes and types of organizations using their own life cycle models are relevant
to this standard.
4.3. Research Challenges on 6G AI ethics
Based on the survey and investigation above, some opportunities on AI ethics and 6G vision could be
suggested as below.
• 6G-AI ethical requirements
• (Re)configuration of data sets to achieve the required level, labeling for ethical responses
• AI ethics learning model(closed Loop system)
• Conducted research to predict the level of AI implementation supportable in 6G network infrastructure and
establish requirements for AI ethics in 6G environment based on global AI ethics principles and guidelines
• In relation to the AI ethics function, it derives requirements that reflect detailed standards such as publicity,
responsibility, and data management, and reflects AI learning content storage and transmission, routing
functions, etc. from the network function point of view
• Through a comprehensive analysis of 6G-AI application fields and ethical issue candidate fields through
literature research, etc., 6G major use cases including AI ethics functions are derived
• List and catalog major decision-making or judgment issues on related field issues
• Define the characteristics of the data set for the required ethical level such as fairness, personal information
protection, and reliability, and establish conformity standards accordingly
• Define fairness so that race, gender, regional and cultural characteristics are taken into account
• Analyze data fairness and labeling methods for existing data sets, and analyze whether the required ethical
standards have been reached
• Investigate and analyze existing research to achieve the required ethical level through data (re)composition
and suggest improvement plans in a more stable way
• Propose a new learning model for data set reconstruction including reinforcement learning in a post-
processing method that applies a learning model for data set fairness and conduct a performance verification
study
• A study on the dataset labeling technique for the output of AI learning to be a situation in which various
ethical answers or responses are possible.
• In relation to data set (re)composition, conduct research to improve the precision of recognition and response
of ethical situations for multi-mode data set composition that multiplexes the form of learning data such as
visual and auditory data
• Study on the characteristics of a learning model suitable for each use case for 6G-AI ethical requirements
• Establishment of suitability standards according to the characteristics of the learning model, such as
explainability, various ethical responses, answers, and accuracy.
• Investigate existing AI ethical learning models and analyze them in terms of complexity, explainability,
１３８
8th ICAEIC-2022
ethical level of output, etc.
• Research on new learning model techniques to achieve 6G-AI ethical requirements
• Proposal of techniques to increase the reliability of the learning model by applying a one-way learning model
and introducing a closed-loop control system to the flow of performance presentation to improve ethical
response or response performance at the output stage
• Completing a 6G-AI ethical closed loop learning model that includes the Recommander element to overcome
and improve the ethical level limitations in the existing model, and propose an algorithm to determine key
parameters to optimize the ethical level
• Develop a performance evaluation methodology for the (re)construction technique and learning model of the
dataset to achieve 6G-AI ethical standards and levels, and perform performance verification on the proposed
technique
• Whether or not the conformity criteria are met according to the definition of the characteristics of the data
set for the required ethical level such as fairness, personal information protection, and reliability
• Develop a methodology so that race, gender, and regional/cultural characteristics can be considered for
fairness verification
5. Conclusion
6G is expected to appear around 2030, in which more advanced AI seems to be embedded. This study presents
a volume of researches on design principles and implementation guidelines for resolving ethical issues in AI
applications, and some potential research issues are also discussed.
Acknowledgement
This work was supported by Basic Science Research Program through the National Research
Foundation of Korea (NRF) funded by the Ministry of Education (2018R1D1A1B07050418) and X-
mind Corps program of National Research Foundation of Korea(NRF) funded by the Ministry of
Science, ICT(NRF-2017H1D8A1029391).
References
[1] G. Eason, B. Noble, and I. N. Sneddon, “On certain integrals of Lipschitz-Hankel type involving products
of Bessel functions,” Phil. Trans. Roy. Soc. London, vol. A247, pp. 529–551, April 1955.
[2] https://futureoflife.org/2017/08/11/ai-principles/
[3] 6G wireless aims at bridging the “physical world” and the “cyber world”; it is about a new paradigm shift:
from connected people and things (information world) to connected intelligence (intelligent world).
[4] N. Ranjan, S. Bhandari, H. P. Zhao, H. Kim and P. Khan, ”City-Wide Traffic Congestion Prediction based
on CNN, LSTM and Transpose CNN,” IEEE Access, Vol. 8, pp. 81606-81620, Apr. 2020.
１３９
AEICP Vol. 5, No. 1
[5] I. Lana, J. Del Ser, M. Velez, and E. I. Vlahogianni, “Road Traffic Forecasting: Recent Advances and
New Challenges,” IEEE Intelligent Transportation Systems Magazine, vol. 10, no. 2, pp. 93-109, 2018.
[6] X. Ma, Z. Tao, Y. Wang, H. Yu and Y. Wang, “Long short-term memory neural network for traffic speed
prediction using remote microwave sensor data,” Transportation Research Part C: Emerging Technologies,
vol. 54, pp. 187-197, 2015.
[7] W. Wei, H. Wu, H. Ma, “An Auto Encoder and LSTM-Based Traffic Flow Prediction Method,” Sensors,
vol. 19, no. 13, Article ID 2946, 2019.
[8] S. Zhang, Y. Yao, J. Hu, Y. Zhao, S. Li, and J. Hu, “Deep autoencoder neural networks for short-term
traffic congestion prediction of transportation networks”, Sensors, Vol. 19, no. 10, 2229, May 2019. DOI:
10.3390/s19102229
[9] Y. Tian, K. Zhang, J.; Li, X. Lin, and B. Yang, “LSTM-based traffic flow prediction with missing data”,
Neurocomputing, Vol. 318, pp. 297-305, Nov. 2018. DOI: 10.1016/j.neucom.2018.08.067
[10] X. Ma, Z. Dai, Z. He, and Y. Wang, “Learning traffic as images: A deep convolutional neural network for
large-scale transportation network speed prediction,” Sensors, vol. 17, no. 4, 818, Apr. 2017. DOI:
10.3390/s17040818
[11] Z. Lv, J. Xu, K. Zheng, H. Yin, P. Zhao, and X. Zhou, “LC-RNN: A deep learning model for traffic speed
prediction”, in International Joint Conferences on Artificial Intelligence Organization, pp. 3470-3476, Jul.
2018. DOI: 10.24963/ijcai.2018/482
[12] H. Yu, Z. Wu, Y. Wang, and X. MA, “Spatiotemporal recurrent convolutional networks for traffic
prediction in transportation networks”, Sensors, Vol. 17, no. 7, 1501, Jun. 2017. DOI: 10.3390/s17071501
[13] C. Meng, X. Yu, and Y.Liu, “PCNN: Deep convolutional networks for short-term traffic congestion
prediction”,IEEE Transactions on Intelligent Transportation Systems, Vol. 19, no. 11, pp. 3550-3559, Nov.
2018.
[14] S. Bhandari, H. P. Zhao and H. Kim, ”An efficient scheduling scheme for fronthaul load reduction in fog
radio access networks,” China Communications, Vol. 16, No. 11, pp. 146-153, Nov. 2019.
[15] S. Bhandari, H. Kim, N. Ranjan, H. P. Zhao and P. Khan, ”Optimal Cache Resource Allocation Based on
Deep Neural Networks for Fog Radio Access Networks,” Journal of Internet Technology, Vol. 21, No. 4,
pp. 967-975, Jul. 2020.
[16] S. Bhandari, H. P. Zhao, H. Kim, and J. Cioffi, ”An Optimal Cache Resource Allocation in Fog Radio
Access Networks,” Journal of Internet Technology, Vol. 20, No. 7, pp. 2063-2069, Dec. 2019.
[17] S. Bhandari, H. P. Zhao, H. Kim, P. Khan, and S. Ullah, ”Packet Scheduling Using SVM Models in
Wireless Communication Networks,” Journal of Internet Technology, Vol. 20, No. 5, pp. 1505-1512, Sep.
2019.
[18] MIT Moral Machine(https://www.moralmachine.net/)
[19] AI Fairness 360(https://aif360.mybluemix.net/)
[20] https://www2.eecs.berkeley.edu/Pubs/TechRpts/2021/EECS-2021- 170.html
１４０
8th ICAEIC-2022
Demonstration Project of Air Pollution Monitoring System in Urban
Environment using LoRa : LoRa Topology Improvement
Ji-Seong Jeong1, Jeong-Gi Lee2, Chul-Seung Yang3, Gi-won Ku4
1,2,3,4Korea Electronics Technology Institute
{json, jklee, yangcs, giwon9}@keti.re.kr
Abstract
This paper proposes an LoRa-based communication system for transmitting and collecting air pollution
measurement data in an urban environment. Currently, we are conducting a pilot project to demonstrate the
air pollution monitoring system in Gwangju, Korea. The system used a Long-Term Evolution (LTE)
communication module. But to keep using it, we have to pay the communication cost periodically. Therefore,
we have developed LoRa-based communication system to solve this cost problem and are currently conducting
research to install them in urban areas. In order to install the LoRa communication module in a total of 181
places, the area where communication is possible was divided through an experiment. The nodes were
constructed by connecting communicationable points, and a topology and communication algorithm suitable
for use in the Gwangju-si environment in Korea were designed.
Keywords: LoRa, LPWAN, Topology, IoT, UHF
1. Introduction
These days, many people's quality of life has improved and they have become more concerned about
environmental issues. In particular, there is a growing interest in the importance of air quality in parks where
they stay with children for a long time or in industrial complexes where air pollution is suspected. Accordingly,
many studies on environmental monitoring are being conducted, and the IoT-based RF communication system
plays an important role in the center.[1]-[3] In addition, many researchers are demonstrating cost-effective
LoRa communication systems for countries and cities.[4],[5] Gwangju Metropolitan City, Republic of Korea
also uses an RF communication system to collect air pollution measurement data. The communication system
used before was an LTE (Long-Term Evolution)-based system, and to use it, you have to pay a fee periodically.
To solve this cost problem, we developed the LPWAN LoRa module. Fig. 1 shows the location of the LoRa
module installed in the demonstration area: Gwangju, South Korea. In order to install the LoRa communication
module in a total of 181 places, the area where communication is possible was divided through an experiment.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１４１
AEICP Vol. 5, No. 1
The nodes were constructed by connecting communicationable points, and a topology and communication
algorithm suitable for use in the Gwangju-si environment in Korea were designed.
Figure 1.
Sample.
2. LoRa-based Communication System
2.1. Proposed System
The configuration diagram of the proposed system is shown in Fig. 1. First of all, LoRa Node board has
several environmental sensors connected to receive environmental sensor data: GPS Sensor, Temperature and
Humidity sensor, Wind Direction Anemometer, Optical Particle Counter and Air Quality (VOC) Sensor. And
the data is transmitted from LoRa Node to LoRa Host. In other words, 22 LoRa Hosts collect environmental
data of 170 LoRa Nodes. We built the tree network topology, because the network is easy to expand and
manage. The topology also has the advantage of saving the number of communication lines. Data collected by
LoRa Host is sent to the server through the Android host.
Figure 2.
Block diagram of proposed system.
１４２
8th ICAEIC-2022
2.2. LoRa Module
LoRa nodes and hosts have a built-in commercial LoRa Module. The LoRa Module is shown in Fig. 3 was
designed. LoRa Module is designed to be able to set spread spectrum modulation bandwidth, spreading factor
and error correction rate. The benefit of the spread modulation is that each spreading factor is orthogonal.
Thus, multiple transmitted signals can occupy the same channel without interfering.
Figure 3.
Block diagram of LoRa Module.
2.3. Specification of LoRa Module
Table. 1 shows that the LoRa module of the proposed system operates at a center frequency of 915MHz and
a sensitivity of -139dBm. And the maximum output power is 27dBm and the receive power is 13mA. The
Received Signal Strength Indicator (RSSI) is -127dB, which is enough for the LoRa host to receive data from
the LoRa Node. The data transfer rate that one LoRa Node must satisfy is 368bps. The specification of the
adopted LoRa Module is 0.018~37.5kbps, which is suitable for our system.
Table 1. LoRa module specification table.
Specification
Value
Frequency Range
915Mhz
Sensitivity
-139 dBm
Maximum Output Power
27 dBm
Rx current
RSSI
Data Transfer Rate
13 mA
-127 dB
0.018~37.5kbps
１４３
AEICP Vol. 5, No. 1
2.4. Bit Error Rate of Proposed System
We conducted an experiment to verify the Success Rate of Data Transmission of the proposed system. The
data transferred from the LoRa Node to the Server was compared with the data stored in the memory directly
from the LoRa Node. Assuming that the data storage success rate of the LoRa node is 100%, the Success
Rate of Data Transmission of this system is shwon in Table 2. 1024(bit)*170(nodes)*60(sec)
Table 2. Success Rate of Data Transmission.
Specification
Value
Total transmitted data size
783,360 kbyte
wrong data size (including missing data)
781,009 kbyte
Success rate of data transmission
99.7%
3. Conclusion
The current research stage is a pilot test, focusing only on whether the environment data of all LoRa Node
boards is transmitted to the server well. Therefore, the LoRa Node of proposed system, signal transmitter,
outputs the signal at full power. In fact, all RF network system need to be considered energy efficieny, because
as the number of nodes increases, the power consumption increases. Overall, we need to calculate the bit error
rate for all nodes and correct the power so that each transmitter outputs a signal of appropriate strength.
Acknowledgement
This research was supported by a subsidy from The Regional Development Investment Agreement Pilot
Project(B0070510000127) supported by the Ministry of Land, Infrastructure and Transport, Gwangju
Metropolitan City and Gwangsan-gu.
References
[1] Grigoryev, V., Khvorov, I., Raspaev, Y., Aksenov, V., & Shchesniak, A. (2016). Pilot Zone of Urban
Intelligent Transportation System Based on Heterogeneous Wireless Communication Network., Internet
of Things, Smart Spaces, and Next Generation Networks and Systems. Springer, Cham., 479-491.
Sukuvaara, T., Nurmi, P., Hippi, M., Autio, R., Stepanova, D., Eloranta, P., ... & Kauvo, K. (2011, November).
Wireless traffic safety network for incident and weather information, Proceedings of the first ACM
international symposium on Design and analysis of intelligent vehicular networks and applications, 9-14.
Chodorek, A., Chodorek, R. R., & Yastrebov, A. (2021). Weather Sensing in an Urban Environment with the
Use of a UAV and WebRTC-Based Platform, A Pilot Study. Sensors, 21(21), 7113.
Zhou, Q., Zheng, K., Hou, L., Xing, J., & Xu, R. (2019). Design and implementation of open LoRa for
IoT, IEEE Access 7, 100649-100657.
１４４
8th ICAEIC-2022
Paredes, Miryam, et al. Propagation measurements for a LoRa network in an urban environment., (2019), Journal of
Electromagnetic Waves and Applications, 33(15), 2022-2036.
１４５
AEICP Vol. 5, No. 1
Bibliometric Analysis of Blockchain and The Survey Development
Hu Chenxi1,Ren Chen1, and Sang-Joon Lee*
1School of Economics and Business Administration, Hefei University, China
*Interdisciplinary Program of Digital Future Convergence Service, Chonnam National University, Korea
hucx@hfuu.edu.cn1, s-lee@jnu.ac.kr*
Abstract
With the development of blockchain systems, more and more studies have been published in various journals.
Our research using CiteSpace and bibliometric analysis, through the comparative study of domestic and
literature samples, analyze the research status of the blockchain field, summarize the evolutionary path of
domestic and topic research, and explore research hotspots. Our research data investigate 500 articles from
web of science database 2018-20211. Also, according to analysis result of indexing terms by using the
CiteSpace, we find that trust is the most indexing variables in blockchain studies, then we are indexing the
factors of trustworthiness to proposed model of trust to support our survey development base on focus group
interview of trust and its reflective factors in blockchain characteristics.
Keywords: Bibliometric analysis, CiteSpace, Blockchain, Trust, Survey.
1. Introduction
As a new technology, Blockchain will be widely used in various industries in the future. Blockchain related
research fields are complex, and there are few related researches about the quantitative studies of blockchain,
so develop the survey base on the decentralization characteristics is necessary. From [1] the applications of
blockchain in business process management, sharing economy and so on are reviewed, but more of them are
qualitative research, the selection of literature is subjective and the sample size of literature is often small. In
order to more clearly explore and analyze the current status of research in this field, a more comprehensive,
efficient and scientific approach to bibliometrics is needed.
Bibliometric analysis can be quantitative analysis of all aspects of written communication materials. In
2003, Van Raan further deepened the bibliometric analysis method. In his understanding, the title, year,
keywords and abstracts of documents are not only bibliographic information, but also the labels of libraries or
database administrators for the classification and collate on of documents, but also the sociological
characteristics of documents in the process of dissemination [2]. In blockchain system academy filed, it does
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１４６
8th ICAEIC-2022
not have a lot of bibliometric analysis studies because it is still a total new topic. So this research try to use
CiteSpace and bibliometric analysis, through the comparative study of domestic and literature samples, analyze
the research status of the blockchain field. We do related work on the blockchain system bibliometric analysis
study and summarized.
2. Methods and Operational definition
In our research we combine two methods include bibliometric analysis by CiteSpace to find most indexing
variables in blockchain characteristics and develop its reflective factors by qualitative study as FGI (focus
group interview) then confirm its measurement items by quantitative study as EFA (exploratory factors
analysis).
2.1. Bibliometric analysis by CiteSpace
We collected about 500 blockchain article information from web of science from 2018 to 2021, then use
CiteSpace to do bibliometric analysis.
2.2. Proposed model of trust
From bibliometric analysis by CiteSpace, we find that trust is the most indexing variable in blockchain, so
we bring the concept of proposed model of trust [3] which has become one of the most frequently cited
concepts in the study of online trust. The original model of this theory is shown in Figure 1.
Figure 1. Proposed Model of Trust.
2.3. The focus group interview (FGI)
Step 1, From [4], we give the ability, benevolence and integrity topic of blockchain, as the trust of this
blockchain what functions and things can provide will fit to ability, benevolence and integrity. We do FGI with
14 persons from the blockchain company Huobi. We get data from Wechat-video-chatting with them. The
results have been recorded in Record 1. Base on this data we proposed confidentiality, availability and
reputation as the factors of trustworthiness of blockchain system. Step 2, From Ricci [5], we believe the
interview result also can give us a very valuable measurement items for our research variables. For extend the
１４７
AEICP Vol. 5, No. 1
blockchain characteristics measurement items, we summarized all non-repetitive talking substance to fit on
each variable. Then we summarized the discussions and make to the measurement items.
2.4. Operational definition
2.4.1. Availability
All public attributes of any profile must be available at any time. Additionally, it always must be possible
to deliver a message to any user [6]. In [7] argues that the blockchain system is maintained by a network of
nodes, who fulfill approved read/write transactions. Data are sufficiently randomized across the nodes and
replicated to ensure high availability.
2.4.2. Confidentiality
Confidentiality defined as the assurance that data will be disclosed only to authorized individuals or systems
[8]. Blockchain provides a neutral platform where all participants can see the published data. With all the
published information, transactions can be validated by all processing nodes.
2.4.3. Reputation
Reputation is defined as information about an individual’s past performance [9]. In blockchain, to add a
new block, a participant must show a certain amount of currency or reputation, which is lost if that block is
not accepted by consensus [10].
2.4.4. Trust
In a blockchain system, the trust means useful and good reputation. Also, in [11] describe the trust in
blockchain community as the disintermediation trust means no need of the third part just trust the technology
by its decentralization characteristics.
3. Results
We combined our FGI results together with the measurement items. We adopted from literatures then
operate the exploratory factor analysis, for easy to compares the exploratory factor analysis results we gives
each items the number of the specified variable. we did it in www.wjx.cn get 333 respondents to do exploratory
factor analysis. And all items were measured on a five-point Likert-type scale, ranging from 1 (strongly
disagree) to 5 (strongly agree).
The EFA has been held twice to figure out the most valuable measurement items for these four variables
that has decentralized characteristics. For first EFA analysis we find out that A.NO2 has explained Trust as
well, but from our FGI review and the operation definition between Trust and Availability, delete this item is
better than change it to trust. And the C.NO4 has explained the Availability as well, but compared the operation
definition of Availability and Confidentiality delete this item is more meaningful. The first EFA results are
deleted the A.NO2 and C.NO4 to do EFA analysis then get a valuable result that every item explain its variables
very well, the final EFA results are shown in Table 1. From these analysis results we have developed the
measurement items of decentralization characteristics for blockchain area.
１４８
8th ICAEIC-2022
Table 1. EFA Results.
variable
item Availability Confidentiality Reputation Trust
NO1
0.904
0.837
0.906
0.676
NO2
0.878
0.874
0.826
NO3
0.78
0.843
0.839
0.828
NO4
0.884
0.917
0.757
NO5
0.828
0.84
0.892
4. Conclusion
For bibliometric analysis in blockchain system, we integrate previous study of each categories such as
CNKI, EI, Web of Science and Scopus and summarize the contributions of each bibliometric research to help
the future researchers integrate relation literatures and use CiteSpace to position research issues.
We provide scalability value for the follow-up research on blockchain system through the development of
questionnaire and the measurement and verification of latitude. The real blockchain enterprises can refer to the
questionnaire developed in this paper to predict the user behavior.
For future research, we dedicated to integrate the characteristic of the consortium blockchain with private
blockchain and compare the three types blockchain from participants. And base on these three types to find
the different between the socialization behavior in each community. With the development of the blockchain
system and people know more about its technology, the social influence factors will bring to analysis to the
participant intention of blockchain system.
References
[1] Yli-Huumo, J., Ko, D., Choi, S., Park, S., & Smolander, K. (2016). Where is current research on
blockchain technology? —a systematic review. PloS one, 11(10), e0163477.
[2] Van Raan, A. F. (2003). The use of bibliometric analysis in research performance assessment and
monitoring of interdisciplinary scientific developments. Technology Assessment-Theory and Practice,
1(12), 20-29.
[3] Mayer, R. C., et al. (1995). "An Integrative Model of Organizational Trust." Academy of Management
Review 20(3): 709-734.
[4] Choi, K., Wang, Y., & Sparks, B. (2019). Travel app users continued use intentions: it’sa matter of value
and trust. Journal of Travel & Tourism Marketing, 36(1), 131-143.
[5] Ricci, L., Lanfranchi, J. B., Lemetayer, F., Rotonda, C., Guillemin, F., Coste, J., & Spitz, E. (2019).
Qualitative methods used to generate questionnaire items: A systematic review. Qualitative health
research, 29(1), 149-156.
[6] Cutillo, L. A., et al. (2009). Privacy preserving social networking through decentralization. 2009 Sixth
International Conference on Wireless On-Demand Network Systems and Services, IEEE.
[7] Zyskind, G., et al. (2015). Decentralizing Privacy: Using Blockchain to Protect Personal Data. 2015 IEEE
１４９
AEICP Vol. 5, No. 1
Security and Privacy Workshops: 180-184.
[8] Komninos, N., et al. (2014). "Survey in Smart Grid and Smart Home Security: Issues, Challenges and
Countermeasures." IEEE Communications Surveys & Tutorials 16(4): 1933-1954.
[9] Podolny, J. M. (1994). Market uncertainty and the social character of economic exchange. Administrative
science quarterly, 458-483.
[10] Sharples, M., & Domingue, J. (2016, September). The blockchain and kudos: A distributed system for
educational record, reputation and reward. In European Conference on Technology Enhanced Learning
(pp. 490-496). Springer, Cham.
[11] Casey, M. J., & Vigna, P. (2018). In blockchain we trust. Technol. Rev, 121, 10-16.
１５０
8th ICAEIC-2022
Enhanced PID Control of UVC 222nm Sterilization Lamp Combining Machine
Learning based on Independent Edge Computing Device
Eon-Uck Kang 1, Duc-hwan Ahn 2
1Chief Technology Officer; Resco system Laboratory Changwon, Gyeongnam, 51391 Korea
2Professor, Department of Electircal Engineering, Masan University, Gyeongnam, 51217 Korea
eu-kang@resco.kr1, dhahn@masan.ac.kr2
Abstract
Regarding the risk of COVID-19 Recently, the development of UVC 222nm light irradiation equipment to
suppress and prevent infection and sterilize is being actively conducted. In particular, research on ICT
convergence technology devices that can prevent the spread of infection in physical space is required. Most
UVC 250 ~ 275nm (λ) wavelength irradiation sterilization recommends sterilization of objects, but it is known
that sterilization of living body and human body is possible only through a lamp with a wavelength of 222nm
(λ). This paper designs and implements a UV system that can irradiate not only objects but also living bodies.
In particular, the proposed irradiation system destroys the DNA Thymine (T) base in viruses and bacteria,
thereby reducing the viability of viruses, reducing infection and preventing spread, and sterilizing the human
body. The amount of energy emitted by the UVC 222nm lamp (mJ) to the human body is limited due to the risk
of irradiation distance, radiation intensity, time domain, and excimer temperature (Temperature) for safety
reasons. Analyze and provide variable dependencies. The amount of energy emitted (mJ) that UVC 222nm
lamps irradiate to the human body is limited by the risk of irradiation distance, radiation intensity, time domain,
and excimer temperature (Temperature) for safety reasons. Dependency is analyzed and provided. If the
irradiation distance is too close or the output amount is increased, the sterilization power is high, but it causes
damage to the human body. Although output control by applying independent PID technology is possible, PID
output control and motion distance control was predicted and presented by machine learning through the
optimal algorithm of UV LAMP device combined with AI technology to reduce “deterministic loss” in PID.
Through the proposed system, it is possible to reduce the risk of printing for human body irradiation, and by
securing the safety of the printing system, sterilization and prevention can be expected in the coexistence of
the virus in the era of with corona.
Keywords: Artificial Intellegenc, UV 222 Excimer Lamp, Irradiation, Iintensity, DNA, PID Control, COVID-
19, Machine Learning.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１５１
AEICP Vol. 5, No. 1
1. Introduction
Many social changes have occurred in the two years since the outbreak of COVID-19. IT convergence
technologies have evolved together in the era of Ontek, which has changed from face-to-face culture to non-
face-to-face culture and is connected without contact. Among them, non-face-to-face remote treatment,
education, telecommuting, and unmanned automatic vending (POS, Bending Machine) were particularly active.
The UVC 222 light irradiation method used in the sterilization device of the COVID-19 virus is an autonomous
irradiation method, not a direct irradiation method. devise and explain the technique.
In order to conduct light irradiation, the International Commission on Illumination (CIE) previously defined
the UV-C biological spectrum band, which is a short-wavelength ultraviolet region, as 100-280 nm. The
wavelength of 100nm-180nm, called vacuum ultraviolet, is attenuated in the air, so actual exposure of the skin
and eyes does not occur, and ACGIH and ICNIRP limit exposure only at 180nm-280nm. These guidelines
were officially adopted and used as guidelines in 1973 as the threshold limit value (TLV) of the eye and skin.
Also, in the early 1990s, the American Illuminating Engineering Society adopted ACGIH ® TLV ® as the
first photobiological safety standard for lamps. The CIE then adopted the same emission limits for lamp safety
standards in 2002, and in 2006 the International Electrotechnical Commission (IEC) adopted the CIE standard
as a joint logo standard, IEC 62471:2006. By applying this standard, the PID output control was
enhanced with a TLV threshold of 3mJ/cm2 for UVC222 light irradiation for limited radiation dose.
2. Back ground
Figure 1 shows the envelope operating spectra formulated for UV-C and UV-B in the 1970s (absolute). The
histogram of the threshold data, along with the uncertainty, shows the wide bandwidth of some data that had
to be adjusted for a set of spectrally resolved limit and hazard functions S(λ). (Sliney, 1972 2).
Figure. 1.1
Figure. 1.2
Figure. 1.3
Spectral band weights
In order to apply a UV limit directed to the spectral distribution for a specific lamp (UV222), it is necessary
to apply the envelope action spectrum S (λ ) weight in Figure 2 and daily limit 23 mJ cm-2 (for S ( λ ) = 222
nm) 0.013) at 222 nm is required.
= · () · 3.
(1)
In Fig.2 ,2020 ACGIH ® Relative Spectral Effect Function S (λ) for UVR (solid line). At wavelengths
１５２
8th ICAEIC-2022
from 180 to 400 nm, S (λ) has not changed for several decades. The 1987 CIE erythema action spectrum of
UV (dotted line) is close to the ACGIH ® S (λ) function at 250-400 nm. The CIE erythema action spectrum
was specified only for wavelengths between 250 and 400 nm. CIE non-melanoma skin cancer (NMSC)
action spectra are shown for UV wavelengths greater than dotted 250 nm. The spectrum of action for both
erythema and NMSC decreases sharply at UV-C wavelengths below 300 nm. ACGIH's Relative Spectral
Effect® Laser TLV ® is shown with dashed lines for comparison, adopted in 2021
Eeff = effective irradiance relative to a monochromatic source at 270 nm [W cm−2].
Eλ = spectral irradiance at a center wavelength [W cm−2 nm−1].
S(λ) = relative spectral effectiveness at the center wavelength [unitless].
Δλ = bandwidth around the center wavelength [nm].
and the above effective irradiance Eeff must be integrated over time by time-weighted averaging (TWA)
to remain below the daily exposure limit Heff-TLV = 3 mJ cm−2 (30 J cm−2)
eff–TLV =
ef f·
≤ 3mJ·cm−2 = 30J·m-2
H eff-TLV = Heff-TLV = 3 mJ cm−2 (30 J cm−2)
Therefore, in Figure 3, the exposure limit is displayed as a threshold for the eyes and skin based on the
number of exposures and the output of the constraint.
2.1. UV-C 222nm KrCl Excimer Lamp Exitation
For the KrCl excimer lamp for testing, exitation is performed at High Voltage DC 6kV. And the output
can be controlled by the high voltage frequency and duty ratio of the 222 excimer lamp in the figure..
Figure. 2
For the KrCl excimer lamp for testing, exitation is performed at High Voltage DC 6kV. And the output can
be controlled by the high voltage frequency and duty ratio of the 222 excimer lamp in the figure. The output
(intensity) of the above aximerram is output by two independent variables of frequency and duty cycle, and
another distance control output is required to move the distance. For optimal output, the output values for three
variables are designed to be optimally controlled according to the wavelength intensity data value of the sensor
being measured. For the KrCl excimer lamp for testing, exitation is performed at High Voltage DC 6kV. And
the output can be controlled by the high voltage frequency and duty ratio of the 222 excimer lamp in the figure.
１５３
AEICP Vol. 5, No. 1
2.2. UV-C Intensity Sensor and Distance sensor
The sensor with UV222nm wavelength is GFUV-T10GD-L(Sharp Japan), and UV output (Intensity)
according to the sensing value must be realized by satisfying two output variables along with the control
independent variables Frequency and Duty Cycle.In addition, the motion PWM output for future use should
also be considered. In addition, an ultrasonic sensor capable of measuring a displacement of 800 mm was
used to satisfy the distance value by the distance sensor. The figure below shows the values of the UV sensor
with guaranteed linearity.
Table1. Field of View(FOV Power analog Output Voltage) .
UV Source
UV
Power(mW/Cm2) Vout(V)
0
0.0
1
0.5
2
1.0
3
1.5
4
2.0
5
2.5
6
3.0
7
3.5
8
4.0
9
4.5
10
5.0
2.3. Structure of Classical PID and Enhanced PID on Control System
The theoretical axiom was first described by Maxwell in 1868 in Minorsky in his Seminal paper “On
Governors”. PID controllers have been widely used in closed loop systems for a long time and have their origins
in speed control in the 19th century. Since then, over time, it has been utilized throughout the industry and has
been proposed as a number of advanced control algorithms. Most industrial controllers can implement the PID
algorithm together with PID and are used easily and conveniently. The PID controller can converge the target
value by receiving feedback of the error e(t) of the continuous sensor and reflecting the error along with the
calculation. The desired value is u(t), and e(t) of the system is expressed as proportion(P), integral(I) and
derivative(D) Term. here mathematically
there is:
u(t) = ()+ ()+ ()
,
１５４
8th ICAEIC-2022
Fig .3 The structure of a classical PID Controller
Fig .4 Loop tuning in genetic
algorithm(GA) for enhanced PID Controller
Fig. 3 shows the feedback structure for the error control of the gain loss (KP, KI, KD) as a typical closed-
loop in the PID system. Fig. 4 s on the right shows the structure of PID for machine learning. The difference
in Fig. 4 is a proposal to converge the gain to 0 with a machine learning algorithm to minimize the gain loss.
indicated. There is a motion control linear rail that can control the distance between the lamp of the upper 20W
UVC 222nm wavelength and the lamp, and the photodiode sensor is designed to measure the intensity of the
UV output. It is also possible to measure the distance through the ultrasonic distance sensor.
Additional peripheral devices for the experiment include a high voltage (HV) power supply and
microcontroller, and a laptop for measurement and monitoring is constructed as a structure of AL6063S T5
heat-treated aluminum profile frame.
Fig .5
１５５
AEICP Vol. 5, No. 1
2.4. Enhanced PID Controller(EPID) Proposed by Genetic Algorithm
3. Result
In Table 2., data of output value convergence optimized through iterative processor and EPID according
to the value of the time function are presented. It can be seen that the average value performed in comparison
through EPID is clearly different.
Table 2. Performance indices of Tuning Optimization PID and EPID(by GA).
Delay(t)
Iterative Method
Optimized by PID
Optimized by EPID
0.025
3.471926
0.089213
0.038659
0.05
6.567769
0.318439
0.155625
0.075
9.507315
0.642704
0.350289
0.1
12.34746
1.035597
0.622276
0.25
28.6514
4.507248
3.854827
0.5
56.44149
16.79849
15.15463
0.75
86.14606
42.31687
33.54903
Initialize population
Measure fitness
Selection
Mutation
Crossover
Optimum
Not an
optimization
EPID Genetic Algoritm Steps
Step 1. Initialize the parameter with a
population (Gain loss error) of
randomsolutions, such as crossover
rate, mutation rate, number of clusters,
and number of generations. Determine
the coding mode.
Step 2. Compute and evaluate the value of the
fitness Value and function.
Step 3. Proceed with crossover and mutation
operation and make up the new
cluster.
Step 4. Repeat Step 2, till the best value is
obtained.
Step 5. Gain Loss Optimization(by EPID)
Step 6. Optimum
１５６
8th ICAEIC-2022
4. Conclusion
It was observed that overshoot and gain loss (Kp, Ki, Kd) can be reduced by more than 26% through machine
learning through GA algorithm for the amount of light irradiated to the human body by applying the typical PID
system and EPID. With GA machine learning, it can be applied to PID optimization with more flexible scalability
and precision, and it is possible to determine and predict in the small embedded-based control field that can
correct errors through EPID (Enhanced PID). Although there is a problem in minimizing the error criterion and
convergence to the optimal target value, it is expected that it will help to find the optimal result by supplementing
the tuning method by optimizing the time delay function, and stabilization of the small PID control system in the
future do.
Acknowledgement
This thesis was researched with the support of the Gyeongsangnam-do Regional Innovation Platform
Smart Manufacturing ICT Technology Joint Research and Development Center, and I would like to pay
tribute to all those who are working hard to sterilize and prevent the COVID-19 virus.
References
[1] A. Jayachitra and R. Vinodha. “Genetic Algorithm Based PID Controller Tuning Approach for
Continuous Stirred Tank Reactor.” Volume 2014, Article ID 791230, 8 pages
[2] K. H. Gharib O. S. Ebrahim H. K. Temraz M. A. Awadalla “Application of the genetic algorithm to design
an optimal PID controller for the AVR system” Proceedings of the 6th ICEENG Conference, 27-29 May,
2008
[3] P.B. de Moura Oliveeira and J. Boaventura Cunha. “Blending Artificial Intelligence into PID Controller
Design IFAC-Papers ONline 49-6(2016)
[4] Voratas Kachitvichyanukul, Comparison of Three Evolutionary Algorithms: GA, PSO, and DE” Vol 11,
No 3, September 2012, pp.215-223 http://dx.doi.org/10.7232/iems.2012.11.3.215
[5] Stanislava Soro. (2020) TinyML for Ubiquitous Edge AI Approved for Public Release; Distribution
Unlimited. Public Release Case Number 20-2709
[6] A. A. Salem Prof. Dr. M. Mustafa Dr. M. E. Ammar (2014) Tuning PID Controllers Using Artificial
Intelligence Techniques 9th International Conference on Electrical Engineering ICEENG 2014
[7] Bong-Hyun Back, Il-Kyu Ha (2021) Development of artificial intelligence-based air pollution analysis
and prediction system using local environmental variables Journal of the Korea Institute of Information
and Communication Engineering Vol. 25, No. 1: 8~19, Jan. 2021
[8] Anna Miller and Szymon Walczak (2020) Maritime Autonomous Surface Ship’s Path Approximation
Using Bézier Curves, Symmetry 2020, 12, 1704
[9] Kikun Park, Sunghyun Sim, Hyerim Bae (2021) Vessel estimated time of arrival prediction system based
on a path-finding algorithm, Maritime Transport Research 2 (2021) 100012
１５７
AEICP Vol. 5, No. 1
[10] Ziegler J. G. and Nichols N. B. (1942), Optimum Settings for Automatic Controllers, Transaction of the
ASME, pp. 759-768.
[11] O’Dwyer A., (2006), Handbook of PI and PID Controller Tuning Rules (2nd Edition), Imperial College
Press, ISBN 1-86094-622-4.
[12] P. V. Gopi Krishna Rao, M. V. Subramanyam, and K. Satyaprasad,Model based Tuning of PID
Controller”, Journal of Control & Instrumentation, Vol. 4, Issue 1.
[13] A. Visioli, “Tuning of PID controllers with fuzzy logic”, IEE Proceedings on Control Theory and
Applications, vol. 148, issue 1, pp. 1-8, 2001.
[14] Khedr, S.F.M.; Ammar, M.E. and Hassan, M.A.M. “Multi objective genetic algorithm controller’s Tuning
for non-linear automatic voltage regulator” Proceedings of the 2013 International Conference on Control,
Decision and Information Technologies (CoDIT), pp. 857 – 863, May 2013.
[15] S. F. Kheder, “Genetic Algorithm Based Controller’s Tuning for Linear and Nonlinear Automatic Voltage
Regulator in Electrical Power System”, M.Sc. thesis, Faculty of Engineering, Cairo University 2013.
[16] Jarboui, B., Damak, N., Sirry, P., and Rebai, A. (2008), A combinatorial particle swarm optimization for
solving multi-mode resource-constrained project scheduling problems, Applied Mathematics and
Computation, 195(1), 299-308.
[17] Kachitvichyanukul, V. and Sitthitham, S. (2011), A twostage genetic algorithm for multi-objective job
shop scheduling problems, Journal of Intelligent Manufacturing, 22(3), 355-365.
[18] Lova, A., Tormos, P., Cervantes, M., and Barber, F. (2009), An efficient hybrid genetic algorithm for
scheduling projects with resource constraints and multiple execution modes, International Journal of
Production Economics, 117(2), 302-316.
[19] Marinakis, Y. and Marinaki, M. (2010), A hybrid genetic-particle swarm optimization algorithm for the
vehicle routing problem, Expert Systems with Applications, 37(2), 1446-1455.
[20] Wisittipanich, W. and Kachitvichyanukul, V. (2012), Two enhanced differential evolution algorithms for
job shop scheduling problems, International Journal of Production Research, 50(10), 2757-2773.
[21] David H. Sliney, Bruce E. Stuck. (2021) A Need to Revise Human Exposure Limits for Ultraviolet UV-
C Radiation. Wiley Online Library
[22] Ulrich Kogelschatz, (2003) Dielectric-barrier Discharges: Their History, Discharge Physics, and
Industrial Applications. Plasma Chemistry and Plasma Processing, Vol. 23, No. 1, March 2003
[23] David Florez, Member, IEEE, Rafael Diez, Senior Member, IEEE, and Hubert Piquet (2016) Optimizing
the Operation of DBD Excilamps (2016) IEEE Transactions on Plasma Science, vol. 44 (n° 7). pp. 1160-
1168. ISSN 0093-3813
[24] N. Thomas, Dr. P. Poongodi, “Position Control of DC Motor Using Genetic Algorithm Based PID Controller”,
Proceedings of the World Congress on Engineering 2009 Vol II, London, U.K., 2009.
１５８
8th ICAEIC-2022
Potential Scholarship Recipients with Simple Additive Weighting and Technique
for Order of Preference by Similarity to Ideal Solution
Qurrotul Aini1, Nurbojatmiko2, and Mega Ayu Silvianingsih3
Dept. of Information Systems, Faculty Science and Technology
UIN Syarif Hidayatullah Jakarta
qurrotul.aini@uinjkt.ac.id1, nurbojatmiko@uinjkt.ac.id2, silvianingsihmega@gmail.com3
Abstract
Awarding scholarships at MTs Bina Insan Kamil Foundation (YABIKA) aims to motivate students to always
study hard and ease parental costs. Nowadays, the awarding of scholarships is carried out a discussion among
teachers concerned to recommend students who are entitled to an outstanding scholarship, not yet using a
measurable method, and scholarship acceptance is undertaken by holding a meeting for one week or more.
The decision is considered not objective and not right on target. Therefore, the authors proposed both methods,
the Simple Additive Weighting (SAW) method and the Technique for Order of Preference by Similarity to Ideal
Solution (TOPSIS) to solve the problem. The purpose of this study is to investigate the selection of the
scholarship recipients with applying SAW and TOPSIS methods with sensitivity tests. The results show that
the decision support system can help the school determine scholarship recipients and the performance of the
sensitivity test both shows that SAW is better because it had changed 6.251%, than TOPSIS only -0.028%.
Keywords: Scholarship recipients, simple additive weighting, SAW, TOPSIS, decision making, MCDM.
1. Introduction
MTs Bina Insan Kamil Foundation (YABIKA) is one of the foundations that provides merit scholarships
for its students. The purpose of this scholarship is to motivate students to always study hard and to ease the
burden on their parents. The scholarship program provided is a merit scholarship in the form of a waiver of
monthly fees for the next three months. The scholarship should be awarded to deserving students. The
scholarship will only be awarded to 3 students selected from 6 classes. Only students ranked 1 to 3 are eligible
to take part in the selection. The selection process is still manual that is by using student report cards as a
reference and by holding meetings with each class teacher and other subject teachers to recommend each other.
The assessment of scholarship recipients is carried out for one week and is discussed together with the teachers.
If it has not been determined who the recipients of the scholarship will be, it will be continued the next day.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１５９
AEICP Vol. 5, No. 1
Determination of scholarship recipients using the SAW method has proven useful in the selection process
with the following criteria: GPA, liveliness, parents' income, student achievement, and parents’ responsibility
[1]. Selection of merit scholarship recipients uses the Analytical Hierarchy Process (AHP) and Simple Additive
Weighting (SAW) models. The decision support system using the AHP and SAW methods is very important
in determining who is entitled to receive the scholarship, with the criteria for the number of parents’
responsibility, parents' income, written math test scores, written English test scores, written general knowledge
test scores, written civics test scores, interview test scores and average high school report cards [2].
TOPSIS and SAW methods have been used in the Selection of Tourist Destinations in Indonesia with 5
criteria, namely location, cost, transportation, distance, and time of visit. These two methods can also be used
to complete the selection of a number of alternatives based on several predetermined criteria. The results of
the comparison of TOPSIS and SAW methods showed that the SAW method was better than TOPSIS where
the SAW method got a value of 0.98 while the TOPSIS method was 0.62. This method can be used to complete
the selection of a number of alternatives based on several predetermined criteria [3]. Other research revealed
that obtaining criteria for scholarship recipients that could be implemented with TOPSIS [4]. The aims of
current study is to investigate selection of the scholarship recipients with applying SAW and TOPSIS methods
with sensitivity test.
2. Fuzzy Multi-Attribute Decision Making
Fuzzy Multi-Attribute Decision Making could be a strategy utilized to discover the finest elective among a
arrangement of options with certain criteria. The substance of FMADM is to decide the weight of value of
each attribute, then perform a classification process to select the given alternative. There are three approaches
to discover trait weights, specifically: subjective, objective and a combination of subjective and objective
approaches. Each approach has focal points and impediments. Within the subjective approach, the weight is
decided based on the subjectivity of the choice producer, therefore, the a few variables within the positioning
of options can be decided freely [5].
2.1. Simple Additive Weighting (SAW)
The basic concept of SAW is to obtain the value from the weighted sum of the performance ratings for each
alternative on all attributes. The SAW requires the process of normalizing the decision matrix (X) to a scale
that can be compared with all existing alternative ratings [6]. The SAW steps are as follows: [6, 7]
a) Determine the criteria that will be used as a reference for decision making ().
b) Determine the weight of each criteria.
c) Determine the suitability rating of each alternative () on each criteria.
d) Normalizing decision matrix
=
()
()
(1)
１６０
8th ICAEIC-2022
with is rating value of normalized performance; express the maximum value of each row and
column, while express the minimum value of each row and column, and is decision matrix of each
row and column.
e) Multiplying weight of criteria by normal value of each criteria of every alternatives and Sum up the values
created in the last step and make the point of each alternatives.
=
(2)
with states final value of relative, states predefined weight. Value of larger indicates that the
alternative is preferred.
f) Choose the alternative that has the maximum point.
2.2. Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS)
TOPSIS is a method for solving multi-criteria decision-making problems based on the concept the best
chosen alternative not only has the shortest distance from a positive ideal solution but also has the longest
distance from a negative ideal solution [8]. The alternatives that have been ranked are then used as a reference
for decision makers to choose the best desired solution, in general the TOPSIS procedure follows the following
steps [6].
a) Calculate normalized decision matrix =
(3)
with i = 1, 2, …, m and j = 1, 2, …, n.
b) Calculate a weighted normalized decision matrix = . (4)
with is weighted normalized matrix row i, column j.
c) Determine the positive ideal solution matrix and a negative ideal solution matrix
= (,, … , ) (5)
= (,, … , ) (6)
where is maximum when j is benefit attribute and minimum when j is cost attribute. Meanwhile,
is minimum when j is benefit attribute and maximum when j is cost attribute.
d) Determine the distance between the values of each alternative with the positive ideal solution
matrix and the negative ideal solution matrix.
= (
) (7)
with is the alternative distance to the positive ideal solution.
= (
) (8)
with is the alternative distance to the negative ideal solution.
１６１
AEICP Vol. 5, No. 1
e) Set a preference value for each alternative and select the maximum preference value.
=
(9)
with express the proximity of each alternative to the ideal solution. A higher value indicates that the
alternative is preferred.
2.3. Related Work
Many previous studies have used SAW to determine scholarship recipients. Ref. [9] revealed that the SAW
method was able to determine the ranking and the best students who received scholarships. However, this
study has not shown the results of the performance of SAW itself. Meanwhile, SAW and TOPSIS methods
were applied in the ranking of alternative recipients of rice aid from the government. This research builds a
decision support system with alternative results getting the highest score of 0.900 on SAW and 0.877 with
TOPSIS. System testing with black box testing and the level of respondents' satisfaction with the system
reached 86.67% [10]. Other approaches, such as the entropy method used for objective weighting of the criteria
were combined with TOPSIS to search for scholarship recipients based on 10 criteria in junior high school.
The output of this system is the ranking of students based on their preference scores [11]. SAW has the ability
to make a more precise assessment because it is based on predetermined criteria and preference weights,
besides that SAW can also select the best alternative from a number of alternatives because of the ranking
process after determining the weight for each attribute. TOPSIS is one method that can help the optimal
decision-making process to solve practical decision problems. This is because the concept is simple and easy
to understand, computationally efficient and has the ability to regulate the relative performance of decision
alternatives in a simple mathematical form [3].
3. Research Method
The stages of this research are composed of two parts, namely: data collection and decision-making
modeling. Data collection consisted of interviews with the school principal and observations of the process of
determining the award of scholarships. Based on interview results, authors determine five criteria, i.e.:
academic rank, average score, absence list, parent’s income, and extracurricular activities (EA). Each criteria
is weighted. The weights are made in order of importance on fuzzy numbers, namely: very low (SR), low (R),
moderate (C), high (T) and very high (ST), then fuzzy numbers can be converted to crisp numbers (Fig. 1).
Also, authors determine the list of potential students for scholarship.
Figure 1.
The weight chart [6].
１６２
8th ICAEIC-2022
The sample data of prospective scholarship recipients from grade 7 consists of 9 students. The first
prospective student is assumed to be first student A1, second student A2, third student A3, and so on. In decision
making modeling, the current research applies SAW and TOPSIS. The results visualize the rank of students
that recommended for scholarship recipients. To compare both method’s performance, the authors conduct the
sensitivity test.
4. Results
Based on the research method that has been described, the determination of scholarship recipients uses SAW
and TOPSIS which the results are as follows:
4.1. SAW
The first step is to determine the criteria used to select scholarship recipients. Next, to determine weight of
each criteria based on Fig. 1. There are 5 criteria based on interview and its weight are shown in Table 1. The
criteria that have been determined from MTs YABIKA as consideration for receiving scholarships (Table 2
and 3). Data on scholarship recipients are assumed to be the first student = A1, the second student = A2, the
third student = A3, and so on. Therefore, the candidate students are listed in Table 4 and expressed at matrix
X.
Table 1. Criteria and its weight of recipients scholarship.
Criteria
Description
Attribute
Weight
Weight value
C
1
Academic ranking
Benefit
Very high
10
C
2
Average score of semester
Benefit
High
7.5
C
3
Absence list
Cost
Low
2.5
C
4
parent’s income
Cost
High
7.5
C
5
extracurricular activities
Benefit
Moderate
5
Table 2. Criteria of academic ranking and average score weight
Academic ranking
Weight
Weight value
Average score
Weight
Weight value
1
Very high
10
< 7.0
Low
2.5
2
High
7.5
7.0 8.0
Moderate
5
3
Moderate
5
8. 0 < 9.0
High
7.5
> 9.0
Very high
10
１６３
AEICP Vol. 5, No. 1
Table 3. Criteria of absence, parent’s income, and extracurricular weights
Absence
Weight
Weight
value
Parent’s income
(million IDR)
Weight
Weight
value
EA
Weight
Weight
value
0
High
7.5
< 2.0
Very high
10
0
Very low
0
1-2
Moderate
5
2.0 3.0
Tinggi
7.5
1
Low
2.5
3-4
Low
2.5
3.0 < 4.0
Moderate
5
2
Moderat
e
5
5
Very low
0
> 4.0
Low
2.5
3
High
7.5
4
Very
high
10
Table 4. Match rating alternatives
Alternatives
Criteria
=
10 7.5 7.5
7.5 5 2.5
5 5 5
5
2.5
5
7.5
2.5
7.5
10 10 5
7.5 7.5 7.5
5 5 5
2.5
5
7.5
5
2.5
7.5
10 10 2.5
7.5 7.5 5
5 5 5
7.5
5
7.5
2.5
5
7.5
C
1
C
2
C
3
C
4
C
5
A
1
10
7.5
7.5
5
7.5
A
2
7.5
5
2.5
2.5
5
A
3
5
5
5
5
7.5
A
4
10
10
5
2.5
5
A
5
7.5
7.5
7.5
5
2.5
A
6
5
5
5
7.5
7.5
A
7
10
10
2.5
7.5
2.5
A
8
7.5
7.5
5
5
5
A
9
5
5
5
7.5
7.5
SAW requires the process of normalizing the decision matrix (X) to a scale that can be compared with all
alternative ratings which can be described as follows:
For example, the value of academic ranking is included in the benefits, then calculate until . This
calculation is done for the next 4 criteria. Eample: :
(;.;)=
= 1
Then the normalized matrix is created, =
1 0.75 0.33
0.75 0.5 1
0.5 0.5 0.5
0.5
1
0.5
1
0.67
1
1 1 0.5
0.75 0.75 0.33
0.5 0.5 0.5
1
0.5
0.33
0.67
0.33
1
1 1 1
0.75 0.75 0.5
0.5 0.5 0.5
0.33
0.5
0.33
0.33
0.67
1
Based on weight of criteria Table 1, = {10 ; 7,5 ; 2,5 ; 7,5 ; 5}, some examples of the results are
described as follows and the complete recommendation is shown in Table 5.
= (1 × 10) + (0.75 × 7.5) + (0.33 × 2.5) + (0.5 × 7.5) + (1 × 5) = 25.208
= (0.75 ×10) + (0.5 × 7.5) + (1 × 2.5) + (1 × 7.5) + (0.67 × 7.5) = 24.583
１６４
8th ICAEIC-2022
Table 5. Recommendation of scholarship recipients with SAW
Alternatives
Result of calculation
Rank
A
1
25.208
2
A
2
24.583
3
A
3
18.75
7
A
4
29.583
1
A
5
19.375
6
A
6
17.5
9
A
7
24.167
4
A
8
21.458
5
A
9
17.5
8
4.2. TOPSIS
To make a normalized decision matrix, refer to Table 5. The calculation is as follows:
||=10+ 7.5+ 5+10+ 7.5+ 5+10+ 7.5+ 5=23.31844
=
||=
.= 0.42885, =
||=.
.= 0.32163. This calculation is carried out up to
|| and . Therefore, the complete matrix R as follows.
=
0.42885 0.34641 0.47434
0.32163 0.23094 0.15811
0.21442 0.23094 0.31622
0.29814
0.14907
0.29814
0.42426
0.28284
0.42426
0.42885 0.46188 0.31622
0.32163 0.34641 0.47434
0.21442 0.23094 0.31622
0.14907
0.29814
0.44721
0.28284
0.14142
0.42426
0.42885 0.46188 0.15811
0.32163 0.34641 0.31622
0.21442 0.23094 0.31622
0.44721
0.29814
0.44721
0.14142
0.28284
0.42426
The stage of the ranking process uses the Table 1 as the weight value, = {10 ; 7,5 ;2,5 ; 7,5 ; 5}. The
result example of ranking process as follows.
= (0.42885 ×10); (0.34641 × 7.5); (0.47434 × 2.5); (0.29814 × 7.5); (0.42426 × 5)
= (0.32163 ×10); (0.23094 × 7.5); (0.15811 × 2.5); (0.14907 × 7.5); (0.282846 × 5)
The complete matrix, =
4.2885 2.59808 1.18585
3.21634 1.73205 0.39528
2.14423 1.73205 0.79055
2.23605
1.11803
2.23605
2.12132
1.41421
2.12132
4.2885 3.4641 0.79055
3.21634 2.59808 1.18585
2.14423 1.73205 0.79055
1.11803
2.23605
3.35408
1.41421
0.70711
2.12132
4.2885 3.4641 0.39528
3.21634 2.59808 0.79055
2.14423 1.73205 0.79055
3.35408
2.23605
3.35408
0.70711
1.41421
2.12132
Positive and negative ideal solution: =max(4.2885; 3.4641; 0.39528; 1.11803; 2.12132); =
min(2.14423; 1.73205; 1.18585; 3.35408; 0.70711)
１６５
AEICP Vol. 5, No. 1
The distance between the weighted value of each alternative to the positive ideal solution:
=(4.2885 4.2885)+ (2.59808 3.4641)+ (1.18585 0.39528)+
(2.23605 1.11803)+ (2.12132 2.12132)= 1.62017
The distance between the weighted value of each alternative to the negative ideal solution:
=(4.2885 2.14423)+ (2.59808 1.73205)+ (1.18585 1.18585)+
(2.23605 3.35408)+ (2.121320.7011)= 2.93221
Preference value for each alternative: =.
..= 0.6441; =.
..= 0.5556 …
Table 6. Recommendation of scholarship recipients with TOPSIS
Alternatives
Result of calculation
Rank
A
1
0.6441
2
A
2
0.5556
3
A
3
0.3808
7
A
4
0.8179
1
A
5
0.4247
6
A
6
0.2908
9
A
7
0.5201
4
A
8
0.4999
5
A
9
0.2913
8
To carry out the sensitivity test, the weight on each criteria is increased by 0.5 and 1, starting from C1, the
weight becomes 10.5, 7.5, 2.5, 7.5, 5 (Table 7). Also add 1 to C1 add 1 and Table 9 for C2 add 0.5. The
final rank change from sensitivity test is shown in Table 10.
Table 7. Sensitivity test C1 add 0.5 Table 8. Sensitivity test C1 add 1 Table 9. Sensitivity test C2 add 0.5
Alternatives
SAW
TOPSIS
Alternatives
SAW
TOPSIS
Alternatives
SAW
TOPSIS
V
1
25.708
0.65
V
1
26.208
0.656
V
1
25.583
0.641
V
2
24.958
0.555
V
2
25.333
0.554
V
2
24.833
0.545
V
3
10
0.375
V
3
19.25
0.369
V
3
19
0.376
V
4
30.083
0.821
V
4
30.583
0.823
V
4
30.083
0.82
V
5
19.75
0.427
V
5
20.125
0.429
V
5
20.125
0.427
V
6
17.75
0.288
V
6
18
0.284
V
6
19.75
0.288
V
7
24.667
0.527
V
7
25.167
0.534
V
7
17.75
0.526
V
8
21.833
0.5
V
8
22.208
0.5
V
8
24.667
0.5
V
9
17.75
0.288
V
9
18
0.284
V
9
21.833
0.288
MAX
30.083
0.821
MAX
30.083
0.823
MAX
30.083
0.82
Changes (%)
0.5
0.003
Changes (%)
1
0.005
Changes (%)
0.5
0.002
１６６
8th ICAEIC-2022
Table 10. Result of ranking change sensitivity test
Criteria
SAW
TOPSIS
C
1
+ 0.5
0.5%
0.003%
C
1
+ 1
1%
0.005%
C
2
+ 0.5
0.5%
0.002%
C
2
+ 1
1%
0.005%
C
3
+ 0.5
0.25%
-0.007%
C
3
+ 1
0.5%
-0.015%
C
4
+ 0.5
0.5%
0.004%
C
4
+ 1
1%
0.007%
C
5
+ 0.5
0.334%
-0.011%
C
5
+ 1
0.667%
-0.021%
Total
6.251%
-0.028%
4.3. Discussion
Several previous studies did not explore the performance of the proposed method with sensitivity tests
calculated to determine which method is the best. In addition, several criteria can be added according to the
requirements of each institution, so that the results obtained are more objective. The weighting is done using
a fuzzy table with a value of 0-10. Meanwhile, the process for modeling scholarship acceptance is by
determining the criteria that will be used as a reference, assigning a weighted value for each criterion, providing
a suitability rating value for each alternative, and making a decision matrix based on the criteria. The practical
implications are certainly a tool for MTs YABIKA to determine scholarship recipients and the implications of
the methodology as a reference in similar research.
5. Conclusion
The criteria for determining scholarship recipients have been previously determined by the school and the
author adds one proposed criterion so that the assessment is more objective. The process of determining
scholarship recipients is weighted using a fuzzy table with a value of 0-10. While the modeling of scholarship
acceptance by determining the criteria that will be used as a reference, assigning a weighted value for each
criterion, assigning a suitability rating value for each alternative, and making a decision matrix based on the
criteria. The final result is the ranking of students as a recommendation for scholarship recipients. Referring
to the sensitivity test results, SAW has a change value of 6.251%, and TOPSIS is -0.028%, therefore, it can be
said that SAW is the most relevant method in this case study.
The constraints of this research is there is no system has been built that supports the process of the two
proposed methods so that policy makers in schools can easily determine which students will receive
scholarships. Therefore, future work can be done by developing a decision support system based on dynamic
platform; various criteria need to be added such as: parents' occupation, number of children, or student status.
１６７
AEICP Vol. 5, No. 1
Scholarship awarding can also be explored with other methods such as: Analytic hierarchy process, Weighted
Product, and Elemination Et Choix TRaduisant la realite (ELECTRE).
References
[1] Kurniawan, Y. I. (2015). Decision support system for acceptance scholarship with simple additive
weighting method. In Proceeding of the 1st International Conference on Science, Technology and
Humanity (pp. 99-108). UMS.
[2] Pratama, R. N. & Marhusin. (2016). The decision support system for evaluating the teaching and learning
process evaluation (PBM) for lecturers at the Hasur Banjarmasin Polytechnic uses AHP and SAW
methods. Phasti, 2(1), 37-43.
[3] Sunarti, Sundari, J., Anggraeni, S., Siahaan, F. B., & Jimmi. (2018). Comparison TOPSIS and SAW
method in the selection of tourism destination in Indonesia. In Proceedings of the 3rd International
Conference on Informatics and Computing (ICIC) (pp. 1-6). doi:10.1109/iac.2018.8780550
[4] Tuslaela, T. (2020). The scholarship awarding decision support system uses the topsis method. Jurnal
Riset Informatika, 2(4), 201-207.
[5] Deni, W., Sudana, O., & Sasmita, A. (2013). Analysis and Implementation Fuzzy Multi-Attribute
Decision Making SAW Method for Selection of High Achieving Students in Faculty Level. IJCSI
International Journal of Computer Science, 10(1), 674-680.
[6] Kusumadewi, S., Hartati, S., Harjoko, A., & Wardoyo, R. (2006). Fuzzy MultiAttribute Decision Making
(Fuzzy MADM), Graha Ilmu.
[7] George, J., Badoniya, P., & Naqvi, H. A. (2018). Integration of simple additive weighting (SAW) and
Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) for Supplier Selection.
International Journal for Science and Advance Research in Technology, 4(8), 18-22.
[8] Rahmayani, A. & Irawan, M. I. (2016). Design and implementation of multicriteria decision support
system software using TOPSIS method. Jurnal Sains dan Seni ITS, 5(2), A37-A42.
[9] Putra, E., Hidayatuloh, S., Nguyễn, P., Sasmita, K., & Wibowo, C. (2020). Decision Support System for
Proposing Scholarship Recipients to Best Students using SAW. International Journal of Control and
Automation, 13(2), 103-109.
[10] Sari, H. N. & Fatmawati, A. (2019). Decision support system recommendations for poor rice using SAW
and TOPSIS method (case study: semagar girimarto wonogiri village). Jurnal Mitra Manajemen, 3(1), 96-
108.
[11] Siregar, M. U., Nasiroh, T., & Mustakim, M. (2021). A hybrid approach using entropy-TOPSIS to
determine merit scholars based on objective criteria. Jurnal Teknologi Informasi dan Ilmu Komputer
(JTIIK), 8(1), 167-176.
１６８
8th ICAEIC-2022
Identify the CARLA Simulator Element of the RSS Model for Variable Angle
Camera Application
Min Joong Kim1, and Young Min Kim*
1Candidate, Department of Systems Engineering, Ajou University, Korea
*Department of Systems Engineering, Ajou University, Korea
aquamjkim@ajou.ac.kr1, pretty0m@ajou.ac.kr*
Abstract
Today, it is important to secure safety and reliability as the spread of vehicles with autonomous driving
functions increases. Recently, Mobile Eye proposed the RSS (Responsibility Sensitive Safety) model, a white
box mathematical model that secures the safety of autonomous vehicles and standardizes the minimum
requirements that all autonomous vehicles must meet in the event of an accident. In addition, as a cognitive
sensor, a camera capable of covering the cognitive area of an existing radar or lidar with a single camera
was used. This paper discusses the identification of necessary elements to apply CARLA, an open-source
simulator for autonomous vehicle research, to verify RSS models suitable for variable focus function cameras
with varying focus and angle of view to improve cognitive ability.
Keywords: Automotive Vehicle, Responsibility Sensitive Safety (RSS), Variable focus function camera,
CARLA, Simulation.
1. Introduction
Today, research on autonomous driving is underway, and vehicles with autonomous driving functions are
rapidly becoming common [1]. Therefore, the safety of self-driving cars is becoming important, and efforts to
increase reliability are essential [2]. As one of the Advanced Driver Assistance (ADAS) systems that assist
drivers while driving in a way that ensures safety and reliability of self-driving cars, Adaptive Cruise Control
(ACC), a car control algorithm that maintains distance from the car ahead, is widely used [3]. Recently, a
Responsibility Sensitive Safety (RSS) model was presented in Mobileye to prevent accidents in autonomous
vehicles [4].
Previous studies on the RSS model are as follows. Xu, X. et al. (2021) extracted the safety risk situation
based on the vehicle tracking scenario of SH-NDS data and corrected the RSS model using the NSGA-II
algorithm [5]. Li, L. et al. (2018) proposed a new anti-collision strategy for automobile tracking methods to
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１６９
AEICP Vol. 5, No. 1
maintain traffic safety and efficiency [6]. Liu, S. et al. (2021) confirmed that RSS is a safety guarantee model
and can be applied to ensure safety performance of various autonomous driving algorithms. The impact of the
RSS model in the vehicle interruption situation was evaluated based on the interruption scenario in which the
time-to-collision (TTC) was less than 3 seconds. It was confirmed that the RSS model is superior to the human
driver and ACC [7]. Kim, M.J. et al. (2021) derived and presented an RSS model suitable for a variable focus
functional camera.
In this paper, we present the necessary elements identified to validate the RSS model derived for application
to variable focus functional camera using CARLA Simulation.
2. Backgound
2.1. RSS (Responsibility Sensitive Safety)
In the RSS model, the safety distance is represented as shown in the following equation. Figure 1 represents
the longitudinal safety distance schematic diagram, Equation 1 represents the longitudinal safety distance in
the RSS model, Figure 2 represents the transverse safety distance schematic diagram, and Equation 2 represents
the transverse safety distance [4].
min
long =+
max, accel+,
min, brake
max, brake (1)
Here, []=max{, 0}, and +max, accel
lat , ,=max, accel
lat is assumed in Equation (2) of
calculating the lateral RSS safety distance. In Equation (1), , represent longitudinal speeds of the
preceding vehicle and the following (ego), respectively, and represents the response time. max, accel
represents acceleration during the reaction time of the following vehicle, and max, brake, min, brake represent
deceleration of the preceding vehicle and the following vehicle, respectively.
Figure 1.
Longitudinal safe distance.
2.2. Variable focus function camera
As its name suggests, a variable angle of view camera is a single camera whose angle of view changes.
Figure 2 shows a schematic diagram of the concept of a variable focus function camera.
１７０
8th ICAEIC-2022
Figure 2.
Schematic of variable focus function.
By changing the angle of view and the focal length, the variable angle camera may cover the area
recognized by the existing radar and lidar with a single camera. In addition, using a single camera
benefit in terms of space compared to using three cameras according to cognitive distance.
3. Simulation engine
CARLA is a simulator for autonomous vehicle learning and is built for the flexibility and realism of
rendering and physics simulations [9]. In addition, it is implemented as an open source using Unreal Engine 4
(UE4) [10] and has various scalability.
4. Identification of elements
4.1. Camera
In order to test by applying a variable focus function camera to an autonomous vehicle, a camera element
must be defined. The camera setting elements are shown in Table 1. Among these variables, we reproduced
the function of variable angle of view by changing fov.
Table 1. Camera setting elements.
Blueprint attribute
Type
Default value
Description
image_size_x
int
800
Image horizontal size (pixel)
image_size_y
int
600
mage vertical size (pixel)
fov
float
90.0
The field of view (degree)
１７１
AEICP Vol. 5, No. 1
4.2. Wheather
The variables related to the weather are shown in Table 2. Various weather environments may be created
by combining the values of Table 2, and predefined weather may be set for convenience.
Table 2. Weather elements.
Weather parameters
Type
Range
Description
cloudiness
float
0 ~ 100
It represents the amount of clouds in the sky
0: Clear sky, 100: Completely cloudy sky
precipitation
float
0 ~ 100
It represents the intensity of rain
0: Sunny, 100: Heavy rain
precipitation_deposits
float
0 ~ 100
It is a variable that makes a puddle of water on the road
0: No puddle at all.
100: Completely covered with water
5. Conclusion
As the supply of vehicles with autonomous driving functions increases, the safety and reliability of
autonomous vehicles have emerged. Recently, Mobile Eye proposed an RSS model based on human judgment
and common sense. In our previous study, we proposed a reconfigured RSS model suitable for variable focus
function cameras. In this paper, necessary elements were identified in CARLA Simulator to confirm the
suitability of the proposed model.
Acknowledgement
This work was supported by a grant from R&D program of the Korea Evaluation Institute of Industrial
Technology (20014470).
References
[1] Hörl, S., Ciari, F., & Axhausen, K. W. (2016). Recent perspectives on the impact of autonomous
vehicles. Arbeitsberichte Verkehrs-und Raumplanung, 1216.
[2] Dixit, V. V., Chand, S., & Nair, D. J. (2016). Autonomous vehicles: disengagements, accidents and
reaction times. PLoS one, 11(12), e0168054.
[3] Magdici, S., & Althoff, M. (2017). Adaptive cruise control with safety guarantees for autonomous
vehicles. IFAC-PapersOnLine, 50(1), 5774-5781.
[4] Shalev-Shwartz, S., Shammah, S., & Shashua, A. (2017). On a formal model of safe and scalable self-
driving cars. arXiv preprint arXiv:1708.06374.
１７２
8th ICAEIC-2022
[5] Xu, X., Wang, X., Wu, X., Hassanin, O., & Chai, C. (2021). Calibration and evaluation of the
Responsibility-Sensitive Safety model of autonomous car-following maneuvers using naturalistic driving
study data. Transportation research part C: emerging technologies, 123, 102988.
[6] Li, L., Peng, X., Wang, F. Y., Cao, D., & Li, L. (2018). A situation-aware collision avoidance strategy for
car-following. IEEE/CAA Journal of Automatica Sinica, 5(5), 1012-1016.
[7] Liu, S., Wang, X., Hassanin, O., Xu, X., Yang, M., Hurwitz, D., & Wu, X. (2021). Calibration and
evaluation of responsibility-sensitive safety (RSS) in automated vehicle performance during cut-in
scenarios. Transportation research part C: emerging technologies, 125, 103037.
[8] Kim, M. J., Yu, S. H., Kim, T. H., Kim, J. U., & Kim, Y. M. (2021). On the development of autonomous
vehicle safety distance by an RSS model based on a variable focus function camera. Sensors, 21(20), 6733.
[9] Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., & Koltun, V. (2017, October). CARLA: An open urban
driving simulator. In Conference on robot learning (pp. 1-16). PMLR.
[10] Epic Games. Unreal Engine 4. https://www.unrealengine.com.
１７３
AEICP Vol. 5, No. 1
Building New Words Validation and Sentiment Analysis Model through AI
Technique
Dong Hyeon Kim1, Da Bin Park2, Seung Ri Park3, Se Jong Oh4, and Ill Chul Doo5
1Global Business and Technology, Hankuk University of Foreign Studies, Yongin, Korea
2 Department of Industrial and Management Engineering, Hankuk University of Foreign Studies, Yongin,
Korea
3 English Linguistics, Hankuk University of Foreign Studies, Seoul, Korea
4,5 Artificial Intelligence Education, Hankuk University of Foreign Studies, Yongin, Korea
donghyunkim1217@gmail.com, dabin3178@hufs-gsuite.kr, qkrtmdfl1103@naver.com, tbells@hufs.ac.kr,
dic@hufs.ac.kr
Abstract
This paper aims to construct models that verifying Korean new words and sentiment analysis. The purpose is
to improve the performance of Korean NLP. New words were extracted and verified from text data to be
analyzed and organized into a neologism dictionary, then used for morphological analysis using deep learning
models. Moreover, we utilized the words for sentiment analysis. To build the models, we collected article titles
from some Internet communities that are particularly active in using new words. In addition, we constructed
a validation model using logistic regression analysis and XG BOOST and used CNN and BERT for the
sentiment analysis model. Consequently, we made it possible that extract new words from a new text and better
sentiment analysis.
Keywords: Korean NLP, New Words, Morphological Analysis, User Dictionary, Validation of New Words,
Sentiment Analysis, AI Technique
.
1. Introduction
New words give convenience in conveying a meaning of words and are very important factors in text
analysis as they often contain social aspects. Besides, the process of creation and extinction is very fast, so
there is a difference between when a word is created and built into a dictionary. However, they now use
neologism dictionaries already created when analyzing texts [1]. As a result, if there is a new word that is not
found in the dictionary in text data to be analyzed, it would be ignored, and the text could be analyzed wrong.
Thus, after extracting and verifying new words from data to be analyzed, we constructed a neologism
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１７４
8th ICAEIC-2022
dictionary and used it for morphological analysis. There are studies to extract new words, but the step of
verifying whether they are real neologisms or not is insufficient [2]. Therefore, developing existing research,
we built a model that verifies new words using AI-based logistic regression analysis and XG BOOST.
Particularly, XG BOOST is proper to verify new words because it is suitable for classifying words and is a
highly scalable technology [3]. Furthermore, we deduced a convincing result by comparing and using CNN
[4] and BERT model for sentiment analysis.
2. Methodology
2.1. New Words Validation Model
First of all, we gathered 2,000,000 post titles which range from 2019 to 2021 from the Internet communities
(DCInside, Youtube, Ppomppu, Everytime, Naver news, Natepann).
Especially, criteria derived from the prior study, we established four criteria suitable for new words analysis.
First, we eliminated special letters such as '@' or '!' and incomplete letters such as 'ㅋ(k)' or 'ㅠ(yu)'. Second,
we generated the partial words adopting the N-gram methodology, so that we can deal with all possible cases.
Third, we excluded words that have only one syllable or more than seven syllables and words that have a
frequency lower than 0.01% out of the total word segments number [5].
Additionally, we set up independent and dependent variables. As dependent variables, we employed 5
variables (Cohesion_forward, Left branching_entropy, Right_branching_entropy, Left_accessor_variety,
Starts_ratio) provided by SOYNLP WORDEXTRACTOR and 2 variables (frequency and length). To establish
the dependent variable, we labeled the words manually checking whether the words are used in reality by
searching them on the Internet [5]. And we also utilized SMOTE Oversampling to resolve the asymmetry of
the dependent variable. Going further from the prior study, we made it possible to extract new words based on
the variables above utilizing XGBOOST and Logistic regressions.
2.2. New Words Sentiment Analysis Model
Figure 1.
Method of New Words Validation & Sentiment Analysis Model.
Moreover, we customized our morphological analyzer dictionary with the final new words extracted from
the models. And we tested whether the morphological analysis goes with good performance. Also, we
１７５
AEICP Vol. 5, No. 1
conducted the sentiment analysis to demonstrate the utilization of new words. Accordingly, we established
CNN model with one layer. And we trained and tested the model with data derived from Naver movie review
data and data from the Internet communities labeled manually [4]. To analyze with more sophistication, we
also adopted the BERT model, so that we compared the performance between CNN and BERT.
3. Result
We demonstrated all the variables have a p-value lower than the significance level(0.05) using Logistic
regressions before new words validation. And we extracted the new words from the models and compared the
performance between them.
Table 1. Outcomes of Logistic Regression and XG BOOST.
Evaluation Metrics
Logistic Regression
XG BOOST
Accuracy
0.7228
0.8486
Precision
0.7019
0.8280
Recall
0.7706
0.8849
Sensitivity
0.7480
0.8729
F1
0.7346
0.8555
AUC
0.7230
0.8481
XGBOOST outperformed Logistic regressions with every metric shown above. Therefore, we finally
adopted XGBOOST and extracted about 4,500 new words(Table 2). Considering accuracy and sensitivity are
81.94% and 81.2% from the prior study[5], we can say that our metrics that scored 84.86% and 87.29%
respectively from XGBOOST are the outstanding point in this study.
Table 2. Extracted New Words by XG BOOST
No.
New Words
No.
New Words
No.
New Words
No.
New Words
No.
New Words
1
결정장애
[kjəldʒəŋdʒɑŋeɪ]
11
잡아떼
[dʒɑpɑteɪ]
21
팬싸인회
[fænsɑɪnhœ]
31
직장상사
[dʒɪkdʒɑŋsɑŋsɑ]
41
녀언
[nyən]
2
팔아먹
[pɑɹɑmək]
12
싸이
[sɑɪ]
22
우병우라인
[upyəŋʊlɑɪn]
32
더쿠
[dəkʊ]
42
가짜뉴스
[kɑdʒɑnyʊz]
3
키차이
[kɪtʃɑɪ]
13
민주당
[mɪndʒʊtɑŋ]
23
라이센
[lɑɪsɛn]
33
차명계좌
[tʃɑmyəŋkyɛdʒw
ɑ]
43
빼박
[pæpɑk]
4
심리테스트
[sɪmlɪtɛst]
14
기상캐스터
[kɪsɑŋkæstəɹ]
24
몇일전
[myətʃɪldʒən]
34
임진왜란때
[imdʒɪnwærɑntɛ]
44
앱스토어
[æpstoəɹ]
5
놀이기구
[nɔɹɪkɪkʊ]
15
밀어주
[mɪɹədʒʊ]
25
에르메스
[ɛɹmɛs]
35
몸상태
[mɔmsɑŋtæ]
45
넘사벽
[nəmsɑpyək]
6
탄핵가자
[tɑnhækkɑdʒɑ]
16
마이크로소프트
[mɑɪkɹɔsɔft]
26
국정농단
[kʊkdʒəŋnɔŋtɑn]
36
논란중인
[nɔnɹɑndʒʊŋɪn]
46
구원파
[kʊwənpɑ]
7
온도차
17
페미
27
메가박스
37
역사왜곡
47
치트키
１７６
8th ICAEIC-2022
[ɔntɔtʃɑ]
[fɛmɪ]
[mɛgəbɑks]
[yəksɑwækɔk]
[tʃɪtkɪ]
8
법정구속
[pəpdʒəŋkʊsɔk]
18
북한인권문제
[pʊkhɑninkənmʊ
ndʒeɪ]
28
살인사건
[sɑɹɪnsɑkən]
38
종북좌빨
[dʒɔŋpʊkdʒwɑpɑ
l]
48
닭근혜
[tɑkkeʊnhyɛ]
9
즈그
[dʒeʊkeʊ]
19
돈까스
[tɔnkɑs]
29
악플
[ɑkpl]
39
북한군
[pʊkhɑnkʊn]
49
서울경기
[səʊlkyəŋkɪ]
10
측근정리
[tʃeʊkeʊndʒəŋɹɪ]
20
탄핵집회
[tɑnhækdʒɪphœ]
30
예능프로그램
[yɛneʊŋpɹɔgɹæm]
40
고민중이
[kɔmɪndʒʊŋɪ]
50
여자배구
[yədʒɑpækʊ]
We also compared the CNN and BERT by the sentiment analysis. We trained and tested the model with data
derived from Naver movie review data and data from the Internet communities. And CNN and BERT scored
about 89% and about 78% respectively for the accuracy. Moreover, we tested the sentiment analysis with
sentences that included new words. We checked out that CNN outperformed BERT as well. The result accords
with many former studies that point out CNN is suitable for sentiment analysis of natural languages [4].
4. Conclusion
This study started with the question of how to identify and analyze new words. Paying attention to the
difference between the creation time of new words and neologism dictionaries, we verified new words and
applied them to sentiment analysis. In addition, we tried to find models suitable for Korean NLP by comparing
various methods of verifying new words and sentiment analysis. Especially, our models can be a great tool for
tracking malicious comments that have recently become a social problem. In fact, when you input a sentence
used on the website, it shows excellent performance in filtering out negative sentences. In addition, many
companies use consumer reviews well in marketing, and it could play an important role in this position. As a
result, we expect this study to be a turning point in Korean NLP research and further practical fields.
Furthermore, if the process focusing on Korean is developed in the case of English data, it is expected to have
great utility.
Table 3. Outcomes of Sentiment Analysis by CNN
No.
Input Sentence
Outcomes of Sentiment Analysis
Probability
1
개누리쉐이들
사무총장에
정문헌
이혜훈
수준이
딱
이거지
..
싀발 합리적 보수는 개뿔
[kænʊɹɪʃeɪdeʊl sɑmʊtʃɔŋdʒɑŋɛ dʒəŋmʊnhən ihyɛhʊn sʊdʒyʊnɪ tɑk
ikədʒɪ suɪpɑl hɑpɹɪdʒək pɔsʊneʊn kæpʊl]
Negative
95.93%
2
난
정말
한남들이
팔로잉과
알티에
이렇게
큰
의미를
부여하는지 처음 알았다.
[nɑn dʒəŋmɑl hɑnnɑmdeʊlɪ fɑloʊɪŋkwɑ ɑɹtɛ iɹətkyɛ keʊn uɪmɪreʊl
pʊyəhɑneʊndʒɪ tʃəʊm ɑɹɑtɑ]
Negative
63.21%
3
페미
나치가
이
세상에
한
번이라도
등장했는진
모르겠는데
한남은 나치가 맞다
[fɛmɪ nɑtʃɪkɑ i sɛsɑŋɛ hɑn pənɪɹɑtɔ teʊŋdʒɑŋhætneʊndʒɪn
mɔɹeʊkɛtneʊntɛ hɑnnɑmeʊn nɑtʃɪkɑ mɑtɑ]
Negative
75.09%
１７７
AEICP Vol. 5, No. 1
Acknowledgement
Funding: This research was supported by Hankuk University of Foreign Studies Research Fund of 2021.
Also, This research was supported by the MIST (Ministry of Science, ICT), Korea, under the National
Program for Excellence in SW), supervised by the IITP(Institute of Information & communications
Technology Planing & Evaluation) in 2021"(2019-0-01816), This work was supported by the Ministry of
Education of the Republic of Korea and the National Research Foundation of Korea (NRF-
2021S1A5A8065934)
References
[1] Shin, P. S. (2020). Emotional analysis system for social media using sentiment dictionary with newly-
created words. The Korean Society of Computer and Information, 25(4), 133-140
[2] Kim, J. W., Jeong, J. W., & Cha, M. Y. (2020). Automatic New Korean Words Extraction Using Portal
News Headlines. The HCI Society of Korea, 163-166
[3] Chen, T., Guestrin, C. (2016). XG BOOST: A Scalable Tree Boosting System. KDD 16: Proceedings of
the 22nd ACM SIGKDD International Conference On Knowledge Discovery and Data Mining, 785-794
[4] Kim, H. J. (2018). Extraction Method of New Word from Online Community: The Application of New
Method to Morphological Analysis. Graduate School of Information Yonsei University, 4-5
[5] Kim, Y. (2014). Convolutional Neural Network for Sentence Classification. The 2014 Conference on
Empirical Methods In Natural Language Processing, 1746-1750.
１７８
8th ICAEIC-2022
Fabrication of Activated Carbon from Areca Nuts and Its Application for
Methylene Blue Dye Removal
Pramila Joshi1, and Sahira Joshi2
1Department of Applied Sciences and Chemical Engineering, Nepal
2Pulchowk Campus, IOE, Tribhuvan University, Lalitpur, Nepal
Corresponding email: sjoshi61@hotmail.com
Abstract
In this study methylene blue dye removal was carried out using areca nut activated carbon to purify waste
water. Activated carbon were prepared by using H3PO4 as an activating agent at different carbonization
temperature 500°C and impregnation ratio of H3PO4 to areca nut 1:1 by weight by batch method. The
comparative study between plain carbon and activated carbon and adsorption isotherm experiment were
carried out by batch method. The results showed that ANAC percentage removal of MB by ac is much higher
than that of plain carbon. The adsorption equilibrium data follows Langmuir isotherm model compared to
Freundlich isotherm which indicate the monolayer adsorption of MB dye to the homogenous AC surface. The
adsorption capacity of activated carbon prepared from Areca nut for methylene blue was found to be 298 mg/g.
Thus, highly efficient AC as an adsorbent could be prepared from areca nut for the adsorptive removal of MB
dye to purify water.
Keywords: Activated Carbon, areca nut, Methylene Blue dye, Preparation, Adsorption
.
1. Introduction
Methylene blue (MB) is a basic or cationic dye with commonly used in various applications such as
colouring and dyeing cotton, wool and silk. Its long-term exposure can cause vomiting, nausea, anemia and
hypertension [1]. Therefore, removal of color from water/wastewater becomes environmentally important.
Adsorption of dye onto activated carbon (AC) has been widely used as a facile and effective technique to
remove a large variety of dyes from aqueous solutions.
Activated carbon is the synonym for charcoal or activated coal and characterized by well-defined porosity.
Activated carbon is available in granular and pellet and powder forms. According to IUPAC, pores of AC are
categorized according to their pore size in three groups micropores (diameter <2nm), mesopores (<2nm to
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１７９
AEICP Vol. 5, No. 1
<50nm) and macropores (diameter, >50nm). For adsorption of small molecules micropores having small area
>2nm are used while the larger pore sizes of <2nm are used for adsorption of larger molecules [2]
The preparation of AC can be done from carbonaceous materials by carbonization in inert atmosphere
which is later followed by activation of carbonized product. Carbonization is resulted in the formation of fixed
carbon with rudimentary pore structure [3-5]. Activation enlarges the pores diameter and also creates new
pores. Activation methods are of two types: physical and chemical. Physical activation consists of two steps
i.e., carbonization in inert atmosphere. Major advantages of the chemical activation over the physical activation
are lower carbonization temperatures, better pore structure and high product yield. The well–known chemical
agents in chemical activation processes are ZnCl2, HPO4, H2SO4, K2S, KCNS, HNO3, H2O2, KMnO4,
(NH4)2S2O8, NaOH, KOH and K2CO3 are used to activate carbons, resulting in a high surface area and
appropriate porous structure. The minimum carbonization temperature of 4000°C is required to allow the
complete chemical transformation of lignocellulosic precursors into graphene structure Carbonization time
must be optimized in order to obtain the maximum porosity development Activated carbon is used as adsorbent
for many different purposes and can be prepared from numerous different raw materials this makes the
production of activated carbon incredibly versatile. In this present study areca nut is used due to its wide
availability in south east Asia. The areca nut is the seed of the areca palm mainly found in tropical part.
Southeast and South Asia and east Africa. The aim of present study was to find the adsorptive property of
Areca Nut Activated Carbon (ANAC) for methylene blue dye [6].
2. Experimental methods
The Areca nut sample were collected from local area of Lalitpur Nepal. These nuts were than were than
washed frequently using distilled water and left to dry in electric oven 110°C for 6 hrs. The dried nuts were
crushed grinded and sieved in size of 212micrometers. All chemicals were used of analytical grade and
purchased from qualigens, India. Distilled water was used for all water requiring solutions. Nitrogen used for
pure atmosphere during carbonization process was Ultra high pure (UHP) nitrogen.
2.1. Preparation of activated carbon
AC samples were prepared by mixing H3PO4 and areca nut powder with impregnation ratio 1:1 by weight
and carbonized at temperature of 500°C and for 3hrs. Then all the prepared ACs were treated with 0.1M HCl
and then washed with warm distilled water to remove different residual organic and mineral matter. Finally,
the samples were dried for 24 hours at temperature of 100°C inside an air oven.
2.2. Adsorption experiment
A stock solution of MB was prepared by dissolving 1 g of MB dye in 1000 mL distilled water, which was
diluted to desired concentrations. 0.025 g of the AC was added separately to 25 ml of MB solution of 350
mg/L in 50 ml conical flask. The suspensions were agitated in an electric shaker at constant speed of 120 rpm
for 3 h. After agitating the solution, the reaction suspensions were filtered. Then, the concentration of dye in
the filtrate was determined by a spectrophotometer using 665 nm as the maximum absorbance wavelength.
The percentage removal of dye by the AC was calculated, using following equation:
１８０
8th ICAEIC-2022
(%) Removal= (Co-Ce ) *100 / Co
where, Co and Ce are the initial and equilibrium of dye concentration (mg/L) in solution, respectively.
3. Result and discussion
3.1. Comparision of %removal of mb dye
Figure1. Comparison on Percentage Removal of MB dye onto different Carbon.
The above figure1 shows the comparative study of adsorption of activated carbon and commercial, plain
carbon. In figure 2 the result showed that ANAC removed MB dye around 336 while plain carbon removed
almost 86.
Adsorption increased from 0-60min and became unsteady upto 180 and became stable at 210 according to
the graph shown in figure 3. While the stable adsorption of dosage was in 3.6g/L.
3.2. Adsorption isotherm
An adsorption isotherm is the relationship between the adsorbate in the liquid phase and the adsorbate
adsorbed on the surface of the adsorbent at the equilibrium at constant temperature.
3.3. Langmuir Isotherm
This model stated that the monolayer adsorption of the adsorbate occurs on a homogenous surface of the
adsorbent without any intereaction between the adsorbate and adsorption. A Langmuir isotherm for the
adsorption of the MB dye on the ANAC with Cₑ/qe vs Ce is shown in figure 5.
Figure 5. Langmuir isotherm Figure 6. Freundlich isotherm
１８１
AEICP Vol. 5, No. 1
3.4.
Freundlich Isotherm
Freundlich isotherm assumes that the uptake of metal ions occurs on a heterogenous surface by multilayer
adsorption and the amount of adsorbate adsorbed increases infinitely with an increase in concentration.
Freundlich isotherm for the adsorption of MB on the ANAC was plotted with logqe vs log Ce which is shown
as in figure 6.
The applicability of isotherm experiment was completed by comparision of coefficient of determination
(R2=0.9977). The coefficient of determination of R2 value is higher as compared to Freundlich isotherm. It
indicates that adsorption of mb on ac carbon follows more to langmuir isotherm. It indicated monolayer
adsorption.
4. Conclusions
In present investigation preparation of AC from areca nut using H3PO4 as an activating agent are 1:1
impregnation ratio at 500°C carbonization temperature. The adsorption equilibrium data follows Langmuir
isotherm model compared to freundlich isotherm which indicate the monolayer adsorption of MB dye to the
homogenous AC surface. Thus, highly efficient AC as an adsorbent could be prepared from areca nut for the
adsorptive removal of MB dye from water. The adsorption capacity of activated carbon prepared from Areca
nut for methylene blue was found to be 298 mg/g.
Acknowledgement
I express my sincere gratitude to Prof. Dr. Bindra Shrestha and Prof. Dr. Sahira Joshi for their constant
supervision and guidance in accomplishment of this research.I convey my heartfelt gratitude to Prof. Dr.Hem
Raj Pant for his support on the accomplishment of this research. for providing lab for research purpose. I would
like to thank Institute of Engineering, Pulchowk and Tri Chandra College, Tribhuwan University for providing
me this platform to conduct research and accomplish this endeavor.
References
[1] Pathania, D., Sharma, S., & Singh, P. (2017). Removal of methylene blue by adsorption onto activated
carbon developed from Ficus carica bast. Arabian Journal of Chemistry, 10, S1445-S1451.
[2] Joshi, S., & Homagai, P. L. (2017). Influence of Activating Agents on the Adsorptive Properties of Betel
Nut Based Activated Carbon. Journal of Nepal Chemical Society, 37, 20-26.
[3] Cao, Y. (2017). Activated carbon preparation and modification for adsorption. South Dakota State
University.
[4] Ma, H. T., Ly, H. C., Pham, N. B., Nguyen, D. C., Vo, K. T. D., & Tuan, P. D. (2017). Effect of the
carbonization and activation process on the adsorption capacity of rice husk activated carbon. Vietnam
Journal of Science and Technology, 55(4), 494-494.
[5] Ademiluyi, F. T., & David-West, E. O. (2012). Effect of chemical activation on the adsorption of heavy
１８２
8th ICAEIC-2022
metals using activated carbons from waste materials. International Scholarly Research Notices, 2012.
[6] Leimkuehler, E. P. (2010). Production, characterization, and applications of activated carbon. University
of Missouri-Columbia.
[7] Standard, A. S. T. M. Designation D4607-94, 2000. Standard Test Method for Determination of Iodine
Number of Activated Carbon, 15, 1-5.
１８３
AEICP Vol. 5, No. 1
Pest Prediction and Diagnosis According to Deep Learning-based
Environmental Change
Man-Ting Li1, Gan Liu 2, Hyun-Tae Kim 3, Sang-Bum Kim4, Eun-Seo Song5, Kyu-Ha Kim6, and
SangHyun Lee7
1,2,3,4,6,7 Department of Computer Engineering, Honam University, Korea
4Department of Electronic Engineering, Honam University, Korea
6LINC+ (Leaders in Industry-university Cooperation) Honam University, Korea
1873648246@qq.com1, 1060110388@qq.co2, gusxo0560@naver.com3, ssdoooo@naver.com 5, {2021115,
kim0659, leesang64}@honam.ac.kr4,6,7
Abstract
The occurrence and spread of diseases and pests in farms are related to many factors such as climate,
environment, and soil. This is paper on according to the environmental characteristics of the smart farm, IoT
technology is used to monitor the environmental variables of the smart farm, and the YOLOX model is used to
detect pests. Through this, we want to predict the occurrence of diseases and pests in smart farms. The
experimental results show that when the IoU is 0.5, all ten types of mAP are 0.853, and when the IoU is 0.95,
the mAP can reach 0.637.
Keywords: Smart Farm, Pests, YOLOX model, IoT, Pest Prediction.
1. Introduction
Crop diseases and pests have always been a major problem in agricultural production management, not only
causing farmers a lot of losses, but also urging farmers to increase their use of pesticides, presenting food
safety concerns where pesticide residues exceeded standards [1]. To maintain the continuous development of
smart farms, it is necessary to monitor and predict the type and distribution characteristics of pests in real time
to establish prevention and control strategies [2].
The development and spread of diseases and pests in smart farms are associated with many factors such as
climate, environment, soil, etc. [3]. In the past, orchard pest monitoring and control was difficult to meet time,
space, and cost requirements [4]. Therefore, utilizing artificial intelligence technology and IoT technology to
monitor diseases and pests in smart farms plays an important role in understanding the rules of disease and
pest occurrence in smart farms, providing real-time alerts, reducing production costs, and improving fruit
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１８４
8th ICAEIC-2022
quality [5].
IoT technology is used to monitor environmental variables in smart farms based on the environmental
characteristics of smart farms, and YOLOX models are used to detect pests. Through this, we want to predict
the disease and pest outbreak of smart farms. Once the high probability of pests and diseases is predicted,
managers can develop a plan to curb pests to reduce the significant damage to the smart farm, which is of great
help to the ecological environment of the smart farm.
2. Related research
2.1. Smart Farm
The smart farm uses IoT technology to analyze Humidity·Sunlight·CO2·Measure and soil, etc., and drive
the control device according to the results of the analysis to change it to the appropriate state [6]. It can also
be managed remotely via mobile devices such as smartphones. Production of agriculture with smart farms ·
Distribution · It can create high value-added value, such as increased productivity, efficiency, and quality
throughout the consumption process. It is a concept that can be appliedto land, greenhouses, plant plants, etc.,
because it refers to agricultural methods using ICT technology [7].
2.2 Recurrent Neural Network
RNN (Recurrent Neural Network) is a network that changes with time like time series data. It is an artificial
neural network for learning data. It is difficult to process continuous data because traditional neural networks
only work with input data [8].
Figure 1.
RNN Basic Structure.
The basic architecture of a circulating neural network consists of three layers: the output layer, the input
layer, and the hidden layer. The data are weighted u through the input layer x, and u input the hidden layer s
for the new weight. There are two new weight outputs, and the weight v input is output on the output layer o
tostore the weight w and use it for the next periodic calculation.RNN is more suitable for text regression
problems because it has the characteristic of sending the results from the node of the hidden layer to the input
of the next calculation of the hidden layer node, sending the results from the node of the hidden layer in the
direction of the output layer.
１８５
AEICP Vol. 5, No. 1
3. Design of pest prediction system according to the proposed IoT and deep learning-
based smart farm environment.
Figure2 is a schematic designed to build a platform to predict pests due to changes in the IoT and deep
learning-based smart farm environments proposed in this paper. IoT can be used to monitor the humidity,
temperature, carbon, etc. of orchards, detect pests through cameras, and is designed to utilize RNN models to
automatically collect and analyze relevant data in the form of big data and predict pests according to the
environment.
Figure 2.
Pest prediction system structure according to proposed smart farm
environment.
The structure of the pest prediction system according to the proposed smart farm environment in this paper
consists of smart farm sensors (Smart IoT), data communication server modules (Brokers), distributed storage
modules (Big Data), RNN data analysis prediction modules (AI Centre), and administrator modules
(Adminzone). Smart IoT allows real-time monitoring of humidity, temperature, and carbon in smart farms,
and can detect pests through cameras. The collected data is stored in a distributed storage module (Big Data)
via a data communication server (Biket). The system is designed to learn past data stored in Big Data through
RNN models, and the current environment predicts the likelihood of a parallel insect exploding.
4. Realization and result
4.1. Dataset structure
The dataset used in this paper is an image of crop pest diagnosis provided by AI hub. Among them, the data
has 10 categories, agrotis ipsilon: 1,037pieces, oriental tobacco budworm: 28 pieces, mamestra brassicae 926
pieces, cotton caterpillar: 3 pieces, athalia rosae ruficornis: 1,141 pieces, diamondback moth 1,141 pieces,
pieris rapae: 1,276 pieces, striolata: 305 pieces, green peach aphid: 212 pieces, silk norijae: 1,285 pieces.
１８６
8th ICAEIC-2022
4.2. Result
Table 2.
Results of each detection model of YOLOX.
Pest
AP(IoU=0.5)
Pest
AP(IoU=0.5)
Athalia rosae ruficornis
0.873
Agrotis ipsilon
0.884
Pieris rapae
0.863
striolata
0.897
Silk Norijae
0.909
Green peach aphid
0.680
Mamestra brassicae
0.689
Cotton caterpillar
1.00
Diamondback moth
0.900
Oriental tobacco budworm
1.00
mAP(0.5)
mAP(0.95)
0.853
0.637
The results of each type of detection model using YOLOX are shown in table 2. When IoU is 0.5, all 10
types of mAP are 0.853, of which type 8 and type 4 have the highest recognition accuracy (100%), and type
14 has the lowest accuracy (68%). Also, when the IoU is 0.95, the mAP can reach 0. 637.
As shown in Figure 5, a high-definition camera is used to collect images and transmit real-time images to
the pest detection model to accurately mark the target area. The accuracy of detecting pests can reach 70% to
90%.
Figure 5.
Detection results.
5. Conclusion
To implement a pest prediction system, the humidity and temperature of the smart farm can be monitored
in real time through Smart IoT, and the image is collected through the camera to detect pests using YOLOX.
The collected data is stored in a distributed storage module (Big Data) through a data communication server
(Broket). The system learns the data stored in Big Data through the RNN model and predicts pests and diseases
through the current environment.
When detecting pests using YOLOX, when the IoU is 0.5, the mAP of all 10 types is 0.853, of which type
8 and type 4 have the highest recognition accuracy (100%), and type 14 has the lowest accuracy (68%). Also,
when the IoU is 0.95, the mAP can reach 0.637. In addition, the image collected from the camera can be
１８７
AEICP Vol. 5, No. 1
transmitted to the pest’s detection model to accurately mark the target area and detect pests with high accuracy
(70% to 90%).
Acknowledgement
Following are results of a study on the "Leaders in INdustry-university Cooperation +" Project, supported
by the Ministry of Education and National Research Foundation of Korea.
References
[1] Kim Yeon-jung, (2017). The 4th Industrial Revolution and the Future of Our Agriculture, Korea Farm
Economic Research Institute, World Agriculture, 202(202). 121-137.
[2] Euncheon Lim, Shin Chang-sun, Shim Chunbo, (2007). Designing and Implementing Multi-Me Deer Pest
Prediction Management System using Wireless Sensor Network, Korea Computer Information Society,
12(3), 27-35.
[3] Seo Dae-seok, Kim Yeon-jung, (2016). Prioritizing Policy Measures to Expand The Spread of Smart
Farms, Journal of the Korean Society of Computer Science and Technology, 17(11), 348-354.
[4] Yang Hee-chan, Lee Jae-so, Lee Hyun-dong, Hyung-Seok Kim, (2018). Automatic detection of pests
caused by cultivation of paprika indoor nutrients using artificial intelligence, Journal of Korea Journal of
Pharmaceutical Bot Systems, 24(11), 1020-1024.
[5] H. Kim, (2012). An Industry ICT Utilization Strategy of Smart Convergence Age, In Proc. Conf. of The
Korea Society of Management Information Systems, Seoul, Korea, 143-151.
[6] Kim Kwan-jung, He Ki-Woong, (2015). Smart Farm Technology Trends and Prospects, Electronic
Communications Research Institute Electronic Communications Research Institute, 30(5).
[7] Kim, Bo-rim, (2016). Convergence of Agriculture and ICT - Smart Farm, Convergence Research Policy
Center, 50, 2-11.
[8] Mikolov, Tomas, et al. (2010). Recurrent neural network-based language model. Interspeech., 2(3).
[9] Min Kum-young, Jung Deok-hoon, (2013). Kite District on The Impact of Big Data Properties on Disaster
Response Decisions, Korea Electronics Trading Society.
[10] https://www.oracle.com/kr/big-data/what-is-big-data/
１８８
8th ICAEIC-2022
A Study on the Anti-wrinkle Effect of Hydrogen Water on Mice’s Wrinkles
Cause by UV and Its Toxicity
Eun-Suk Lee1, Ji-Ung Yang 1, In-sang Lee2, Ki-jin Kwon3, and Dae-Gyeom Park1
1
Department of Bio-industrial Machinery Engineering Pusan National University, Korea
2Divine Research Institute Co., Ltd., Korea, 3 Hunature, Korea
suk099@hanmail.net1, jiung20@pusan.ac.kr 1, saupil@hotmail.com2, dajokkkkk@naver.com3,
dypark@pusan.ac.kr1
Abstract
Water is the source of life. It decomposes food and carries the waste to keep the blood and vessels clean. The
lack of water that makes up 70% of the human body can lead to various diseases as well as aging. Usually,
oxygen exists with two atoms joined together in a colorless, odorless, and tasteless state. Almost all oxygen
inhaled by a person becomes water, but 1~2% turns to active oxygen. (Gale, 2008) On the other hand, wrinkles
and freckles that are part of the aging process are produced by excessive active oxygen, particularly
photoaging caused by ultraviolet (UV) creates singlet oxygen, a type of active oxygen. (Eun-sang Ji) The
characteristic of hydrogen water is that it removes hydroxyl radical and per hydroxyl radical light, which are
types of strong active oxygen that other antioxidants cannot remove. In this study, the toxicity of hydrogen
water was assessed, and the anti-wrinkle effect of hydrogen water on skin aging was investigated. The subject
animal's skin was exposed with UV to cause skin aging and wrinkles, and the effect of the hydrogen water was
confirmed through oral administration and skin application of hydrogen water. As a result, there was no
significant change in blood and serum, and no toxicity was observed. In the case of the anti-wrinkle effect of
hydrogen water on skin aging and wrinkles, the level of melanin and erythema was significantly reduced by
hydrogen water, and oil-water content in the group has also reduced but the water content has significantly
increased. The thickness of the skin was also significantly reduced in the group. This result presents that
hydrogen water has no toxicity and has the effect of anti-wrinkle on skin aging and wrinkles caused by UV
exposure
Keywords: Hydrogen water. Oral administration, toxicity, Melanin, Skin aging.
1. Introduction
Various diseases in the modern society of longer life expectancy usually occur due to the complicated social
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１８９
AEICP Vol. 5, No. 1
structure and stress. So, studies are being made in various areas to improve and eliminate chronic diseases caused
by work stress, lifestyle and habits, social anxiety, and many smart devices in daily life. Oxygen is essential in
producing body energy and an important element for survival. Most oxygen humans inhale changes into water, but
1-2% changes into active oxygen (Gale, 2008). Normal oxygen exists as a form of oxygen molecule consisting of
two oxygen atoms as one pair. However, active oxygen is an unstable element that one of its electrons is separated
due to several side effects, wandering to find the lost electron (In-hyeuk Kim, 2011). Meanwhile, wrinkles and
freckles that are usually a part of the aging process are also produced by excessive active oxygen. Particularly,
photoaging that UV causes produces singlet oxygen, which is a type of active oxygen. (Eun-sang Ji, 2009) This
study, therefore, assesses the harmfulness(toxicity) of hydrogen water to blood and tissue by conducting oral
administration to hairless mice and assesses the anti-wrinkle effect by applying hydrogen water on its skin where
UV-B was treated through biopsy.
2. Theoretical background
Hydrogen water has a form that hydrogen molecules (H2) are between the water molecules. Scientifically,
hydrogen water refers to the water containing more than 80ppb of hydrogen, and generally, it refers to the
water containing more than 300ppb of hydrogen dissolved (Seo-kon Kim, Dong-soo Im 2015). A technology
was developed to achieve saturated dissolved hydrogen in water to 1,600ppb at 20℃ room temperature in
atmospheric pressure, which allows dissolving a maximum of 1,500ppb in various ways (Seo-kon Kim, 2015).
In 1800, British Nicholson and Carlyle discovered electrolyzed reduced water artificially generated from water
electrolysis in the development process of hydrogen water. It had proved that hydrogen water was like the
miracle water that cured incurable diseases. The four-miracle water –the water of Lourdes, France (1958), the
water of Tlacote, Mexico (1991), the water of Nordenau, Germany (1991), and the water of Nadana, India
(1992) contain lots of hydrogen with the dissolved amount of 800 - 1200ppb and are known as miracle water
that cured leukemia, diabetes, polio, and skin disease (In-hyeuk Kim, 2011). Later, in 1958,
an electrolytic water purifier (hydrogen water) was manufactured in Japan, so its history has continued for
more than 50 years. Moreover, in 1997, a study made by Professor Shirahata, Kyushu University, proved that
the hydrogen generated in electrolysis of water (hydrogen water) near a cathode could remove active oxygen,
and its study has been continued. The best feature of hydrogen is that it removes hydroxyl radical and
peroxynitrile radical, strong active oxygen that cannot be removed by many other antioxidants (doctoral thesis
by Kim Eun-joo, Konkook University).
3. Research method
The hydrogen water used in this experiment was provided by H, S, and KAIST in Korea. And it was
measured for 30 days using a pH meter (pH/ISE Meter pH250L) and ORP meter (RUSTLEX ENH-1000,
Japan). Moreover, through scientific and ethics review by Pusan National University Institutional Animal Care
and Use Committee (PUNU-IACSU), this experiment was approved (approval number: 2021-1365) and
performed. The animals used in the experiment were the 7-week old Cr1/Or1: SKH1-hairless mice, bought
from OrientBio in Seongnam, Korea. The mice spent 1 week of adaptation period in the animal laboratory,
１９０
8th ICAEIC-2022
and the common symptoms were observed to check whether they were normal before being used in the
experiment. The animals were fed freely with irradiated food (Purian Mills Inc) and were bred in PNU-
Laboratory Animal Resources Center (Temperature 38±2℃, relative humidity 50±10%), which was approved
by the Ministry of Food and Drug Safety (certification number: 00231) and AAALAC (certification number:
001525) as specified pathogen-free (SPE) laboratory with an illumination cycle of 12 hours (08:00 - 20:00).
For the subject animals to stay comfortably without stress, a covered environment from the outside was
maintained without noise and stimulating smell. Each cage had 5 mice, respectively. The animals were divided
into a group (NO group) where it is not treated with UV and a group treated with UV. And the group with UV
treatment was divided into a group treated by physioligical saline (UV+Vehicle group, n=5), a group in which
hydrogen was administered orally and applied on the skin (UV+Oral administration+Skin application group
n=5), a group which electrolyzed hydrogen water was administered orally and applied on the skin (UV+Oral
administration+Skin application group n=5), and a group which oxidanium was administered orally and
applied on the skin (UV+Oral administration+Skin application group n=5). The oral administration for the
animals was done by 6cc once a day, and UV treatment was done 3 times a week for 4 weeks by fixing the
subject mouse on a tray. After the 4-week experiment, all animals were euthanized by using carbon dioxide,
and their blood and tissues were collected.
4. Research results and considerations
The harmfulness (toxicity) to blood and tissue was assessed by oral administration of hydrogen water. The
analysis result of skin wrinkle after oral administration and skin application of hydrogen water to the hairless
mice treated by UV-B did not show any statistically significant difference within the groups of weight and
internal organs. Also, the fact that hydrogen water does not cause any organ failure or toxicity reactions was
confirmed. The result of hematological examination and serum biochemical analysis shows that hydrogen
water does not cause any toxicity reaction to the liver and kidneys. The analysis result of the changes of skin
wrinkle presents that the wrinkles significantly increased in the UV+vehicle group compared to the group
without UV treatment, and it also shows the tendency that the wrinkles are reduced in UV+HW, UV+EHW,
and UV+OIW groups. This result presents that the long-term treatment by hydrogen water has an effect of
reducing the wrinkles caused by UV (Figure 1).
Figure 1. Anti-wrinkle effect on the skin, which is not damaged by UV when AGM was applied on the
hairless mice. (A) The wrinkles were measured by reproduction level during the final week.
１９１
AEICP Vol. 5, No. 1
5. Conclusion
To check the toxicity of water, the change of pH, ORP(mV) and the weight of the hairless mice were checked,
and a hematological examination, serum biochemical analysis, and analysis of the liver and tissue were
performed. As a result, pH level of HW, EHW, and OIW was HW 7.49±0.94, EHW 7.22±1.74, and
OIW6.52±1.33, so the water is potable and close to neutral with pH6.52-8.9. The result of ORP(mV) shows
that HW 1,505±6.8, EHW 1,073±38.7, OIW 0, and the dissolved hydrogen of HW has increased 24.8%
compared to EHW. There was no change of weight, and the weight of internal organs such as the thymus,
lungs, heart, kidneys, spleen, and liver did not show any statistically significant difference. In the
hematological examination, WBC, RBC, HGB, HCT, MCV, MCH, MCHC, PST, LYM, and RDW did not
have any significant difference between treated sample groups and NO group. The serum biochemical analysis
test showed that ALP, ALT, ASP GGT, BUN, ALB, and LDH had no toxicity, and the tissue analysis test of
the liver and kidney shows that the toxicity level of the liver and kidney in serum did not have any significant
change. Also in the liver tissue, significant pathological changes such as inflammatory response, meronecrosis,
apoptosis, and liver fibrosis were not observed, and in the kidney tissue, pathological changes such as
meronecrosis and atrophy of glomerulus and renal tubule were not observed. In the result of hematoxylin and
eosin(H&E), the dermal thickness of UVB group has significantly increased compared to the control group.
And the hydrogen oral administration group’s dermal thickness was significantly decreased compared to the
UV group. Also, the result of measuring the thickness of skin and the stratum corneum shows that the skin
thickness that was increased by UV treatement has significantly decreased by taking 3 kinds of samples, and
the thick stratum corneum got significantly thinner. These results present that hydrogen water has an effect to
reduce skin thickness, the indicator of aging due to exposure to UV.
Acknowledgment
This paper was prepared with the support of Hunature in Changnyeong-gun, and Divine Research Institute
Co., Ltd., in Seoul.
References
[1] Petrucci, Harwood, Herring, Madura (2007). General Chemistry Principles & Modern Applications.
Prentice Hall. New Jersey.
[2] Marx, Tuckerman, Hutter, J. & Parrinello, M. (1999) The nature of the hydrated excess proton in water.
[3] Gadek Z, Hamadaju T, shirahata S, Nordenau Phenomenon-Application of Natural Reduced Water to
Therapy Cell Tdchnology, Basid & Applirf Aspects, Vol, 15:279-285, 2008
[4] Eun-sang Ji, Now Is the Era of Hydrogen Water, Seoul Health Newspaper: 42- 68, 2009
[5] Seo-kon Kim. 2015 The Answer Is Hydrogen Water Gyeonggi-do Sangsang Tree, pp. 22-56
[6] In-hyeuk Kim, (2015) Hydrogen water that saves human, Gyeonggi-do, Pyeungdan, pp. 28- 47
[7] 2016, The effect of drinking hydrogen water on the face skin of the middle-aged women, Konkook
University graduate school, doctoral thesis.
１９２
8th ICAEIC-2022
A Novel Scheme of an Object-Preserving GAN Architecture for Data
Augmentation
Juwon Kweon1, Jaekwang Oh2, and Soonchul Kwon*
1, 2Department of Electronic Engineering, Kwangwoon University, Seoul, Korea
*Graduate School of Smart Convergence, Kwangwoon University, Seoul, Korea
02kjw0203@gmail.com1, dhworhkd11@gmail.com2, ksc0226@kw.ac.kr*
Abstract
Recent studies such as object detection, tracking, and semantic segmentation using deep learning show
remarkable performance, but have data-dependent limitations. Therefore, most studies necessarily use data
augmentation techniques. GAN-based data augmentation is also receiving a lot of interest, and we present a
novel data augmentation scheme using GAN. Our research aims to preserve the area of objects and generate
bird synthetic data. In the experimental part, birds images synthesized using the Caltech-UCSD Birds 200
dataset are shown as results, and the conclusions suggest the limitations of our research and future works.
Keywords: GAN, Data augmentation, Generative model architecture.
1. Introduction
Recently, many applications such as object detection, tracking, and semantic segmentation using deep
learning have been developed. Modern methods yield impressive performance with large datasets. However,
many studies show results are dependent on the amount of training data. Therefore, data augmentation methods
are indispensable in most deep learning works like object detection and segmentation. A GAN-based method
is also used as one of data augmentation [1, 2, 3].
Here, we present a novel data augmentation scheme using GAN. Our research provides a scheme for
preserving the masked part from the reference images and generating synthetic data for easy use in future
research, such as object detection or segmentation.
In this paper, we describe the methods, experiments, and experimental results of the proposed scheme.
Finally, in the conclusion, we discuss the limitations of our study and suggest future directions.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１９３
AEICP Vol. 5, No. 1
2. Related Work
Since many tasks using deep learning show results dependent on the amount of data, GAN-based data
augmentation techniques are receiving a lot of attention. Frid-Adar, et al [1] used generated data using GAN
for liver lesion classification task. They trained 182 liver lesions images to generate synthetic data. And after
synthetic images are added to the dataset, 78.6% sensitivity and 88.4% specificity are improved to 85.7% and
92.4%, respectively.
Waheed, et al [2] presents a method to generate X-ray images by developing a model CovidGAN based on
Auxiliary Classifier Generative Adversarial Network (ACGAN) [3]. Synthetic images generated using
CovidGAN increased the detection accuracy of COVID-19 in chest x-ray images from 85% to 95%.
Huang, et al [4] take issue with the lack of practicality in preserving objects and generating data for other
domains, although GAN-based methods have shown amazing results. Therefore, they present an integrated
network consisting of encoders, generators, and discriminators in each of the different domains. As a result,
they produce visually excellent synthetic images and improve object detection accuracy.
Figure 1.
The architecture of our proposed Generator.
3. Method
Our proposed method uses a simpler pyramid structure proposed by Shocher, et al [5]. Figure 1. shows the
generator architecture of our proposed novel scheme. Our generator maps a 512-length latent vector as an input
to an image with 256x256 resolution through 6 residual blocks. The generator has a structure in which the
output features of the first 5 residual blocks are added to the features obtained from the masked images. Our
scheme aims to generate synthetic images while preserving the object in the reference image.
As shown in Figure 2. (a), we operate element-wise multiplication of the reference image and the
corresponding mask to obtain the masked image. The masked image is passed to the convolution block.
The mask image consists of 0,1, each determining whether to block or pass pixels of the reference image.
So our mask image has a value of 1 in the area of the object. This means that only the object region of the
reference image is passed to the generator. We rescale the masked image to 8x8, 16x16, 32x32, 64x64,
128x128 resolutions. The rescaled images are passed through a convolution block and added to the residual
１９４
8th ICAEIC-2022
block output features that match the scale level of the generator.
Figure 2. (b) shows the structure of the internal layers of the residual block constituting the generator. It has
a residual structure through Up-Sampling, Convolution layer, Batch Norm layer, etc. with the output of the
previous scale as input. Activation of each layer uses leakyReLU and only the last layer uses tanh. The
discriminator is designed to have a symmetrical structure with the generator.
Figure 2.
(a) Detail structure of generator blocks. (b) The layer structure of the residual block.
4. Experiments
4.1. Dataset
We train our networks with the dataset Caltech-UCSD Birds 200. This dataset is a birds photo dataset with
a total of 6033 images with 200 categories. The dataset provides bounding box and segmentation information
with annotations corresponding to each image. We remove the images where the object-occupied region is too
large as data preprocessing.
And we crop the remaining data with the corresponding segmentation image in 256x256 size. As a result, a
total of 4925 images are used as the training dataset.
4.2. Loss Function
We train our networks generator and discriminator simultaneously, with a loss function consisting of the
following two terms given by:
min
max
(,)+() (1)
The first term,, is an adversarial loss from the LSGAN [7]. The second term, , is a masked
reconstruction loss according to the following equation:
=(,,) (2)
Where is the set of reference images, is the set of masks. We set = 0.1 during training.
１９５
AEICP Vol. 5, No. 1
4.3.Results
We generate synthetic images which have a resolution of 256x256. The bird image samples in Figure 4 are
generated images from our generator. Visually, we observe the object in the synthetic data preserve the details
well. This is because the generator is computed with the features of the objects in the reference image during
the training process. However, compared to the object, the details of the background are relatively poor.
Figure 3.
Samples of generated bird images.
Figure 4.
Artifacts in the generated image after training is complete (upper) / Samples showing the influence
of artifacts in the early stages of training(lower).
5. Discussion and Conclusion
We aimed to generate new synthetic images while preserving the properties of the object. As a result, we
１９６
8th ICAEIC-2022
generate an object of high quality. However, there is a limitation of the quality of the background. Also, some
of the generated images show some artifacts around the objects. It is judged that this is because of the roughly
masked segmentation provided by the dataset. As shown in Figure 4, the upper image shows an example of
the image in which the effect of the mask remains after learning is completed. The two images at the bottom
show that the artifacts strongly influence early training. Removal of these artifacts remains limited.
In future research, we want to improve the abnormal parts around the object to generate more natural images.
In addition, we would like to present an improved method of GAN-based data augmentation by synthesizing
high-quality backgrounds.
Acknowledgments
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea
government (MSIT) (No. NRF-2020R1F1A1069079).
References
[1] Frid-Adar, M., Klang, E., Amitai, M., Goldberger, J., & Greenspan, H. (2018, April). Synthetic data
augmentation using GAN for improved liver lesion classification. In 2018 IEEE 15th international
symposium on biomedical imaging (ISBI 2018) (pp. 289-293). IEEE.
[2] Waheed, A., Goyal, M., Gupta, D., Khanna, A., Al-Turjman, F., & Pinheiro, P. R. (2020). Covidgan: data
augmentation using auxiliary classifier gan for improved covid-19 detection. Ieee Access, 8, 91916-91923.
[3] Odena, A., Olah, C., & Shlens, J. (2017, July). Conditional image synthesis with auxiliary classifier gans.
In International conference on machine learning (pp. 2642-2651). PMLR.
[4] Huang, S. W., Lin, C. T., Chen, S. P., Wu, Y. Y., Hsu, P. H., & Lai, S. H. (2018). Auggan: Cross domain
adaptation with gan-based data augmentation. In Proceedings of the European Conference on Computer
Vision (ECCV) (pp. 718-731).
[5] Shocher, A., Gandelsman, Y., Mosseri, I., Yarom, M., Irani, M., Freeman, W. T., & Dekel, T. (2020).
Semantic pyramid for image generation. In Proceedings of the IEEE/CVF Conference on Computer
Vision and Pattern Recognition (pp. 7457-7466).
[6] Welinder, P., Branson, S., Mita, T., Wah, C., Schroff, F., Belongie, S., & Perona, P. (2010). Caltech-
UCSD birds 200.
[7] Mao, X., Li, Q., Xie, H., Lau, R. Y., Wang, Z., & Paul Smolley, S. (2017). Least squares generative
adversarial networks. In Proceedings of the IEEE international conference on computer vision (pp. 2794-
2802).
１９７
AEICP Vol. 5, No. 1
Fake News Detection using Hybrid Neural Network
Niroj Ghimire1, Tanka Raj Pandey2, and Surendra Shrestha3
1,2Nepal Telecom, Provincial Directorate, Lumbini Province, Bhairahawa, Nepal
3Corresponding Author, Department of Electronics & Computer Engineering, Pulchowk Campus
Institute of Engineering, Tribhuvan University, Lalitpur, Nepal, 44700
niroj.ghimire@ntc.net.np, tankar.pandey@ntc.net.np, surendra@ioe.edu.np
Abstract
News is an effective technique for disseminating information. It is also one of the most effective ways for
individuals to communicate with the rest of the world. The spread of fake news is becoming more with the
advancement in technology, which sometimes misleads the readers and leads to inaccurate social opinions.
Fake news may be found on the Internet, news sources and social media platforms. The spread of low-quality
news has harmed both individuals, and society. In this research work, we analyze three hybrid models,
CNN+simple RNN, CNN+GRU, and CNN+BiLSTM in encoder-decoder architecture to predict the fake news
based on the relationship between headline and article of the news. Pre-trained GloVe word embedding is
used for the word to vector representation as it can capture the inter-word semantic information. The CNN-
RNN combination had been shown efficient in deep learning applications because it can capture sequential
and local features of input data. The models were successfully trained and tested on ISOT fake news dataset.
It is found that the CNN+BiLSTM model had better results than the other two hybrid models in binary
classification tasks for the fake news detection system.
Keywords: Fake news; Hybrid Neural Network; Encoder decoder architecture; NLP; GloVe.
1. Introduction
News is a very effective technique for disseminating information. It is an excellent source of information. It
is also one of the most effective ways for individuals to communicate with one other and with the rest of the
world. In the past, the common individual would wait until the next day to discover what had occurred in the
world the day before. This is not the case in today’s society when news moves nearly at the speed of light.
News and posts, both in physical and digital form, are one of the most common methods of knowledge sharing.
Fake News represents false news or propaganda comprising disinformation transmitted via classical media
outlets like newspapers, TV, and modern media sources like social media [1]. There are two aspects to the idea
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
１９８
8th ICAEIC-2022
of fake news: authenticity and intent. Authenticity refers to the fact that fake news contains incorrect material
that can be recognized. The second, intent, refers to the fake content that was written to deceive the reader.
Now, it is essential to know that what features may be utilized to categorize them.
Detecting fake news manually is a subjective and tedious task. Evaluating the veracity of a news article, also
for professional experts, is complicated. News, no longer only circulated via traditional media channels, but
also through the new platforms of social media. Based on multiple neural network experimental approaches,
this study presents a neural network model that can detect false news by correctly predicting the relationship
between the title and the news content.
2. Dataset
The ISOT Fake News Dataset comprises both real and fake contents, published by ISOT research lab,
University of Victoria [2]. The real articles were taken from reuters.com, a well-known news website, whereas
the false stories came from a variety of sources, primarily from websites identified by politifact.com. The data
include 44,898 different pairs of article titles, article text, date, and article label (real or fake). The dataset
contains 53.3 % fake news and 47.7 % real news.
3. Methodology
The methodology for the fake news detection system consists of a hybrid neural network in encoder-decoder
architecture is shown in Figure 2. The study is based on the text features contained in the headline and article,
to find the stance between them. The headline and news content need data preprocessing steps to remove
unnecessary words and symbols. The preprocessed words are applied to GloVe word embeddings to represent
a word in a numerical vector with capturing their semantic and syntactic information. We create three distinct
hybrid neural network models for this research, each with a separate encoder and decoder network that employs
a combination of CNN and RNN.
Figure 1. Method for Fake news Stance Classification.
Word vector representation is used to give the numerical vector for words that can represent the meaning of
a word. The 100-dimensional pre-trained GloVe embedding for the word vector representation [3] is used. The
output from the embedding layer is fed to the convolutional layers which read the inputs as one-dimensional
１９９
AEICP Vol. 5, No. 1
sequential data. The role of the convolutional layer is to automatic feature extraction from the input fed into
the network. The pooling layers employ maximum pooling operations to derive the most important features
from the generated feature maps. The pooling layer’s output is sent to the RNN layer, which learns long-term
dependency and sequential information. RNNs have a unique characteristic that sets them apart from other
deep neural networks in that they include feedback mechanisms. Three distinct models used as the hybrid
network in both encoder and decoder are, CNN+SimpleRNN, CNN+GRU, and CNN+BiLSTM.
The CNN-RNN combinations have been used to recognize emotion [4] and sign language detection from
video streams [5], due to their capacity to learn spatial characteristics with the CNN and sequential features
with the RNN. RNN learns temporal and context characteristics from text, as well as capture long-term
relationships between text entities.
4. Encoder-decoder Architecture
Figure 2. Encoder-decoder architecture for the sequence to sequence model.
The encoder-decoder model has the potential to train a single end-to-end model directly on input and target
phrases, as well as the capacity to handle variable-length input and output text sequences. The encoder, context
vector, and decoder are the three components of the model. Encoder and decoder, both consist of the first CNN
network with convolution and max pool layer and then RNN layer. The encoder takes the input sequence and
summarizes it in internal state vectors. We preserve the internal states of the encoder. The starting states of the
decoder module are set to the encoder module’s final states. The decoder begins producing the output sequence
using these initial states and input sequence to the decoder.
5. Hybrid Neural Network Models
Figure 3. Layers in Hybrid (CNN RNN) Neural Network.
２００
8th ICAEIC-2022
The encoder takes the input sequence from the article body embedding vector and summarizes it in internal
context vectors, using a hybrid neural network model. Context vector tries to capture all input element
information to support the decoder in making correct predictions. In our study, we employ the decoder to take
inputs as context vectors from the encoder and word embedding vector of the news headlines from the
embedding layer. The decoder begins producing the output sequence based on the context vectors from the
encoder and headline embedding vectors. The hidden unit uses relu as an activation function and the prediction
layer uses the softmax activation function.
(a) (b)
(c)
Figure 4. Implementation of Hybrid Neural Network for fake news stance detection (a) CNN + Simple RNN
model (b) CNN+ GRU model (c) CNN + BiLSTM model.
5. Result and Discussion
Table 1. Model building parameters and values
Parameters
Value
Number of pairs (headline and article) for training
25143
Number of pairs (headline and article) for validation
6286
Number of pairs (headline and article) for testing
13469
Optimizer
Loss function
Batch size
Epoch
Rmsprop
Binary cross-entropy
64
10
２０１
AEICP Vol. 5, No. 1
Three different hybrid neural network models that use a combination of CNN and RNN in a separate encoder
and decoder network are implemented using python programming. Figure 6 shows the confusion matrix of our
three models on the testing dataset of the ISOT dataset. It shows the accuracy on ISOT testing dataset 91.74%,
99.86%, and 99.9% using CNN+Simple RNN, CNN+GRU, and CNN+BiLSTM hybrid models on our
encoder-decoder architecture respectively.
(a) (b) (c)
Figure 6. Confusion matrix for three hybrid models for fake news stance detection (a) CNN + Simple RNN
model (b) CNN+ GRU model (c) CNN + BiLSTM model.
Table 2. Evaluation Metrics Comparison of Three models on ISOT Dataset.
Model
Accuracy(%)
Precision(%)
Recall(%)
F1-Score(%)
CNN+ Simple RNN
91.74
91.62
91.63
91.62
CNN+ GRU
99.86
99.86
99.86
99.86
CNN+ BiLSTM
99.9
99.9
99.9
99.9
6. Conclusion
Fake news is usually created to confuse and attract audiences for commercial and political benefit. In this
research work, we analyze three hybrid models, CNN+simple RNN, CNN+GRU, and CNN+BiLSTM in
encoder-decoder architecture to predict the stance between headline and article of the news. The models were
successfully trained and tested on a binary ISOT fake news dataset. The accuracy on the ISOT testing dataset
using three models CNN+Simple RNN, CNN+ GRU, and CNN+ BiLSTM is 91.74%, 99.86%, and 99.9%
respectively. Also on comparing other evaluation metrics precision, recall, and F1-score, the CNN+ BiLSTM
performed well. Hence, it is concluded that the CNN+ BiLSTM model had better results than the other two
hybrid models in binary classification tasks for the fake news detection system.
References
[1] A. Thota, P. Tilak, S. Ahluwalia, and N. Lohia, “Fake news detection: A deep learning approach,” SMU Data Science
Review, vol. 1, no. 3, p. 10, 2018.
[2] H. Ahmed, I. Traore, and S. Saad, “Detecting opinion spams and fake news using text classification,”
Security and Privacy, vol. 1, no. 1, e9, 2018. 51 [26] J. Camacho-Collados and M. T. Pilehvar, “On the
２０２
8th ICAEIC-2022
role of text preprocessing in neural network architectures: An evaluation study on text categorization and
sentiment analysis,” arXiv preprint arXiv:1707.01780, 2017.
[3] J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors for word representation,” in
Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), 2014,
pp. 1532–1543.
[4] D. Kollias and S. P. Zafeiriou, “Exploiting multi-cnn features in cnn-rnn based dimensional emotion
recognition on the omg in-the-wild dataset,” IEEE Transactions on Affective Computing, 2020.
[5] S. Masood, A. Srivastava, H. C. Thuwal, and M. Ahmad, “Real-time sign language gesture (word)
recognition from video sequences using cnn and rnn,” in Intelligent Engineering Informatics, Springer,
2018, pp. 623–632.
２０３
AEICP Vol. 5, No. 1
Efficient Semantic Segmentation by Using Down-Sampling and Subpixel
Convolution
Young-Man Kwon1, Sung-Hoon Bae2, Dong-Keun Chung3, and Myung-Jae Lim*
1,2,3,*Department of Medical IT, Eulji University, 553, Sanseong-daero, Sujeong-gu, Seongnam-si, Gyeonggi-
do 13135, Korea
ymkwon@eulji.ac.kr1, 2017162030@g.eulji.ac.kr2, tchung@eulji.ac.kr3, lk04@eulji.ac.kr
*
Abstract
Recently, Semantic Segmentation has been widely applied to image processing and scene understanding,
medical image, robotic perception, video surveillance and among many others. The U-Net has the good
performance except disadvantage of losing location information and detailed content on the image. To solve
this problem, this paper proposed the U-Net model by using down-sampling and Subpixel convolution layer
used in ESPCN. We also measured the performance of U-Net and proposed system. As a result, our proposed
model achieved 83.8% mean-IoU and U-Net achieved 77.3% mean-IoU, confirming that the proposed model
achieved better performance.
Keywords: Semantic Segmentation; U-Net; Down-Sampling; Subpixel Convolution Layer.
1. Introduction
Recently, as artificial intelligence research has been actively conducted, research on semantic segmentation
was conducting in various fields. Among various studies, research on deep learning-based technology has
developed rapidly in the semantic segmentation. For deep learning-based semantic segmentation, methods
using Convolutional Neural Network (CNN) are being studied. Representative methods include Fully
Convolutional Networks (FCN) [1], U-Net [2], and SegNet [3]. Important problems in semantic segmentation
are recognition of small objects, maintenance of detailed content of images, and localization of objects in
images. Several methods have been studied to solve these problems [4].
These problems increase errors in the location and accuracy of objects in semantic segmentation
applications. To solve these problems, in this paper, the pooling layer of U-Net is changed to down-sampling
and the transposed convolution layer is changed to the Subpixel Convolution Layer proposed by ESPCN.
Through this, research is conducted to solve problems and increases segmentation accuracy.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２０４
8th ICAEIC-2022
2. Related work
2.1. U-Net
U-Net is a Fully convolutional network-based model proposed for image segmentation in the biomedical
field [2]. U-Net consists of a network for obtaining the overall context information of an image and a network
for accurate localization in a symmetrical form. Contracting path aims to capture context while reducing the
size of the input image and Expanding path aims to constitute detailed localization while increasing the size of
the feature map.
2.2. Maxpooling and Down-Sampling
Since maxpooling reduces the number of parameters in the training of the model, the expressive power of
the network decreases and overfitting is suppressed [5]. And the computation decreases in proportion to the
decreasing parameter, saving hardware resources and increasing training speed. However, the detailed content
of the image is lost as much as the image size that decreases in maxpooling.
Downsampling is to reduce the size of the image, and in this paper, four feature maps are created in which
only the height and width of the input image are reduced by half. Each feature map consists of pixels located
in indexes which are even rows and columns, even rows and odd columns, odd rows and even columns, and
odd rows and columns.
2.3. Subpixel Convolution Layer
The task of ESPCN is to estimate a SR image like HR image given a LR image downscaled from the
corresponding original HR image [6]. The LR image is upscaled to the original HR image size using the
Subpixel convolution layer. Subpixel convolution layer increases the number of channels of LR by r², and then
combines the feature maps to create an HR image. LR and created HR image are represented as real-valued
tensors of size H × W × C and rH × rW × C, respectively. r is the upscaling ratio.
3. Proposed Network Architecture
Figure 1 shows the Proposed model structure. Proposed model constructs a model with Encoder-Decoder.
The encoder reduces the size of the image and expands the image to the original size through the decoder. The
contracting path serves to reduce the size of the image [2].
The contracting path repeats two 3x3 convolutions and one down-sampling. Each iteration doubles the
number of channels in the image. Down-sampling reduces the size of the image by half and creates four feature
maps. Two out of four feature maps generated are input to the next layer, and the remaining two feature maps
are delivered through skip connection to the symmetrical layer based on the blue arrow. The process of
expanding the size of the image is repeated by two 3x3 convolution, Concatenation and Subpixel convolution
layer. The concatenation connects the feature map entered as an input and the feature map transmitted through
the skip connection in down-sampling of the contraction path. When concatenating, a skip-connected feature
map is connected in the middle of the channel of the input feature map. Subpixel convolution layer extends
２０５
AEICP Vol. 5, No. 1
the concatenated feature maps to the size of the feature map prior to down-sampling of the contracting path.
In the last layer, each channel is mapped to the suitable label classes using 1x1 convolution. There are 20
convolution layers in the network.
Figure 1. Maxpooling and Transposed convolution of the U-Net model is changed to Down-Sampling and
Subpixel convolution layer. Down-Sampling has the advantage of Pooling, and there is no computational
process. Subpixel convolution layer restores the feature map before Down-Sampling of the symmetrical
layers.
4. Experiment
4.1. Dataset
During the evaluation, Cityscapes, a publicly available benchmark dataset, was used. Cityscapes is a large-
scale database with a focus on semantic understanding of urban street scenes [1]. It contains a diverse set of
stereo video sequences recorded in street scenes from 50 cities, with high quality pixel-level annotation of 5k
frames, in addition to a set of 20k weakly annotated frames. It includes semantic and dense pixel annotations
of 30 classes, grouped into 8 categories—flat surfaces, humans, vehicles, constructions, objects, nature, sky,
and void. LeftImg8bit and gtFine are used among the Cityscapes datasets for evaluation. It includes 2975 train
images and 500 validation images.
4.2 Data Augmentation
The data used in the evaluation have the same trainId for several classes. Labels with trainId of 255 or less
than 0 are ignored so that they are not used for evaluation. Additionally, various random transformations
２０６
8th ICAEIC-2022
(random color/brightness change, random horizontal flip, random crop/scaling, etc.) are applied to increase the
data to be used for training. Some transformations should be applied to both the input image and the target
label map. When the image is turned over or cut, the same should be performed on the label map.
4.3 Results
Figure 2 shows the measurement results of mean-IoU and validation mean-IoU performance for the
proposed model. The blue and green line represents mean-IoU, and the orange and red line represents
validation mean-IoU. As a result of measuring to U-Net and proposed model up to 1000 epoch and epoch 1600,
U-Net achieved 75.8% mean-IoU and 77.3% validation mean-IoU, and proposed model achieved 84.3% mean-
IoU and 83.8% validation mean-IoU. The proposed model recorded better performance, and the segmentation
accuracy continued to increase.
Figure 2. Result of the mean-IoU performance of the proposed model.
5. Conclusion
In this paper, a method using down-sampling and Subpixel Convolution Layer was proposed to improve
accuracy in the U-Net based semantic segmentation. As a result, the segmentation accuracy of the proposed
method was improved. And we confirmed that the performance of the proposed model can be improved
without the loss of detailed content and localization of the original image. Additionally, as a future task, find
２０７
AEICP Vol. 5, No. 1
the hyperparameter through experiments, and confirm whether the difference in performance is significant
through T-test based on the results obtained through experiments several times.
References
[1] Jonathan Long, Evan Shelhamer, Trevor Darrell(2015), “Fully Convolutional Networks for Semantic
Segmentation“ -3431-3440
[2] Olaf Ronneberger, Philipp Fischer, Thomas Brox(2015), “U-Net: Convolutional Networks for Biomedical
Image Segmentation” -8.
[3] V. Badrinarayanan, A. Kendall, and R. Cipolla, "Segnet: A deep convolutional encoder-decoder
architecture for image segmentation," IEEE transactions on pattern analysis and machine intelligence, vol.
39, no. 12, pp. 2481–2495, 2017.
[4] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri
Terzopoulos(2020), “Image Segmentation Using Deep Learning: A Survey” -22
[5] Sara Sabour, Nicholas Frosst, Geoffrey E. Hinton(2017), “Dynamic Routing Between Capsules” -11.
[6] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew, P. Aitken, Rob Bishop, Daniel
Rueckert, Zehan Wang(2016), “Real-Time Single Image and Video Super-Resolution Using an Efficient
Sub-Pixel Convolutional Neural Network”
２０８
8th ICAEIC-2022
A Way of Implementing Brain Password System based on Graphical Password
Gi-Chul Yang
Department of Convergence Software, Mokpo National University, Chonnam, Korea
gcyang@mokpo.ac.kr
Abstract
This paper introduces a way of Implementing more secure password system by using user's brain waves. Brain
waves are not appeared to outside. So, it can be more secure and hard to steal, yet the signals are weak to
detect. Explained system, BPGP, utilizes the visually related brain wave signals which can detect from
occipital lobe area using visual function in order to figure out the intention of the user through brain wave
signals easily and reliably. BPGP uses brain wave signals not as the pass secrete itself but as a tool to
implement a secure password system.
Keywords: barin passwowd, graphical password system, security problem, security awareness.
1. Introduction
Password systems should have strong security and high usability. Currently the most widely used password
system type is alphanumeric password. Alphanumeric password cause problems from the security and usability
aspects. It is cumbersome to remember and weak against shoulder-surfing attack. Biometric authentication
system can be an alternative system of alphanumeric password system. It is, however requires special devices
to read user's biometric information such as finger prints and iris, so implementation cost of the system will be
increased and creates problems of handling necessary devices. Moreover, it has critical problem of non-
changeability. If you lose your password you need to change your password to a new password. But it is
impossible to change your biometric information. How about a graphical password? Graphical password has
advantages over current alphanumeric password and biometric password in both security and usability aspects.
However, it is also week against shoulder-surfing attack. There are meaningful attempts to overcome this
problem [1].
This paper introduces a way of Implementing more secure password system. The password system using
user's brain waves has been studied in various ways [2,3,4]. Brain wave is a kind of biometric information too.
hence it inherits the same problem of biometric systems. However, the brain waves are not appeared to outside.
So, it can be more secure and hard to steal, yet the signals are weak to detect. It is difficult to figure out user's
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２０９
AEICP Vol. 5, No. 1
intention by detecting weak brain wave signals. That means it is hard to use brain wave signals directly as a
password. The system introduced in this paper is called BPGP (Brain Password system based on Graphical
Password). BPGP uses brain wave signals as a tool for manipulating the graphical password system not as a
password itself. This idea was introduced in [5] by G.-C.Yang. This paper describes an efficient way of
implementing BPGP. Next section presents the way of implementing BPGP. The paper concludes in section
3.
2. BPGP
Brain password system introduced in [5] provides good idea of developing a secure password system.
However it has not presented the implementation details. This section describes the implementation details of
BPGP. Construction idea of BPGP is shown in Fig. 1. [5].
Figure 3. Flow diagram of BPGP.
As shown in Fig.1. The first step of registration or login process is BPGP is getting user's brain wave signals
and figure out user's intention. BPGP use non-invasive dry device for usability though it has low resolution.
BPGP needs to find the features of user's brain wave signal right after detecting the signal to decide what is
user's intention. There are various ways of feature detection techniques [6]. BPGP use simple and well known
technique in order to make BPGP reliable and practical system. BPGP utilizes the visually related brain wave
signals which can detect from occipital lobe area using visual function in order to figure out the intention of
the user through brain wave signals easily and reliably.
With the visually related brain wave signals rather than other brain wave signals, it is relatively easier to
detect the signal and figure out user's simple intentions such as choose one out of few targets. For example,
one can control a robots direction left, right, forward, and backward by watching four different flickering
buttons which are flickering with different frequencies each other. This way of brain wave signal detection is
well studied techniques [7,8] and relatively easy and reliable. The visually related brain wave signal can be
detected in Occipital Lobe (Oz area of a brain) as shown in Fig. 2.
２１０
8th ICAEIC-2022
Figure 2. Oz area in International 10-20 system.
Graphical Password Interface with flickering buttons (i.e., faces in Fig. 3) is shown in Fig. 3.
Figure 3. Example of a Graphical Password Interface with flickering buttons.
As shown in Fig. 3. each faces are flickering with different frequencies and a user selects a registered face
by focusing on that face. When a user focusing on a certain face corresponding brain waive is generated from
a user and select it as a password image. This process can be iterated in multiple times to increase the security
level.
4. Conclusion
This paper introduces a way of Implementing brain password system called BPGP. Like other personal
authentication system based on brain waves BPGP is a personal authentication that use the brain wave signals
but it does not use the brain wave directly as a password unlike other systems. BPGP just use the brain wave
signal to control the password system and the real secrete entities are the graphical information. In this way
we can build a reliable brain password system easily without warring about week brain wave signal power.
２１１
AEICP Vol. 5, No. 1
While BPGP maintains high security as other brain password systems since BPGP can hide the login process
into user's brain as well. The next paper will present the performance of BPGP.
Acknowledgement
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea
government (MSIT) (No. 2021R1F1A1046667).
References
[1] G. Yang, "Development Status and Prospects of Graphical Password Authentication System in Korea,"
KSII Transactions on Internet and Information Systems, vol. 13, no. 11, pp. 5755-5772, 2019. DOI:
10.3837/tiis.2019.11.026.
[2] Abo-Zahhad, M., Ahmed, S.M., Abbas, S.N.: ‘Biometric authentication based on PCG and ECG signals:
present status and future directions’, Signal Image Video Process., 2014, 8, (4), pp. 739– 751 (doi:
10.1007/s11760-013-0593-4)
[3] Delac, K., Grgic, M.: ‘A survey of biometric recognition methods’. Proc. Int. Symp. Electronics in Marine,
Zadar, Croatia, 2004, pp. 184– 193
[4] Klonovs, J., Petersen, C.: ‘Development of a mobile EEG-based biometric authentication system’, MS
thesis, Aalborg University, 2012
[5] G. Yang, ‘Brain Password System based on Graphical Password’, Proc. BIC2021, Jeju, Korea, 2021.08.
[6] Mohammed Abo-Zahhad, Sabah Mohammed Ahmed, Sherif Nagib Abbas, ‘State-of-the-art methods and
future perspectives for personal recognition based on electroencephalogram signals’, IET Biometrics,
Volume 4, Issue 3, pp. 179-190.
[7] Yeom, S.K., Suk, H.I., Lee, S.W.: ‘Person authentication from neural activity of face-specific visual self-
representation’, Pattern Recognit., 2013, 46, (4), pp. 1159– 1169 (doi: 10.1016/j.patcog.2012.10.023)
[8] Devue, C., Collette, F., Balteau, E., Degueldre, C., Luxen, A., Maquet, P., Brédart, S.: ‘Here I am: the
cortical correlates of visual self-recognition’, Brain Res., 2007, 1143, pp. 169– 182 (doi:
10.1016/j.brainres.2007.01.055)
２１２
8th ICAEIC-2022
A Study on Determining Wearing a Hard Hat and Face Identification
NaeJoung Kwak1, KeunWoo Lee2, and DongJu Kim3
1,2Division of AI Software Engineering (Information Security), Pai Chai University, Daejeon, Korea
3Department of the research division, POSTECH Institute of Artificial Intelligence
knj0125@pcu.ac.kr1, gly5937@pcu.ac.kr2, kkb0320@postech.ac.kr3
Abstract
Workers' activities in unsafe industrial sites are always exposed to many risks anytime, anywhere. In addition,
identification of workers by face recognition before they enter the workplace is an important issue in terms of
security. In this study, using Yolov5s and Facenet, we implement and verify a solution that can automatically
recognize whether or not a hard hat among personal safety protection equipment is worn and the identity of
the worker using face recognition. It was tested that the proposed model determinses whether or not to wear
a hard hat in the real environment and also confirms face identification well.
Keywords: helmet, personal protective equipments, safety, object detection, face detection, Identification.
1. Introduction
The injured part among the various body parts of workers on the job site is diverse, but the head accounts
for 48% of the total. Therefore, a major accident can be prevented just by protecting the head when entering
an industrial site. Wearing a hard hat that protects the head safely protects the head of the worker by mitigating
the impact caused by falling accidents, flying accidents, and falling objects and preventing electric shock [1].
Therefore, it is necessary to check the wearing of personal protective equipment for personal safety before the
worker enters the work site, and the identification of the worker must be simultaneously confirmed for security.
In this paper, we implement a system that checks whether workers are wearing a hard hat in real time before
entering the workplace and identifies their identity using face recognition. The hard hat is automatically
detected using the s-model of YOLOv5. For face identification, transfer learning was performed using facenet
[3], a face recognition model, and applied to real-time data. The proposed model determines whether or not to
wear the helmet well in the real environment, and also performs face identification well.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２１３
AEICP Vol. 5, No. 1
2. The proposed Method
For the safety and security of workers, we implements a model that checks whether a hard hat is worn before
entering the workplace and confirms the identity of the worker using facial recognition.
Using YOLOv5s, transfer learning was performed to detect whether a hard hat was worn, and the facenet,
a face recognition model, was trained with Masked VGG2 among the datasets provided in [4], and then the
identity was identified through transfer learning.
Figure 15.
System structure of the proposed method.
The experiment was conducted in the following environment on Ubuntu 16.04.7 LTS.
Table 2. Test environments.
OS
GPU
CPU
RAM
CUDA
cuDNN
software
Deep Learnig model
Ubuntu 18.04.5 LTS
Tesla V100-SXM2
16 core Intel(R) Xeon(R)
Gold 5120 CPU @ 2.20GHz
177GB
10.1
7.6.0
python/pytorch
YOLOv5 S model/Facenet
As the data set, Kaggle's 'YOLO helemt/head'[5] was used for determining whether or not to wear a hard
hat. All heads not wearing a hard hat were labelled as head, and heads wearing hard hats were labelled as
helmets. As the dataset for face recognition, Masked VGG2 among the datasets provided in [3] was used as
Facenet training data, and 9 classes of same persons taken with a smartphone were used as data for transfer
learning.
２１４
8th ICAEIC-2022
Figure 16.
The result of Helmet/Hed detection.
(a) Original (b) Result of (a)
Figure 3.
The result of
Helemet detection and face Identification.
Figures 2 and figure 3 show that using the proposed model, the Helmet/Head detection result and the face
identification result. The results show that wearing or not wearing of the helmet is well checked and face
identification is good.
3. Conclusion
In this paper, we implemented and verified a deep learning model that can determine whether a hard hat is
worn or not using transfer learning of the Yolov5s model and automatically recognize the identity of a worker
using transfer learning of Facenet. The proposed model checks whether or not to wear a hard hat and performs
woker’s identification well.
Acknowledgement
This work was supported by Institute of Information & communications Technology Planning & Evaluation
(IITP) grant funded by the Korea government(MSIT) (No.2020-0-02029-002)
References
[1] J. Constr. Eng. Manag.(2015), Hardhat-wearing Detection for Enhancing On-site Safety of Construction
Workers. 141
２１５
AEICP Vol. 5, No. 1
[2] ultralytics/yolov5. https://github.com/ultralytics/yolov5
[3] F. Schroff et al.(2015, June), FaceNet: A Unified Embedding for Face Recognition and Clustering.
Procedding of Conf. on Comput. Vision Pattern Recogn. (pp. 815-823). IEEE.
[4] Masked dataset :https://github.com/SamYuen101234/Masked_Face_Recognition
[5] YOLO helmet/head : https://www.kaggle.com/vodan37/yolo-helmethead
２１６
8th ICAEIC-2022
A Study on CNN-based Object Recognition Technology for Autonomous Vehicle
Seung-Yeon Hwang1, Dong-Jin Shin2, Mun-Yong Park3, Dae-Sun Yum4, and Sung-Youn Cho*
1,2,3,4,*Novacos Co., Ltd., Anyang, South Korea
{syhwang1, djshin2, pmy3, daesun12144, scho*}@novacos.co.kr
Abstract
Recently, a new virus disease called COVID-19 shas emerged, and much effort has been spared in developing
technologies to prevent the spread of the disease worldwide. In particular, with the development of big data
and artificial intelligence technologies, attempts to incorporate deep learning into technologies to prevent the
spread of diseases have increased. The government mandates people to wear indoor and outdoor masks to
prevent the spread of COVID-19. However, it is often the case that masks are not worn or worn properly.
Therefore, in this work, we want to study the Convolutional natural network (CNN)-based mask wear
recognition techniques to identify whether masks are worn or not. First, Data augmentation is performed to
secure sufficient datasets. Then, in order to analyze the influence of the contour image of the mask dataset on
neural network learning, the performance of the case where the contour image is trained and the case where
the contour image is not trained is compared.
Keywords: Convolutional neural network, COVID-19, Deep learning, Mask recognition
.
1. Introduction
The recent spread of the COVID-19 virus worldwide is causing a lot of social and economic damage. The
World Health Organization (WHO) declared a pandemic on COVID-19. Therefore, each country is making
great efforts to prevent the spread of the COVID-19 virus. In particular, research related to virus spread
prevention technologies that combine big data and artificial intelligence, which are rapidly developing in the
era of the 4th Industrial Revolution, is actively underway. The government implements social distancing
policies and mandates wearing masks to prevent the spread of the COVID-19 virus through respiratory organs.
However, when wearing a mask, it is often seen that the respiratory system is not completely covered or that
the mask is not worn. Therefore, in this work, we study CNN-based mask-wear recognition techniques to
recognize and classify whether masks are worn correctly or not.
In this work, we utilize a set of mask-wearing data collected from the web for experimental performance
evaluation. We also utilize ResNet among several CNN models (VGGNet [1], GoogleNet [2], ResNet [3],
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２１７
AEICP Vol. 5, No. 1
DenseNet [4], EfficientNet [5]) that have achieved superior performance in image recognition and
classification recently. In order to analyze the influence of the contour images on neural network learning in
mask datasets, performance comparisons are carried out depending on whether the contour images are learned.
Model learning strategies use transfer learning strategies to overcome overfitting that may occur during the
learning of models and to speed up learning.
This paper describes the CNN architecture and the outline image extraction techniques used in section 2 and
describes the datasets used in section 3 for experimental performance evaluation. Section 4 describes the
experimental results and analysis, concludes the paper in Section 5, and presents future research directions.
2. Related Works
2.1 Convolutional Neural Networks
CNN is a neural network model that is primarily used to process image and image data. In general, deep
natural network (DNN) uses one-dimensional data. Processing two-dimensional shapes using DNN requires a
process of transforming them into one-dimensional shapes, which are inefficient in feature extraction and
learning and limitations in increasing accuracy due to the loss of spatial/local information on the images.
However, in the case of CNN, the dimensions of the image are utilized without deformation, enabling learning
while maintaining spatial/local information, enabling efficient image feature extraction and learning.
2.2 ResNet101
ResNet was developed by Microsoft as a neural network that won the ImageNet Large Scale Visual
Registration (ILSVRC) competition in 2014. ResNet uses residual blocks to add the input value of each layer
to the output value when training the neural network. In this study, ResNet101, which consists of 101 layers,
is used among the 5 models of ResNet. Figures 1 and 2 visualize the Residual block and ResNet architecture.
Figure 1.
Sample images of mask dataset for 3 classes.
２１８
8th ICAEIC-2022
Figure 2.
Sample images of mask dataset for 3 classes.
2.3 Canny Edge Detection [6]
Canny Edge Detection was developed by John F. Canny in 1986 and is one of the most popular edge
extraction algorithms. The canny edge detection process consists of five stages: noise reduction, gradient
calculation, non-maximum suppression, double threshold, and edge tracking by hysteresis. Table 1 shows the
details for Canny edge detection.
Table 1. Details of Canny edge detection.
Step
Description
Noise reduction
If there is noise in the image, it is difficult to find the edge properly, so use
Gaussian filters to reduce the noise in the image.
Gradient calculation
To find the point where the pixel value changes rapidly, calculate the gradient
to check the strength and direction of the edge.
Non-maximum suppression
To remove pixels that did not contribute to Edge, scan the entire image and
convert the value of those pixels to zero.
Double threshold
By specifying the maximum and minimum thresholds, we identify the strong
and weak pixels of the edge.
Edge tracking by hysteresis
Determine whether it is egde or not for pixels between the maximum and
minimum thresholds.
3. Datasets
In this study, the Mask dataset collected from the web is used. The mask dataset consists of a total of 14,771
images and is divided into three classes: ‘incorrect mask’, ‘with mask’, and ‘without mask’. In order to secure
sufficient data, the amount of data was increased by rotating, zooming, flipping, and shifting the collected
images. Figure 3 shows sample images for three classes of mask dataset.
２１９
AEICP Vol. 5, No. 1
Figure 3.
Sample images of mask dataset for 3 classes.
4. Experimental Results and Discussion
4.1 Performance Evaluation Environment and Hyper-parameter Settings
Experiments performed in this study were performed on the Windows 10 OS and implemented using the
Keras library. he hardware of the PC used in the experiment was NVIDIA GeForce RTX 2070 SUPER 8GB
GPU, AMD Ryzen 73700X 8-Core CPU and 32GB RAM.
In neural network learning, the ratio of training data to test datasets was set to 7:3. In addition, the learning
rate was set to 0.0001 due to the use of the pre-trained model, and the epoch was set to 10 due to the fast
convergence rate of the model using the Adam [7] optimizer.
4.2 Experimental Results and Analysis
Figure 4.
Results of performance comparison between learning and not learning contour images.
２２０
8th ICAEIC-2022
Figure 4 shows results of performance comparison between learning and not learning contour images. As a
result of training the model with the dataset not including the contour image, an accuracy of 95.09% was
achieved, and as a result of training the model on the dataset including the contour image, the accuracy of
98.85% was achieved. Since the accuracy of the case of learning the contour image is 3.76% higher than that
of the case of not learning the contour image, it is judged that learning the contour image is an important factor
in improving the performance of the neural network.
5. Conclusion
In this study, CNN-based mask wearing recognition technology was studied to classify whether the mask is
worn or not. And in order to analyze the influence of the contour image of the mask dataset on neural network
learning, performance comparison was performed with and without contour image learning. Through the
performance evaluation result, it was proved that learning the contour image is an important factor in
improving the performance.
As a future study, we plan to conduct experiments using contour images of famous benchmark datasets such
as CIFAR10, CIFAR100, and MNIST to check whether performance improvement can be achieved regardless
of the type of dataset as well as the mask dataset.
Acknowledgement
This work was supported by Korea Institute of Police Technology (KIPoT) grant funded by the Korea
government (KNPA) (No. 1325163981, Development of Advanced Technology for Perception Improvement
on Traffic Entities and Risk Mitigation on Adverse Driving Conditions for Lv.4 Connected Autonomous
Driving) in 2021.
References
[1] Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556.
[2] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. (2015). Going
deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 1-9).
[3] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings
of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
[4] Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convolutional
networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700-
4708).
[5] Tan, M., & Le, Q. (2019, May). Efficientnet: Rethinking model scaling for convolutional neural networks.
In International Conference on Machine Learning (pp. 6105-6114). PMLR.
２２１
AEICP Vol. 5, No. 1
[6] Canny, J. (1986). A computational approach to edge detection. IEEE Transactions on pattern analysis and
machine intelligence, (6), 679-698.
[7] Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
２２２
8th ICAEIC-2022
Anti-Collision Control System for Unmanned Aerial Vehicle based on
Reinforcement Learning
Chung-Pyo Hong
Division of Computer Engineering, Hoseo University, Republic of Korea
cphong@hoseo.edu
Abstract
Unmanned aerial vehicles (“UAV”) have become popular for industrial, public, and other domains in recent
years. Based on these circumstances, there is a need for control methods for preventing UAV quadcopters
from colliding with nearby obstacles by stopping rapidly. This study proposed and tested a system to reduce
braking distance by switching proportional, integral, and derivative parameters dynamically based on sensor
data.
Keywords: Quadcopter; Artificial intelligence; PID Control; Parameter Tuning.
1. Introduction
With the growth of the popularity of unmanned aerial vehicles (“UAV”) and the development of UAV
technology, demand for micro UAVs, such as quadcopters, is replacing demand for large, fixed-wing UAVs
[1]. The development of autonomous flight systems that automatically and safely fly micro UAVs is a
significant focus of research. Therefore, this study proposed an emergency obstacle avoidance system for
avoiding collisions with objects that are only detected by short-range obstacle sensors. Our method
dynamically changed proportional, integrative, and derivative (“PID”) parameters to first provide a large
backward thrust and to then stabilize the quadcopter to prevent it from overturning, which ultimately reduced
the braking distance by 60.5%. The remainder of this paper consists of four parts. Chapter 2 explains the trend
of related research, and Chapter 3 explains the contents of the proposed scheme. Section 4 describes the
experimental results, and finally, Section 5 discusses the conclusion.
2. Related Work
Most recent quadcopter control studies have focused on generating robust and precise control systems to
overcome problems such as wind turbulence using mostly PID- or PD-based control systems [2] as well as
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２２３
AEICP Vol. 5, No. 1
fuzzy [3] and MPC control systems [4]. McClamroch et al. [5] proposed a nonlinear robust control system for
tracking applications. Anavatti et al. [6] proposed a robust hybrid nonlinear control system which enabled a
quadcopter to be robust against external disturbances such as wind. However, while these studies’ proposed
systems increased quadcopter robustness, they were not intended to help quadcopters respond to emergency
situations which require more dynamic and aggressive systems. We need to tilt the quadcopter as much as
possible to gain enough braking force while not losing control. Therefore, we need an additional method to
gain control even when the quadcopter is tilted greatly. Thus, this study proposed a method for quickly tuning
PID parameters by focusing on avoiding nearby objects.
3. Proposed Scheme
This study proposes a two-level PID parameter convert system to allow quadcopters to engage in sudden
braking to avoid objects. In the proposed system, whenever sensors, which could be either cameras, obstacle
sensors, or LIDAR sensors, detects any obstacle in front of the quadcopter, the system would attempt to add
force in the opposite direction of the quadcopter’s current direction of travel to minimize braking distance and
avoid a collision. However, this system could cause the quadcopter to crash if it is not accompanied by precise
parameter optimization. Therefore, a gradient evolutionary algorithm was also proposed which used a
reinforcement learning-based method to optimize the parameters required to enable fuzzy quadcopter control.
This method changes PID parameters during runtime with two levels. In the first level, the system significantly
changes PID parameters to add a large negative thrust force on the quadcopter without significant concern for
maintaining balance. The proportion control parameter would be expected to rise significantly because it is
directly associated with PID control response time. In the second level, the PID parameters are changed to
mainly focus on fixing the quadcopter’s unbalanced state. Figure 1 shows a diagram of the system’s functional
flow. During this process, all user control functions are disabled.
Figure 1.
Proposed two-level PID control system overview.
２２４
8th ICAEIC-2022
4. Experimental Result
To evaluate our proposed scheme, we adopted conventional PID scheme and one-level PID parameter
convert system as comparatives. Figure 2 shows the braking distance produced by our proposed scheme and
two comparatives. The PID parameters obtained by sensors are trained using the proposed gradient
evolutionary method to produce the best results possible. This strategy resulted in better performance than
conventional PID scheme and one-level PID parameter convert system. Our proposed scheme shows the
relatevely shorter braking distance by 60.5%, and 86.9% compared with the conventional and one-level scheme,
respectively.
Figure 2.
Braking distance comparison.
5. Conclusion
This study proposed a two-level PID parameter convert system with a corresponding reinforcement learning
method to properly train the system’s parameters. Simulations were performed to determine how the system
performes relative to traditional methods. The proposed scheme shows better braking distance in response to
unexpected obstacles during the flight. The experimental result shows the relatevely shorter braking distance
by 60.5%, and 86.9% compared with the conventional and one-level scheme, respectively.
Acknowledgement
This research was supported by the MIST(Ministry of Science and ICT), Korea, under the National Program
for Excellence in SW(No. 20190018340021001) supervised by the IITP(Institute of Information &
communications Technology Planning & Evaluation)(No. 20190018340021001)
２２５
AEICP Vol. 5, No. 1
References
[1] Beard, Randal W., et al. (2005) Autonomous Vehicle Technologies for Small Fixed-wing UAVs. Journal
of Aerospace Computing, Information, and Communication 2(1), 92–108.
[2] Luukkonen, Teppo.(2011) Modelling and Control of Quadcopters. Independent Research Project in
Applied Mathematics, Espoo 22.
[3] Domingos, Diego, Camargo, Guilherme, and Gomide, Fernando. (2016) Autonomous Fuzzy Control and
Navigation of Quadcopters. IFAC-PapersOnLine 49(5), 73–78.
[4] Santoso, Fendy, et al. (2018) Robust Hybrid Nonlinear Control Systems for the Dynamics of a Quadcopter
Drone. IEEE Transactions on Systems, Man, and Cybernetics: Systems 99, 1–13.
[5] Lee, Taeyoung, Leok, Melvin, and McClamroch, N. Harris. (2013) Nonlinear Robust Tracking Control
of a Quadcopter UAV on SE. Asian Journal of Control 15(2) ,391–408.
[6] Mitsukura, Yasue, Yamamoto, Toru, and Kaneda, Masahiro. (1999) A Design of Self-tuning PID
Controllers Using a Genetic Algorithm. American Control Conference 1999.
２２６
8th ICAEIC-2022
Super Resolution Method using Parallel Attention Mechanism Architecture
Dongwoo Lee1, Kyeongseok Jang2, Chae-bong Sohn3, Bhanu Shretha4, Seongsoo Cho5, and
Kwang Chul Son*
1, 2Department of Plasma Bio Display, Kwangwoon University, 20, Gwangun-ro, Nowon-gu, Seoul, Korea
3Department of Electronics and Communications Engineering, Kwangwoon University, 20, Gwangun-ro,
Nowon-gu, Seoul, Korea
4Department of Electronic Engineering, Kwangwoon University, 20, Gwangun-ro, Nowon-gu, seoul, Korea
5School of Software, Soongsil University, 06978, Seoul, Korea
* Department of Information Contents, Kwangwoon University, 20, Gwangun-ro, Nowon-gu, Seoul, Korea
led0121@kw.ac.kr1, ksjang1234@kw.ac.kr2, bnu@kw.ac.kr4, css3617@gmail.com5, kcson@kw.ac.kr*
Abstract
Super-resolution is a method of upscaling a low-resolution image to a high-resolution image. The conventional
super-resolution method needs to emphasize the high frequency region and limit the low frequency region.
Low resolution images are featureless in restoring to high resolution images. This paper proposes a structure
that combines channel attention and special attention to make up for the missing features and a method of
using subpixel convolution to solve the problem of checkerboard artifacts. The proposed method showed
improved results over traditional Super Resolution.
Keywords: Super-resolution, Channel Attention, Spatial Attention, Sub-pixel Convolution.
1. Introduction
High-resolution monitors, AR, and VR devices require high-resolution content. However, high-resolution
video production with existing content has high costs and low-resolution problems. In order to solve these
problems, a Super Resolution method for enlarging a low-resolution image and various image processing
techniques [1-3] have been proposed. Interpolation is a conventional super-resolution method, and recently, a
method applying the Convolution Neural Network (CNN) has been proposed. In this paper, we improved
feature enhancement and inhibition by combining Channel Attention and Spatial Attention, and proposed
Super Resolution combined with Sub-pixel Convolution.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２２７
AEICP Vol. 5, No. 1
2. Related Works
2.1. Attention Mechanism
Attention Mechanism has been extensively studied in the fields of brain science and Neural Nerowk [4].
When applied in the RNN for the first time, it was used to emphasize the characteristics of the previous step.
Research is underway to apply these features to CNN as well. This is because it is generally more effective to
emphasize the most relevant features than to use all the information to recognize the object. Attention
Mechanism consists primarily of Channel Attention and Spatial Attention based on Channel and Spatial,
respectively. Figure 1 is a flowchart of Channel Attention and Spatial Attention.
(a) Channel Attention (b) Spatial Attention
Figure 1.
Attention Mechanism Flowchart.
3. Proposed Method
In this paper, we proposed Super Resolution, which connects Channel Attention, which is Attention
Mechanism, and Spatial Attention in series, and combines Sub-pixel Convolution to extend the feature map
[5]. Features were extracted from the input image through the first convolution. After then, we used an
Attention Block that combines Channel Attention and Spatial Attention. This supplemented lacks expansion
from low resolution to high resolution through emphasis. After then, the feature map was expanded through
the sub-pixel convolution. At this time, one Sub-pixel Convolution was used to magnify the feature map twice
and thrice, and two Sub-pixel Convolutions were used to magnify the feature map four times. Figure 2 shows
a flowchart of the proposed method.
Figure 2.
Proposed Flowchart.
２２８
8th ICAEIC-2022
A flowchart of the Attention Block of the proposed method is shown in Figure 3. By combining Channel
Attention and Spatial Attention in parallel, we supplemented the missing features through extraction of high-
frequency features.
Figure 3.
Attention Block Flowchart.
The loss function is a measure for learning a Neural Network, and the Network modifies the weights in the
direction of minimizing the loss function. In this paper, learning is advanced through the network using L1
Loss, and equation (1) is the L1 Loss used in the proposed Network. In Equation 1, is a high-resolution
image patch scaled down via Bicubic Interpolation [6], and () is the proposed Network. Also, is the
original patch image and is the total number of embedded patches in the dataset.
()=|()|
(1)
4. Experimental Results
Figure 4.
Result Image.
２２９
AEICP Vol. 5, No. 1
The proposed network used the DIV 2K Dataset for training by dividing it into patches. The DIV 2K Dataset
consists of 800 high resolution images from different environments. Experiments were conducted with Set5,
Set14, B100 and Urban100. A comparison of existing bicubic interactions with the proposed method is
depicted in Figure 4. The structure improved by Attention Block using Attention Mechanism is shown in
Figure 4 and it showed the emphasis on high frequency range.
5. Conclusion
In this paper, we proposed a method of combining Channel Attention and Spatial Attention with Sub-pixel
Convolution to improve super-resolution. The Super Resolution process, which magnifies high-resolution
images from low-resolution images, applied the Attention Mechanisms Channel Attention and Spatial
Attention to compensate for defective features. These supplemented features have been extended to the feature
map by Sub-pixel Convolution. The proposed method showed improved results over traditional super-
resolution methods.
References
[1] Lee, D. W., Lee, S. H., Han, H. H., & Chae, G. S. (2019). Improved Skin Color Extraction Based on Flood
Fill for Face Detection. Journal of the Korea Convergence Society, 10(6), 7-14.
[2] Lee, D. W., Lee, S. H., & Han, H. H. (2020). Deep Learning-based Super Resolution Method Using
Combination of Channel Attention and Spatial Attention. Journal of the Korea Convergence Society,
11(12), 15-22.
[3] Lee, D. W., Lee, S. H. & Han, H. H. (2019, November). A Study on Super Resolution Method using
Encoder-Decoder with Residual Learning. Journal of Advanced Research in Dynamical and Control
Systems, 11(7), 2426-2433.
[4] Woo, S., Park, J., Lee, J. Y., & Kweon, I. S. (2018). Cbam: Convolutional block attention module. In
Proceedings of the European conference on computer vision (ECCV) (pp. 3-19).
[5] Shi, W., Caballero, J., Huszár, F., Totz, J., Aitken, A. P., Bishop, R., ... & Wang, Z. (2016). Real-time
single image and video super-resolution using an efficient sub-pixel convolutional neural network. In
Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1874-1883).
[6] Keys, R. (1981). Cubic convolution interpolation for digital image processing. IEEE transactions on
acoustics, speech, and signal processing, 29(6), 1153-1160.
２３０
8th ICAEIC-2022
Sentiment Analysis of Public Opinion on COVID-19 Vaccines on Twitter Social
Media using Naive Bayes Classifier Method
Arini1, Saepul Aripiyanto2, Alvian Aristya3, and Iik Muhamad Malik Matin4
123Informatics Department, UIN Syarif Hidayatullah Jakarta, Jakarta, Indonesia
4Electrical Engineering, Garut University, Garut, Indonesia
arini@uinjkt.ac.id1, saepul.aripiyanto@uinjkt.ac.id2, Alvian.aristya2017@mhs.uinjkt.ac.id,
iikmuhamadmalikmatin@uniga.ac.id4
Abstract
A year since the Sars-Cov-2 virus began to enter Indonesia, many opinions have been developed en masse
regarding the Covid Vaccine through social media networks, one of which is Twitter. Due to the large number
of public opinions about the good and bad of the vaccine. In sentiment analysis research on the Covid Vaccine,
a strong classifier is needed in order to get a high accuracy value. One of the classifiers is the Naïve Bayes
Classifier (NBC) Algorithm. This study aims to implement the Naive Bayes Classifier Algorithm to classify
positive and negative sentiments on Covid Vaccine tweets. Several stages were carried out to analyze sentiment,
namely the data collection stage, data preprocessing, and classification using Naive Bayes Classifier. Using
3 scenarios for training data, the first scenario uses 400 data, the second scenario uses 600 data and the
scenario uses 800 data. The results of this study show that the naive bayes classifier can be implemented for
sentiment analysis and the highest accuracy value is obtained in the third scenario, namely 90% accuracy, 86%
precision, 90% recall and 89% f1-score.
Keywords: Naive bayes Classifier, Classification, Sentiment Analysis, covid-19, vaccine.
1. Introduction
The development of the internet allows the dissemination of information quickly, widely, and easily. This
is used by the mass media to meet the public's need for information. Based on the results of research from the
Association of Indonesian Internet Service Providers in 2016 on the behavior of internet users in Indonesia,
internet users in Indonesia have reached 132.7 million of the total population of Indonesia which reaches 256.2
million people. A total of 97.4 percent of people use the internet to access social media, and 96.4 percent, or
equivalent to 127.9 million users also use it to access news. [1]. Among the various motives of internet users
in accessing social media, one of them is the existence of supporting facilities that allow users to access and
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２３１
AEICP Vol. 5, No. 1
interact in social media. [2]. Social media can provide information that can be used to predict and explain
information about health [3]. Data on social media can be extracted to obtain health information, one of which
is on social media platforms such as twitter [4]. Twitter allows researchers to obtain user-generated samples
to track initial response information and understand community assessments of health issues [5].
Covid-19 has infected millions of people in the world, including Indonesia [6]. At that time the world worked
together to create a vaccine. Currently, vaccines are the only solution to reduce the spread of COVID-19.
Various communities support this vaccine program. However, the effectiveness of the vaccine is doubted by
some people, causing concern. This is because the Covid vaccine is relatively new. Misinformation also
increases anxiety, fear, confusion, anger and a variety of negative emotions [7].
Public perception is very important in measuring the success of vaccination. For this reason, sentiment
analysis is needed to assess public perceptions in response to the presence of the COVID-19 vaccine. In this
paper, we analyze public sentiment regarding the presence of the COVID-19 vaccine. We analyze sentiment
on Twitter social media. We determine the class of information based on negative and positive sentiments,
then classified to find out public sentiment.
2. Research Method
2.1. Gathering Data
We use data from Twitter using the Twitter API about netizens' tweets against the Covid Vaccine in the
period 5 to 11 June 2021, with the Twitter developer feature accessed on the website https://developer.twitter.
com/. The data we use is 850 tweets and then stored in CSV form.
2.2. Split Data
We divided the data into two parts, namely training data and test data. In this study, we divide into 3
scenarios so that each scenario has training data of 400, 600, and 800 respectively with 50 test data.
2.3. Preprocessing
Preprocessing is used to change the form of raw data into data that is suitable for analysis. The purpose of
this pre-processing is to remove noise, uniform word form and reduce word volume.
2.4. Classification
Classification is used to analyze sentiment. Classification consists of 2 classes of sentiment, namely
negative and positive. We use the Naïve Bayes algorithm to classify each class. Nave Bayes classifies positive
and negative sentiment in scenario 1, scenario 2, and scenario 3.
2.5. Output Analysis
Output analysis is done by recapitulating the classification results and mapping the number of classification
results in the confusion matrix table. The confusion matrix consists of true positive, true negative, false positive
and false negative. The results of the confusion matrix can be calculated for accuracy, precision, recall, and
f1-score with the following formula:
２３２
8th ICAEIC-2022
Akurasi =
(1)
Precission =
(2)
Recall =
(3)
F1 Score = ()
() (4)
3. Result and Disscussion
3.1
.
Sentiment Classification Result
We classify sentiment based on data from 50 tweets of test data taken from Twitter. We use the Covid
Vaccine Keyword to determine the tweets to be taken. Then they are classified using the Naïve Bayes Classifier
Algorithm and the Lexicon Based method.
In this test, we divide it into 3 scenarios with each scenario having a varying amount of test data. the
following is the amount of test data for each scenario:
1. Scenario 1 with 400 training data
2. Scenario 2 with 600 training data
3. Scenario 3 with 800 training data
In table 1. shows the results of the classification and compares them with the actual sentiment.
Table 1. The classification results generated by Naïve Bayes.
Data
Scenario 1
Scenario 2
Scenario 3
True Class
Data
Scenario 1
Scenario 2
Scenario 3
True Class
1
Positive
Positive
Positive
Positive
26
Positive
Positive
Positive
Positive
2
Negative
Negative
Positive
Positive
27
Negative
Negative
Negative
Negative
3
Negative
Negative
Positive
Positive
28
Negative
Negative
Negative
Negative
4
Positive
Positive
Positive
Positive
29
Negative
Negative
Negative
Negative
5
Positive
Negative
Positive
Negative
30
Positive
Positive
Positive
Positive
6
Positive
Negative
Positive
Positive
31
Negative
Positive
Positive
Negative
7
Negative
Negative
Negative
Negative
32
Negative
Negative
Negative
Negative
8
Negative
Positive
Negative
Positive
33
Negative
Negative
Negative
Negative
9
Positive
Positive
Positive
Positive
34
Negative
Negative
Negative
Negative
10
Negative
Negative
Negative
Negative
35
Positive
Positive
Positive
Positive
11
Positive
Positive
Positive
Positive
36
Negative
Negative
Negative
Negative
12
Negative
Negative
Negative
Negative
37
Negative
Negative
Negative
Negative
13
Positive
Positive
Positive
Positive
38
Positive
Positive
Positive
Positive
14
Positive
Negative
Negative
Negative
39
Negative
Negative
Negative
Negative
２３３
AEICP Vol. 5, No. 1
15
Negative
Negative
Negative
Negative
40
Negative
Negative
Negative
Negative
16
Negative
Negative
Negative
Negative
41
Negative
Negative
Negative
Negative
17
Positive
Positive
Positive
Positive
42
Negative
Negative
Negative
Negative
18
Positive
Positive
Positive
Positive
43
Negative
Negative
Negative
Negative
19
Negative
Negative
Positive
Negative
44
Negative
Negative
Negative
Negative
20
Positive
Positive
Positive
Positive
45
Negative
Negative
Negative
Negative
21
Positive
Positive
Negative
Negative
46
Positive
Positive
Positive
Positive
22
Negative
Negative
Negative
Negative
47
Positive
Positive
Positive
Positive
23
Positive
Positive
Positive
Positive
48
Positive
Positive
Negative
Negative
24
Negative
Negative
Negative
Negative
49
Positive
Positive
Positive
Positive
25
Positive
Negative
Negative
Positive
50
Negative
Negative
Negative
Negative
The results of sentiment classification using the Naïve Bayes Algorithm in scenario 1 obtained 22 test data
with positive sentiment and 28 test data with negative sentiment. While the classification results in scenario 2
obtained 20 test data with positive sentiment and 30 test data with negative sentiment. In the last scenario,
there are 22 test data with positive sentiment and 28 test data with negative sentiment. Fig. 1 shows the
difference in the number of positive and negative sentiments in scenario 1, scenario 2, and scenario 3.
Figure 1.
The difference in sentiment for each scenario.
Based on Fig. 1. It can be seen that the number of negative sentiments has more than positive sentiments.
While the difference in sentiment for each scenario does not have a significant difference.
3.2 Accuracy
We use a multiclass confusion matrix to determine the group of sentiment classification results for each
feature selection (prediction class) and compare it with the results of the Lexicon Based classification (real
class). We calculated the number of values in the TPos, FPos, FNeg, TNeg categories. Here are the results of
the level of accuracy. Table 2 is the number of identified sentiments mapped based on the confusion matrix.
0
10
20
30
40
Scenario 1 Scenario 2 Scenario 3
Sentiment Result
Positive Negative
２３４
8th ICAEIC-2022
Table 2. confusion matrix on scenario 1.
Sentiment
Prediction Class
Positive
Negative
True Class
Positive
18
4
Negative
4
24
Table 2 shows scenario 1 getting match values of 18 true positives, 24 true negatives, 4 false positives and
4 false negatives. So it can be calculated:
Accuracy =
=84%
while the number of sentiments identified in scenario 2 is mapped based on the confusion matrix in table 3.
Table 3. confusion matrix on scenario 2.
Sentiment
Prediction Class
Positive
Negative
True Class
Positive
17
3
Negative
4
26
Table 3 shows scenario 2 getting match values of 18 true positives, 26 true negatives, 3 false positives and
4 false negatives. So it can be calculated:
Accuracy =
=86%
Table 4 shows the number of sentiments identified in scenario 3 mapped based on the confusion matrix.
Table 4. confusion matrix on scenario 3.
Sentiment
Prediction Class
Positive
Negative
True Class
Positive
19
3
Negative
2
26
Table 4 shows scenario 3 getting a match value of 19 true positives, 26 true negatives, 3 false positives and
2 false negatives. So it can be calculated:
Accuracy =
=86%
２３５
AEICP Vol. 5, No. 1
In addition to accuracy, we calculated precision, recall and F1-Score based on the confusion matrix in the
three scenarios. Table 5 shows the comparison of accuracy, precision, recall and f1 scores in scenario 1,
scenario2 and scenario 3.
Table 5. The classification results generated by Naïve Bayes.
Scenario
Accuracy
Precision
Recall
F1-Score
Scenario 1
84
81
81
81
Scenario 2
86
85
86
83
Scenario 3
90
86
90
89
Table 5 shows that scenario 1 has a score of 84% on the accuracy, 81% on precision, 81 on recall, and 81%
on f1-score. In scenario 2, it shows a score of 86% on the accuracy, 85% on precision, 86 on recall, and 83%
on the f1-score. While scenario 3 gets the highest score, which is 90% on the accuracy, 86% on precision, 90%
on recall, and 89% on the f1-score.
4. Conclusion
In this study, we conducted a sentiment analysis of the COVID-19 vaccine on Twitter social media using
the Naive Bayes classifier. We use Twitter data with training data variations of 400, 600, and 800 with 50 test
data. The results show that public sentiment related to the covid vaccine tends to be negative with the highest
accuracy achieved in scenario 3. In scenario 3 the accuracy is achieved with a score of 90%. the precision of
86%, recall of 90%, and f1-score of 89%.
References
[1] Association of Indonesian Internet Service Providers, “Penetrasi & Perilaku Pengguna Internet Indonesia,”
APJII, pp. 31–48, 2016.
[2] B. Nugraha and M. F. Akbar, “Perilaku Komunikasi Pengguna Aktif Instagram,” J. Manaj. Komun., vol.
2, no. 2, p. 95, 2019, doi: 10.24198/jmk.v2i2.21330.
[3] J. Ginsberg, M. H. Mohebbi, R. S. Patel, L. Brammer, M. S. Smolinski, and L. Brilliant, “Detecting
influenza epidemics using search engine query data,” Nature, vol. 457, no. 7232, pp. 1012–1014, 2009,
doi: 10.1038/nature07634.
[4] K. Jahanbin and V. Rahmanian, “Using twitter and web news mining to predict COVID-19 outbreak,”
Asian Pac. J. Trop. Med., vol. 13, no. 8, pp. 378–380, 2020, doi: 10.4103/1995-7645.279651.
[5] L. E. Charles-Smith et al., “Using social media for actionable disease surveillance and outbreak
management: A systematic literature review,” PLoS One, vol. 10, no. 10, pp. 1–20, 2015, doi:
10.1371/journal.pone.0139701.
[6] Worldometers, “COVID-19 Coronavirus Pandemic,” 2020.
２３６
8th ICAEIC-2022
[7] G. S. Putri, “Keraguan pada Vaksin Covid-19, Bagaimana Masyarakat Harus Bersikap?,” Kompas.com,
2020.
２３７
AEICP Vol. 5, No. 1
Design of BoP Big Data Processing System for BoP Data Analysis
Sun Park*, Byung-joo Chung1, ByungRea Cha2, and JongWon Kim3
*,2,3AI Graduate School, GIST
1Affiliated research institute, GSPIN
sunpark@gist.ac.kr*, jayeon9@gspin.co.kr1, brcha@gist.ac.kr2, jongwon@gist.ac.kr3
Abstract
The world is trying to reduce carbon dioxide emissions to solve the problem of global warming. Interest in
hydrogen fuel cells as an alternative energy source to reduce carbon emissions is growing. The biggest
influence on the efficiency of hydrogen fuel cells is BOP (Balance of Plant). BOP consists of fuel, air, heat
recovery pumps, blower, and sensor except the body of fuel cells (stack). In this paper, in order to increase
BoP efficiency, a system of the BoP big data processing based on IoT-Edge-BigData is proposed to store and
process data collected from BoP for BoP data analysis.
Keywords: BoP big data processing system, IoT-Edge-BigData, hydrogen fuel cell.
1. Introduction
As an alternative to solving energy and environmental problems such as climate change and resource
depletion, the importance of the hydrogen industry with high eco-friendliness and energy efficiency is
increasing. Hydrogen power generation (fuel cell) is an eco-friendly power source with high energy conversion
efficiency and less pollutants, noise and vibration. Hydrogen fuel cells produce electricity by electrochemically
reacting hydrogen and oxygen. The power generation unit (stack) of a hydrogen fuel cell is composed of two
poles and an electrolyte membrane between the poles. In the fuel cell, the auxiliary device except the stack is
the BoP. It plays the role of thermal management, water management, and air supply of the BoP. BoP consists
of eBOP (power converter, etc.) and mBOP (pump, compressor, etc.) [1-3].
This paper designed a BoP big data processing system that can analyze BoP data to increase BoP efficiency
of hydrogen fuel cells. The proposed BoP big data processing system can wirelessly collect data from BoP
devices and store them in edge devices for visualization. Data collected on each edge device can be stored and
analyzed with big data. The proposed system consists of an IoT sensor module, an edge module, and a big data
processing module.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２３８
8th ICAEIC-2022
2. BoP Big Data Processing System
2.1. IoT sensor module
The IoT sensor module collects data generated from the BoP and transmits it wirelessly to the edge module.
The IoT sensor module collects data from eBOP and mBOP. It collects temperature, water and air related data
from the mBoP. It also collects data related to voltage and current in eBoP.
2.2. Edge module
The Edge module stores data collected from IoT sensors. The module supports message queue to safely
store various types of sensor data transmitted at the same time. The data stored in the edge module is delivered
to the big data processing model at regular intervals.
2.3. BoP Big Data Processing module
The big data processing module receives data from each edge module, builds big data, and supports big data
analysis. The big data module also supports message queues to safely store various types of edge data delivered
at the same time. It can also monitor the data collected from each edge device in real time and support the
visualization of accumulated big data.
Figure 1.
BoP Big Data Processing Module.
２３９
AEICP Vol. 5, No. 1
Figure 2.
Real Time Monitoring and Visualization of BoP data.
3. Conclusion
In this paper, to improve BoP efficiency, we propose an IoT-Edge-BigData-based BoP big data processing
system that stores and processes data collected from BoP for BoP data analysis. The proposed system consists
of an IoT sensor module, an edge module, and a big data processing module. The IoT sensor module collects
data from the eBOP and mBOP of the BoP and sends it wirelessly to the edge module. The Edge module stores
data collected from IoT sensors, and the stored data is periodically delivered to the big data processing model.
The big data processing module receives data from each edge module, builds big data, and supports big data
analysis. Also, it can monitor in real time and support visualization of accumulated big data.
Acknowledgement
This research was supported by the National Research Council of Science & Technology(NST)
grant by the Korea government (MSIT) (No. CCL-20-28-GIST), This work was supported by Institute
of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the
Korea government (MSIT) (No. 2019-0-01842, Artificial Intelligence Graduate School Program
(GIST)), This work was supported by the Technology development Program(S3147824) funded by
the Ministry of SMEs and Startups(MSS, Korea), This work was supported by Institute of Information
& communications Technology Planning & Evaluation (IITP) grant funded by the Korea
government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub).
２４０
8th ICAEIC-2022
References
[1] Economy in hydrogen, http://www.h2news.kr/mobile/article.html?no=8693, 2021
[2] Gook-Hyun Y., Technology Development Trend of Renewable Energy Hydrogen Fuel Cell, Konetic
Report, Vol. 2014-42, 2016.
[3] Byonggil, K., Jeongbin, K., Han-Sang K., Sensitivity Analysis of Net System Power with the Variations
in Modeling Parameters of Automotive PEMFC Systemincluding MBOP (Mechanical Balance of Plant),
The conference of Korean Society of Automotive Engineers, pp.1189-1189, 2018
２４１
AEICP Vol. 5, No. 1
A Study of Estimate Information Service using DID (Distribute IDentification)
Jeong-Kyung Moon1, and Jin-Mook Kim*
1Department of Innovation and Convergence, Hoseo University, KOREA
*Division of IT Education, Sunmoon University, KOREA
jkmoon@hoseo.edu1, calf0425@sunmoon.ac.kr*
Abstract
In recent years, O2O services for real estate sales are widely distributed in web platforms and apps. This
allows sellers, buyers, and real estate brokers to quickly and conveniently conduct real estate sales and charter
contracts. However, in the O2O-based real estate sales information system, it wastes time and money for real
estate buyers due to the posting of fake information, partial correction of the sales information, and intentional
non-posting of the sales information. Therefore, we propose a method of detecting the false or not of real estate
property information that can occur on the web platform, and design and implement a proposal system for this.
To this end, we propose a method of detecting personal identity and property information based on DID, a
distributed identity authentication protocol. The false real estate sales information detection system proposed
by us can determine the existence of real estate sales information, partially correct the false sales information,
or prove whether or not intentionally unpublished in three steps.
Keywords: Write more than 4-6 Keywords..
1. Introduction
In order to reduce the damage caused by false sales of real estate, the Ministry of Land, Infrastructure and
Transport established a reporting system for counterfeit sales in August 2020. In addition, intensive
improvement efforts are being made, such as establishing a real estate market monitoring center. However,
cases of victims due to false sales continue to occur [1]. It is necessary to secure the credibility of the real
estate O2O market by reducing the inconvenience of consumers who are exposed to various fake products
such as no listings, different prices, different floors or different options [2].
As a result of previous studies, various efforts are being made to prevent damage caused by false sales of
real estate sales information, but security problems are still highly likely to occur. There was still the fact that
time and cost overload occurred in the work of brokers. To this end, this study proposes a method for detecting
false property information using distributed identity verification technology and block chain technology [4-6].
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２４２
8th ICAEIC-2022
2. Related works
The Blockchain-based decentralized identity authentication (hereafter DID, Decentralized Identity) is an
electronic identity authentication technology that stores personal information in the user's terminal and selects
and submits only the necessary information for personal information authentication. Individuals manage their
own data, and verification is possible even without a verified organization, so it can be used in various ways
in situations where authentication is required [3].
Blockchain technology is conducting research to maintain security and reliability in various fields such as
documents, proofs, authentication, verification, and storage using the consensus ledger and smart contracts. A
block in the blockchain refers to a ledger in which individual and individual transaction (P2P) data is recorded.
After these blocks are formed, they have a chain structure that is sequentially connected according to the
passage of time [4].
3. Proposed sytem design and procedures
When using a real estate brokerage site or app, the proposal system periodically performs filtering of false
property information by using the personal identification (DID) for the false property. In addition, if it is
determined that the sale is false, you can request the deletion of the sale by using your personal identification.
In this way, it is possible to block the time and economic waste of general users in advance through filtering
of false property information. When registering a property, it is possible to prevent double registration of a
single property by registering personal identification and property authentication information. At this time, big
data analysis is performed so that it can be determined whether or not duplicate entries are made.
When registering the certification information for sale, double registration of the property is
prevented by comparing the price or location information for the property information. At this time,
information is compared by issuing a blockchain-based unique identification code.
Figure 17. Design of Proposed System.
The unique identification code uses the province, city, dong, road address, and ledger number to ensure that
there is no overlapping property, and the number of floors or the number of the same building is entered. Data
２４３
AEICP Vol. 5, No. 1
for judging whether or not to sell and options for sale are entered so that it can be compared with other items
for sale. The registered information can become reliable information for sale after determining fake
information about real estate for sale information through big data analysis. If the authenticated user determines
that the registered item is fake, it is allowed to be modified as a duplicate item.
Figure 18. Procedures of Proposed System.
4. Conclusion
In this study, using Distributed Identity Identification Technology (DID), it is possible to provide clarity to
buyers, sellers, and real estate agents through the individual identity authentication process, and property
information through the blockchain distributed ledger record mechanism for the property information itself.
Among them, a method for detecting false information was proposed. The research method proposed in this
study has been implemented as a system on a web platform based on HTML, CSS, and Spring Framework so
far, and we plan to develop it as an app for user convenience in the future. However, in this study, studies on
the availability of web platforms or accessibility to real estate sales information were excluded. Duplicate
listings were filtered using big data analysis for additionally registered data. In the future, we plan to conduct
research on how to apply it not only to real estate but also to the sale of used products such as automobiles.
References
[1] C. G. Hwang (2020), “Implications and limitations of advertising for fake real estate on the Internet.” The
Journal of KISO, Vol. 40, pp. 18-22.
[2] S. S. Lee (2018), “Consumer Protection Plan in Mobile Real Estate Brokerage.” Ministry of Land,
Infrastructure and Transport, Vol. 438, pp. 16-22.
[3] J. H. Hwang (2020), “A Study on the Application of DID Digital Forensic Framework for Traffic Accident
Sensor Data Analysis”, Journal of Digital Forensics, Vol. 13, no. 3, pp.221-238.
[4] Y. H. Jung (2018), “Blockchain-based new identification system”, Korea Academy Industrial Cooperation
Society, Vol. 22, no. 2, pp.452-458.
２４４
8th ICAEIC-2022
[5] Jin-Mook Kim, Jeong-Kyung Moon, Sung-Woo Park (2020), “An Real Estate Platform Service using
Two-factor Authentication Techniques”, KIICE Conference Proceeding, 24(1), 508-510.
[6] Jeong-Kyung Moon (2021), “A Design of Estimate-information Filtering System using Artificial
Intelligent Technology”, KIICE Conference Proceeding, 21(1), 115-120.
２４５
AEICP Vol. 5, No. 1
Suggestion of Maintenance Criteria for Electric Railway Facilities System based
on Fuzzy TOPSIS Method
Sunwoo Hwang1, Joouk Kim1, and Youngmin Kim*
1Innovative Transportation and Logistics Research Center, Korea Railroad Research Institute, Uiwang,
16105, Korea
*Department of Systems Engineering, Ajou University, Suwon, 16499, Korea
marcell93@krri.re.kr1, jookim@krri.re.kr1, pretty0m@ajou.ac.kr*
Abstract
Railway facilities need for long-term operation as the initial acquisition cost required to build infrastructure
is high. This paper is on the suggestion of maintenance items for electric railway facility systems. With the
recent increase in the use of electric locomotives, the utilization and importance of railroad electrical facility
systems are also increasing, but the railroad electrical facility system in Korea is rapidly aging. To solve this
problem, various methodologies are applied to ensure operational reliability and stability for railroad
electrical facility systems, but there is a lack of detailed evaluation criteria for railroad electrical facility
system maintenance. Also, maintenance items must be selected in a scientific and systematic method. Therefore,
regular maintenance of railway facilities is essential, and furthermore, reliability needs to be achieved through
systematic performance assessment. Therefore, in this paper, railway electrical facility systems are selected
for study. This paper conducted a study that utilizes Fuzzy-TOPSIS to calculate weights, a mulit-criteria
decision-making problem in the facility of interest. The results of this may be contributed to the underlying
research in carrying out maintenance activities to ensure the reliability and safety of railway electrical facility
systems.
Keywords: Electrical Facility System; Railway Maintenance; Fuzzy Theory; Fuzzy TOPSIS; Railway
Maintenance.
1. Introduction
With the recent increase in the use of electric railroads, the utilization and importance of railroad electrical
facility systems is also increasing. Railway facilities need for long-term operation as the initial acquisition cost
required to build infrastructure is high. Therefore, regular maintenance of railway facilities is essential, and
furthermore, reliability needs to be achieved through systematic performance assessment. Therefore, in this
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２４６
8th ICAEIC-2022
paper, railway electrical facility systems are selected for study. The ratio of Korea's railway facilities below
grade C is 76.9% for subway power facilities, 51.4% for signal control facilities, 66.4% for information and
communication facilities, 11.7% for subway power facilities, 33.9% for signal control facilities, and 25% for
information and communication facilities. Also, after five years, the rate below grade C is expected to be 83.0%
and 93.4% by 2030. The performance evaluation score is also expected to be 3.37 in 2025, five years later, and
2.31 in 2030, 10 years later. Rapid changes can be seen from 2025.
2. Fuzzy-TOPSIS
Efficient and systematic standards are needed to perform efficient maintenance of aging railroad electrical
facility systems, taking into account the limitations of economic and temporal costs, and the establishment of
maintenance assessment criteria involves the decision maker's subjective judgment. Because decision-makers'
judgments are subjective and ambiguous, a methodology is needed to quantitatively derive opinions from
experts in the field. Chen presents a methodology that extends the Technique for Order Performance by Ideal
Solution (TOPSIS) logic in a Fuzzy environment to solve the MCDM problem [2]. Pushpendra Kumar et al.
approached the predict liver disease problem using fuzzy sets and boosting techniques [3], and Yoosef
B. Abushark et al. conducted a usability evaluation study on security requirement perspective using
the fuzzy AHP-TOPSIS approach [4]. Regarding COVID-19, Rupkumar Mahapatra et al. conducted
a study to distinguish COVID-19 detected regions by fuzzy directed graphs [5]. This paper conducted
a study that utilizes Fuzzy-TOPSIS to calculate weights, a mulit-criteria decision-making problem in the
facility of interest. We find that the Fuzzy TOPSIS theory is effective in processing ambiguous data. The
Fuzzy-TOPSIS technique uses the fuzzy set, fuzzy number’s data, and presents the closest alternatives, FPIRP
(Fuzzy Positive Ideal Reference Point) and FNIRP (Fuzzy Negative Ideal Reference Point), respectively. The
equation for calculating the Fuzzy matrix is (1-4), and the equation for caculating Closeness Coefficient is (5-
7).
= {} (1)
=
(2)
=
(3)
= {} (4)
(,)=1/[()] (5)
(,)=1/[()] (6)
= /(+) (7)
２４７
AEICP Vol. 5, No. 1
3. Facility of interest
The criteria for selecting the interest of facility are Safety(), Durability(), and Usability(). In this
paper, we select Degradation and Insulation(), Abrasion Strength and Noise() and Corrosion, Crack, Oil
leak, Slope and Subsidence(). Figure 1 represents the result of the selection.
Figure 1.
Criteria and Facility of interest.
4. Caculation
In this paper, we gather opinions from 3 experts using 7 scales of criteria and alternatives. With reference
to the survey data results, we compute distances for FPIRP (Fuzzy Possible Ideal Reference Point) and
FNIRP(Fuzzy Neggative Ideal Reference Point). Hereby, the results of the maximum and minimum fuzzy data
of the alternatives for each evaluation criteria can be found. Normalization can be performed according to the
Closeness Coefficient value to determine the percentage value for each alternative. By normalizing the results
for this, the facility of interest weights of alternatives for each railroad electrical facility maintenance are
presented as Percentage. As a result, Degradation and Insulation were weighted the highest at 44.40%, and
Corrosion, Crack, Oil leak, Slope and Subsidence were weighted the lowest at 27.25%. Results for this are
shown in Table 1.
Table 1. Main parameters.
Alternative
Closeness Coefficient
Weight
0.624
44.40
0.399
28.35
0.383
27.25
5. Conclusion
This paper conducted a weight determination study on the maintenance facility of interest with the aim of
ensuring reliability and safety of railroad electrical facility systems in South Korea. The performance
２４８
8th ICAEIC-2022
evaluation score is also expected to be 3.37 in 2025, five years later, and 2.31 in 2030, 10 years later. Rapid
changes can be seen from 2025. Efficient and systematic standards are needed to perform efficient maintenance
of aging railroad electrical facility systems, taking into account the limitations of economic and temporal costs,
and the establishment of maintenance assessment criteria involves the decision maker's subjective judgment.
Because decision-makers' judgments are subjective and ambiguous, a methodology is needed to quantitatively
derive opinions from experts in the field. We find that the Fuzzy TOPSIS theory is effective in processing
ambiguous data. The Fuzzy-TOPSIS technique uses the fuzzy set, fuzzy number’s data, and presents the closest
alternatives, FPIRP (Fuzzy Positive Ideal Reference Point) and FNIRP (Fuzzy Negative Ideal Reference Point),
respectively. The criteria for selecting the interest of facility are Safety (), Durability (), and Usability
(). In this paper, we select Degradation and Insulation (), Abrasion Strength and Noise () and Corrosion,
Crack, Oil leak, Slope and Subsidence (). Figure 1 represents the result of the selection. To determine the
weighting of the maintenance detail basis alternatives by the evaluation criteria, the weighting of the
assessment criteria and the weighting of the basis alternatives were represented in Fuzzy matrix, and the
Closeness Coefficient values were calculated by determining FPIRP and FNIRP. By normalizing the results
for this, the weights of alternatives for each railroad electrical facility maintenance detail criteria item are
presented as Percentage As a result, As a result, Degradation and Insulation were weighted the highest at
44.40%, and Corrosion, Crack, Oil leak, Slope and Subsidence were weighted the lowest at 27.25%. The results
of this may be contributed to the underlying research in carrying out maintenance activities to ensure the
reliability and safety of railroad electrical facility systems.
Acknowledgement
This research was supported by a grant from R&D Program of the Korea Railroad Research Institute,
Republic of Korea.
References
[1] Ministry of Land, Infrastructure and Transport, “Basic Plan for Maintenance and Maintenance of Railway
Facilities”, Korea: Ministry of Land, Infrastructure and Transport, 2020.
[2] Chen, C. T., “Extension of the TOPSIS for Group Decision-making Under Fuzzy Environment”, Fuzzy
Set and Systems, Vol. 114, Issue 1, pp. 1-9, 2000.
[3] Pushpendra Kumar and Ramjeevan Singh Thakur, “An Approach Using Fuzzy Sets and Boosting
Techniques to Predict Liver Disease”, Computers, Materials & Continua, Vol.68, No.3, pp. 3513-3529,
2021.
[4] Yoosef B. Abushark, Asif Irshad Khan, Fawaz Jaber Alsolami, Abdulmohsen Almalawi, Md Mottahir
Alam, Alka Agrawal, Rajeev Kumar and Raees Ahmad Khan, “Usability Evaluation Through Fuzzy AHP-
TOPSIS Approach: Security Requirement Perspective”, Computers, Materials & Continua, Vol.68, No.1,
pp. 1203-1218, 2021.
２４９
AEICP Vol. 5, No. 1
[5] Rupkumar Mahapatra, Sovan Samanta, Madhumangal Pal, Jeong-Gon Lee, Shah Khalid Khan, Usman
Naseem and Robin Singh Bhadoria, “Colouring of COVID-19 Affected Region Based on Fuzzy Directed
Graphs”, Computers, Materials & Continua, Vol.68, No.1, pp. 1219-1233, 2021.
２５０
8th ICAEIC-2022
Synthesis and Characterization of the Phosphoric Acid Activated Carbon
Prepared from Giant Cane
Sahira Joshi1, and Shishir Baral*
1Department of Applied Sciences and Chemical Engineering, Nepal
*Pulchowk Campus, IOE, Tribhuvan University, Lalitpur, Nepal
Corresponding email: sjoshi61@hotmail.com
Abstract
The aim of this study is to synthesize and characterize of activated carbon derived from Giant cane using
phosphoric acid. Activated carbon (AC) was synthesized by activation using H3PO4 at ratio of 1:1
carbonization temperatures 400 °C. The prepared activated carbon (AC) was characterized by iodine number
(IN), methylene blue number (MBN), SEM and XRD spectroscopy. The results showed that AC was a porous
material with microporous and mesoporous structures and had a high specific surface area of 936 m2/gm.
SEM analysis of activated carbon illustrates a heterogeneous surface morphology with a developed porous
structure of various sizes. The absence of a sharp peak reveals a predominantly amorphous structure of ACs.
The results indicated that the activated carbon developed in this study has the potential to be a promising
adsorbent in water purifications.
Keywords: Giant cane, activated carbon chemical activation, phosphoric acid.
1. Introduction
Activated carbons with abundant micropores (0–2 nm width), mesopores (2–50 nm width) and macropores
(> 50 nm width) together with a high specific surface area are widely used as versatile adsorbents for the
adsorption of gaseous and liquid phases. For the preparation of activated carbons, physical and chemical
activation are normally used. Physical activation involves a carbonization step followed by a stage of
controlled oxidation to activate the carbon in the presence of an activating agent such as steam or carbon
dioxide. Chemical activation provides a single-step method of activated carbon preparation in that the
precursors are carbonized in the presence of chemical agents such as phosphoric acid, zinc chloride, potassium
hydroxide etc. Out of these, ZnCl2 and H3PO4 are preferred when activating lignocellulosic material like
agricultural materials. H3PO4 activation is recently preferred over ZnCl2 because of the environmental and
economic concerns.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２５１
AEICP Vol. 5, No. 1
Activated carbons can be produced from a variety of raw materials. Of these, coal and agricultural wastes
such as sawdust and coconut shell are the most commonly used precursors. Now days, the production of
activated carbon from agricultural and industrial by-products is encouraging since they are inexpensive and
abundantly available. These agricultural materials include nut shells (Hayashi et al., 2002), corncob (El-
Hendawy et al., 2001) coconut shell (Cazetta et al., 2011), pistio-nut shell (Lua et al., 2005), Dates stone
(Ahmed et al., 2012) etc. Giant reed (Arundo donax L.), is a tall perennial grass, is a perennial grass with little
economic value. It is found in wetlands and riparian natural surroundings and is believed to be local to eastern
Asia. So preparation of activated carbon using Giant cane as a precursor has been of particular interest. The
aim of the present work was to synthesize and characterize the H3PO4 activated carbon derived from Giant
cane.
2. Material and method
Precursor used for preparation of activated carbon (AC) was giant cane collected from Sundarijal,
Kathmandu. 20 gm of dried, crushed giant cane powder was mixed with phoshoric acid (H3PO4) in the ratio of
1:1 by weight. The mixture was warmed at 70oC with constant stirring with glass rod until partly dried. Then,
product was oven dried at 110 °C for 24 hours. The sample was carbonized in a horizontal tubular furnace at
400°C for 3 hrs under a constant flow of nitrogen (75 ml/min). The AC so obtained was treated with 1%
NaHCO3 solution and then, subsequently washed with warm distilled water until the pH of washing became
neutral. The sample was then dried at 110o C for 24 hrs and sieved to get the particles of size 106 μm. For the
preparation of plain carbon, similar procedure was carried. In this case, activating agent was not used. The
chemicals used for this investigation and analysis were all analytical grade purchased from Qualigen, India
were used.
Characterization of ACs was performed by adsorption of iodine and methylene blue. Iodine number (IN) is
the amount of iodine adsorbed (in milligrams) by 1g of carbon (3). Iodine number (IN) of AC was determined
according to ASTM D4607-94 method (ASTM Standards, 2006). Methylene blue number (MBN) is defined as
the milligram of methylene blue adsorbed onto 1.0 gm of adsorbent (Raposo et al., 2009). Methylene blue
number (MBN) of AC was determined according to the Method (Raposo et al., 2009). The surface areas of
ACs were estimated by iodine and methylene blue numbers using multiple regressions (Cleiton and Guerreiro,
2011). The surface morphology was studied by scanning electron microscope (SEM) Hitachi S-4800 analysis
using JEOL JEM-2100F operating at 200 kV. X-ray diffraction (XRD) measurements were carried out on
Rigaku X-ray diffractometer, RINT, Japan and operated at 40 kV and 40 mA with Cu-Kα radiation at room
temperature.
3. Result and discussion
3.1. Iodine number and methylene blue number of activated carbon
IN, MBN and surface area of plain carbon, commercial AC and prepared ACs using H3PO4 are presented in
Figure: 1.
２５２
8th ICAEIC-2022
Figure 1. IN, MBN and surface area of plain carbon, commercial AC and prepared AC using H3PO4.
Iodine number and methylene blue number are useful indicators to evaluate the adsorptive capacity of AC.
IN is a measure of micropore content of the AC and relates to the ability of AC to adsorb low molecular weight
substances. MBN is a measure of mesopore content and indicates ability of AC to adsorb high molecular weight
substances. IN, MBN and surface area of Giant cane AC were found to be 784 mg/g, 555 mg/g, and 970 m2/g.
It suggested that, the microporosity, mesoporosity of AC were better developed with high surface area and
theses values are also comparable to commercial AC.
3.2. Scanning electron microscopy (SEM) image
Surface morphology of carbons was studied by Scanning Electron Microscopy (SEM). SEM images of plain
carbon and AC are shown in Figure: 2.
(A) Plain carbon (B) Activated carbon
Figure 2. SEM images of the plain carbon and activated carbon.
0
200
400
600
800
1000
1200
Plain Carbon Commercial AC H3PO4 AC
I
N
, MB
N
and surface area
Carbon
MBN IN Surface Area
２５３
AEICP Vol. 5, No. 1
In the SEM image of plain carbon, pores are not clearly seen on the external surface while the external
surface of AC is filled with pores of different shape and size. The pores are resulted by activation with H3PO4
as an activating agent. The activating reagent may promote the formation of cross-links, leading to the
formation of rigid matrix, less prone to volatile loss and volume contraction upon heating to high temperature.
H3PO4 could function both as an acid catalyst to promote bond cleavage reactions and the formation of cross-
links via processes such as cyclization, and condensation. It could also combine with organic species to form
phosphate and polyphosphate bridges that connect and crosslink biopolymer fragments. The addition or
insertion of phosphate groups drives a process of dilation that, after removal of the acid, leaves the matrix in
an expended state with an accessible pore structure (Luo et al., 2016).
3.3. X-Ray Diffraction
X-ray diffraction pattern of plain carbon and activated carbon are presented in Figure 3.
Figure 3. X-ray diffraction patterns of plain carbon and activated carbon.
In Figure 3, all the ACs exhibited two broad diffraction peaks in XRD patterns. The peaks positioned around
2θ = 24° and 43°, corresponds to the diffraction of (002) and (100) planes, respectively. The AC exhibit very
broad diffraction peaks and the absence of a sharp peak reveals a predominantly amorphous structure. The
XRD pattern of the AC was similar to that of commercial AC (Shrestha et al., 2018).
4. Conclusion
Activated carbon was produced using giant cane by chemical activation by H3PO4. The result of iodine and
methylene blue adsorption has shown that, the microporosity, mesoporosity of AC were better developed with
high surface area. SEM images showed an irregular and heterogeneous surface morphology with a developed
and fragmented porous structure in various sizes. X-Ray spectroscopy indicates that the AC has an amorphous
２５４
8th ICAEIC-2022
structure. The adsorption properties of the AC is comparable to the commercial AC, demonstrating that the
Giant cane derived AC material would have potential as a high efficiency adsorbent in water purification.
References
[1] Annual Book of ASTM Standards, 2006, Standard Test Method for Determination of Iodine Number of
Activated Carbon, ASTM D4607-94, Philadelphia PA, United State of America.
[2] Cazetta A.L, Vargas M.M., Nogami E.M., Kunita M.H., 2011, “NaOH-activated carbon of high surface
area produced from coconut shell: Kinetics and equilibrium studies from the methylene blue adsorption”,
Chemical Engineering Journal, 174, 117-125.
[3] Cleiton N. A. and Guerreiro M. C., 2011, “Estimation of Surface Area and Pore Volume of Activated
Carbons by Methylene Blue and Iodine Numbers”, Quimica Nova, 34(3), 472-476.
[4] El-Hendawy A.N.A., Samra S.E., Girgis B.S., 2001, “Adsorption characteristics of activated carbons
obtained from corncobs”, Colloids and Surfaces, 180, 209-221.
[5] Hayashi J., Horikawa T., Takeda I., Muroyama K., Ani F.N., 2002, “Preparing activated carbon from
various nutshells by chemical activation with K2CO3”, Carbon, 40, 2381-2386.
[6] Lua AC, Yang T., 2005, “Characteristics of activated carbon prepared from pistachio-nut shell by zinc
chloride activation under nitrogen and vacuum conditions”, Journal of Colloid and Interface Science,;
290: 505-513.
[7] Luo Y., Street, J., Steele P., Entsminger E. and Guda, V. 2016, "Activated carbon derived from pyrolyzed
pinewood char using elevated temperature, KOH, H3PO4 and H2O2," Bio Reseources 11(4), 10433-
10447.
[8] Raposo F., De La Rubia M. A. and Borja R., 2009, “Methylene Blue Number as useful Indicator to
Evaluate the Adsorptive Capacity of Granular Activated Carbon in Batch Mode: Influence of
Adsorbate/Adsorbent Mass Ratio and Particle Size”, Journal of Hazardous Materials, 165(1-3), 291–299.
[9] Shrestha D., Gyawali G.and Rajbhandari (Nyachhyon) A., 2018, “Preparation and Chracterization of
activated carbon from waste saw dust, from saw mill”, Journal of Institute of Science and Technology,
22, (2), 2467-9240.
２５５
AEICP Vol. 5, No. 1
Restaurant Skyline for Users with Physical Disabilities
Chae-Eun Lim1, Eun-Young Park2, Jong Sup Lee3, and Sun-Young Ihm*
1, *Dept. of Computer Engineering, Pai Chai University, Korea
2Dept. of Visual Formative Arts Design, Hyupsung University, Korea
3Development of smart community policing system(Googi) Center, Dongguk University-Seoul, Korea
lcy_0907@naver.com1, pey54@naver.com2, jsleearmy@dongguk.edu3, sunnyihm@pcu.ac.kr*
Abstract
This study constructs a skyline for recommending restaurants for users with physical disabilities by
considering attributes such as distance, area and star rating. Also convenience attribute is defined with
parking space and accessibility. We construct restaurant skyline based on these attributes and we expect that
it is helpful for users with physical disabilities.
Keywords: Skyline, Restaurant Recommendation
.
1. Introduction
Various websites or applications introduce restaurants. In these applications, people can select desired
information, mainly attributes such as ratings, food types, and themes. However, some users may find it
difficult to go to the recommended restaurant due to their physical condition. For example, if a restaurant can
only go by stairs, a user who has to move in a wheelchair cannot visit.
The skyline method [1] helps a user to find a desired item by considering various attributes when there are
many attributes. The skyline method is a layer-based indexing technique, which builds a skyline as an index
for retrieving only a part of the entire data to find the result. When there are objects with various properties, a
set of objects that do not dominate each other is defined as a skyline. Here, 'dominant' means that when two
objects are compared, the values of all properties of one object dominate over the values of all properties of
the other object [2, 3]. Therefore, in this paper, we construct a skyline for recommending restaurants for users
with physical disabilities. We first collect data related to restaurants and defined basic properties. In addition,
we additionally define the 'convenience' attribute for users with disabilities. Next, we check the results by
building a skyline based on the collected data.
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２５６
8th ICAEIC-2022
2. The Proposed Method
In this section, we propose a skyline for users with physical disabilities. We first collected public data related
to restaurants [4, 5]. We define attributes such as distance, star rating, and area based on the collected data.
Here, the distance is calculated assuming that the starting point is 'Pai Chai University'. Next, in this study, we
define the 'convenience' attribute for users with physical disabilities. Convenience is calculated considering
the number of parking facilities and parking spaces, and accessibility. Parking facilities should be considered
for users with disabilities often travel by car rather than on foot. In addition, accessibility was investigated
whether the road to the restaurant was wide and whether there were stairs at the entrance.
Next, we construct a skyline based on the collected data. At this time, we determined that the closer the
distance and the higher the star rating, the more dominant. We also determined that the larger the area and the
higher the convenience, the better. Figure 1 shows the restauruante skyline result considering two attributes,
star rating and distance.
Figure 1.
Restaurant skyline result for users with physical disabilities.
3. Conclusion
In this paper, we constructed a skyline for recommending restaurants for users with physical disabilities.
For this, convenience such as distance, area, star rating, and parking space and accessibility were considered.
By utilizing the proposed skyline, it is expected that not only general users but also users with disabilities will
be able to find the restaurant they want in consideration of various factors.
Acknowledgement
This work was suported by the National Research Foundation of Korea(NRF) grant funded by the Korea
government(MSIT) (No. 2021R1C1C2011105).
２５７
AEICP Vol. 5, No. 1
This research was supported by National Research Foundation of Korea (NRF) grant funded by the Korea
government (MSIT) (No. 2018R1A5A7023490).
References
[1] Borzsony, S., Kossmann, D., & Stocker, K. (2001, April). The skyline operator. In Proceedings 17th
international conference on data engineering (pp. 421-430). IEEE.
[2] Ihm, S.Y., Lee K.E., Nasridinov, A., Heo J.S., & Park, Y.H. (2014). Approximate convex skyline: A
partitioned layer-based index for efficient processing top-k queries. Knowledge-Based Systems, 61, 13-
28.
[3] Kodama, K., Iijima, Y., Guo, X., & Ishikawa, Y. (2009, November). Skyline queries based on user
locations and preferences for making location-based recommendations. In Proceedings of the 2009
International Workshop on Location Based Social Networks (pp. 9-16). ACM.
[4] Public Data Portal, https://www.data.go.kr/data/15063351/fileData.do
[5] Public Data Portal, https://www.data.go.kr/data/15008957/fileData.do
２５８
8th ICAEIC-2022
Creation of a Mesh by Applying Poisson Disk Sampling to PointCloud :
Comparison of RTAB-Map and VisualSFM
Yujin Yang1, Sohee Kim2, Jungsuk Park3, and Dong Ho Kim*
1,2,3,*Seoul National University of Science and Technology
{2216056, kksshh0327, whiter1234}@naver.com1,2,3, dongho.kim@seoultech.ac.kr*
Abstract
In this paper, we propose to apply Poisson Disk Sampling, which has been mainly applied in computer
graphics, to 3D image modeling. This sampling technique uses the lidar sensor, which is used as the core
technology of the 3D image sensor, to ensure that the PointClouds collected are uniformly distributed without
clustering in a specific location, and also removes image noise. PointCloud extracted using Intel's LiDAR
sensor L515 and Apple's tablet LiDAR sensor goes through Poisson Disk Sampling process to create a mesh
and compare it with the case of extracting it with SLAM technique. A better mesh can be generated when the
number of PointClouds is reduced.
Keywords: Poisson Disk Sampling, Pointcloud, Mesh, SFM, RTAB-MAP.
1. Introduction
Recently, with the entry of a non-face-to-face society, the consumption of virtual reality (VR) and
augmented reality (AR: Augmented Reality) content is rapidly increasing. Virtual reality (VR) refers to a
technology to experience the real world as if it were real by stimulating the five senses of the human body by
providing an experience or environment that is difficult or impossible to obtain in reality using artificial
technology. Augmented reality (AR), a system that augments the real experience by adding additional virtual
information to the real-world information, is being used in various fields [1].
However, using 3D spatial modeling for VR and AR content, especially 3DoF+ or 6DoF 360 video or AR
service, requires more data than conventional audio, voice, and video. Therefore, it is necessary to reduce the
capacity itself or to implement similarly to the original data through a post-processing process that selects and
simplifies the acquired information.
Poisson Disk Sampling provides a good distribution of sample points and is known to be useful in many
applications [2]. In computer graphics, the randomness of the distribution enables anti-aliasing [3] [4], and the
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２５９
AEICP Vol. 5, No. 1
number of samples can be reduced through the process of determining whether adjacent samples are not too
close by randomly taking a point corresponding to a fixed section.
In this paper, the PointCloud collected using the lidar sensor (Intel lidar sensor L515, Apple tablet lidar
sensor) used as the core technology of the 3D image sensor is relocated and the number of points is reduced
using Poisson Disk Sampling. Create a mesh. Also, after extracting feature points based on photos taken from
various angles, meshing is performed in the same way to compare the two cases.
Section 2 of this paper examines RTAB-MAP (Real-Time Appearance Based Mapping) [5], VisualSFM [6],
and Poisson Disk Sampling, and Section 3 compares the results of applying Poisson Disk Sampling. Section 4
discusses the conclusion of this paper and future research directions.
2. Point Cloud Meshing Techniques
Before applying Poisson Disk Sampling, we introduce RTAB-Map and VisualSFM, tools that can extract
PointCloud, and explain Poisson Disk Sampling.
2.1. RTAP-MAP
RTAB-Map is a program of RGB-D (Depth) stereo and Lidar (Light Detection And Ranging) graph-based
Simultaneous Localization And Mapping (SLAM) approach using a loop closure detector, where the loop
closure detector is a previously measured position. It refers to the process of detecting whether it has returned
to we use a bag-of-words [7] approach that compares the frequency of occurrence of feature points in pictures
to determine the likelihood that a new image will come from an old or new location.
Figure 1. RTAP-Map Example.
２６０
8th ICAEIC-2022
The blue circle in the upper part of Figure 1 shows the loop closure detection that estimates the new location
compared to the previous location, and in the lower right picture of Figure 1, you can see the PointCloud
matched to the 3D space and the path the user moved.
2.2. VisualSFM
VisualSFM is a system that analyzes feature points and reconstructs three-dimensional space according to
movement using structure from motion (SFM), and it runs fast utilizing multi-core parallelism for feature
detection, feature matching, and bundle coordination. PointCloud is extracted by estimating depth information
by capturing feature points based on the image. Figure 2 shows the overall process of VisualSFM, and a
PointCloud is created by extracting feature points from several photos.
Figure 2. Using Visual SFM Example.
２６１
AEICP Vol. 5, No. 1
2.3. Poisson Disk Sampling
Poisson Disk Sampling is an improvement on clusters, which is a phenomenon in which samples are
agglomerated near a specific point during the sampling process. Therefore, the final result will have a uniform
but random distribution. A dimension grid is used, and the size of one grid is =
. You can only have
one point per grid. Initially, an active point is randomly generated, and then new active points are created. If
no such point is found in the adjacent grid, the active is terminated. Through this process, the samples can be
adjusted so that they are not too close [8].
3. Point Cloud Mesh Method with Poisson Disk Sampling
3.1. Implementation environment
RTAB-Map was tested with iPad LiDAR sensor and Intel L515, respectively. Beta version 0.20.10 of
RTAB-Map for IOS and iPad Pro 12.9 type were used, and Intel L515 was connected to a computer with 64-
bit operating system and Windows 10 to use RTAB-Map version 0.20.8. The resolution of L515 is RGB
1920x1080, Depth 640x480, and iPad is set to RGB Maximum (1920x1440), Depth Maximum (256x192) as
HD Mode. In addition, both RTAB-Maps proceed with the same point size 5 and detection rate 1Hz.
VisualSFM used version 0.5.26 on a computer with the same specifications as RTAB-Map.
3.2. Implementation result
Figure 3. Mesh Generating Results.
２６２
8th ICAEIC-2022
Both RTAB-Map and VisualSFM data were meshed through Poisson disk sampling and ball pivoting, a
surface reconstruction method in MeshLab [9, 10]. When performing Poisson Disk Sampling, in the case of
RTAB-Map, select Base Mesh Subsampling, fix the number of samples to 1000, set the minimum radius to
0.01, 0.007, 0.004, 0.001 (0.003 interval, unit []), and then mesh carried out as the sample's minimum radius
value is larger, many PointClouds are removed and simplified. On the other hand, if the value is set small,
detailed sampling is possible, but if there is no RGB value of the image in the sample, it is processed as a preset
value, so as shown in Figure 3, the portion treated as a dummy value increases. Rows 1 to 4 of Figure 3 are
the results of setting 0.01, 0.007, 0.004, and 0.001 in sequence.
On the other hand, in the case of VisualSFM, since the PointCloud data was measured 1.7 times larger than
the RTAB-map, 0.017, 0.0119, 0.0068, and 0.0017, which are 1.7 times the RTAB-Map raidus value, were
used to maintain the ratio. Similar to RTAB-Map, rows 1 to 4 of Figure 3 are the results of setting the values
to 0.017, 0.0119, 0.0068, and 0.0017 in sequence, and here, ball pivoting is all applied with a radius of 1%.
Figure 4. When the mesh of RTAB-Map (PC) is enlarged.
２６３
AEICP Vol. 5, No. 1
As a result, sparser PointCloud (going to row 1, the higher it is) is more suitable it is to simplify the mesh,
so there are fewer holes. In the range smaller than 0.01 m, both RTAB-Map and VisualSFM show better results
than the naked eye because the larger the Poisson Disk Sampling is, the smaller the hole is. Therefore, if
Poisson Disk Sampling is applied by adjusting to an appropriate value, a mesh with fewer holes and reduced
capacity can be obtained. Here, the number of each PointCloud is as shown in the Table 1.
Table 1. Number of pointcloud.
RTAB-Map (PC)
RTAB-Map (iPad)
VisualSFM (x1.7)
0.01
260233
207879
56103
0.007
385329
364503
102095
0.004
434296
423597
246388
0.001
438885
427209
1147697
Total PointCloud
438902
427214
1373849
VisualSFM creates a PointCloud by extracting feature points from photos, so the depth information of thin
books or bookshelves appears relatively accurately. However, in the case of a face without a feature point, a
hole is formed instead of being expressed as a PointCloud. Therefore, compared with RTAB-Map, the more
complex the image, the fewer holes, but the more holes in the simple plane.
4. Conclusion
If the number of PointClouds is adjusted to an appropriate value using Poisson Disk Sampling, a better mesh
than raw data can be extracted, and the capacity can also be reduced. The method using the lidar sensor that
directly extracts depth information with RTAB-Map shows better results than the method of extracting
PointCloud by estimating depth by extracting image-based feature points with VisualSFM.
Acknowledgement
This work was supported by Institute of Information & communications Technology Planning & Evaluation
(IITP) grant funded by the Korea government (MSIT) (No.2020-0-00994, Development of autonomous VR
and AR content generation technology reflecting usage environment) in 2021
References
[6] Kook Gyeongwan, “Recent Trends of VR/AR Systems, Application Cases and Prospects,” Korea Institute
of Science and Technology Information, Sep. 2018.
[7] Lagae, Ares, and Philip Dutré. "A comparison of methods for generating Poisson disk distributions."
Computer Graphics Forum. Vol. 27. No. 1. Oxford, UK: Blackwell Publishing Ltd, 2008.
２６４
8th ICAEIC-2022
[8] Dippé, Mark AZ, and Erling Henry Wold. "Antialiasing through stochastic sampling." Proceedings of the
12th annual conference on Computer graphics and interactive techniques. 1985.
[9] Ebeida, Mohamed S., et al. "Efficient maximal Poisson-disk sampling." ACM Transactions on Graphics
(TOG) 30.4 (2011): 1-12.
[10] Labbé M “Real-Time Appearance-Based Mapping” 2018 [Online]. Available:
http://introlab.github.io/rtabmap/. [Accessed 11 05 2021]
[11] “VisualSFM : A Visual Structure from Motion System,”Changchang Wu,last modified Sep 25.2013,accessed May
11.2021,http://ccwu.me/vsfm/.
[12] Kejriwal, Nishant, Swagat Kumar, and Tomohiro Shibata. "High performance loop closure detection using
bag of word pairs." Robotics and Autonomous Systems 77 (2016): 55-65.
[13] Bridson, Robert. "Fast Poisson disk sampling in arbitrary dimensions." SIGGRAPH sketches 10 (2007):
1.
[14] Corsini, Massimiliano, Paolo Cignoni, and Roberto Scopigno. "Efficient and flexible sampling with blue
noise properties of triangular meshes." IEEE transactions on visualization and computer graphics 18.6
(2012): 914-924.
[15] Bernardini, Fausto, et al. "The ball-pivoting algorithm for surface reconstruction." IEEE transactions on
visualization and computer graphics 5.4 (1999): 349-359.
２６５
AEICP Vol. 5, No. 1
Determination of Methylene Blue Number, Iodine Number, SEM and XRD to
Characterize Activated Carbon Prepared from Rudraksha (Elaeocarpus
ganitrus) Bead
Bishwas Pokharel, Rinita Rajbhandari (Joshi), Rajeshwar Man Shrestha
Department of Applied Sciences and Chemical Engineering, Pulchowk Campus,
Institute of Engineering, Tribhuvan University, Nepal
rajeshwar@ioe.edu.np
Abstract
Activated carbon prepared from Rudaksha (Elaeocarpus ganitrus) beads by chemical activation with zinc
chloride was characterized by chemical methods such as Iodine number and Methylene blue number to identify
micropore and mesopore content. Methylene blue number and Iodine number are found to be 499.6 mg/g and
920.6 mg/g respectively. The activated carbon was also analyzed by instrumental techniques like XRD (X-ray
Diffraction) and SEM (Scanning Electron Microscope). The SEM of the activated carbon indicates the pores
of different diameters whereas XRD profile exhibited two broad diffraction peaks indicating the amorphous
nature of the activated carbon.
Keywords: Activated carbon, Rudraksha bead, Chemical activation, Zinc chloride
.
1. Introduction
Activated carbon, also called activated charcoal, is a form of carbon processed to have small, low-volume
pores that increase the surface area. Activated carbon is a highly adsorbent powdered or granular carbon made
usually by carbonization and chemical activation and used chiefly for purifying by adsorption. Because of
extended surface area, micro-pore structures, high adsorption capacity and high degree of surface reactivity,
activated carbon is used for removal of pollutants from water and air. However, commercially available
activated carbon is very expensive and has high regeneration cost while being exhausted. Furthermore,
generation using solution produces a small additional effluent while regeneration by refractory technique
results in a 10–15 % loss of adsorbent and its uptake capacity. This has led to search for cheaper substances
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２６６
8th ICAEIC-2022
for the preparation of adsorbents. Activated carbon is produced from carbonaceous source materials such as
mushroom waste [1], peach stone [2], date stone [3], corn stalk [4], coconut shell [5]and [6], Elaeagnus stone
[7], Lapsi Seed Stone [8] etc.
In this study activated carbon has been prepared from Rudraksha beads. The application of many waste
materials in the preparation of activated carbon has been reported in literature. No study has been reported in
literature Rudraksha beads as waste materials for preparation of activated carbon. The Rudraksha bead acts like
a protective guard that safeguards its wearer from negative energies.
The Rudraksha beads as shown in Fig. 1 are abundantly available in Nepal. In season, most of the beads
were wasted because of unable to gain market. Rudraksha refers to a stone fruit, the dried stones of which are
used as prayer beads by Hindus (especially Shaivas), as well as by Buddhists, Sikhs, and Muslims.
Figure 1. Rudraksha beads.
2. Experimental
2.1. Materials
Rudraksha bead is precursor used in this study for the preparation of activated carbon. The precursor was
washed with tap water and then with distilled water. The precursor was crushed and sieved to a size range of
300 μm and was soaked with zinc chloride in the ratio of 1:1. The materials were dried in an oven at 110 o C.
The dried Walnut shell particles were carbonized under high purity nitrogen flow of 75 ml min−1 by raising
the temperature in tubular furnace at a rate until 500 o C and kept at this temperature for three hours. The
prepared activated carbon was then cooled at room temperature and washed with dil. HCl and then with warm
distilled water until the pH of the washing reached 6 –7.
2.2. Chemicals and Equipment
The chemicals and reagents of analytical grade (Merck and Qualigens Company) have been used in this
study. Digital pH meter was used to adjust pH of solutions. The adsorption experiments for the determination
of Methylene blue number were carried out by using Shaker (Digital VDRL Rotator RPM-S) and UV –Visible
Spectrophotometer.
２６７
AEICP Vol. 5, No. 1
2.3. Methylene blue number and Iodine number
The amount of micropore and mesopore content in activated carbon has been determined by Methylene blue
number and Iodine number. Methylene blue number [mg/g] indicates the extent of meso pore distribution in
the carbon. The Methylene blue (MB) is the maximum amount of dye adsorbed on 1.0 gm of adsorbent while
the iodine number indicates the number of milligrams of iodine adsorbed by one gram of carbon. It gives rough
measurement of the micropores content of the activated carbon by adsorption of iodine from solution.
Methylene blue number can be calculated by the following formula [9]
=()×
(1)
where Co and Ce = initial and equilibrium concentration of MB (mg / L) respectively, M = the mass of adsorbent
in gram and V = the volume of the solution in liter. Iodine number can be calculated by the equation as follows
[10].
=
(2)
3. Characterization of Activated carbon
3.1. XRD (X-ray diffraction)
XRD analysis of the activated carbon exhibits two broad diffraction peaks located near at 2θ = 25.5 o and
43o reflected from 002 and 001 planes shown in Fig.2. The two broad diffraction peaks indicate the amorphous
nature of the activated carbon.
Figure 2. XRD of Activated. Figure 3. IN and MBN of Activated.
２６８
8th ICAEIC-2022
Figure 4. SEM of Activated Carbon.
3.2. Methylene blue number (MBN) and Iodine number (IN)
Methylene blue number (mg/g) and Iodine number (mg/g) indicate the extent of meso pore and micro pore
distribution in the carbon respectively. The mesopores and micropore content in the Activated carbon have
been determined by the chemical methods like Methylene blue number and Iodine blue number have as
presented in Fig. 3. Methylene blue and iodine numbers are found to be 499.6 mg/g and 920.6 mg/g
respectively.
3.3. SEM (Scanning Electron Microscope)
SEM image of activated carbon as shown in Fig.4 exhibits the pores with different diameters. The formation
of porous structure in the activated carbon may be due to the dehydrating action of zinc chloride. Zinc
chloride due to the dehydrating action, removes oxygen and hydrogen from the precursor as water and pores
are formed.
4. Conclusion
The current study explores the characterization of activated carbon prepared from Rudraksha beads by
chemical activation with zinc chloride by chemical methods like Methylene blue number and Iodine number
and instrumental techniques such as SEM and XRD. SEM of activated carbon shows the pores of different
diameters. XRD analysis indicated two broad diffraction peaks located near at 2θ = 25.5 o and 43 o reflected
from 002 and 100 planes. Methylene blue number and Iodine number are found to be 499.6 mg/g and 920.6
mg/g respectively. The results observed from chemical methods and instrumental techniques indicates that the
activated carbon prepared can be used for the removal of pollutants.
Acknowledgement
The authors would like to thank Dr. Lok Kumar Shrestha, National Institute for Materials Science
(NIMS), Tsukuba, Japan for XRD and SEM of activated carbon. The authors also acknowledge The University
Grants Commission, Nepal, for financial support (Award No: FRG 75/76-S &T-4).
２６９
AEICP Vol. 5, No. 1
References
[1] Ahmad Alhujaily, Hongbo Yu,Xiaoyu Zhang, Fuying Ma,(2020), Adsorptive removal of anionic dyes from
aqueous solutions using spent mushroom waste, Applied Water Science,10:183.
[2] Silvia Álvarez-Torrellas, Rafael García-Lovera, Araceli Rodríguez, Juan García, (2015), Removal of
Methylene Blue by Adsorption on Mesoporous Carbon from Peach Stones, Chemical Engineering
Transactions, Vol.43, 1963-1968.
[3] Abdelali G., Abdelmalek C.1, Réda Y.A., Ammar S. and Boubekeur N. (2019), Removal of methylene
blue using activated carbon prepared from date stones activated with NaOH, Global Nest Journal, Vol.
21, 1-7.
[4] Malik D.S., Jain C.K. Anuj K. Yadav, Richa Kothari, Vinayak V. Pathak, (2016), Removal of methylene
blue dye in aqueous solution by agricultural waste- corn stalk, International Research Journal of
Engineering and Technology.Vol.3, 864-880.
[5] Deshpande, D.P. Kirti Zare, Pankaj Vardhe and Utkarsh Maheshwari, (2017), Removal of dye from
aqueous solution using activated carbon from coconut shell, Research Journal of Chemical Sciences, Vol.
6(7), 20-24.
[6] Shrestha Rajeshwar Man and Joshi Sahira, (2019), Application of Coconut Shell for the Preparation of
Activated Carbon to Remove Heavy Metal from Aqueous Solution, International Journal of Advanced
Engineering, Vol.2(2), 1-10.
[7] Gecgel Unal,Uner Osman, Gokara Guney,(2016),Adsorption of Cationic dyes on activated carbon obtained
from waste Elaeagnus Stone, Adsorption Science and Technology,34 (9-10),512-525.
[8] Shrestha Rajeshwar Man, (2018), Adsorption Isotherms and Kinetics Studies of Ni (II) Adsorption from
Aqueous Solution onto Activated Carbon Prepared from Lapsi Seed Stone, Indian Journal of Engineering,
Vol.15, 291-298.
[9] Raposo, F., De La Rubia M.A., Borja, R., (2009), Methylene blue number as useful indicator to evaluate
the adsorptive capacity of granular activated carbon in batch mode: Influence of adsorbate/adsorbent mass
ratio and particle size”, Journal of Hazardous Materials, 165, 291–299.
[10]Shrestha Rajeshwar Man, (2018), Adsorption Isotherm and Kinetic Studies of Cd(II) from Aqueous
Solution using Activated Carbon Prepared from Lapsi Seed Stone by Chemical Activation, International
Journal of Advanced Engineering, Vol.1 (1), 16-22.
２７０
8th ICAEIC-2022
Isolation of Cellulose from Sisal and Its Application for Removal of Arsenic
from Aqueous Solution
Manish Shrestha, Sanju Khatri, and Bindra Shrestha
Department of chemistry, Tri-chandra Multiple campus Kathmandu, Nepal
binraghu@yahoo.com
Abstract
People are becoming more aware of the organic value of eco-friendly products, which has excited their interest
in the use of eco-friendly natural fibers. In this work, cellulose was isolated from sisal (Agave haradiana) and
then modified by iron oxide (Fe(NO3)3.9H2O). Cellulose isolation was performed by extraction followed by
delignification process. Isolated cellulose fiber (CF) and modified cellulose fiber (MCF) were used as bio-
adsorbents and were analyzed by using FTIR, XRD, and EDX. The present study showed that CF and MCF
are an effective adsorbent for removal of As (III) from aqueous solution. The ability of cellulose fiber to adsorb
arsenic from aqueous solution has been investigated through batch experiments. The arsenic adsorption was
found to be dependent on pH, contact time and initial concentration. The maximum adsorption was obtained
at pH 9 for both CF and MCF. The equilibrium time for sorption of As (III) onto CF and MCF were 180 min
and 210 min, respectively. Adsorption isotherm test showed that equilibrium adsorption data were better
represented by Langmuir isotherm model than Freundlich isotherm model for both CF and MCF and maximum
adsorption capacity of CF and MCF were found to be 21.42 mg/g and 26.35 mg/g, respectively. Kinetic
modelling studies displays that the experimental data best fitted to pseudo second order kinetic model for both
CF and MCF.
Keywords: arsenic, cellulose fiber, bio-adsorbent, adsorption isotherm, kinetics.
1. Introduction
Biopolymers are naturally occurring macromolecules that are typically produced by living systems such as
plants, animals, and microorganisms. These biopolymers may contain various functional groups such as
hydroxyl, amino, amide, carboxyl, phosphate, phenolic, and others, which contribute to their various biological
activities. Biopolymers are typically divided into three categories: polysaccharides (Starch, Cellulose, Agar,
Alginate, Carrageenan, Pectin), proteins (silk, Collagen/Gelatin, Elastin Polyaminoacid, Soy, Zein, Wheat
Gluten, Casein) and polynucleotides(DNA and RNA). (P. Yadav, 2015) (Qureshi et al., 2020) Polysaccharides
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２７１
AEICP Vol. 5, No. 1
are polymers that are biodegradable and found in all living organisms. These polymers are made up of
monosaccharide units that are linked together by glycosidic linkages, which are a type of ether bond. They are
very promising natural biomaterials due to their biodegradability, processability, and bioactivity. They are
biopolymers that are abundant, renewable, biocompatible, nontoxic, hydrophilic, and photostable.
Polysaccharides can be derived from a variety of natural sources. (García, 2018) Natural saccharides are
primarily composed of simple carbohydrates known as monosaccharides with the general formula (CH2O)n,
where n is three or more. Glucose, fructose, and glyceraldehyde are examples of monosaccharides.
Polysaccharides, on the other hand, have a general formula of Cx(H2O)y, where x is basically a large number
between 200 and 2500. (Morrison R T & Boyd R N.)
Cellulose, hemicellulose, and lignin are the three primary components of plant fibers. Cellulose, which
provides the mechanical properties of the entire natural fiber, is organized in micro-fibrils and is surrounded
by the other two main components: hemicellulose and lignin. (Moran et al., 2008) Natural cellulosic fibers are
composed of 60-95 % cellulose. The remaining constituents are hemicellulose, lignin, pectin, waxes, and
proteins, with their proportions varying depending on growth conditions, fiber source, and fiber isolation
method. (Reddy & Yang, 2005) Cellulose is most abundant natural polymer on the planet with molecular
formula (C6H10O5)n. Its molecular chains are typically arranged into larger structures known as cellulose,
which is made up of glucose units connected by β-1-4-linkages. (Martinez-Sanz et al., 2017) (Hon, 1994)
Cellulose is an unbranched polymer composed of β (1 - 4) D-glucopyranosyl units.
O
HO
OH
OH
OO
HO
OH
OH
O
n
Figure 1. Molecular structure of (C6H10O5)n.
Industrial pollution levels have been gradually growing in many nations as a result of fast growth and
industrialization. As a result, the world's industrial wastewater pollution problem is becoming very serious.
(Hashem, 2007) With the fast expansion of agriculture, industry, commerce, hospitals, and healthcare facilities,
many activities are consuming enormous amounts of harmful chemicals and producing large amounts of
hazardous waste. There are about 110000 different types of hazardous compounds commercially accessible
today. Every year, a thousand new chemicals for industrial and other purposes are introduced to the market
Heavy metals are one of the most dangerous contaminants in the environment. (Feng et al., 2004).
Many human activities leak heavy metals into the atmosphere. Parent rocks and metallic minerals are the
most common natural sources, whereas agricultural practices, such as the use of heavy metal-containing
fertilizers, animal manures, and pesticides, metallurgical activities, such as mining, smelting, metal finishing,
２７２
8th ICAEIC-2022
and other energy processing, microelectronic goods, and transportation, are the most common anthropogenic
sources. (Bradl, 2005) Heavy metals in the environment as aerosols, gases and particulates can be harmful to
a wide range of living species. Heavy metals are distinguished from other toxic pollutants by their inability to
biodegrade and their affinity to accumulate in living matter. Therefore, removal of heavy metals from
wastewater is a critical public health issue. Many methods for removing heavy metals have been proposed.
Some of the most common processes are chemical precipitation, membrane filtration, ion exchange, alum
coagulation, iron coagulation, and adsorption. (Orhan & Buyukgungor, 1993) Adsorption is thought to be less
expensive than membrane filtration, easier and safer to handle than precipitated contaminated sludge, and more
versatile than ion exchange among possible treatment processes. (Guo & Chen, 2005).
2. Materials and Methods
2.1.
Materials
Sisal from Snkhuwasabha and Bhaktapur Districtwere used in this work.Other reagents used were: Nitric
acid (HNO3, 99.9%), Sodium Hydroxide (NaOH) , sulphuric acid (95-98%), hydrogen peroxide (40% solution)
and distilled water.
2.2.
Cellulose isolation
The SF was used as a source of cellulose for the study. Fiber of sisal plant was washed with distilled water
and then filtered. The washed SF was dried in sunlight for 2-3 days. The fiber of sisal was treated with toluene
and ethanol (2:1) for removal of waxes. After dewaxing, sample was washed by distilled water and dried in
sunlight. The dewaxed sample was put in a container and 1000 mL of 1 M HNO3 was added to remove pectic
polysaccharides. Then solution was filtered and dried in oven and then left at room temperature. After that,
obtained residue was treated with 1000 mL of 6 M NaOH and left for 4 hrs. It was then filtered to remove
impurities like hemicellulose and lignin. The solution was filtered and residue was dried in sun until complete
dryness. The obtained dry residue was bleached using 4 % H2O2 solution and then washed many times by
distilled water and dried on sunlight and finally pure cellulose was obtained. (Szymarska-Chargot et al., 2017)
3. Characterization Techniques
The isolated samples were characterized via X-ray diffraction (XRD) and Fourier transform infrared
spectroscopy (FTIR) and Energy-dispersive X-ray (EDX).
3.1. X-ray diffraction (XRD)
X-ray diffraction (XRD), also known as x-ray crystallography, is a technique used to identify the phase of
a material and provide information on and unit cell dimensions. The angle between the incident and diffracted
rays is a critical component of all diffraction. At the Nepal Academy of Science and Technology (NAST),
Lalitpur, Nepal, the crystal phase and structure of the samples were determined using an X-ray diffractometer
(Bruker D2 Phaser) with a monochromatic
２７３
AEICP Vol. 5, No. 1
3.2. Fourier transform infrared spectroscopy (FTIR)
The FTIR analysis was performed using an IR tracer-100 FTIR Spectrometer (SHIMADZU) at the NAST,
Nepal to identify the species, functional groups, and vibration modes associated with each peak. Spectra were
obtained using the KBR pellet method in the spectral range of 4000-400 cm-1
3.3. Energy Dispersive X-ray (EDX) Analysis
Energy Dispersive X-ray (EDX) Analysis was performed using an EDX-8000 (SHIMADZU at Central
Customs Laboratory, Kathmandu, Nepal to reveals the presence of elements present in the specimens. EDX
was used to examine surface morphology, microstructure and elemental composition of Fe loaded and
unloaded adsorbents after arsenic adsorption.
4. Conclusions
In this work, cellulose fiber was isolated from sisal fiber and obtained cellulose was modified by
Fe(NO3)3.9H2O. These two cellulose fiber and modified cellulose fiber were characterized by using FTIR,
XRD, and EDX. FTIR data concluded that the isolated cellulose was found to be type II. XRD experiments
confirms the cellulose fiber was amorphous in nature. EDX analysis of both CF and MCF suggested the
existence of Fe(OH)3 and adsorption of As (III) more on MCF. Isolated cellulose fiber and modified cellulose
fiber were investigated as low-cost bio-adsorbent for removal of As (III) ion from aqueous solution. Batch
adsorption technique was used to compare efficiency of adsorbents using different parameters.
The adsorption of arsenic was found to be strongly dependent upon pH of solution and adsorption was
maximum at the optimum pH of 9 for both CF (81 %) and MCF (88 %). The equilibrium time for adsorption
of As (III) onto CF and MCF were found to be 180 min and 210 min, respectively. The maximum adsorption
capacity of CF and MCF were found to be 21.42 mg/g and 26.35 mg/g, respectively. The adsorption isotherm
study revealed that Langmuir adsorption isotherm can best explain the adsorption of the CF and MCF for As
(III) the kinetic data was analyzed using pseudo first and pseudo second order kinetic models. It was found
that the obtained data were best fitted using pseudo second order kinetic model. Therefore, cellulose fiber can
be used as potential bio-adsorbent for removal of As (III) from aqueous solution and can be used for water
treatment. From above all information modified cellulose was more effective than unmodified cellulose for
arsenic removal from aqueous solution.
References
[1] Bradl, H. B. (2005). Chapter 1 Sources and origins of heavy metals. In Interface Science and Technology
(Vol. 6, pp. 1–27). Elsevier. https://doi.org/10.1016/S1573-4285(05)80020-1
[2] Feng, Q., Lin, Q., Gong, F., Sugita, S., & Shoya, M. (2004). Adsorption of lead and mercury by rice husk
ash. Journal of Colloid and Interface Science, 278(1), 1–8.
[3] García, M. C. (2018). Drug delivery systems based on nonimmunogenic biopolymers. In Engineering of
Biomaterials for Drug Delivery Systems 317–344. https://doi.org/10.1016/B978-0-08-101750-0.00012-X
２７４
8th ICAEIC-2022
[4] Guo, X., & Chen, F. (2005). Removal of Arsenic by Bead Cellulose Loaded with Iron Oxyhydroxide from
Groundwater. Environmental Science & Technology, 39(17), 6808–6818. https://doi.org/10.1021/
es048080k
[5] Hashem, M. A. (2007). Adsorption of lead ions from aqueous solution by okra wastes. International
Journal of Physical Sciences, 2(7), 178–184.
[6] Hon, D. N.-S. (1994). Cellulose: A random walk along its historical path. Cellulose, 1(1), 1–25.
https://doi.org/10.1007/BF00818796
[7] Martínez-Sanz, M., Pettolino, F., Flanagan, B., Gidley, M. J., & Gilbert, E. P. (2017). Structure of cellulose
microfibrils in mature cotton fibres. Carbohydrate Polymers, 175, 450–463. https://doi.org/10.1016/
j.carbpol.2017.07.090
[8] Morán, J. I., Alvarez, V. A., Cyras, V. P., & Vázquez, A. (2008). Extraction of cellulose and preparation
of nanocellulose from sisal fibers. Cellulose, 15(1), 149–159. https://doi.org/10.1007/s10570-007-9145-9
[9] Morrison R T, Boyd R N 9780205084548: Study Guide to Organic Chemistry: Amazon.com: Books.
Retrieved September 6, 2021, from https://www.amazon.com/Study-Guide-Organic-Chemistry
Morrison/dp/0205084540
[10] Orhan, Y., & Büyükgüngör, H. (1993). The Removal of Heavy Metals by Using Agricultural Wastes.
Water Science and Technology, 28(2), 247–255. https://doi.org/10.2166/wst.1993.0114
[11] Qureshi, D., Nayak, S. K., Anis, A., Ray, S. S., Kim, D., Hanh Nguyen, T. T., & Pal, K. (2020).
Introduction of biopolymers. In Biopolymer-Based Formulations (pp. 1–45). Elsevier.
https://doi.org/10.1016/B978-0-12-816897-4.00001-1
[12] Reddy, N., & Yang, Y. (2005). Properties and potential applications of natural cellulose fibers from
cornhusks. Green Chemistry, 7(4), 190.
[13] Szymarska-Chargot, M., Chylińska, M., Gdula, K., Kozioł, A., & Zdunek, A. (2017). Isolation and
Characterization of Cellulose from Different Fruit and Vegetable Pomaces. Polymers, 9(12), 495.
https://doi.org/10.3390/polym9100495
[14] Yadav, P. (2015). Biomedical Biopolymers, their Origin and Evolution in Biomedical Sciences: A
Systematic Review. Journal of clinical and diagnostic research. https://doi.org/10.7860/JCDR/2015/
13907.6565
２７５
AEICP Vol. 5, No. 1
Analysis of the Movement Path of Micro Dust in a Simulation-based Urban
Environment for Specifying the Installation Location of the Optical Particle
Counter
Jeong-Gi Lee, Chul-Seung Yang, Gi-won Ku, Ji-Seong Jeong
Korea Electronics Technology Institute
{jklee, yangcs, giwon9, json }@keti.re.kr
Abstract
Hundreds of optical particle counters are planned to be installed in urban areas to measure air pollution in
urban environments. Before that, this study was conducted to analyze the movement path of micro dust. In this
paper, 3D modeling of Gwangju City in Korea is performed and simulation analysis is performed to confirm
the movement path of micro dust. As a result, several points where a group of micro dust gather were revealed,
and based on the results of this study, we plan to demonstrate a micro dust monitoring system in the city.
Keywords: microdust, analysis, simulation, particle, tracking, modeling.
1. Introduction
Nowadays, many people are concerned about air pollution. In particular, air pollution particulate matter is
very fatal to the human body. [1]-[3] Accordingly, many studies such as air pollution modeling analysis and
air pollution monitoring are being conducted. [4]-[6] In this study, the movement path of micro dust was
analyzed by 3D modeling of Gwangju Metropolitan City in Korea. Simulation was performed using the particle
tracking analysis method based on the Fluent analysis system. The governing equations were analyzed by
applying the k-ε turbulence analysis of the flow analysis.
2. Particle Tracking Analysis
2.1. Boundary Conditions
As shown in Table 1., it was set as an environmental boundary condition based on the 1991 ~ 2020 normal
year climate data of the Korea Meteorological Administration. PM is an abbreviation of Particulate Matter,
and PM2.5 refers to dust with a diameter of 2.5 ㎛ or less. In order to express PM2.5 in the simulation, the
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２７６
8th ICAEIC-2022
simulation was performed by setting particles with diameters of 0.1, 0.3, and 1 ㎛.
Table 1. Environmental boundary conditions.
Specification
Value
Temperature
14.1 ℃
Humidity
68.1 %
Precipitation
1380.6 mm
Wind speed
Wind direction
PM2.5
2.0 m/s
West
24 ㎍/m3
2.2. 3D Modeling and analysis methods
Based on the 2D contour data, the topography of Gwangju, such as mountains and rivers, was constructed
in 3D. The result can be seen in Figure 1. It is configured in the form of a mesh so that the flow of particles
can be easily seen. The size of the model is about 1.2 x 0.8m because the model was reduced to a size of
500,000:1 compared to the actual model. Additionally, for mesh quality, the element size is 50mm and the
mesh is configured. Finally, the scale of the reduced model is 1.2(W) x 0.8(L) m, and the z-axis of the flow
field is 0.05 m: the actual z-axis is 25 km. Simulation was performed using the particle tracking analysis
method based on the Fluent analysis system. The governing equations were analyzed by applying the k-ε
turbulence analysis of the flow analysis. The formula for turbulent kinetic energy is (1).
()
+()
=
+ (1)
Figure 1.
3D Modeling: Gwangju, Korea.
2.3. Particle tracking Analysis Result
Simulation analyzes were performed 100 times. Through this analysis, the flow of particles could be seen,
and finally, particles remaining in a specific area could be identified. In other words, It can be seen that particles
remains on the back side of most mountainous terrain. As a result, Since it is efficient to collect residual
２７７
AEICP Vol. 5, No. 1
particles when installing an optical particle counter, the particle residual area is a suitable place to measure
micro dust.
Figure 2.
Block diagram of proposed system.
3. Conclusion
In this study, to carry out a project to demonstrate hundreds of optical particle counters, the movement path
of micro dust was analyzed by 3-dimensional modeling of Gwangju Metropolitan City in Korea. 100
simulations were performed using the particle tracking analysis method based on the Fluent analysis system.
As a result, it was found that the particles were gathered at three points. In order to intensively monitor the air-
polluted areas, it is planned to install optical particle counters intensively at those three points.
Acknowledgement
This research was supported by a subsidy from The Regional Development Investment Agreement Pilot
Project(B0070510000127) supported by the Ministry of Land, Infrastructure and Transport, Gwangju
Metropolitan City and Gwangsan-gu.
References
[1] Becker, S., Soukup, J. M., & Gallagher, J. E. (2002). Differential particulate air pollution induced oxidant
stress in human granulocytes, monocytes and alveolar macrophages. Toxicology in vitro, 16(3), 209-218.
[2] Gurgueira, S. A., Lawrence, J., Coull, B., Murthy, G. K., & González-Flecha, B. (2002). Rapid increases
in the steady-state concentration of reactive oxygen species in the lungs and heart after particulate air
pollution inhalation. Environmental health perspectives, 110(8), 749-755.
[3] Utell, M. J., & Frampton, M. W. (2000). Acute health effects of ambient air pollution: the ultrafine particle
hypothesis. Journal of aerosol medicine, 13(4), 355-359.
[4] Zannetti, P. (Ed.). (2013). Air pollution modeling: theories, computational methods and available
software. Springer Science & Business Media.
[5] Vallero, D. A. (2016). Air pollution monitoring changes to accompany the transition from a control to a
systems focus. Sustainability, 8(12), 1216.
２７８
8th ICAEIC-2022
[6] Xie, X., Semanjski, I., Gautama, S., Tsiligianni, E., Deligiannis, N., Rajan, R. T., ... & Philips, W. (2017).
A review of urban air pollution monitoring and exposure assessment methods. ISPRS International
Journal of Geo-Information, 6(12), 389.
２７９
AEICP Vol. 5, No. 1
Design and Development of Fan Boss to Reduce Noise and Vibration of Air
Purifier Fan
Hyeong-Sam Park1, Doo-Sik Kim 2, and Sang-Hyun Lee 3
1,2Institute of Technology MIDO Co., Ltd, Korea
3Department of Computer Engineering, Honam University, Korea
postmido@gmail.com1, entlr103@naver.com2, leesang64@honam.ac.kr3
Abstract
Recently, technology development for home appliances such as air purifiers is being developed in
consideration of environmental issues as well as high efficiency and multifunctionality. Vibration and noise
are generated as a problem with air purifiers, which are recognized as environmental issues rather than
performance. These vibration and noise issues cause stress and fatigue when exposed to the vibration and
noise generated by the fan for a long time. The cause of noise is fan vibration, which not only reduces energy
efficiency but also adversely affects product life. Therefore, the purpose of this paper is to reduce the noise by
reducing the vibration of the fan used in the air purifier and dehumidifier. Therefore, it is intended to develop
a fan boss assembly that reduces noise and vibration, and a unit that reduces noise and vibration while making
the operation of a fan that sucks in and exhausts wind flexibly.
Keywords: Fan Boss, Air Purifier, volatile organic compounds.
1. Introduction
Recently, air pollution in the atmosphere has become serious due to fine dust, pollutants and viruses. As a
result, air purifier-related air appliances are being launched in each business or home, and related companies
are making and selling air purifiers, air conditioners, and dehumidifiers to satisfy various consumer needs.
These air cleaning-related air home appliances are being released in a small, customized form with specialized
functions according to various purposes and uses, and the demand is also increasing significantly.
According to the announcement by the Ministry of Environment in 2017, the number of people who died
prematurely due to the increase in ultrafine dust increased to 120,000 people, and the public health threat posed
by fine dust increased. As part of the roadmap to reduce domestic fine dust emissions by 30% by 2022 with
the government’s announcement of comprehensive measures for fine dust management, as preemptive and
proactive measures are being taken, measures to reduce fine dust are urgently needed from the point of view
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２８０
8th ICAEIC-2022
of protecting the weak [1].
As such, the problem of indoor air quality that directly affects humans is becoming a significant social issue,
and as the social structure changes, the time spent indoors is getting longer. Accordingly, the impact of the
indoor environment on humans is also increasing. As the public is known about the effects of dust, bacteria,
volatile organic compounds (VOCs), and formaldehyde floating in the indoor space on the human body, there
is a surge in demand for improvement [1].
Recently, technology development for home appliances such as air purifiers is progressing in the direction
of considering environmental issues along with high efficiency and multifunctionality. In addition, among the
environmental problems that people are very interested in, the problem of vibration and noise is also recognized
as an environmental problem rather than being recognized in terms of performance.
These vibration and noise problems cause stress and fatigue when people are exposed to vibration and noise
generated by the fan for a long time. The cause of noise is fan vibration, which not only reduces energy
efficiency but also adversely affects the life of the product.
Therefore, the purpose of this paper is to find ways to reduce the vibration of fans used in air purifiers and
dehumidifiers to reduce noise. Therefore, we are going to develop a fan boss assy that reduces noise and
vibration and a unit that makes the operation of the fan used to suck in and discharge wind flexibly and reduces
noise and vibration.
2. Research Contents
As shown in Fig. 1, it was set centering on road and non-road mobile pollution sources such as urban multi-
living spaces, bus stops/road areas, and areas with a concentration of living pollutants, as well as living
pollutants that generate non-industrial combustion, biological combustion, and scattering dust [2].
Figure 1. Maps of PM2.5 areas in Seoul.
Air purifiers are showing the fastest growth among single home appliances, and according to the home
appliance industry, the domestic air purifier market is estimated to be worth 1.5 trillion won in 2018. In addition,
the number of air purifiers supplied in Korea exceeded 2.5 million units last year and is expected to reach 3
million units in 2021, but the penetration rate is only less than 40% as of 2020.
As shown in Fig 2, the fan boss for generating blow among the important internal parts of the air purifier is
generally manufactured by casting, and it should be as small or light as possible to reduce the power
consumption of the fan. It must be able to withstand vibration and vibration. If the drive shaft of the motor is
２８１
AEICP Vol. 5, No. 1
directly connected to the center of the rotating body where the fanboss is not coupled, a large load is applied
from the drive shaft during the initial operation of the motor, and the installation part of the drive shaft is easily
damaged. This fan injection molding process has the characteristics of being very superior in moldability,
productivity and economic feasibility compared to other production processes. And it is used in the production
of various products such as household goods.
Figure 2. Fan Boss Assy.
3. Designed by Fan Boss Assy
FAN BOSS ASSY
Plate-Fanboss
Mount-Fanboss
Shaft-Fanboss
Figure 3. Designed by Fan Boss Assy.
Figure 3 is newly designed by supplementing the existing fan boss assy problem. Due to the existing problem,
if the drive shaft of the motor is directly connected to the center of the rotating body where the fanboss is not
coupled, a large load is applied from the drive shaft during the initial operation of the motor, and the installation
part of the drive shaft can be easily damaged. Also, even if the fanboss is closely coupled to the rotating body,
the outer peripheral surface of the fan and the inner peripheral surface of the rotating body are separated from
each other by the force in the circumferential direction (centrifugal force) and the axial force (thrust) acting as
the rotating blade rotates and slips. As this occurs or the coupling part is damaged, there is a problem in that
the rotational force provided from the drive shaft of the motor is not properly transmitted to the fan.
To compensate for these existing problems, it is installed between the drive shaft of the motor and the fan
to transmit the rotational force provided by the motor to the fan, while alleviating the shock or vibration
２８２
8th ICAEIC-2022
generated by the rotational force of the motor, and a firm bond can be maintained so that slip and breakage do
not occur due to circumferential and axial forces.
4. Result
FAN BOSS ASSY(a) Plate-Fanboss(b) Mount-Fanboss(c)
Shaft-Fanboss(d)
Figure 4. Produced by fan boss assy.
Figure 4 is the result of improving the problems of the existing FAN BOSS ASSY. (a) of Figure 4 shows
the improved fan boss assy, and (b) is a Plate-Fanboss - several structures were used to increase the friction
area, but the hardness of the TPE was increased to secure the strength. In addition, in (c), the moldability is
very good among the materials, the weight width of the pan balance is small, and strength security is possible
according to hardness, and in (d), the extrusion molding is larger than the original size to facilitate post-
processing.
5. Conclusion
Recently, there is a high interest in vibration and noise problems in home appliances such as air purifiers,
and these problems are recognized as environmental problems rather than product performance.
These vibration and noise issues cause stress and fatigue when exposed to the vibration and noise generated
by the fan for a long time. The cause of noise is fan vibration, which not only reduces energy efficiency but
also adversely affects product life.
Therefore, the purpose of this paper is to find a way to reduce the noise by reducing the vibration of the fan
used in the air purifier. Therefore, we developed a fan boss assembly that reduces noise and vibration and a
unit that reduces noise and vibration while making the operation of a fan that sucks in and exhausts wind
flexibly.
Through the static structural analysis of the fan boss assembly, the deformation and stress distribution of the
fan and boss during operation were identified, and the vibration characteristics, the resonance frequency, and
the resulting deformation were relatively compared through the modal analysis, and the performance was
improved through the parameter study. The optimal shape was derived in the direction.
Although fandml rotation is made at the center of the rotating body, vibration and noise are avoided due to
tolerances that inevitably occur during product manufacturing and installation (assembly), fluid flow
characteristics (speed, turbulence, etc.), and friction between parts. To reduce this, a fan boss was installed in
the center of the rotating body, and as a result, the vibration and noise characteristics were changed according
to the shape of the fan boss or the characteristics of the material, thereby reducing vibration and noise.
２８３
AEICP Vol. 5, No. 1
Acknowledgement
This work was supported by the Technology development Program(S2948636) funded by the Ministry of
SMEs and Startups(MSS, Korea).
References
[1] Elkamhawy, A., Lee, S. -M., & Jang, C.-M. (2020). Development of tower-type air purifier for
atmospheric air purification. Proceedings of the Korean Society for Fluid Machinery, 213-214.
[2] Kim, D., Choi, M. & Yoon, B., (2019). Analysis of PM hot-spot Emission Zone in Seoul Metropolitan
area, Journal of Korean Society for Atmospheric Environment, 35(4), pp. 476-501.
２８４
8th ICAEIC-2022
Characterization of Activated Carbon Prepared from Walnut (Jaglans regia)
Shell
Srijan Adhikari1, Rinita Rajbhandari (Joshi) 1, and Rajeshwar Man Shrestha*
1Department of Applied Sciences and Chemical Engineering, Pulchowk Campus, Nepal
*
Institute of Engineering, Tribhuvan University, Nepal
rajeshwar@ioe.edu.np
*
Abstract
Activated carbon prepared from Walnut (Jaglans regia) shell by chemical activation with zinc chloride at 400
o C was characterized by chemical methods such as Iodine number and Methylene blue number to identify
micropore and mesopore content. Methylene blue and iodine numbers are found to be 499.5 mg/g and 1053.4
mg/g respectively. The activated carbon was also analyzed by instrumental techniques like XRD and SEM.The
SEM of the activated carbon indicates the pores of different diameters whereas XRD profile exhibited two
broad diffraction peaks indicating the amorphous nature of the activated carbon..
Keywords: Activated carbon, Walnut shell, Chemical activation, Zinc chloride.
1. Introduction
Activated carbon is highly porous carbonaceous material with high surface area. Because of extended
surface area, micro-pore structures, high adsorption capacity and high degree of surface reactivity, activated
carbon is used to purify liquids and gases in a variety of applications, including municipal drinking water, food
and beverage processing, odor removal, industrial pollution control. However, commercially available
activated carbon is very expensive and has high regeneration cost while being exhausted. Furthermore,
generation using solution produces a small additional effluent while regeneration by refractory technique
results in a 10–15 % loss of adsorbent and its uptake capacity. This has led to search for cheaper substances
for the preparation of adsorbents. Activated carbon is produced from carbonaceous source materials such as
peach stone [1], date stone [2], coconut shell [3] and [4], date stone [5] cornstalk [6], mushroom waste [7],
Elaeagnus stone [8], Lapsi Seed Stone [8]and [9] etc.
Present study deals with Peach stone as shown in Fig. 1 for the preparation of activated carbon. Peach stone
is the waste materials left after using fleshy part of peach fruits as shown in Fig. 2. Every year large amount of
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２８５
AEICP Vol. 5, No. 1
peach stones has been left as waste materials. The application of the waste materials for preparation of activated
carbon minimizes environmental pollution.
Figure 1. Peach fruits. Figure 2. Peach stones.
2. Experimental
2.1. Materials
The precursor used in this study is Walnut shell. The precursor was washed with tap water and then with
distilled water. The precursor was crushed and sieved to a size range of 300 μm and was soaked with zinc
chloride in the ratio of 1:1. The materials were dried in an oven at 110 o C. The dried Walnut shell particles
were carbonized under high purity nitrogen flow of 75 ml min−1 by raising the temperature in tubular furnace
at a rate until 400 o C and kept at this temperature for three hours. The prepared activated carbon was then
cooled at room temperature and washed with dil. HCl and then with warm distilled water until the pH of the
washing reached 6 –7.
2.2. Chemicals and Equipment
The chemicals and reagents used are of analytical grade (Merck and Qualigens Company. To adjust pH of
solutions Digital pH meter was used. The adsorption experiments for the determination of Methylene blue
number were carried out by using Shaker (Digital VDRL Rotator RPM-S) and UV –Visible Spectrophotometer.
2.3
.
Methylene blue number and Iodine number
Methylene blue number and Iodine number have been determined to determine amount of micropore and
mesopore content in activated carbon. Methylene blue number is the amount of the dye adsorbed by one gram
of the adsorbent while the iodine number indicates the amount of iodine adsorbed per gram of activated carbon
at an equilibrium concentration. Methylene blue number can be calculated by the following formula [10].
=()×
(1)
where Co and Ce are initial and equilibrium concentration of MB (mg / L) respectively, M is the mass of
adsorbent in gram and V is the volume of the solution in liter. Iodine number can be calculated by the following
equation [11].
=
(2)
２８６
8th ICAEIC-2022
3. Characterization of Activated carbon
3.1. XRD (X-ray diffraction)
Two broad diffraction peaks located near at 2θ = 25.5 o and 43o reflected from 002 and 001 planes have been
revealed by XRD analysis of the activated as shown in Fig.3. The two broad diffraction peaks indicate the
amorphous nature of the activated carbon which is one of the good properties of activated carbon for the
adsorption.
Figure 3. XRD of Activated Carbon. Figure 4. IN and MBN of Activated Carbon.
Figure 5. SEM of Activated Carbon.
3.5. Methylene blue number (MBN) and Iodine number (IN)
The chemical methods like Methylene blue number and Iodine blue number have been determined to analyze
the mesopores and micropore content in the activated carbon and are presented in Fig. 4. Methylene blue and
iodine numbers are found to be 499.5 mg/g and 1053.4 mg/g respectively.
3.6. SEM (Scanning Electron Microscope)
SEM image of activated carbon as shown in Fig.5 exhibits the pores with different diameters. The formation
of pores in the activated carbon may be due to the dehydrating action of zinc chloride. Owing to dehydrating
action it removes water and pores are formed.
２８７
AEICP Vol. 5, No. 1
4. Conclusion
This study demonstrates the characterization of activated carbon prepared from Walnut shells by chemical
activation with zinc chloride by chemical methods like Methylene blue number and Iodine number and
instrumental techniques such as SEM and XRD. SEM of activated carbon exhibits the pores of different
diameters. XRD analysis indicates two broad diffraction peaks located near at 2θ = 25.5 o and 43 o reflected
from 002 and 100 planes. Methylene blue number and Iodine number are found to be 499.5 mg/g and 1053.4
mg/g respectively. The results observed from SEM and XRD. MBN and IN indicates that the activated carbon
prepared can be used to purify liquids and gases in a variety of applications, including municipal drinking
water, food and beverage processing, odor removal, industrial pollution control.
Acknowledgement
The authors would like to thank Dr. Lok Kumar Shrestha, National Institute for Materials Science (NIMS),
Tsukuba, Japan for XRD and SEM of activated carbon. The authors also acknowledge The University Grants
Commission, Nepal, for financial support (Award No: FRG 75/76-S &T-4).
References
[1] Silvia Álvarez-Torrellas, Rafael García-Lovera, Araceli Rodríguez, Juan García, (2015), Removal of
Methylene Blue by Adsorption on Mesoporous Carbon from Peach Stones, Chemical Engineering
Transactions, Vol.43, 1963-1968.
[2] Ahmad Alhujaily, Hongbo Yu,Xiaoyu Zhang, Fuying Ma,(2020), Adsorptive removal of anionic dyes from
aqueous solutions using spent mushroom waste, Applied Water Science,10:183
[3] Deshpande, D.P. Kirti Zare, Pankaj Vardhe and Utkarsh Maheshwari, (2017), Removal of dye from
aqueous solution using activated carbon from coconut shell, Research Journal of Chemical Sciences, Vol.
6(7), 20-24.
[4] Shrestha Rajeshwar Man and Joshi Sahira, (2019), Application of Coconut Shell for the Preparation of
Activated Carbon to Remove Heavy Metal from Aqueous Solution, International Journal of Advanced
Engineering, Vol.2(2), 1-10.
[5] Abdelali G., Abdelmalek C.1, Réda Y.A., Ammar S. and Boubekeur N. (2019), Removal of methylene
blue using activated carbon prepared from date stones activated with NaOH, Global Nest Journal, Vol.
21, 1-7.
[6] Malik D.S., Jain C.K. Anuj K. Yadav, Richa Kothari, Vinayak V. Pathak, (2016), Removal of methylene
blue dye in aqueous solution by agricultural waste- corn stalk, International Research Journal of
Engineering and Technology.Vol.3, 864-880.
[7] Ahmad Alhujaily, Hongbo Yu,Xiaoyu Zhang, Fuying Ma,(2020), Adsorptive removal of anionic dyes from
aqueous solutions using spent mushroom waste, Applied Water Science,10:183
[8] Gecgel Unal,Uner Osman, Gokara Guney,(2016),Adsorption of Cationic dyes on activated carbon obtained
from waste Elaeagnus Stone, Adsorption Science and Technology,34 (9-10),512-525.
２８８
8th ICAEIC-2022
[9] Shrestha Rajeshwar Man, (2018), Adsorption Isotherms and Kinetics Studies of Ni (II) Adsorption from
Aqueous Solution onto Activated Carbon Prepared from Lapsi Seed Stone, Indian Journal of Engineering,
Vol.15, 291-298.
[10] Raposo, F., De La Rubia M.A., Borja, R., (2009), Methylene blue number as useful indicator to evaluate
the adsorptive capacity of granular activated carbon in batch mode adsorbate/adsorbent mass ratio and
particle size”, Journal of Hazardous Materials, 165, 291–299.
[11] Shrestha Rajeshwar Man, (2018), Adsorption Isotherm and Kinetic Studies of Cd (II) from Aqueous
Solution using Activated Carbon Prepared from Lapsi Seed Stone by Chemical Activation, International
Journal of Advanced Engineering, Vol.1 (1), 16-22.
２８９
AEICP Vol. 5, No. 1
A Machine Learning Model for Unintentional Anomaly Detection and Control
on TinyML Lite Platform
Eon-Uck Kang1, and Woo-Sang Hwang2
1Chief Technology Officer; Resco system Laboratory Changwon, Gyeongnam, 51391 Korea
2Professor, Division of Judicial Affairs Business, Korea Maritime & Ocean University, Busan, 49112 Korea
eu-kang@resco.kr1, hws225@nate.com2
Abstract
In the ultra-small Tiny ML platform, the AI learning model can be applied to various models such as waveform
prediction model, voice recognition model, gesture recognition model, and notification service by
microcontroller-based Tensorflow LITE. It is designed in a lightweight learning algorithm with a portable
structure so that machine learning can be performed lightly at 160MHz of ARM Coretax M4 specification in
limited RAM memory and flash memory based on a microcontroller. In this paper, machine learning based on
deep learning collects the dangers of the surrounding operating environment with sensors and extracts
unintended and abnormal characteristics of sensing information to learn the model to determine the
unintended specific location and location of equipment during operation. Infer, reason, and control the
progression of an abnormal operating environment. Through this, inference prediction technology of real-time
progress path for sensing information is possible, and if it is put to practical use in the future, it is possible to
secure and guarantee stability from additional risks required for autonomous system driving such as
autonomous driving based on inference engine and autonomous navigation. will be able.
Keywords: Artificial Intellegenc, TinyML, ARM Cortex-M4 MicroController, Machine Learning Model,
Autonomous, Abnomal Position, Operating Environment.
1. Introduction
Due to the remarkable speed improvement of microprocessors and the advent of big data in the ICT industry
over the past 10 years, it is possible to design artificial intelligence systems that are capable of machine learning.
Machine learning is now possible on distributed and independent devices without accessing vast, huge cloud-
based server data and processing. This is also possible through local system-based embedded devices or
industrial PCs. In addition, by interworking with the cloud system, more flexible network safety can be ensured
against security risks from the network. The study of machine learning was conducted on an independent high-
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２９０
8th ICAEIC-2022
speed CPU as a small artificial intelligence machine learning on a micro board in the Tiny ML platform. By
learning the route from sea and land, recognizing and predicting the route risk, and guiding the risk through
real-time prediction so that safe access to the destination route is possible, it is implemented and analyzed so
that it is possible to prevent from the risk of the unintended route. possibility was presented. In addition, the
autonomous navigation system plans and controls the optimal route and progress trajectory by grasping
information about the autonomous navigation system's own location, surrounding environment, and obstacles
on the ground, through the sensing information of the location, direction, and speed of the moving object. As
a safe system, it is expected to be used for future navigation. For this purpose, ESP32 32bit microcontroller
was mounted and implemented by porting to gcc Compiler.
2.
TinyML(Tiny Machine Leanrning) Framework Growing on Small Embedded Devices
In the embedded market with limited memory capacity, small independent TinyML (Tiny Machine Learning)
is a fast-growing field, and according to a 2020 report by IC Insights, there are more than 250 billion
microcontrollers in the world. At the same time, new application products are being produced. It is
implemented through a neural network with small power consumption and small memory of a portable battery
such as a microcontroller, DSP-based processor, and FPU. TinyML is an inference engine with a small neural
network. It can be applied to embedded systems and implemented as industrial applications that can be
predicted through analysis in real time by interworking with sensors in devices in edge computing
environments. Among them, TensorFlow is the name of TinyML for small machine learning.
The major issues facing the frameworks
• Inability to easily and portably deploy models across multiple
embedded hardware architectures
• Lack of optimizations that take advantage of the underlying
hardware without requiring framework developers
to make
platform-specific efforts
• Lack of productivity tools that connect training pipelines to
deployment platforms and tools
• Incomplete infrastructure for compression, quantization, model
invocation, and execution
• Minimal support features for performance profiling, debugging,
orchestration, and so on
• No benchmarks that allow vendors to quantify their chip’s
performance in a fair and reproducible manner
• Lack of testing in real-world applications.
Figure 1. above is a TensorFlow Lite Micro (TF), which plays the role of an interpreter that supports
upper application applications and lower client APIs, and performs operation processing at the bottom lower
stage to perform inference.
２９１
AEICP Vol. 5, No. 1
2.1. Single-model and Multiple models On TensorFlow Lite Micro(TFLM)
TFLM is thread-safe as long as there is no state corresponding to the model that is kept outside the interpreter
and the model’s memory allocation within the arena. The interpreter’s only variables are kept in the arena, and
each interpreter instance is uniquely bound to a specific model. Therefore, TFLM can safely support multiple
interpreter instances running from different tasks or threads. TFLM can also run safely on multiple MCU cores.
Since the only variables used by the interpreter are kept in the arena, this works well in practice. The executable
code is shared, but the arenas ensure there are no threading issues.
Figure 2. shows the memory allocation of TensorFlow Lite Micro Interpreter by dividing it into a single
model and a multi- model. Executable code is shared in the model repository and execution is performed.
Although there are two methodologies, a single model was applied to the study to detect and predict the
abnormal environment of the present environment.
3. Result
Since TFML can be ported to Arm Cortex-M and Xtensa-based processors by performing the tests below,
this study is conducted by applying the framework of TFML and quantifying the memory overhead in the
Arduino-based c/c++ compiler. It can predict the values of non-ideal sensors and external environments. It was
also possible to perform in the basic framework. Table.1 shows the operation cycle and interpreter overhead
in the TFLM platform.
Table 1. Performance results for TFML target platforms.
Model
Total Cycles
Calculation Cycles
Interpreter Overhead
VWW Reference
387,341.8K
387,330.6K
< 0.1%
VWW Optimized
49,952.3K
49,952.3K
< 0.1%
Google Hotword
Reference 990.4K 987.4K 0.3%
Google Hotword
Optimized 88.4K 84.6K 4.3%
２９２
8th ICAEIC-2022
4. Conclusion
Machine Deep learning can be applied to embedded systems through the TFLM platform that can perform
machine learning in several kilobytes of memory and cycles. It can be applied to the recognition of
unrecognized danger, environment awareness, and path recognition.
References
[1] Yao Qin, Dongjin Song, Haifeng Cheng, Wei Cheng, Guofei Jiang, and Garrison W Cottrell. A dual-stage
attention-based recurrent neural network for time series prediction. In International Joint Conference on
Artificial Intelligence (IJCAI), pages 2627–2633. AAAI Press, 2017
[2] Souhaib Ben Taieb and Amir F Atiya. A bias and variance analysis for multistep-ahead time series
forecasting. IEEE Transactions on Neural Networks and Learning Systems, 27(1):62–76, 2016.
[3] Nguyen Hoang An and Duong Tuan Anh. Comparison of strategies for multi-step-ahead prediction of
time series using neural network. In International Conference on Advanced Computing and Applications
(ACOMP), pages 142–149. IEEE, 2015.
[4] Arun Venkatraman, Martial Hebert, and J Andrew Bagnell. Improving multi-step prediction of learned
time series models. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015.
[5] Vincent Le Guen and Nicolas Thome, Shape and Time Distortion Loss for Training Deep Time Series
Forecasting Models 33rd Conference on Neural Information Processing Systems (NeurIPS 2019),
Vancouver, Canada.
[6] Agarwal Hitesh 2 and King Ho Holden Li, Application of Machine Learning Algorithm on MEMS-Based
Sensors for Determination of Helmet Wearing for Workplace Safety, Micromachines ,2021
[7] Voratas Kachitvichyanukul, Comparison of Three Evolutionary Algorithms: GA, PSO, and DE” Vol 11,
No 3, September 2012, pp.215-223 http://dx.doi.org/10.7232/iems.2012.11.3.215
[8] Stanislava Soro. (2020) TinyML for Ubiquitous Edge AI Approved for Public Release; Distribution
Unlimited. Public Release Case Number 20-2709
[9] A. A. Salem Prof. Dr. M. Mustafa Dr. M. E. Ammar (2014) Tuning PID Controllers Using Artificial
Intelligence Techniques 9th International Conference on Electrical Engineering ICEENG 2014
[10] Robert David, Jared Duke, Advait Jain, Vijay Janapa Reddi, TENSORFLOW LITE MICRO:
EMBEDDED MACHINE LEARNING ON TINYML SYSTEMS, 2021
[11] Shlomi Regev 1 Rocky Rhodes 1 Tiezhen Wang 1 Pete Warden 1Anna Miller and Szymon Walczak (2020)
Maritime Autonomous Surface Ship’s Path Approximation Using Bézier Curves, Symmetry 2020, 12,
1704
[12] Kikun Park, Sunghyun Sim, Hyerim Bae (2021) Vessel estimated time of arrival prediction system based
on a path-finding algorithm, Maritime Transport Research 2 (2021) 100012
[13] Ziegler J. G. and Nichols N. B. (1942), Optimum Settings for Automatic Controllers, Transaction of the
ASME, pp. 759-768.
[14] O’Dwyer A., (2006), Handbook of PI and PID Controller Tuning Rules (2nd Edition), Imperial College
Press, ISBN 1-86094-622-4.
２９３
AEICP Vol. 5, No. 1
[15] P. V. Gopi Krishna Rao, M. V. Subramanyam, and K. Satyaprasad, Model based Tuning of PID
Controller”, Journal of Control & Instrumentation, Vol. 4, Issue 1.
[16] A. Visioli,”Tuning of PID controllers with fuzzy logic”, IEE Proceedings on Control Theory and
Applications, vol. 148, issue 1, pp. 1-8, 2001.
[17] Khedr, S.F.M.; Ammar, M.E. and Hassan, M.A.M. “Multi objective genetic algorithm controller’s Tuning
for non-linear automatic voltage regulator” Proceedings of the 2013 International Conference on Control,
Decision and Information Technologies (CoDIT), pp. 857 – 863, May 2013.
[18] S. F. Kheder, “Genetic Algorithm Based Controller’s Tuning for Linear and Nonlinear Automatic Voltage
Regulator in Electrical Power System”, M.Sc. thesis, Faculty of Engineering, Cairo University 2013.
[19] Jarboui, B., Damak, N., Sirry, P., and Rebai, A. (2008), A combinatorial particle swarm optimization for
solving multi-mode resource-constrained project scheduling problems, Applied Mathematics and
Computation, 195(1), 299-308.
[20] Kachitvichyanukul, V. and Sitthitham, S. (2011), A twostage genetic algorithm for multi-objective job
shop scheduling problems, Journal of Intelligent Manufacturing, 22(3), 355-365.
[21] Lova, A., Tormos, P., Cervantes, M., and Barber, F. (2009), An efficient hybrid genetic algorithm for
scheduling projects with resource constraints and multiple execution modes, International Journal of
Production Economics, 117(2), 302-316.
[22] Marinakis, Y. and Marinaki, M. (2010), A hybrid genetic-particle swarm optimization algorithm for the
vehicle routing problem, Expert Systems with Applications, 37(2), 1446-1455.
２９４
8th ICAEIC-2022
A Study on the Influence of Psychological Factors of Parents on Children's
Behavior
Jung-Eun Lee1, and Chun-Ok Jang2
1Essel Tree Art Psychology Center, Co. Korea
2Department of Social Welfare, Honam University, Korea
meek0713@naver.com1, spring@honam.ac.kr2
Abstract
Recently, interest in children's various psychological and behavioral problems is also increasing, parents
cannot afford to take care of their children's emotions due to an increase in overall stress due to problems
such as double-income parents, and parents have a difficult social life. The time to communicate with children
is decreasing due to the problem of lack of mental space due to stress. In this study, among various variables
of parents on their children's psychology, improve the effects on children with a focus on psychological factors
and policy and practical suggestions in the field of social welfare.
Keywords: Parents' psychological emotions, Children's externalizing behaviors, Children's intrinsic
behaviors, Parental roles.
1. Introduction
It can be said the importance of the role of parents in the family does not change even in the present age
when the perception of the family is changing a lot. In order to discuss the importance of the role of parents, it
can be seen that "the psychological state of the parents is important first. Recently, interest in children's various
psychological and behavioral problems is also increasing, and parents can't afford to take care of their
children's emotions due to socially increasing overall stress and problems such as double-income parents,
parents also have a difficult society Due to the stress of life, there is not enough room for mind, so the time to
communicate with children has decreased. In addition, it can be seen that various problems are occurring in
relationships with children. In order to solve this problem, opinions on various policy methods have been
proposed, and as a result, it has been shown that the number of children receiving psychological treatment is
increasing [1]. The increase in children's psychotherapy can be a problem due to the various stresses arising
from parenting and the child's temperament.
According to the pew research Center, an american public opinion polling agency, conducted a survey on
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２９５
AEICP Vol. 5, No. 1
the 'value of life' of 17 developed economies in 17 countries ranked as 'material affluence', 'health' as the 2nd
priority, and 'family' as the 3rd priority.
According to previous studies, children's problem behaviors were low when parents' lifestyle and daily
satisfaction were high, and children's problem behaviors were high when parents' parenting and educational
stress were high. It has a negative effect on the reaction, which can make it difficult for the child's ability to
adapt to social and emotional development [3, 4, 5]. And a study on father's parenting behavior and
psychological and social factors to solve the behavioral problems of early school-age children [6] through a
study on the mediating effect of child-accompanying self-worth and difficulty in emotional regulation due to
the mother's implicit narcissism [2], it can be seen through previous studies that the psychological factors of
parents have an effect on the children's behavior. Therefore, in this study, among various variables of parents
that affect their children's psychology, the effect on children is investigated, focusing on psychological factors,
and policy and practical suggestions are made in the field of social welfare.
2. Related Works
2.1. Research subject
The analysis was conducted using data from the 8th year among the data from the Panel Study on Korean
Children [PSKC] of the Parenting Policy Research Institute [8].
In this study, SPSS 26.0 and AMO19.0 were used for analysis. First, a frequency analysis was performed to
calculate the general characteristics of the study subject and the reliability coefficient of the test tool.
Second, to calculate the mean and standard deviation between the variables, and to examine the normality,
the skewness and kurtosis values were confirmed by performing descriptive statistical analysis, and the
correlation was analyzed. Third, the validity of each factor was reviewed through confirmatory factor analysis,
and the absolute fit indices x2 and RMSEA and the incremental fit indices NFI, TLI, and CFI were examined
to verify the fit of the model.
2.2. Analysis method
Study Hypothesis 1: Parental psychological factors will affect children's externalizing behavior.
Study Hypothesis 2: Parental psychological factors will influence children's internalization behavior.
The basic architecture of a circulating neural network consists of three layers: the output layer, the input
layer, and the hidden layer. The data are weighted u through the input layer x, and u input the hidden layer s
for the new weight. There are two new weight outputs, and the weight v input is output on the output layer o
tostore the weight w and use it for the next periodic calculation.RNN is more suitable for text regression
problems because it has the characteristic of sending the results from the node of the hidden layer to the input
of the next calculation of the hidden layer node, sending the results from the node of the hidden layer in the
direction of the output layer.
２９６
8th ICAEIC-2022
Figure 1. Research model.
3. Results and Analysis
The data used for the descriptive statistics and correlation analysis results of this study are data from the 8th
year, and a total of 1,598 participants were involved. 1,577 people who responded to the behavior were selected
for data analysis.
The correlations with the mean, standard deviation, kurtosis, and skewness of the measured variables are
shown in table 1. Parental psychological factors were found to have a significant positive correlation with the
child's externalizing behavior, and showed a significant positive correlation with the child's intrinsic behavior.
Children's externalizing behaviors showed a significant positive correlation with children's intrinsic behaviors.
Table 1.
Correlation analysis.
Variable
1
2
3
psychological factors of parents
-
Your child's externalizing behavior
.213***
-
Your child's intrinsic behavior
.248***
.6678***
-
M
1.870
4.176
3.721
SD
.4596
.4.569
4.240
Skewness
.665
3.252
3.330
Kurtosis
.711
29.29.
27.814
N = 124. *p<.05, **p<.01, ***p<.001.
２９７
AEICP Vol. 5, No. 1
In this study, the effect of parental psychological factors on children's externalizing and intrinsic behaviors
was empirically analyzed. Through Confirmatory Factor Analysis (CFA), the validity of the measurement
variables that explain the psychological factors of parents, the child's externalizing behavior, and the child's
intrinsic behavior is secured, and structural model analysis is conducted based on this. The fitness of the
research model was verified using the χ2 statistic, the absolute fit index, RMSEA, and the incremental fit index,
NFI, TLI, and CFI.
4. Conclusion
This study used the 8th data of the Korean Children's Panel to confirm and confirm "the effect of
psychological factors on children's externalizing and intrinsic behaviors" through structural equation model
analysis.
Since the results of this study cannot be generalized, follow-up studies are suggested as a limitation. First,
this study investigated the effects of parents' psychological factors on their children's externalizing and
internalizing behaviors, focusing on depression. Therefore, it has limitations in not examining various variables.
Subsequent studies should look at the factors affecting various parental psychological variables by
supplementing these limitations. Second, follow-up studies are needed to more clearly reveal the causal
relationship between children's behavioral factors under the influence of parental psychological factors through
a longitudinal approach.
Despite these limitations, this study is expected to contribute to enhancing the effectiveness of various policy
and practical measures for parental psychological health as the psychological factors of parents have an
important effect on the children's externalization and intrinsic behavior.
References
[1] https://www.ibabynews.com/news/articleView.html?idxno=46633.
[2] https://blog.naver.com/lavid/222580310470.
[3] Bang, So Young, Choi, Sun Hee, Lee, Soo Hyun, & Hwang, Hye Jung. (2013). Impact of Psychological
and Behavioral Variables of Parents from Low Income Families upon Children’s Problematic Behaviors :
Comparison between Two-parent Families and Single-parent Families. Korean Journal of Childcare and
Education, 9(5), 157–179.
[4] Kim，Yeong-hee & Moon, Jeong-sook. (2007). Effects of mother's depression and sleep quality, and
marital conflict on children's sleep problem. The Journal of Play Therapy, 11(1), 1-19..
[5] Kim，Yeong-hee & Kim, D. G (2017). A Meta-Regression Analysis of Parental Factors on Adjustment
and Maladjustment of Preschool Children, Journal of Early Childhood Education & Educare Welfar, 21(3).
pp. 255-2.
[6] Lee, Woo-Kyung. The Effects of Mother's Covert Narcissism on Psychological Control of a child in Late
Childhood: the Mediating Effect of Child-based Self‐-Worth and the Difficulties in Emotion Regulation.
Korea Master's Thesis, Catholic University of Korea, Graduate School of Counseling Psychology, 2021.
２９８
8th ICAEIC-2022
Implementation of Support Vector Machine (SVM) Method to Determine the
Response of the Indonesian People to the Administration of the Sinovac Vaccine
Siti Ummi Masruroh1, Obey Al Farobi 2, Khodijah Hulliyah3, Husni Teja Sukmana4, Yusuf Durachman5, and
Nanda Alivia Rizqy Vitalaya6
1,2,3,4,5,6Department of Informatics, UIN Syarif Hidayatullah Jakarta, Indonesia
ummi.masruroh@uinjkt.ac.id1, obey.al16@mhs.uinjkt.ac.id 2, khodijah.hulliyah@uinjkt.ac.id3,
husniteja@uinjkt.ac.id 4, yusuf_durachman@uinjkt.ac.id5, nanda.vitalaya17@mhs.uinjkt.ac.id 6
Abstract
This study aims to implement the Support Vector Machine and Word Embedding in the case of sentiment
analysis about public responses to the administration of the Sinovac vaccine uploaded on Twitter and 3 classes:
positive, negative, and neutral. The method chosen is the Support Vector Machine classification method.
Before doing the classification, the preprocessing in this study includes tokenization, normalization, removing
emoticons, Convert Negation, Stemming, Stopword Removal, and Word embedding. The dataset used is 30000
records. The program is designed using the python programming language with several libraries such as hard,
TensorFlow and pandas, NLTK, Skitlearn. The results showed that the accuracy obtained in training using the
Support Vector Machine was 85% and said to be good.
Keywords: Word embedding, Support Vector Machine, Sentiment analysis.
1. Introduction
Sinovac became the first vaccine to be used by the Indonesian government, at a full cost [1]. Information on
the existence of a vaccine has caused some people to get disinformation, in this case, the manipulation of
information carried out by the Anti Sinovac vaccine movement which has not been proven to spread in the
community [2]. Nowadays, users tend to spread information using short 140-character messages called
"tweets" [3]. Twitter is a popular social network on the Internet with hundreds of millions of users, where
people are not an important part of an account [4]. Based on data from we are social and Hootsuite, the number
of active social media users in Indonesia has reached 160 million users. This amount is 59% of the total
population in Indonesia. Furthermore, from this data, the increase that occurred from 2019 for active users was
8.1 percent or 12 million users [5].
An opportunity that can be captured from these facts and data is that there is more information that can be
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
２９９
AEICP Vol. 5, No. 1
extracted from activities that occur on social media. There are two ways to collect tweets: 1) using the API
provided by Twitter, and 2) crawling the webpage of tweets [6]. The crawling process can be done to retrieve
data or content on social media [7]. The data taken is still raw and dirty data that needs to be carried out in the
pre-processing stage and through many stages, so that new and useful information can be produced. This is
often referred to as sentiment analysis [8]. Sentiment analysis is analyzed the natural language meaning of the
free text, the words, and symbols used in a document, and the strength of positive and negative opinions and
feelings [9]. Sentiment analysis is a text analysis technique used to understand the opinions, sentiments, and
subjectivity of text [10]. This study aims to implement the Support Vector Machine and Word Embedding in
the case of sentiment analysis about the public's response to the administration of the Sinovac vaccine uploaded
on Twitter and 3 classes: positive, negative, and neutral.
2. Related Work
Research by [11], shows that the results of SVM accuracy are higher with an accuracy of 96.88% and Naïve
Bayes with an accuracy of 89.63%. Furthermore, research by [12] concluded that the Naïve Bayes Algorithm
and Support Vector Machine have the same accuracy of 0.6758 or a percentage of 67%. Further related
research by [13], the application of SVM and Lexicon Based Features on sentiment classification of cellular
telecommunications services on Twitter social media resulted in the greatest accuracy with the scheme without
Lexicon Based Features with results of 84%, 76% precision, 86% recall and F-Measure 80%. The fourth
reference study by [14], which applied the Support Vector Machine method by retrieving data on Twitter social
media using the Twitter API in the python programming language resulted in an accuracy of 87%, precision
of 86%, recall of 95%, error rate of 13%, and f1- score 90%.
Furthermore, research by [15] which applies the KNN and SVM methods, can build a simple model, several
features such as n-gram features, pattern features, punctuation features, and based on keywords. The results of
the study by [15] showed that the comparison classification algorithm (SCA) performed better than SVM.
Research by Ahmad and Ali, by applying the SVM method, for the first dataset file resulted in an average
precision, recall, and f-measure are 55.8%, 59.9%, and 57.2%. For the second data set, the mean Precision,
Recall, and F-Measure were 70.2%, 71.2%, and 69.9%, respectively. From these studies, the accuracy
produced by the Support Vector Machine (SVM) method is quite good with fairly high accuracy. So that
researchers will use the Support Vector Machine (SVM) method in this study to detect responses to the Sinovac
vaccine. The focus in this study is only on the Sinovac vaccine because the data are taken from Kaggle in
October 2020-January 2021 only appeared the Sinovac vaccine.
3. Methodology
3.1. Research Flow
The literature study process is carried out to find references related to research that has been done before, as
well as references in the form of books and information on related cases. Observations were made by observing
３００
8th ICAEIC-2022
the trend of the object of research on social media. Data collection is done by collecting data on social media,
especially Twitter and Kaggle here which are taken as case studies. In the data processing itself, a data cleaning
process is carried out, so that the data processing is really clean data, not raw data (dirty data). Presentation of
data is done to display the results of data processing. After all stages of the research have been carried out, the
last step is to compile it into a research report.
Figure 1.
Research Flow.
3.2. Proposal Method
Figure 2.
Proposed Method.
The initial data used is commentary data on Twitter social media such as sentiment on vaccine use, sentiment
on use or injection of vaccines as the theme or topic to be explored. Split data aims to divide the dataset into
training data and testing data. The preprocessing used is tokenization, normalization, removing emoticons,
converting negations, stemming, and stopword removal. Classification is done by analyzing word for word or
in full sentences to get results from existing sentiment classes and labeling positive and negative data.
Furthermore, classification is carried out using the Support Vector Machine (SVM) method. This method will
find the accuracy of how much accuracy is obtained from the grouping carried out by the method. The last step
is testing the Twitter sentiment analysis system by testing the accuracy where the value from the classification
test will be sought regarding the F-Measure, Precision, and Recall values. In addition, testing uses cross-
３０１
AEICP Vol. 5, No. 1
validation which is an additional method of data mining techniques that aims to obtain maximum accuracy
results.
4. Result and Discussion
4.1. Data Retrieval
Data retrieval is done by taking data from Twitter social data through a public source, namely the Kaggle
repository. The dataset used consists of 30,000 records with 16 attributes such as Id, Date, Text, Hashtag,
user_name, user location, user description, user created, user followers, user friends, user favorites, retweet,
favorite, source, is retweet and reply to status.
4.2. Initial Processing
4.2.1 Data Transformation
The dataset used has 16 original attributes from the source, but not all of these attributes or features will be
used, the features that will be used are 2 attributes as shown in the table below.
Table 2. Description of Dataset Attributes.
No
Attribute
Description
1
Id
Id tweet
2
Date
Fill in the narrative response about the Sinovac vaccine
4.2.2. Sampling
The dataset used has 16 original attributes from the source, but not all of these attributes or features will be
used, the features that will be used are 2 attributes as shown in the table below. For testing the model used, the
data will be divided into two parts, including training data and testing data using Split Validation. The amount
of data sharing is 70% for training data and 30% for testing data.
4.3. Software Used
In making simulations for implementing SVM in the case of sentiment analysis, it requires some software
including the Windows 10 Pro Operating System, Jupyter Notebook, Browser (Google Chrome), Python
programming language, as well as libraries such as Numpy, NLTK, Keras, Tensorflow, Sastrawi, Seaborn and
Gensim.
4.4. How Models Work
4.4.1. Dataset
The dataset used in this study is stored in Comma Separated Values (CSV) format and totals 30,000 datasets,
as an example, several datasets are shown in the table below.
３０２
8th ICAEIC-2022
Table 3. Dataset.
Attribute
Description
1348286903527768065
#vaksin untuk #indonesia
4.4.2. Preprocessing Step
The dataset that we take from the repository in the original form is certainly not ready to be used to
distinguish between positive, negative, and neutral tweets, so it is necessary to clean the data.
a. Tokenizing
This process is to separate each word in a sentence with a separator which can be a space, comma, semicolon,
or period. The following are some of the tokenizing data presented in the table below.
Table 4. Tokenizing Result.
Id
Text
Tokenizing Result
1348286903527768065
#vaksin untuk #indonesia J
#vaksin | untuk | Indonesia | J
b. Normalization
Text normalization is a text processing process that aims to change the structure or form of text that was
originally difficult for computers to understand until finally, it is easy to understand and further processed. The
following is an example of a normalized result table.
Table 5. Normalization Result.
Id
Text
Normalization Result
1348286903527768065
#vaksin | untuk | Indonesia | J
#vaksin | untuk | Indonesia | J
c. Remove Emoticons
In the case of sentiment analysis, emoticons cannot negatively, neutrally, or positively affect a response, so
we need to remove emoticons. The following table shows the results of removing emoticons.
Table 6. Results of Removing Emoticons.
Id
Text
Removing Emoticons Result
1348286903527768065
#vaksin | untuk | Indonesia | J
#vaksin | Untuk | Indonesia
d. Convert negation
The step in converting negation is to combine the word negation with the preposition, for example, there is
the sentence "Tidak Sembuh" which will change to " TidakSembuh ".
３０３
AEICP Vol. 5, No. 1
Table 7. Results of Convert Negation.
Id
Text
Convert Negation Result
1348286903527768065
#vaksin | untuk | Indonesia
#vaksin | untuk | Indonesia
e. Stemming
Steps to remove affixes or return a word to its root word. In this study, to do stemming required a library
with the name Sastrawi where this library contains all the basic words and affixes.
Table 8. Results of Stemming.
Id
Text
Stemming Result
1348286903527768065
#vaksin | untuk | Indonesia
#vaksin | untuk | Indonesia
f. Stop Word Removal
Stop word removal is a filtering process, selecting important words from the token results, namely what
words are used to represent documents. In NLP (Natural Language Processing) stop words are words that are
ignored in processing, these words are usually stored in stop lists.
Table 9. Stop Word Removal.
Id
Stemming Result
Stop word Removal Result
1348286903527768065
#vaksin | untuk | Indonesia
#vaksin | Indonesia
g. Word Embedding
Word embedding is a term used for the technique of converting a word into a vector or array of numbers.
Table 10. Word Embedding Result.
Id
Text
Word Weight in the Dictionary
1348286903527768065
#vaksin | Indonesia
0.22 | 0.031
To simplify the classification, the text must be changed to a one-hot matrix form, this stage is done because
each word has one different dimension.
Table 11. Word Embedding Result
Id
Text
One hot encoding
1348286903527768065
#vaksin | Indonesia
10 | 01
３０４
8th ICAEIC-2022
4.3. Classification with SVM
At this stage, the training document is used as an input document. The text of each sentence about the
Sinovac covid vaccine has previously been transformed into a vector representation of words using word
embedding then becomes an input in the first layer with a maximum filter length of 1000, so the input will be
a matrix measuring 1000 x 300. The input description in the process SVM is as follows.
Table 12. SVM Input.
Word
One hot encoding
Input Layers
Information
#vaksin | Indonesia
10 | 01
2 kernel and 4 filter
The training data and test data used are data that already has a class label, with a comparison of training data
and test data being 80%: 20%.
Table 13. The proportion of Sentiment Class Result of Labeling with Sentiment Scoring on Training Data
and Test Data.
Classification
Positive
Negative
Amount
Training Data
1058
293
1351
Test Data
114
35
149
Amount
1172
328
1500
In this research, the Support Vector Machine (SVM) method is used with the kernel functions used are the
linear kernel and the RBF kernel. In the linear kernel, there is one parameter that is tested, namely the value of
Cost with the parameter value of Cost (C): 0.01; 0.1; 1; 10; 100; 1000 for training data.
Table 14. Overall Accuracy and Kappa Accuracy Values on the Linear Model of Labeling Results with
Sentiment Scoring.
Model Evaluation
Cost (C)
0,01
0,1
1
10
100
1000
Overall Accuracy
0,7919
0,7718
0,7584
0,7651
0,7651
0,7651
Kappa accuracy
0,21
0,1673
0,0948
0,0833
0,0571
0,0571
The most optimal C values in Table 14 are 0.01 and 0.1 because the overall accuracy and kappa accuracy
are 79.19% and 11.05%, respectively. Meanwhile, the table shows that the most optimal value of C is 0.01,
because the overall accuracy and kappa accuracies are 79.19% and 21%, respectively. This study uses 10-cross
validation to test machine performance in forming classifications. After performing the analysis using the
linear kernel and the RBF kernel, the best model from each model is obtained as follows:
３０５
AEICP Vol. 5, No. 1
Table 15. The Best Model for Linear Kernel and Radial Kernel from Data Labeling Results with Sentiment
Scoring.
Model Evaluation
Linear Kernels (C=0.1)
RBF kernel (C=1000 and =0.00026)
Overall Accuracy
0,7919
0,7919
Kappa accuracy
0,21
0.21
Based on Table 15, the linear kernel model and the RBF kernel from the results of manual data labeling and
sentiment scoring have the same highest overall accuracy of 79.19%. In manual data labeling, the kappa
accuracy value in the RBF kernel is greater than the kappa accuracy value in the linear kernel, while data
labeling with sentiment scoring produces the same highest kappa accuracy, which is 21%. This shows that the
model with the RBF kernel has a correct match between the sentiment classification results compared to using
a linear kernel.
4.5. Result Interpretation
After the proposed model is created and training is carried out, there are results from the performance of the
model created as follows:
Table 16. The proportion of Sentiment Class Result of Labeling with Sentiment Scoring on Training Data
and Test Data.
No
Number of Epoch
Split
Accuracy
1
10
70:30
72%
2
10
80:20
74%
3
15
70:30
82%
4
15
80:20
83%
5
20
70:30
84%
6
20
80:20
85%
7
25
70:30
85%
8
25
80:20
85%
5. Conclusion
This study applies the Support Vector Machine (SVM) algorithm in classifying sentiments on social media
for giving Sinovac vaccines properly and can help detect sentiment. The accuracy produced by the Support
Vector Machine (SVM) method in classifying sentiment is 85% and is said to be good. SVM managed to auto-
detect sentiment on Twitter about the Sinovac vaccine with the help of word embedding and weighting of each
word. The suggestion for further research is the development by trying to do a comparison or try other
classification methods and further development can be done in the form of a prototype or prototype program.
３０６
8th ICAEIC-2022
References
[1] S. Tani, “Jokowi pledges free COVID vaccinations for all Indonesians.,” Nikkei Asia, 2021.
[2] F. Belinda and M. Puspitasari, “Sinovac Vaccine Polemic Disinformation in Asymmetrical Warfare
Perspective,” vol. 6, no. 2, pp. 824–835, 2021.
[3] A. Kanavos, N. Nodarakis, S. Sioutas, A. Tsakalidis, D. Tsolis, and G. Tzimas, “Large scale
implementations for twitter sentiment classification,” Algorithms, vol. 10, no. 1, pp. 1–21, 2017, doi:
10.3390/a10010033.
[4] Jorgec Rodríguez-Ruiza, J. I. Mata-Sánchezb, R. Monroyc, O. Loyola-Gonzálezd, and A. López-Cuevase,
“A one-class classification approach for bot detection on Twitter,” Comput. Secur., vol. 91, 2020, doi:
https://doi.org/10.1016/j.cose.2020.101715.
[5] Hootsuite & We Are Social, “We Are Social & Hootsuite (2020),” Digital 2020 Global Digital Overview,
2020.
[6] J. You, J. Lee, and H.-Y. Kwon, “A Complete and Fast Scraping Method for Collecting Tweets,” 2021
IEEE Int. Conf. Big Data Smart Comput., 2021, doi: doi:10.1109.
[7] M.-J. Lee, T.-R. Lee, S.-J. Lee, J. Jin-Soo, and J. E. Kim, “Machine Learning-Based Data Mining Method
for Sentiment Analysis of the Sewol Ferry Disaster’s Effect on Social Stress,” Front. Psychiatry, vol. 11,
2020.
[8] Y. R. A. D. L. Alfantoukh, “Using Twitter trust network for stock market analysis,” Knowledge-Based
Syst., vol. 145, no. 1, pp. 207–218, 2018, doi: https://doi.org/10.1016/j.knosys.2018.01.016.
[9] V. J. Jayalakshmi, V. Geetha, and R. Vivek, “Classification of autism spectrum disorder data using
machine learning techniques,” Int. J. Eng. Adv. Technol., vol. 8, no. 6 Special issue, pp. 565–569, 2019,
doi: 10.35940/ijeat.F1114.0886S19.
[10] S. A. Yan and P. Mawhorter, “Twitter Sentiment Analysis : Fan Engagement In Esports Matches,” Proc.
13th IADIS Int. Conf. ICT, Soc. Hum. Beings 2020, ICT 2020, no. Berntsen 2009, pp. 257–261, 2020.
[11] O. Arifin and T. B. Sasongko, “Analisa Perbandingan Tingkat Performansi Metode Support Vector
Machine dan Naïve Bayes Classifier Untuk Klasifikasi Jalur Minat SMA,” Semnasteknomedia Online,
vol. 6, no. 1, 2018.
[12] J. N. Anisa Eka Puridewi, “Perbandingan metode naive bayes, Support Vector Machine dan ID3 dalam
penetapan status penanganan kecelakaan kerjaPerbandingan metode naive bayes, Support Vector Machine
dan id3 dalam penetapan status penanganan kecelakaan kerja,” Semin. Nas. Mat. Dan Pendidik. Mat., vol.
4, no. 1, 2018, [Online]. Available: http://eproceedings.umpwr.ac.id/index.php/sendika/article/view/286.
[13] U. Rofiqoh, R. S. Perdana, and M. A. Fauzi, “Analisis Sentimen Tingkat Kepuasan Pengguna Penyedia
Layanan Telekomunikasi Seluler Indonesia Pada Twitter Dengan Metode Support Vector Machine dan
Lexion Based Feature,” J. Pengemb. Teknol. Inf. dan Ilmu Komput., vol. 1, no. 12, pp. 1725–1732, 2017.
[14] I. D. S. Rian Tineges, Agung Triayudi, “Analisis Sentimen Terhadap Layanan Indihome Berdasarkan
Twitter Dengan Metode Klasifikasi Support Vector Machine (SVM),” J. Media Inform., vol. 4, no. 3,
2020, doi: http://dx.doi.org/10.30865/mib.v4i3.2181.
[15] M. R. Huq, A. Ali, and A. Rahman, “Sentiment Analysis on Twitter Data using KNN and SVM,” Int. J.
Adv. Comput. Sci. Appl., vol. 8, no. 6, pp. 19–25, 2017, [Online]. Available: https://pdfs.semanticscholar.
org/05a8/78000170abcd0c6f8208080470858422e17c.pdf.
３０７
AEICP Vol. 5, No. 1
A Study on the Combination of Extractive Summary and AI-Based Abstractive
Summary
Woo Won Choi1, Jeong Hyeon Kim2, Min Kyu Park3, Se Jong Oh4, and Ill Chul Doo*
1,2Global Business and Technology, Hankuk University of Foreign Studies, Yongin, Korea
3,4,*Artificial Intelligence Education, Hankuk University of Foreign Studies, Yongin, Korea
dndnjs996@naver.com1, ellie10107@naver.com2, mkpark@hufs.ac.kr3, tbells@hufs.ac.kr4,
dic@hufs.ac.kr*
Abstract
Currently, it is said to be an AI-based big data era. Even at this moment, so much information and text data
are being produced that it is impossible to estimate the scale. As the availability of documents increases, text
auto-summarization fields have been actively studied in the NLP and AI areas, and Extractive and Abstractive
summarizations are representative ways of summaries. Therefore, this paper proposes a new 'Mutual
Summarization' model that combines Extractive and Abstractive to summarize text more quickly and
accurately. To build a model, an Extractive summarization using sentence embedding-based Text-rank and an
Abstractive summarization based on Transformer and BART were combined. The ‘Mutual Summarization’
model enables high-quality text summaries with sophistication, naturalness and demonstrates better summary
performance than using a single existing model.
Keywords: NLP, Extractive summarization, Abstractive summarization, Text-rank, Transformer, BART.
1. Introduction
The 4th Industrial Revolution was the first step in the big data era. So, we are currently living in a flood of
information. According to IDC statistics, data information is expected to reach 175 ZB by 2025. As such, the
need to efficiently select important information from vast amounts of data has increased, and for this reason,
the field of text summary is becoming more important. Text automatic summary is largely divided into
Extractive summarization and Abstractive summarization [1]. This paper focused on improving summary
performance by constructing a new ‘Mutual Summarization’ model that combines the above two summary
models. In a study on the existing text summary, the main sentences were extracted and summarized by ranking
them with weights for each sentence using the Extractive summarization model [2]. In another study, sentences
were generated and summarized through the Abstractive summary modeling using the LSTM (Long Short –
Advanced Engineering and ICT-Convergence Proceedings (AEICP)
ISSN : 2635
-4586
©
ICTAES 2018
３０８
8th ICAEIC-2022
Term Memory) model. Unlike existing papers, this paper complements each of the advantages and
disadvantages of Extractive and Abstractive summaries and presents a new direction of the summary paradigm.
2. Methodology
2.1. Build an Extractive Summarization Algorithm
Extractive summarization is a model of extracting and summarizing sentences from the original text as the
word implies [3]. Each sentence is converted into a vector of a fixed length through the embedding of the pre-
trained Glove (Global Vectors for Word Representation). This is the first process for calculating the relevance
between words and sentences.
In this study, functions for tokenization and preprocessing are defined using Text-rank Algorithm, and
sentence vectorization for all sentences is performed by removing the stop-words provided by NLTK (Natural
Language ToolKit). By creating a "Similarity Matrix", a graph is drawn from a similarity matrix and scored
based on the result values shown. Among them, select the top n sentences with the highest score and use them
as summaries of the sentences.
Table 1. Generating a weighted graph using a similarity matrix.
×
=
2.2. Build an Abstractive Summarization Algorithm
The Abstractive Summarization is a model of generating and summarizing new sentences from the original
text. Unlike the Extractive summarization, it does not simply extract important phrases from the original text
but presents potentially relevant new phrases. In the above algorithm, based on transformers, the BART
(Bidirectional Transformer + Auto-Regressive Transformer) model learned from CNN/Daily Mail News
Dataset is retrieved through the pipe module and the basic parameters are directly used. The reason for adopting
BART in this study is that it is predicted independently of the missing token, which is the problem of BERT
T1
T2
…
Tn
S1
W11
W12
…
W1n
S2
W21
W11
…
W2n
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Sm
Wm1
Wm2
…
Wmn
S1
S2
…
Sm
T1
W11
W21
…
Wm1
T2
W12
W22
…
Wm2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Tn
W1n
W2n
…
Wmn
S1
S2
…
Sm
S1
1
r12
…
r1m
S2
W12
1
…
r2m
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Sn
rm1
rm2
…
1
３０９
AEICP Vol. 5, No. 1
(Bidirectional Encoder Representations from Transformers). In addition, the BART model solved the inability
to learn two-way interactions, which are problems with the GPT (Generative Pre-Training of a language model)
model [4].
2.3. The Progress of the Mutual Summarization Model
The summary model that combines the Extractive Summarization and the Abstractive Summarization model
will be called the ‘Mutual Summarization’ model. The Mutual Summarization model primarily extracts
sentences with high scores from the similarity matrix through Extractive Summarization. Through the first
extraction, it is summarized as 50% text compared to the original text. This is the primary process of removing
unnecessary sentences and less readable sentences. Extractive summarization not only has poor readability by
extracting the contents of the text as it is and connecting the sentences, but also confuses the reader with sudden
changes in viewpoint. Therefore, based on the first summary through the Extractive Summarization, the second
summary aims to change the paradigm for "text summary" that is readable and can be quickly read in short
text, consisting only of the core of key information with the BART model [5]. The accuracy and performance
of the Mutual Summarization model were evaluated using the ROUGE (Recall-Oriented Understudy for
Gisting Evaluation) Score, and for performance comparison, we compared the original text of CNN News, the
sentence summarized by CNN, and the summary of CNN News using the Mutual Summarization model.
Figure1. Process of Mutual Summarization.
3. Results and Analysis
Mutual Summarization is a combination model of Extractive Summarization, which extracts important key
sentences from the original text, and Abstractive Summarization, which creates sentences that can reflect the
contents of the original text. In this paper, experiments on the model were conducted using articles provided
by CNN. First, the original CNN article is summarized through an Extractive Summarization model. The
Abstractive summarization is performed secondarily using the BART model with a summary that has passed
３１０
8th ICAEIC-2022
through the Extractive summarization model. Table 2 shows an example of how the original text to be
summarized is made into a Mutual summarization.
Table 2. Example of Mutual Summarization.
ROUGE evaluation is used as an indicator for evaluating the performance of natural language generation
models such as text automatic summary. The ROUGE score was also used in this paper to evaluate the
performance of the model. Table 3 is the result of the ROUGE evaluation of the 'answer summary' and the
Mutual summary. To objectify the performance of the Mutual Summarization model, the performance of the
existing BART model was compared. When the ROUGE evaluation was conducted with the same data as
BART's ROUGE evaluation, the Mutual Summarization Model showed better performance in all indicators
of ROUGE 1, 2, and L than Lead-3, BERTSUMABS, and BART.
Table3. ROUGE Evaluation Result.
Original
Article
The bishop of the Fargo Catholic Diocese in North Dakota has exposed potentially hundreds of church members
in Fargo, Grand Forks and Jamestown to the hepatitis A virus in late September and early October. The state
Health Department has issued an advisory of exposure for anyone who attend
ed five churches and took
communion. Bishop John Folda (pictured) of the Fargo Catholic Diocese in North Dakota has exposed
potentially hundreds of church members in Fargo, Grand Forks and Jamestown to the hepatitis A. State
Immunization Program Manager Molly Howell says the risk is low, but officials feel it's important to alert people
to the possible exposure. The diocese announced on Monday that Bishop John Folda is taking time off after
being diagnosed with hepatitis A. The diocese says he contracted the infection through contaminated food while
attending a conference for newly ordained bishops in Italy last month. Symptoms of hepatitis A include fever,
tiredness, loss of appetite, nausea, and abdominal discomfort. Fargo Catholic Diocese in North Dakota (pictured)
is where the bishop is located.
Extractive
Summarization
The state Health Department has issued an advisory of exposure for anyone who attended five churches and took
communion. The diocese announced on Monday that Bishop John Folda is taking time off after being diagnosed
with hepatitis A. The diocese says he contracted the infection through contaminated food while attending a
conference for newly ordained bishops in Italy last month. Symptoms of hepatitis A include fever, tiredness, loss
of appetite, nausea and abdominal discomfort.
Mutual
Summarization
Bishop John Folda is taking time off after being diagnosed with hepatitis A. He contracted the infection while
attending a conference for newly ordained bishops in Italy last month. Symptoms of hepatitis A include fever,
tiredness, loss of appetite, nausea, and abdominal discomfort.
Evaluation Index Lead-3
BERTSUMABS
(Liu & Lapata, 2019) BART
Mutual
Summarization
ROUGE
1
40.42
41.72
44.16
46.57
2
17.62
19.39
21.28
29.72
L
36.67
38.76
40.90
46.57
３１１
AEICP Vol. 5, No. 1
4. Conclusion
As vast amounts of text data are generated every day and the importance of data utilization increases, the
importance of research on text processing technology is increasing. The text automatic summary system can
greatly reduce the time required to search and process information, thereby improving efficiency [6]. In recent
years, due to the increase in demand for brief information, the summary industry has developed, and summary
contents have become the main form of content. It is expected that this proposed model will occupy a large
part of the industry and grow steadily while helping individuals consume information quickly.
This paper proposed the 'Mutual Summarization' model to improve the performance of the system by
supplementing the shortcomings of each model of Extractive Summarization and Abstractive Summarization
in text automatic summary. In addition, a document summary experiment was conducted on CNN document
summary datasets to evaluate the performance of the proposed model, and the experimental results confirmed
that the proposed methodology performed better than other models based on ROUGE evaluation indicators.
The Mutual Summarization Model solves the problem of connectivity between summary sentences and
provides more natural summary results. In future research, experiments using more diverse sizes and types of
datasets are expected to be conducted, and if model performance is improved accordingly, it will be a
groundbreaking model that produces more natural sentences.
Acknowledgement
Funding: This research was supported by Hankuk University of Foreign Studies Research Fund of 2021.
Also, This research was supported by the MIST (Ministry of Science, ICT), Korea, under the National
Program for Excellence in SW), supervised by the IITP (Institute of Information &
communications Technology Planing & Evaluation) in 2021"(2019-0-01816), This work was supported by the
Ministry of Education of the Republic of Korea and the National Research Foundation of Korea (NRF-
2021S1A5A8065934)
References
[1] Liu, W., Gao, Y., Li, J., & Yang, Y. (2021). A Combined Extractive with Abstractive Model for
Summarization. in IEEE Access, vol. 9, 43970-43980.
[2] Madhuri, J.N., & Ganesh Kumar, R. (2019). Extractive Text Summarization Using Sentence Ranking.
2019 International Conference on Data Science and Communication (IconDSC), 1-3.
[3] Rani, U., & Bidhan, K. (2021). Comparative assessment of extractive summarization: textrank tf-idf and
lda. Journal of Scientific Research, vol. 65, no. 1, 304-311.
[4] Wang, Z., Duan, Z., Zhang, H., Wang, C., Tian, L., Chen, B., & Zhou, M. (2020). Friendly topic assistant
for transformer based abstractive summarization. In Proceedings of the 2020 Conference on Empirical
Methods in Natural Language Processing (EMNLP), 485-497.
３１２
8th ICAEIC-2022
[5] Lee, N., Park, S., Shin N., & Koo, M. (2021). A Summarizing Application of Technical Papers based on
BART model, Proceedings of the Korean Information Science Society Conference, vol. 48, no. 1, 1756-
1758.
[6] Kim, H., & Lee, S. (2014). Korean Text Automatic Summarization using Semantically Expanded
Sentence Similarity. Proceedings of the Korea Information Processing Society Conference, vol. 21, no. 2,
841-844.
