Content uploaded by Rodolfo Elias Haber Guerra
Author content
All content in this area was uploaded by Rodolfo Elias Haber Guerra on Jun 10, 2021
Content may be subject to copyright.
Annual Reviews in Control xxx (xxxx) xxx
Please cite this article as: Alberto Villalonga, Annual Reviews in Control, https://doi.org/10.1016/j.arcontrol.2021.04.008
1367-5788/© 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
A decision-making framework for dynamic scheduling of cyber-physical
production systems based on digital twins
Alberto Villalonga
a
,
*
, Elisa Negri
b
, Giacomo Biscardo
b
, Fernando Castano
a
, Rodolfo E. Haber
a
,
Luca Fumagalli
b
, Marco Macchi
b
a
Centre for Automation and Robotics, CSIC- Universidad Polit´
ecnica de Madrid (UPM), Arganda del Rey, 28500, Spain
b
Department of Management, Economics and Industrial Engineering, Politecnico di Milano, Milano, Italy
ARTICLE INFO
Keywords:
Digital twin
Decision-making
Cyber-physical systems
Fuzzy logic
Condition-based Monitoring
ABSTRACT
Nowadays, one important challenge in cyber-physical production systems is updating dynamic production
schedules through an automated decision-making performed while the production is running. The condition of
the manufacturing equipment may in fact lead to schedule unfeasibility or inefciency, thus requiring respon-
siveness to preserve productivity and reduce the operational costs. In order to address current limitations of
traditional scheduling methods, this work proposes a new framework that exploits the aggregation of several
digital twins, representing different physical assets and their autonomous decision-making, together with a
global digital twin, in order to perform production scheduling optimization when it is needed. The decision-
making process is supported on a fuzzy inference system using the state or conditions of different assets and
the production rate of the whole system. The condition of the assets is predicted by the condition-based moni-
toring modules in the local digital twins of the workstations, whereas the production rate is evaluated and
assured by the global digital twin of the shop oor. This paper presents a framework for decentralized and in-
tegrated decision-making for re-scheduling of a cyber-physical production system, and the validation and proof-
of-concept of the proposed method in an Industry 4.0 pilot line of assembly process. The experimental results
demonstrate that the proposed framework is capable to detect changes in the manufacturing process and to make
appropriate decisions for re-scheduling the process.
1. Introduction
Currently, a fresh push towards smart manufacturing and cyber-
physical production systems is given to automatically and dynamically
update production by decision-making tools in runtime (Panetto, Iung,
Ivanov, Weichhart & Wang, 2019). This is an open challenge and various
approaches have been proposed in the past for the automated
decision-making in this eld, among which distributed and agent-based
architectures are a rich research stream Chan & Chung (2013). Nowa-
days, the digitization and the Industry 4.0 enabling technologies may
offer new possibilities in this realm (Frazzon, Agostino, Broda & Freitag,
2020).
Reprogramming is in fact necessary, in order to update a production
schedule when a change in the state of the manufacturing system makes
the current schedule unfeasible or inefcient (Ma, Yang, Liu & Wu,
2018). Therefore, rescheduling updates are performed in response to
certain performance indicators, e.g. subsequent to some predictive
maintenance activities leading to the prediction of the remaining useful
life before the failure. Overall, the dynamic scheduling capability aims
at increasing the productivity and reducing the operational costs of the
manufacturing system.
There are many strategies to perform production rescheduling,
especially if the new trend towards Industry 4.0 (I4.0), which brings
Information and communications technology (ICT) to production sys-
tems, is considered. Industrial Internet of Things (IIoT) and Industrial
Cyber-Physical Systems (ICPS) have led to achieve the next level in
smart manufacturing (Iarovyi, Martinez Lastra, Haber & Del Toro,
2015). ICPS have become particularly relevant as the main enablers in
bridging virtual and physical worlds, thanks to their computing and
communication capabilities Wolf (2009) (Villalonga, Beruvides, Castano
& Haber, 2020). An important ICPS-based technology that is fostering
the digital transformation process is the Digital Twin (DT). In order to
* Corresponding author.
E-mail addresses: alberto.villalonga@car.upm-csic.es (A. Villalonga), elisa.negri@polimi.it (E. Negri), giacomo.biscardo@polimi.it (G. Biscardo), fernando.
castano@car.upm-csic.es (F. Castano), rodolfo.haber@car.upm-csic.es (R.E. Haber), luca1.fumagalli@polimi.it (L. Fumagalli), marco.macchi@polimi.it (M. Macchi).
Contents lists available at ScienceDirect
Annual Reviews in Control
journal homepage: www.elsevier.com/locate/arcontrol
https://doi.org/10.1016/j.arcontrol.2021.04.008
Received 21 December 2020; Received in revised form 22 March 2021; Accepted 5 April 2021
Annual Reviews in Control xxx (xxxx) xxx
2
obtain a perfect emulation and mirroring of the operating conditions of
the corresponding real systems, DT uses the best available representa-
tions, physical and virtual models (Rosen, Von Wichert, Lo & Betten-
hausen, 2015)(Guerra, Quiza, Villalonga, Arenas & Castano, 2019), in
order to support an enhanced decision-making. There is no unique
denition of DT. DT can be considered as one or many simulation model
(s) of a real system that is/are always connected with the physical
counterpart (or is/are disconnected only temporarily for specic rea-
sons). This connection enables to gather real-time data from the eld, to
perform simulations (e.g. scenario analyses or other types of analyses)
and to send feedback to the physical world in order to modify the
behavior of the real entity or component. Very often, the DT is connected
to an intelligence layer that contains rules, optimization algorithms, and
decision-making functionalities, in order to make decisions on how to
act on the real system.
Looking at manufacturing operations management, Cyber-Physical
Systems (CPS) are leading to an evolution towards future smart fac-
tories where decentralization and autonomy are two important char-
acteristics (Napoleone, Macchi & Pozzetti, 2020). Some collaborative
CPS (Ivanov & Sokolov, 2012; Nazarenko & Camarinha-Matos, 2017),
specically Cyber-Physical Production Systems (CPPS), behave as
evolving entities with a high degree of autonomy to ensure the adaptive
response to disturbances, while communicating and coordinating their
actions by exchanging information in order to meet common organiza-
tional goals. A key change is driven by the development of advanced and
distributed DT-based frameworks on the basis of engineering methods
focused on condition-based monitoring (CM) strategies (Leit˜
ao,
Colombo & Karnouskos, 2016).
Efcient and reliable decision-making procedures are indeed the
bottleneck at machine/equipment level with two main needs: (i) highly
accurate and (ii) high delity mirroring and efcient communication
between physical and virtual spaces are required. Special relevance is
given to emulate human-in-the-loop socio-cognitive skills (Haber--
Haber, Haber, Schmittdiel & del Toro, 2007). Moreover, DTs are
assuming a leading role for DT-based scheduling frameworks, both at
global level to manage the production process in the factory, and at local
level for generating simulations to address the CM of production
equipment Barricelli, Casiraghi & Fogli (2019).
Nowadays, the availability of data from industrial equipment and the
computing power open up the opportunity of designing and developing
a new framework to carry out production scheduling tasks. This can be
done by exploiting a new type of aggregation of multiple DTs repre-
senting different physical assets. Therefore, the main motivation of the
present work is to better exploit DTs, data gathered from physical assets
and decision-making to improve the scheduling process, thus optimize,
and increase the productivity of manufacturing systems. This leads to
consider new DT-based scheduling methods to reduce scheduling de-
viations, by updating resource parameters from interactive program-
ming strategies(Fang et al., 2019). In addition, a DT model enhances the
ability to digitally simulate how the production line will perform in the
real world, contributing to decision-making when a rescheduling in the
production system is needed on the basis of state-of-art frameworks
(Zhang, Liu, Chen, Zhang & Leng, 2017). DTs for supporting dynamic
and automated decision making also eliminate human errors in gath-
ering data and can respond faster to the changes in a production system,
thanks to the constant update with the eld data (Bevilacqua et al.,
2020; Borangiu et al., 2020). Furthermore, DTs can decentralize the
decision-making activities and new modication are easier to be intro-
duced into the system.
The literature is plenty of new methods to move towards DT-based
scheduling frameworks that exploits CM strategies. On one hand, the
use of architectures seamlessly integrating the production scheduling
and CBM (Condition Based Maintenance) through a DT-based eld
synchronization is proposed by Negri et al. (2020); this leads to a
scheduling optimization method (based on genetic algorithms) and a
eld-synchronized Equipment Health Indicator module, all together
embedded into the DT-based simulation. A step ahead could be
embedding DTs in local controllers, to make the role of predictive
models for CBM easy, as well as to provide an efcient local
decision-making to detect faults and assist the operators. Therefore,
implementing distributed frameworks with embedded DTs in the local
nodes while making decisions based on adaptive thresholds techniques
and local simulations is a promising strategy that could be used in order
to improve the manufacturing operations management and conduct
more efcient scheduling tasks.
On the other hand, the use of distributed architectures in smart
manufacturing contributes to enhance efciency and reliability. In
particular, it is worth remarking that scalability provides robustness
against failures, facilitating reconguration actions without affecting
the production. One strategy is to develop distributed frameworks based
on DTs which, besides the robustness against failures, to enrich the
knowledge about the manufacturing process because of the added value
given by the simulations, close to the local process, and the generation of
useful data. This information allows to take more efcient actions both
globally (at factory level) and locally (at the level of workstations or
single equipment pieces), nally improving the scheduling and the
optimization tasks.
Potentialities of DTs can be better exploited in distributed frame-
works. Nowadays, there are some important gaps and shortcomings. The
lack of well-dened frameworks that combine DTs, the limitation of
methods for aggregating DTs, the limited practical applications and the
poor use of the data gathered from physical counterparts are some
evident weaknesses. Overall, current limitations of DTs can be overcome
by designing a distributed framework where different virtual-physical
embedded nodes at the local and global levels cooperate for achieving
common goals. The work reported in this paper proposes the design and
implementation of a framework based on local and global DTs for smart
decision-making in cyber-physical production system and then validate
the technical viability and the possibility to use it real industrial setups
by a proof-of-concept in an Industry 4.0 pilot line. The proposed
approach is supported on DTs that simulate the performance of each
device/machine of the production system as well as of the whole pro-
duction system. Correspondingly, DTs are adopted at local (e.g., device/
machine) and at global (e.g., production system or plant) levels, com-
bined with local CM for rescheduling actions. This aims to improve ef-
ciency by avoiding the decrease in production performances due to
malfunctioning or components degradation.
In order to illustrate the proposed contribution, the paper is struc-
tured as follows. Section 2 presents a review of the state-of-the-art of
related works about DT-based architectures and shows the main gaps
and shortcomings. Section 3 describes the distributed DT-based frame-
work and the decision-making algorithms. Section 4 presents the case
study and framework proof-of-concept validation in a use case to show
the validity of the proposed framework. The conclusions are provided in
Section 5.
2. Related works
The denition and the concept of DT in smart manufacturing are not
new and not unique in literature. The rst denition of a DT was given
by the NASA, stating that a DT is “an integrated multi-physics, multi-
scale, probabilistic simulation of a vehicle or system that uses the best
available physical models, sensor updates, eet history, etc., to mirror
the life of its ying twin” (Shafto et al., 2012). From this rst descrip-
tion, the concept of DT evolved, and the scientic community provided
several denitions, declined in different elds and applications. How-
ever, in literature the proper characteristics of a DT are common among
the different descriptions given by the authors. Nowadays a DT is
referred as a real-time simulation of a physical asset, capable to
communicate to its physical “twin”, to elaborate data coming from it in a
seamless way, to optimize its performance, to monitor and to predict its
behavior (Fei Tao et al., 2018). Indeed, today the availability of data and
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
3
information coming from industrial equipment is higher than in the past,
and as many studies demonstrate, the intersection between optimization
models and available information, has not been thoroughly explored yet
(Cimino, Negri & Fumagalli, 2019). A DT has to be frequently updated
or tuned, and any change in software or hardware in the physical
counterpart has to be carefully implemented also in the DT (Borangiu
et al., 2020). For this reason, a DT is an evolving virtual entity.
The use of DTs to improve the operational activities in a production
system is worldwide addressed, and many applications have been
already described and reported (Cimino et al., 2019). DTs have been
exploited in real-time reading of machine states (Kritzinger, Karner,
Traar, Henjes & Sihn, 2018), to address some production problems by
interacting with the manufacturing execution system (MES) (Negri,
Berardi, Fumagalli & Macchi, 2020), or by scheduling tasks on the basis
of actual health state of the assets, thus providing a more robust solution
(Negri et al., 2020).
Decision making algorithms for scheduling process in CPPS are
focused on goals such as minimization of the makespan, the work in
progress or the production costs (Hamid, Nasiri, Werner, Sheikhahmadi
& Zhalechian, 2019). One challenge is to determine if the current
decision-making performs well enough in complex systems. An
improvement can be achieved by simulating the effects of
decision-making in scheduling systems (Frantz´
en, Ng & Moore, 2011).
An appropriate system can schedule under consideration of the current
system state and the production goals. Instead of using one single pri-
ority rule focus on one production goal, other rules can be considered
such as the state of the assets or the combination of different production
goals. By introducing simulations, the impact of different rules on the
system the can be compared Güçdemir & Selim (2018). Moreover, the
use of articial intelligence techniques such Fuzzy Logic (Han, Liu, Luo
& Mao, 2020)(Dorfeshan, Tavakkoli-Moghaddam, Mousavi &
Vahedi-Nouri, 2020) and machine learning (Poongothai, Kannan &
Godhandaraman, 2021)(Feng, Li, Cen & Huang, 2003) can improve the
decision-making process in presence of uncertainty.
This work contributes to this challenge by proposing a new frame-
work to carry out production scheduling tasks exploiting a new aggre-
gation in a framework with several DTs representing different physical
assets. The focus of the scheduling process is on the decision-making
framework based on the aggregated DTs and the corresponding data
and information coming from them. For this reason, the literature re-
view is centered on available architectures of DTs, to understand how
information and data can be better exploited to improve schedules and
thus optimize the productivity of the manufacturing system.
This work is the natural evolution and a next step in the research on
dynamic production scheduling with distributed architectures, such as
those based on multi-agent systems (Cavalieri, Garetti, Macchi & Taisch,
2000; Li, Xiao & Yang, 2019). With the advent of Industry 4.0, the use of
CPS in manufacturing, the new levels of autonomy and smartness,
naturally rejuvenated the interest to change the centralized scheduling
methods into decentralized ones. Thus, dynamic decision-making and
scheduling methods have to be re-designed in order to ensure exibility
and adaptation before disturbances in the production systems (Jiang,
Jin, Mingcheng & Li, 2017). This is inuencing on current studies for
production scheduling optimization, putting the focus on the denition
of new methods that more easily describe the complex modern CPS and
change accordingly to their modication or breakdown, leading to
research works on new dynamic scheduling algorithms that elaborate
more robust and reliable solutions Long, Zheng & Gao (2017).
2.1. Literature review on digital twin-based architectures
The review of the state-of-the-art focused the attention on DTs
explicitly considered in distributed architectures. In particular, the main
priority is to study the published works where more than one DTs were
built and merged or connected into a unique architecture or platform for
complex production systems. Works considered relevant for this
literature review are reported in Table 1. They were analysed based on
six criteria, essential to understand the aggregation of DTs into a single
architecture or platform. Columns in the Table 1 are devoted to each
criterion:
•Architectures, in which the described DTs are interacting with each
other;
•Domain refers to whether the DTs are built using a single software
(single-domain scenario) or different DTs are built with different
software tools (multi-domain scenario);
•Machine to model (M2M) communication deals with the IT part of the
architecture, describing the way DTs are connected to their physical
counterparts, the communication protocols and the connected
communication problems;
•Targets considers the application elds for which the proposed
frameworks are thought, namely the main objectives reachable by
the adoption of the studied solutions;
•Practical application takes into account if a purely theoretical
approach is addressed or a practical implementation of the DTs is
present, either in a real or in a lab context;
•Software considers the software tools, when mentioned, used to
develop the DTs; this aspect also helped the authors to understand
which software ts the requirements of the application, according to
its software characteristics;
•Industry 4.0 alignment to identify if the reported work explicitly
mentions the Industry 4.0 paradigm.
This review of a wide spectrum of architecture models reveals that
most of them are hierarchical models (H.M.). Each level of the hierarchy
has its own function and provides data or information to the level above.
Moreover, the upper levels can send commands or information down-
stream the hierarchy to actuate decisions in a specic lower level. The
arrangement in a hierarchy eventually expresses the relative importance
of the decisions/gathered information: a higher level in the architecture
corresponds to a higher priority of the decisions/information. Activity-
Resource-Type-Instance architecture (ARTI), rst proposed by Valck-
enaers (2019) is characterized by a local DT for each equipment piece
and by the fact that the intelligence layer is separate from the DT
simulation modeling (Borangiu et al., 2020); it has been also considered
of interest for the relationship between the DT and the computational
layer where data are processed and decisions are taken. The
co-simulation framework is also quite common due to its ability to
describe complex and heterogeneous systems and its exibility
(Stecken, Lenkenhoff & Kuhlenk¨
otter, 2019). Co-simulation represents a
complex and heterogeneous system implemented in a distributed way.
Different simulation models are built to represent the entire system, but
they can be used in a black-box way. This allows to de-couple the
problems and provides highly exible solutions. The simulation models,
in fact, can be built with different software and run as standalone models
(Stecken et al., 2019). From the literature review, other types of archi-
tecture emerged, however their implementations are at early stages of
development and thus are not relevant yet.
The multi-domain scenario is adopted to describe complex systems
while considering more disciplines in the DT simulation, e.g. thermo-
dynamics, mechanical behavior, operations performance, etc., thus
more requirements are needed to better describe or represent the com-
plex system. On the contrary, when dealing with only one discipline, the
reported works suggested specic architectures that are suitable for a
single-domain scenario.
Machine-to-model communication was thoroughly considered to
understand the current most used communication protocols in smart
manufacturing. The Open Platform Communications Unied Architec-
ture protocol(OPC-UA) (A. Redelinghuys, Basson & Kruger, 2019) and
the cloud-based approach are the most widely adopted. The OPC UA
protocol is well known in the industrial world as one of the standards on
which Industry 4.0 is leveraging. The cloud-based communication
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
4
Table 1
Literature review summarizing table.
REF. Architecture Domain M2M Communication Targets Practical application Software Industry
4.0
H.
M.
ARTI Co
sim
Other Single Multi OPC
UA
Cloud TCP/
IP
Modbus
TCP
Others Industrial
case
Lab
case
(A. Redelinghuys et al., 2019) x x C x
(A. J. H. Redelinghuys, Kruger &
Basson, 2020)
x x C x (cell) x
(Fei Tao et al., 2019) x x S x
(Uhlemann et al., 2017) x X x C x
(Gurjanov, Zakoldaev, Shukalov &
Zharinov, 2019)
x x x
(Qi et al., 2018) x X x x
(Cardin et al., 2020) x x x M, CT x Rock-well Arena, Java x
(Wang & Wang, 2018) x X x T x x
(Beregi et al., 2018) x X x E x Tech. Plant Simulation,
AnyLogic
x
(Alam & Saddik, 2017) x x x(s) Java x
(Bakliwal, Dhada, Palau, Parlikad &
Lad, 2018)
x x x M x (eet) Python x
(Ashtari Talkhestani et al., ) x x x x x x M x Java x
(Jung, Shah & Weyrich, 2018) x x x C(s) Simulink, Modelica x
(Stecken et al., 2019) x x x x C(s) x(s) AutomationML x
(Havard et al., 2019) x x x VR x Catia, Modelica x
(Brandenbourger & Durand, 2018) x x x x x
(Qamsane et al., 2019) x x x x M x(s) Rockwell, AutoMod x
(Catarci et al., 2019) x x x M x(s) x
(F Tao & Zhang, 2017) x x x x x(s) Catia, SolidWorks x
(Fera et al., 2020) x x x VR x Tecnomatix Process
Simulate
x
(Khan, Farnsworth, McWilliam &
Erkoyuncu, 2020)
x x x x
(Morel, Pereira & Nof, 2019) x x CT x
M2M: Machine To Model; C: conceptual contribution, M: monitoring, T: tracking, CT: control, E: error management, VR: virtual reality, S: simulation, (s): simulated application.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
5
approach is not a protocol. However, it has been considered relevant for
the analysis of the state-of-the-art, since many papers deal with DT
developed in cloud ambient, thus often the issue related to their
communication and interoperability is mentioned (Qi, Zhao, Liao & Tao,
2018). TCP/IP (Transmission Control Protocol and Internet Protocol) is
a set of communication protocols widely used today in internet and
similar networks of computers. It has often been mentioned when DTs
were connected via Internet or the data sources were reachable mainly
via network (Beregi, Szaller & K´
ad´
ar, 2018). The last major communi-
cation method described several times among the papers is the Mod-
bus/TCP. This is the most used protocol for communications with the
programmable logic controllers (PLC).
Targets are namely the objectives for the architectures proposed in
the papers. DTs can be used for many different purposes (monitoring the
production, improving maintenance, making decision, etc.). The aim of
this criterion is then to classify the studied papers according to their
main purpose. It is possible to see that the main targets proposed in the
analysed research works on DT-based architectures and platforms are:
control, monitoring, error management and virtual reality. Control
targets, using DTs, mainly deal with the online managing of a production
system, by adopting decisions and changes in real-time. On the other
hand, monitoring regards the tracking and the evaluation of KPIs (Key
Performance Indicators) of a production system, or the state monitoring
of the physical assets (e.g. health states, machine states, etc.). The last
most common target is virtual reality, where the DT-based architectures
is exploited to render the behavior and the state of a whole production
system in real-time or to perform simulations for layout performance
assessment (Havard, Jeanne, Lacomblez & Baudry, 2019).
The study of practical applications of the analysed DT-based archi-
tectures demonstrates a gap of this topic. Few practical implementations
of the analysed DT-based architectures are in fact reported: most of them
elaborate only theoretical solutions and approaches. In addition, simu-
lations only validate most of the practical realizations, and therefore DTs
are not really assessed in real scenarios with a physical counterpart. In
Table 1 this is summarised as follows: lab cases and industrial cases are
ticked if the architecture under analysis is contextualized within an in-
dustrial or academic laboratory or in a real production system, respec-
tively. When the cells in the columns are empty, this means that the
analysed architecture had no practical implementation and the contri-
bution remained at a theoretical level.
The software aspect deals with the specic software tool used or
suggested for developing the architecture proposed in each paper. This
analysis intrinsically helps to understand the characteristics of each
software. Indeed, the architecture’s purpose deeply affects the selection
of the software, since - as emerged from the literature review - some
software tools are apt to address only some specic tasks, and others are
more generic, thus more exible but less focused on specic tasks.
As emerges clearly from the analysed papers in Table 1, Industry 4.0
is explicitly mentioned by all analysed research works, suggesting that
the most recent works on DT-based architectures are developed and
analysed in the context of Industry 4.0 research. This is not surprising
because, as mentioned in the Introduction section, DT can be considered
as hosted in the cyber aspect of CPS, and therefore together with CPS are
among the main concepts related to the Industry 4.0 paradigm.
2.2. State-of-the-art: gaps and shortcomings
The review of the state-of-the-art shows that some key developments
are still required to cover four identied scientic and technical gaps
and to better support operations management of cyber-physical pro-
duction processes. From this analysis emerges that Industry 4.0 para-
digm and in particular the potential of the DTs can be better exploited
when a well-dened framework for multiple DTs connected into a single
architecture is available. This is not evident in architectures analysed so
far, thus leading to a clear shortcoming in the aggregations of several
DTs into a unique architecture to represent and improve smart
manufacturing systems. It is a relevant gap, as ICPS equipped with DTs
have proven to be an alternative approach to conduct optimization and
monitoring tasks in a faster and efcient way Uhlemann, Lehmann &
Steinhilper (2017).
The main gaps of the state-of-the-art are summarised as follows.
•(First gap) A well-dened framework that combines and links many
DTs is not available, and this is the most evident gap. DTs for
modeling different single physical assets at local level, linked into a
single architecture that has also a representation of the whole pro-
duction system at global level, are missing. Besides, there is the need
of a method to describe a system of heterogeneous physical assets
into a single-domain scenario, rather than each single asset into a
multi-domain scenario (which is quite often reported in the literature
review).
•(Second gap) The proposed methods of aggregation of DTs are only
targeting already generated DTs. It results in poverty of studies
related to “green-eld” scenarios, where there is the need to describe
a whole system from scratch.
•(Third gap) There are only few practical applications (either in lab-
oratory or in industrial contexts), and very few models appropriately
and strictly consider smart manufacturing systems. Besides theoret-
ical studies, only single tasks done by DTs are reported, instead of
heterogeneous tasks (Fei Tao, Qi, Wang & Nee, 2019) such as
decision-making, problem resolution, optimization, health assess-
ment, performance evaluations, etc. This is not fullling the expec-
tations in Industry 4.0 paradigm with regard to the role of DTs.
•(Fourth gap) Current reported applications of DTs in state-of-art ar-
chitectures make a poor usage of the data gathered from their
physical counterparts. The reported studies propose solutions limited
to data collection and visualization, i.e., structured and more elab-
orated data analysis and decision-making are not convincingly sup-
ported on the real time data. This does not take advantage of having a
virtual representation, constantly updated with the last available
information from the physical counterpart, which is the improve-
ment of understanding and managing of the physical assets promised
in the Industry 4.0 paradigm.
2.3. Objectives of the work
In order to address the main gaps found in the state-of-the-art review,
the overarching goal of this paper is to design and implement a
distributed DT-based framework to address smart decision-making for
scheduling tasks in CPPS.
The framework is characterized by three fundamental elements: (i)
DTs at different hierarchy levels, (ii) local CM and (iii) a global decision-
making. The approach is centered on the local CM of different work-
stations and the production rate of the production system, in order to
automatically perform the re-scheduling to improve the performance of
the whole system. The local condition of the single workstations is
inferred from the local CM that makes use of local DTs, whereas the
production rate is evaluated and assured from the global DT of the whole
system.
Industrial setups usually depict uncertainty and complexity. Non-
linearities, noise and uncertainty are still limiting the effectiveness and
validity of rst-principle and mathematical models based on differential
equations. Alternative techniques are usually applied to obtain more
reliable and accurate models. Machine learning-based strategies are
becoming the main methods reported to develop models in industrial
environments. Indeed, this is the main rationale for using machine
learning to generate the local DTs and the CM. Likewise, autonomous
decision-making in the industry should be intuitive, user-friendly and
emulate operators’ know-how and their socio-cognitive skills through
verbalization. Therefore, the design of the smart decision-making sys-
tem will be carried out using Fuzzy Logic-based inference systems.
Another important aspect of the work reported in this paper is the aim of
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
6
carrying proof-of-concept and validation to assess feasibility of the DT-
based framework for decision-making to dynamically deal with uncer-
tainty and complexity of industrial setups.
3. The proposed framework
The design and the implementation of a distributed framework based
on DTs and CM facilitate decision-making at each level of the
manufacturing system, with the ultimate goal of increasing efciency of
operations management. From single equipment to the overall shop
oor management, the behaviors are emulated, and decisions are made
according to the condition of different local assets and the global per-
formance of the whole plant or oor shop. The nal target is certainly
reaching higher efciency and productivity at a global factory level.
Fig. 1 illustrates the conceptual diagram of the proposed framework.
The framework is centered on a novel decision-making to improve the
global performance by using scheduling actions through the global DT,
while exploiting the local condition-based monitoring to consider the
currently running conditions of the physical assets, through the local
DTs. The framework takes advantages of the potential of edge
computing in local DTs. In each workstation of the shop oor is located a
local node in charge of gathering, preprocessing and ltering data from
the eld in order to obtain, in real time, the main features/conditions of
the assets. The use of edge computing in local nodes can improve the
real-time response of the framework because these are in constant
interaction with the dynamics of each of the assets of the shop oor.
Local nodes in fact consist of two main modules: one for the local DT
and one for CM algorithm, respectively. The DTs at local workstations
mirror and emulate the behavior of the workstations and the process,
enabling simulations to detect and predict the current and future
behavior of the asset. CM consists of a predictive model based on a
machine learning strategy, to predict the state of the components that
compose each of the assets. By combining both modules, the future
state/condition of the system components or devices can be detected,
enriching the global decision-making of the whole production system.
The global node is composed of three modules: (i) the global DT, (ii)
the global decision-making and (iii) the scheduling optimization mod-
ule. The global DT collects the information from local DTs. Simulations
of Global DT are run with the collected data to predict the behavior of
the production system; it is then possible to analyze future production
and efciency rates. The global DT is also fed with the production goals
of the production system. which are typically an external input coming
from the production management level (see in the Fig. 1 the box, called
“Goals for the production system”). The global decision-making de-
termines which is the best action to be performed to improve the per-
formance of the whole CPPS. In order to deal with uncertainty and
nonlinearities, the literature is plenty of soft-computing technologies
such as Fuzzy and Neuro-Fuzzy Systems, that can be used by global DT
and CM of each local workstation to make the best decision, action or
recommendation, such as triggering the rescheduling action or simply
informing the operator in the shop oor for minor adjustments. The
module for scheduling optimization eventually performs an optimiza-
tion process to reschedule the production system. It can use gradient-
free population-based optimization algorithms to estimate the optimal
sequence of products that minimizes the overall production makespan.
The inputs are the jobs that have to be performed and the processing
times of the worktations involved in the process. The algorithm then
computes different possible schedules, which are later on evaluated by
Discrete Event Simulation (DES) model of the production system. The
best sequence, that accomplishes the stopping criterion, is the feedback
given to the operator or to the manufacturing execution system.
It is important to remark that the proposed framework can be
extended without losing generality to any CPPS with local and global
DTs. This advantage of the framework is endorsed by a data exchange
method, which is independent from the application case. The main
features, specicities or peculiarities of the production system and its
components are fully included into the global and local DT modules that
are developed for the specic application. DTs must be updated or re-
tuned along the time. If a component or sub-system represented by the
DT in the virtual world is modied or changed, either in relation to
software or, hardware, the DT must be updated. This is a crucial aspect
to be considered when building a DT as well as the interoperability
between DTs. Each local DT is independent from the others in the pro-
posed framework, and thererfore it can be managed as a single entity of
the framework without affecting the other entities.
3.1. Local digital twin
The design and implementation of a local DT for representing a local
process is essential to guarantee a good performance of the proposed
framework. A model that appropriately mirrors and emulated the
physical counterpart is essential to detect deviations in KPIs of the local
Fig. 1. Conceptual diagram of the proposed framework.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
7
system and thus clearly contributes to improve the predictive mainte-
nance and the decision-making procedures. modeling an industrial
process is usually a complicated task. Due to the nonlinearities or in
general uncertainties, it is not straightforward to obtain an accurate
mathematical representation. Machine learning-based modeling tech-
niques usually are alternative techniques due to their ability to represent
some nonlinearities and to capture the main characteristics of the
physical processes. Among dozens of available methods for modeling
reported in the literature, the hybrid incremental model (HIM) that
exploits the principle of incrementality was selected. The design and
implementation of the HIM –based method is inspired by Pedrycz and
Kwak (2007).
HIM approach has the advantage of using a general representation to
capture the main characteristics of the physical process in addition to an
incremental representation to catch nonlinearities. Overall, it is rec-
ommended to use generic models, such as linear or polynomial re-
gressions, when there is no prior knowledge of the system to be
modelled. In this study, the general representation is carried out by a
polynomial of degree m tted using the least squares algorithm. The
output has the following expression:
̂yB(xi) = fB(xi,O(xi)) (1)
where xi is the i th input point and O(xi) is the output value of the xi
point.
Likewise, a wide range of techniques is also available for yielding the
incremental part. In this study, Fuzzy k-Nearest Neighbours (F-kNN) is
selected because of its simplicity from the computational viewpoint,
ease of interpretation and good accuracy. F-kNN consists of averaging
the value of the points closest to the objective point weighted by the
similarity of each points. In order to calculate the proximity, the
Euclidean norm, the most widely used approach, was applied.
Finally, the HIM-based approach combines the two above-mentioned
representations (the physical representation and the incremental one).
The tuning parameters of HIM are the degree of the polynomial (m), the
neighborhood size (k), and the fuzzy strength (p). The training of the
general representation consists in tting the polynomial of degree m and
the incremental training to populate the state space of F-kNN. The
evaluation consists in calculating the output, at a point q, by adding the
term from the incremental model to the output of the general repre-
sentation, as follows:
̂y(q) = ̂yB(q) +
̂
l(q)(2)
where ̂
yB(q)is the general model output, ̂
l(q)the term obtained by the
incremental model and ̂
y(q)is the output value at q point.
Once the DT is obtained, it is continuously updated to keep the ac-
curacy in the physical asset emulation and, for that purposes, it is
continuously connected with the physical process. The tuning procedure
is triggered when the accuracy of the DT decreases. Fig. 2 shows the
local DT evaluation and updating diagram. The output, corresponding to
an operation, is compared with the real values to calculate the tting
error. When the error exceeds a threshold, the model is re-trained
through the tuning procedure supported on (Beruvides, Casta˜
no,
Haber, Quiza & Villalonga, 2017). The objective function is the mini-
mization of the tting error.
3.2. Condition based monitoring
The CM in the local workstation takes into consideration the signals
obtained from sensors and the output from the local DT to determine the
current state of the corresponding asset. Thus, the inuence of degra-
dation on the different components of the shop oor can be detected.
Moreover, in combination with the DT, equipment degradation can also
be predicted in advance. In this way, scheduling decisions can be made
not only considering the production KPIs.
Monitoring the state of the assets in a production system is a very
important task. For example, an unexpected failure in one of the ele-
ments of a production system could cause the stop or decrease the
production rate. Certainly, CM is a powerful tool in predictive mainte-
nance because it makes it possible to more accurately schedule pro-
duction and maintenance actions. In addition, CM contributes to
minimize the risk of unexpected failures improving equipment reli-
ability and operators’ safety.
The CM system is composed of two main stages Jardine, Lin &
Banjevic (2006): the data processing and the predictive model. The rst
stage carries out the features extraction of the signal obtained from the
sensors. This stage allows detecting the main characteristics of the sig-
nals measured. Thereby noise is ltered obtaining more precise infor-
mation to generate more accurate predictive models. Furthermore, the
input data volume is reduced thus decreasing the training and evalua-
tion computational cost of the models.
The predictive model is the most important element of the CM. It
evaluates the current condition of the asset analyzing the signals ob-
tained from the sensors. Due to the main characteristics of the industrial
signals, the use of a machine learning strategy is proposed to generate
the predictive models.
3.3. Global digital twin
Until now the local DTs were presented: they can run as stand-alone
models of the single workstations and interact with the CM to predict the
future state of the asset. Local DTs are then connected to a global DT,
that is the DT of the whole production system.
The global DT interrogates the local DTs to acquire the required data
and information. Obviously, the role of the global DT is to merge and to
elaborate data and information coming from the local DTs in a coherent
and efcient way, by only gathering those data and information that are
necessary to the decision-making at this level with the required sampling
frequency to meet specications. This global DT does not interface
directly with the physical world, instead it communicates with the local
DTs that lay on the lower levels to get the information about the shop
oor. It is worth to say that this type of DT can also be seen as a middle
representation, this may happen in situations where the complexity of
the production systems is higher such as in Cyber-Physical Systems of
Systems; in this way, the structure, that here is presented as a double
layer, could be expanded to have three or more layers of aggregation.
The functionality of the global DT is described as follows.
Fig. 2. Diagram of the local digital twin evaluation and updating mechanism.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
8
3- Firstly, it collects data and information from relevant variables in the
lower-level DTs, to give a comprehensive view of the behavior and
performance of a production system. The local DTs collect data from
the eld sensors and IIOT. In addition, they can combine information
by elaborating those coming from the eld using proper algorithms
or models included inside the DT. The large amount of data that can
be generated, for instance in a shop oor, is huge. This potential
overow has thus to be managed by the DT-based framework, by pre-
processing and processing the proper data at each level of the pro-
duction system. Thus, data must be ltered to get the right infor-
mation, volume, sampling frequency, data already elaborated and
aggregated, etc. For this reason, local DTs should share with the
global DT only data and information that are useful to be replicated
and simulated.
3- The global DT has the fundamental role to enable communication
with all the underlying local DTs. This capability will allow the
global DT to control and to manage the local DTs, which in turn
support the actuation on their physical counterparts. The operating
conguration must be correspondingly set.
3- The global DT also supports decisions to be sent to the lower level of
the framework, for instance stopping a machine with deviated
behavior (e.g., malfunctioning, failure, etc.) from the pre-dened
one, according to the CM task run by the DT.
3- The global DT interacts with the production system management on
the shop oor, therefore with the MES, or with the HMI (Human
Machine Interface) to production operators. Indeed, at this point all
decisions and actions will be the ones that have the most inuence on
the whole system. The global DT will host the computational pro-
cedures and interface with tools that elaborate and carry out
decision-making, in order to enhance the performance of the pro-
duction systems replicated in the virtual world. In this sense, it is
work remarking that global decisions will affect the whole system
rather than a single unit and can be sent to a single or multiple local
DTs. These types of decisions are then considered more important,
and they will prevail on the local decisions, since the global optimum
overcomes the local ones. At the end, the communication with MES/
HMI makes possible to leverage the information elaborated by the
global DT, such as a rescheduled production plan or an alarm setting
for the operators or plant workers in case of a failure on the system.
Another important aspect that characterizes the global DT is that
users, such as operators or operation managers, mostly interact only at
this level of the architecture. Thus, the inputs and the outputs of this
level of the framework should be suitable for interactions with humans.
Fig. 3 shows the scheme of the DT-based framework (i.e., case of
double layered architecture). It represents the logical connection of the
global DT to various local DTs, linked to the single workstations as
physical assets composing the production system. The optimization
module is connected to the global DT that is composed of two other sub-
modules, represented by discrete event simulation (DES) model of the
production system and the scheduling optimization algorithm. This
module performs the scheduling optimization triggered by the global
decision-making and will be better described in Section 3.5.
3.4. Global decision-making
In this work, the main target of the decision-making process is to
increase productivity of a production system. For the sake of computa-
tional efciency, a simple approach is proposed using two key variables:
the production rate (PR) and the currents state of each workstation
(ST
1…M
) where M is the total number of local workstations/assets. The
production rate is obtained by running the global DT of the production
system. The condition of workstations is also a very important aspect to
be taken into account because a failure can slow down the production or
lead to stop the whole production system. The condition of workstations
is inferred from CM model embedded in local workstations.
Within the different approaches used for decision-making processes,
fuzzy inference systems (FIS) are among the most reported in the liter-
ature Isermann (2005). FIS are a technology widely applied in industry
since they are very simple, powerful and intuitive (Haber, Alique, Ali-
que, Hern´
andez & Uribe-Etxebarria, 2003)(Ramírez, Haber, Pe˜
na &
Rodríguez, 2004). FIS have the potential to appropriately represent the
non-linear relationship between input and output variables and to sup-
port the decision-making procedure in a robust and exible way based
on the fuzzy sets and the fuzzy rules (Caiado et al., 2021) Haber &
Alique (2007).
Therefore, the selected inputs for the fuzzy inference system are the
global production rate (PR) and the condition of the workstations of the
production system (ST
1…M
). Fuzzy membership functions are dened by
Gaussian-type functions, shown in Fig. 4. The selection of the type of
membership function is not straightforward due to the inuence on the
whole performance of the fuzzy inference system. The rationale for
selecting Gaussian membership functions for the input is twofold.
Firstly, this selection is supported on most contemporary studies re-
ported in the literature. A recent study presents a comparison of fuzzy
inference systems including Gaussian and linear membership functions
for inputs and outputs, respectively. The study demonstrates that
Sugeno FIS outperformed the Mamdani FIS for accurate results Tab-
bussum & Dar (2021). Secondly, both Gaussian and triangular mem-
bership functions have demonstrated to be closely performing well
without remarkable differences. However, the continuity and concise
notation of Gaussian-type functions pave the way of future works in
Fig. 3. Diagram of digital twins at local and global levels.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
9
relation to optimization Tabbussum & Dar (2021) and hybridization
with other techniques such as neuro-fuzzy evolving systems (Gajate,
Haber, Vega & Alique, 2010).
The membership functions of the global production rate were
labelled as LOW, MED and HIGH indicating the three possible states of
the production. The membership functions of the workstation condition
are FAILURE, DEGRADED and GOOD.
Takagi-Sugeno-Kang (TSK) is the choice of fuzzy system because it is
considered more exible and computationally efcient (La F´
e-Perdomo,
Beruvides, Quiza, Haber & Rivas, 2019). Moreover, the outputs in TSK
provide a more explicit relationship with the inputs since a weighted
sum of the data points is used in the defuzzication stage. The output is
dened as singleton membership functions labelled: NO_OPERATION
(1), OPERATOR_ASSISTANCE (0.5) and SCHEDULING (0). The rst
Fig 4. Inputs membership functions. a) Production rate (PR). b) Workstation condition (ST).
Fig. 5. Surfaces resulting from the fuzzy inference system, a) for each ST
1
condition, b) for each ST
2
condition.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
10
membership function represents that no action is needed. The second
one means that the system does not count with enough information to
take an action so the operator attention is mandatory, therefore keeping
the human in the loop. The last membership function represents the
need of automatically triggering the scheduling procedure.
The fuzzy rules are dened establishing a direct relationship between
the production rate and the workstations condition with the action
needed to improve the system. A fragment of the pseudocode of the rules
considering the monitoring of two workstations is shown in Pseudocode
1. In this case, the rule base is composed by 27 rules. The number of rules
are set in two iterations. Firstly, rules are automatically generated on the
basis of input and output variables and the number of membership
functions. Secondly, the pruning of the rule base is done by combining
the know-how of operators and technologists with a verbalization
technique. “Sup-Product” is applied as compositional rule of inference.
In Pseudocode 1, labels ST
1
and ST
2
correspond to the state of the
workstations 1 and 2, respectively and PR is the production rate.
The representation of surfaces resulting from the FIS is not easy when
there are more than two antecedents in the rules. For the sake of clarity
and without losing generality, Fig. 5 shows the surfaces assuming that
each condition (FAILURE, DEGRADED and GOOD) of the workstation
ST
1
is constant (Fig. 5a) and vice versa for ST
2
(Fig. 5b).
3.5. Scheduling optimization
When a decrease in the production rate is detected, the decision-
making module could trigger the scheduling process to solve the prob-
lems that is affecting the production system. The scheduling process has
two main stages: (i) a population-based optimization algorithm and (ii) a
simulation model. The simulation of the production system is used to
replicate the sequence of operations, and their processing times,
occurring during a production for each individual of the population. A
population-based optimization algorithm is selected because it is
appropriate to deal with noise and uncertainty. This strategy, based on
generated populations with a large number of individuals, makes it
possible to try different scenarios for optimization. These elements are
representing the optimization module, already depicted in the diagram
shown in Fig. 3, enabling global DT to perform the scheduling process
described in this section.
The simulation model is a discrete-event simulation model syn-
chronized according to the eld conditions with the global DT model
and then it is used to emulate various alternative scheduling sequences
of the production system to support the scheduling optimization
algorithm. More specically, it is used to replicate the production in
terms of processing times of each workstation. Thus, it is able to deter-
mine the makespan of a single product or an entire batch of products for
each simulation run. The required simulation inputs are the processing
times of each workstation of the production system and the list of jobs
that must be produced. This list considers all the operations that each
single product requires to be produced, and at the same time, it also
considers the priority that each job might have. The DES model is also
part of the DT-based framework. However, it must not be confused with
the global DT module that monitors the operations of the production
system through the connection with the local DTs and alerts on the
global eld conditions.
The scheme shown in Fig. 6 is applicable for any kind of population-
based optimization algorithm (in the picture it is Opt. A). The objective
function of the optimization algorithm is the minimization of the total
makespan and it operates as follows:
1 It receives as inputs: the list of products (later called jobs) to be
produced. Each job is identied with an ID number, a priority index
(the higher the index the sooner the product must be produced) and
its technological route.
2 The optimization algorithm will generate a rst population of in-
dividuals, which are different scheduling alternatives. Each indi-
vidual consists of a unique sequence of jobs which expresses the
order in which the jobs are released into production.
3 The rst population is then simulated in the simulation model. The
single individual is given as input at the simulation model which will
simulate the production. For each individual the optimization algo-
rithm collects the production performance, namely the makespan.
The processing times at this step are xed and set by the simulation
model.
4 The optimization algorithm selects the best individuals (i.e. the ones
with the lowest makespan) and it generates a new population. Then
the discrete simulation model simulates again this population to nd
the most performing individuals. This process is iterative, and it
nishes once the stopping criterion is achieved, and the lowest
makespan sequence will be the new production schedule of the
whole system.
4. Case study on pilot assembly line
The case study aims at demonstrating the suitability of the proposed
DT-based decision-making framework applied to a pilot assembly line at
Fig. 6. Diagram of the scheduling optimization process.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
11
the Industry 4.0 Laboratory of the School of Management of Politecnico
di Milano. The assembly line is composed of an educational set of
modular workstations provided by Festo, that can be arranged to form
different combinations of assembly lines for research and educational
purposes. All workstations are equipped with two PLC from Siemens
(one for actuation control and sensor reading and the other for energy
monitoring). Each workstation is also equipped with sensors for elec-
trical energy and for pneumatic air consumptions. The line has an in-
ternal MES for production progress control and an energy management
MES. Each PLC, both of the operational and energy PLC types, provides
an OPC UA server publishing all the internal variables and sensors states
of the workstation.
The pilot line, represented in Fig. 7a, is composed of seven work-
stations, each one dedicated to one or more tasks, to assemble a
simplied mobile phone.
In Fig. 7b, the “Manual” workstation (1) is the starting point of the
process, where the loading/unloading takes place. The “Front Cover”
workstation (2) is in charge of the positioning of the front cover on the
pallet. The “Drilling” workstation (3) is the workstation where cover
drilling is performed. In the “Robot Assembly” workstation (4), the
Printed Circuit Board (PCB) and the fuses are placed inside the front
cover. The “Camera Inspection” workstation (5) controls the different
components positioned in the inside of the cover. In the “Back Cover”
workstation (6) the back cover is placed over the front cover. The “Press”
workstation (7) presses the two covers to close the product. At the end,
the piece returns to the initial workstation where it is unloaded by the
operator. The position 8 represents a bridge to switch the production
ow either to the robotic cell or to the camera workstation, depending
on the assembly route of the current piece.
The pilot line has a pneumatic compressor, which supplies pressur-
ized air to the seven workstations. Each workstation then operates the
pneumatic actuators through the compressed air. The single workstation
is also responsible for a section of the conveyor. In fact, despite there is
only a single conveyor route on the line, the conveyor system is frag-
mented into eight modules, that allow to stop any of them when not
needed, thus saving energy. Finally, considering the IIOT already
available in the line, it is worth remarking that: i) the path of each
product can be tracked using a RFID-tagged chip embedded on the pallet
that carry the product, ii) the line is equipped with sensors for energy
and condition monitoring; iii) all the elements in the local network
interchange data via OPC UA protocol (i.e. MES, edge computers,
sensors).
Data communications in the assembly line are carried out via OPC
UA protocol to the MES of the line. Moreover, data are also stored in an
online database, which is a MongoDB. This database stores and even-
tually provides data for further utilizations. The process of data sharing
is automatic, and the frequency of acquisition is 1 Hz for all the data
coming from the eld. The only exception is the accelerometer sensor,
which has a frequency of acquisition of 200 Hz.
Overall, the assembly line complexity and the variety of products
that can be produced through different operations performed by the
workstations provide it with the characteristics proper of a real “in-
dustrial like” system. Given this setting, the implementation of the
framework was then performed by rstly developing the local nodes for
two different workstations – the front cover workstation and the drilling
workstation – whereas the global node was nally deployed in the
computer managing the whole line. The results obtained so far (and
presented in next sub-sections) allow validating the framework as an
initial proof-of-concept.
There are several gradient-free and bio-inspired optimization
methods reported in the literature (Haber, Beruvides, Quiza & Her-
nandez, 2017). A full review of most emerging techniques goes beyond
the scope of this paper. Genetic Algorithm (GA), is selected because it is
considered the most promising optimization approach for dynamic
scheduling, widely used simulations and DT modeling (Fumagalli,
Polenghi, Negri & Roda, 2019; Negri et al., 2020). The most important
parameters and their setting such as genetic operators (crossover and
mutation), stopping criteria (maximum number of iterations; maximum
number of iterations in which the makespan does not decrease), amond
others are summarized in Table 2.
4.1. Results: local digital twin and condition-based monitoring
As a rst step, the local DT and the CM models corresponding to each
workstation were generated. The local DTs were trained with data ac-
quired from the real operation of the whole pilot line. Table 3 shows the
parametrization corresponding to each model. As described in Section
3.1, the tuning parameters are the degree of the polynomial (m), the
neighborhood size (k), and the fuzzy strength (p). To analyze the quality
of the local DT, three indexes were considered: root mean square error
(RMSE), the mean absolute error (MAE) and the relative absolute error
(RAE). Table 4 shows the models performance indicators, and Figs. 8
and 9 illustrate the behavior of the DT and the process signals.
Fig. 7. a). Picture of the Industry 4.0 assembly line b) Scheme of the main assets.
Table 2
Parameters of the implemented GA.
Parameter Value
No. of jobs scheduled 50
Population dimension 10
Mutation rate 0.02
No. of children for elitism operator 1
Maximum number of iterations 100
No. of iterations for stall criterion 20
No. of repetitions of each simulation 30
Weighing factor 0.5
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
12
In order to generate the CM models an experimental set-up was
designed to emulate failures in the different workstations. In the front
cover workstation, the monitoring of the pneumatic system was
considered. In the drilling workstation, the pneumatic system and the
drilling process were considered. In the pneumatic system, an exhaust
valve was opened to drop the pressure and simulate a failure. In order to
analyze the pressure signal, the standard deviation and the mean were
chosen as pre-processing analyses. Predictive models were built using
articial neural networks, in particular the multilayer perceptron to-
pology (MLP). This topology is used due to its suitability to emulate the
nonlinear behavior of industrial systems (Alique, Haber, Haber, Ros &
Gonzalez, 2000) because of good tradeoff between generalization ca-
pacity and computational cost. The model corresponding to the front
cover workstation was congured with a hidden layer of 12 neurons,
trained with 5000 epochs. The transfer function of the hidden layer was
hyperbolic tangent, while the linear function was used for the output
layer. The learning rate was 10
−3
and the minimum gradient 10
−7
. The
MLP in the drilling workstation was congured with the same parame-
ters but the hidden layer was populated with 15 neurons. Figs. 10 and 11
show the performance of the models in the validation process. The states
of the workstation were dened as good (2), degraded (1) and failure
(0).
The vibration signal from three accelerometers was also used to
develop the CM of the drilling workstation. In order to emulate the
degradation in the process, a shaker was applied to the drilling tool. The
preprocessing of the vibration signal was carried out by wavelet trans-
form, taking as the mother wavelet (db4) and n =1, combined with the
statistics, i.e. RMS and kurtosis of the obtained wavelet coefcients. The
algorithm used for the predictive model was MLP with backpropagation
with a hidden layer with 22 neurons, 5000 epochs for training, hyper-
bolic tangent hidden layer transfer function, linear output layer, output
layer with one neuron, learning rate µ =10
−3
and minimum gradient of
10
−7
. Fig. 12 depicts the behavior of the CM model for the vibration in
the drilling workstation.
4.2. Results: global digital twin, decision-making and scheduling
optimization
The proof-of-concept validation of the whole developed system is
presented in this section. For the sake of clarity and in order to validate
the framework an experimental set up was designed. The set up consists
of ve stages over time: 1- normal conditions, 2- Shaker turned on,
simulating degradation in drilling workstation, 3-Normal conditions,4-
Leak in pressure valve and 5- Failure in pressure. The main objective is
to evaluate the decision-making capability of the system under different
situations. Having generated the local DTs and CM procedures, the
decision-making module was embedded into the global layer and con-
nected to the global DT as it is shown in Fig. 1. The global layer is the
higher level of the framework, from where the global DT has a complete
overview of the system, thus overall optimization decisions can be made.
More specically, different stages are so arranged, as follows: i) 20
operations in normal condition were initially run; ii) at operation 21 the
Table 3
Digital twin initial parametrization.
Front cover workstation Drilling workstation
Pressure Current Pressure Current
m 2 2 2 2
k 3 5 3 5
p 1.89 2.01 1.93 1.98
Table 4
Digital Twin performance indicators.
Front cover workstation Drilling workstation
Pressure Current Pressure Current
RMSE 0.0430 0.0042 0.0318 0.0037
MAE 0.0335 0.0030 0.0252 0.0040
RAE 0.1621 0.138 0.146 0.1238
Pseudocode 1
Fuzzy rules.
if (PR is HIGH) and (ST
1
is FAILURE) then (ACTION is SCHEDULING)
if (PR is HIGH) and (ST
1
are DEGRADED) and (ST
2
is GOOD) then (ACTION is
OPERATOR_ASSISTANCE)
if (PR is HIGH) and (ST
1
is DEGRADED) and (ST
2
is DEGRADED) then (ACTION is
OPERATOR_ASSISTANCE)
if (PR is HIGH) and (ST
1
is GOOD) and (ST
2
is GOOD) then (ACTION is
NO_OPERATION)
if (PR is MED) and (ST
1
is FAILURE) then (action is SCHEDULING)
if (PR is MED) and (ST
1
is DEGRADED) and (ST
2
is Good) then (ACTION is
OPERATOR_ASSISTANCE)
if (PR is MED) and (ST
1
is DEGRADED) and (ST
2
is DEGRADED) then (ACTION is
SCHEDULING)
if (PR is MED) and (ST
1
is GOOD) and (ST
2
is DEGRADED) then (ACTION is
NO_OPERATION}
if (PR is LOW) and (ST
1
is FAILURE) then (ACTION is SCHEDULING)
if (PR is LOW) and (ST
1
is GOOD) and (ST
2
is GOOD) then (ACTION is
OPERATOR_ASSISTANCE)
if (PR is LOW) and (ST
1
are GOOD) and (ST
2
is DEGRADED then (ACTION is
OPERATOR_ASSISTANCE}
if (PR is LOW) and (ST
1
is DEGRADED) and (ST
2
is DEGRADED) then (ACTION is
SCHEDULING)
Fig. 8. Digital twin in the drilling workstation.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
13
shaker is turned on to simulate degradation in the drilling process; iii) in
operation 26, the shaker is turned off and the production process comes
back to normal conditions; iv) during the operation 31, an exhaust valve
was opened in the front cover workstation to emulate a failure in the
pressure system; v) in operation 43, a failure occurs in pressure in the
front cover workstation.
In order to illustrate the performance of the framework, Fig. 13
shows the results of the global decision-making module. It is worth
pointing out that the system correctly detects the failure in the front
cover workstation, triggering the scheduling algorithm to avoid a
decrease in production. The decision-making module, hosted into the
global layer together with the global DT, triggers the optimization
module for the rescheduling process, which will then start to elaborate
the information gathered from the local DTs to compute the production
schedule that minimizes the overall makespan. The global DT plays then
a key role, merging the information gathered from the CM module of
each local DT, elaborating a decision based on the embedded decision-
making module, and nally activating the scheduling process through
the optimization module.
According to the re-scheduling order sent by the global DT-supported
decision-making module, the scheduling algorithm computes a new
sequence of the remaining products that had to be produced. After nine
generations, the GA found the best sequence and the stopping criterion
was reached. The output is the sequence of products that minimizes the
overall makespan. As already stated, this scheduling activity was trig-
gered by the decision-making module, to consider what happened on the
assembly line and to produce an updated schedule. Fig 14 shows the
output of the scheduling algorithm, where the reaching of an optimal
solution is clearly visible considering the trend of the solution found
after each generation. The sequence provided as the ID of each product
in the proper order, is expected to take around 3.43 min to be fully
produced by the assembly line.
5. Conclusions
The work reported in this paper is fueled by the current research line
aimed at designing and providing more autonomy, decentralization and
more integrated decision-making functionalities on the basis of digital
Fig. 9. Digital twin in the front cover workstation.
Fig. 10. Pressure condition-based monitoring of the drilling workstation.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
14
Fig. 11. Pressure condition-based monitoring of the front cover workstation.
Fig. 12. Vibration condition-based monitoring of the drilling workstation.
Fig. 13. Behavior of the global decision-making.
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
15
twin-based frameworks and architectures, with the nal goal of
achieving automated decision making in cyber-physical production
processes. The presented proof-of-concept demonstrates that the pro-
posed architecture for decision-making supported on local and global
digital twins, and optimization is very promising to face challenges
beyond re-scheduling of cyber-physical production processes. The pro-
posed strategy outperforms available solutions in the market to address
production scheduling in terms of automated decision-making and
higher exibility, taking into account that local and global digital twins
can be also automatically updated.
From the viewpoint of cyber-physical production systems, digital
twins has not only the role of accompanying the physical production
systems in a mirroring simulation mode but has also the role of elabo-
rating advanced data analytics to actively control and perform dynamic
decision-making. The proposed framework combines a fuzzy inference
system and a scheduling procedure, in an efcient decision-making
strategy at the global level of the production system. Furthermore, the
distributed framework exploits digital twins and condition monitoring
for intelligent decision-making by enabling the use of local data to
elaborate the required condition monitoring of cyber-physical produc-
tion processes, which is a key aspect to for automatic dynamic sched-
uling and further steps towards self-reconguration.
The framework is dened as a generic digital twin-based layered
architecture merging local digital twins with a single global digital twin,
capable of representing, assessing, and managing the whole production
system. The proposed architecture is fully independent of type of the
cyber-physical production system, as it denes data ows whereas the
digital twin modules at local and global levels are incorporating the
whole specicity of production system. Moreover, all the bricks that
constitute this structure, namely the single local digital twins, are also
able to work as stand-alone modules for local decision-making support.
The only hierarchical rule is that a global optimization overwrites local
ones.
The experimental study corroborates how local digital twins and
condition monitoring serve to predict the state of the components based
on the information collected by sensors. Then, the global digital twin
estimates the production rate based on the data collected from the shop
oor. This information is processed in the decision-making stage to
determine if re-scheduling of the whole system is required. The imple-
mented modules are assessed and validated in an assembly pilot line,
and the proposed approach serves to automatically detect changes in the
manufacturing process and make appropriate decisions for re-
scheduling accordingly.
This work contributes to the progress of the state-of-the-art lling
some gaps identied in this study. Firstly, the proposed framework in-
tegrates local digital twins into a global one representation. Secondly,
the design of layered digital twins contextually is deployed in experi-
mental setup of a real cyber-physical production process. Moreover, the
proposed distribute framework integrates both digital twins structures
and real-time data gathering and analysis for an accurate dynamic
decision-making. The main contribution is therefore the design and
implementation of a framework that integrates local digital twins and
global digital twin to deal with an automatic re-scheduling process for
cyber-physical production processes.
The main limitations of the proposed framework are the computing
constraints for re-scheduling which relies on the accuracy of the digital
twins and the complexity of the production system in relation to the
computing power nowadays available in cyber-physical production
systems. Secondly, the complexity of data acquisition, gathering and
processing grows with the size of cyber-physical production systems,
especially when the framework is applied to real industrial
environments.
Future works will be conducted to add more functionalities to the
framework, to explore other modeling approaches and the decision-
making strategies, and to extend the proposed strategy beyond the dy-
namic rescheduling. Moreover, further studies should be carried out to
reduce the latency by developing more computational efcient machine
learning algorithms and to improve the communication with the
manufacturing execution systems. Another future work is the develop-
ment of new monitoring method to detect any change both in produc-
tivity and ergonomics of operators in the workstations. This may lead to
design local digital twins of a different nature, as they must be adapted
to the human behaviors and tasks.
Declaration of Competing Interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Acknowledgments
This work was partially supported by the H2020 projects "Platform
enable KITs of Articial Intelligence for an Easy Uptake of SMEs
Fig. 14. Output of the optimization algorithm (genetic algorithm).
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
16
(KITT4SME)" under Grant 952119, and “Power2Power: Providing next-
generation silicon-based power solutions in transport and machinery for
signicant decarbonisation in the next decade (Power2Power) funded
by ECSEL-JU and MICINN under grant agreement No. 826417. More-
over, this research work has been possible thanks to the use of the In-
dustry 4.0 Laboratory of the School of Management of the Politecnico di
Milano (lab site: www.industry40lab.org).
References
Alam, K. M., & Saddik, A. El (2017). C2PS: A digital twin architecture reference model
for the cloud-based cyber-physical systems. IEEE access : practical innovations, open
solutions, 5, 2050–2062. https://doi.org/10.1109/ACCESS.2017.2657006
Alique, A., Haber, R. E., Haber, R. H., Ros, S., & Gonzalez, C. (2000). Neural network-
based model for the prediction of cutting force in milling process. A progress study
on a real case. IEEE International Symposium on Intelligent Control - Proceedings.
Ashtari Talkhestani, B., Jung, T., Lindemann, B., Sahlab, N., Jazdi, N., Schloegl, W. et al.
(n.d.). An architecture of an intelligent digital twin in a cyber-physical production
system. At - Automatisierungstechnik, 67(9), 762–782. https://doi.org/10.1515/auto-
2019-0039.
Bakliwal, K., Dhada, M. H., Palau, A. S., Parlikad, A. K., & Lad, B. K. (2018). A Multi
Agent system architecture to implement collaborative learning for social industrial
assets**this research was funded by the royal academy of engineering under the
newton bhabha scheme (project no. hepi—1516—10). this research was supported
by Sust. IFAC-PapersOnLine, 51(11), 1237–1242. https://doi.org/10.1016/j.
ifacol.2018.08.421
Barricelli, B. R., Casiraghi, E., & Fogli, D. (2019). A survey on digital twin: denitions,
characteristics, applications, and design implications. IEEE access : practical
innovations, open solutions, 7, 167653–167671. https://doi.org/10.1109/
ACCESS.2019.2953499
Beregi, R., Szaller, ´
A., & K´
ad´
ar, B. (2018). Synergy of multi-modelling for process control.
IFAC-PapersOnLine, 51(11), 1023–1028. https://doi.org/10.1016/j.
ifacol.2018.08.473
Beruvides, G., Casta˜
no, F., Haber, R. E., Quiza, R., & Villalonga, A. (2017). Coping with
complexity when predicting surface roughness in milling processes : hybrid
incremental model with optimal parametrization. Complexity, 2017, 1–12.
Bevilacqua, M., Bottani, E., Ciarapica, F. E., Costantino, F., Donato, L. D., Ferraro, A.,
et al. (2020). Digital twin reference model development to prevent operators’ risk in
process plants. Sustainability (Switzerland), (3), 12. https://doi.org/10.3390/
su12031088
Borangiu, T., Oltean, E., Raileanu, S., Anton, F., Anton, S., & Iacob, I. (2020). Embedded
digital twin for ARTI-type control of semi-continuous production processes.
International Workshop on Service Orientation in Holonic and Multi-Agent
Manufacturing, 2019, 113–133.
Brandenbourger, B., & Durand, F. (2018). Design pattern for decomposition or
aggregation of automation systems into hierarchy levels. In 2018 IEEE 23rd
international conference on emerging technologies and factory automation (ETFA), 1 (pp.
895–901). https://doi.org/10.1109/ETFA.2018.8502627
Caiado, R. G. G., Scavarda, L. F., Gavi˜
ao, L. O., Ivson, P., Nascimento, D. L.de M., &
Garza-Reyes, J. A. (2021). A fuzzy rule-based industry 4.0 maturity model for
operations and supply chain management. International Journal of Production
Economics, 231, Article 107883. https://doi.org/10.1016/j.ijpe.2020.107883
Cardin, O., Castagna, P., Couedel, D., Plot, C., Launay, J., Allanic, N., et al. (2020).
Energy-aware resources in digital twin: The case of injection moulding machines.
Studies in Computational Intelligence, 853, 183–194. https://doi.org/10.1007/978-3-
030-27477-1_14
Catarci, T., Firmani, D., Leotta, F., Mandreoli, F., Mecella, M., & Sapio, F. (2019).
A conceptual architecture and model for smart manufacturing relying on service-
based digital twins. In 2019 IEEE international conference on web services (ICWS) (pp.
229–236). https://doi.org/10.1109/ICWS.2019.00047
Cavalieri, S., Garetti, M., Macchi, M., & Taisch, M. (2000). Experimental benchmarking
of two multi-agent architectures for production scheduling and control. Computers in
Industry, 43(2), 139–152. https://doi.org/10.1016/S0166-3615(00)00063-4
Chan, H. K., & Chung, S. H. (2013). Optimisation approaches for distributed scheduling
problems. International Journal of Production Research, 51(9), 2571–2577. https://
doi.org/10.1080/00207543.2012.755345
Cimino, C., Negri, E., & Fumagalli, L. (2019). Review of digital twin applications in
manufacturing. Computers in Industry, 113, Article 103130. https://doi.org/10.1016/
j.compind.2019.103130
Dorfeshan, Y., Tavakkoli-Moghaddam, R., Mousavi, S. M., & Vahedi-Nouri, B. (2020).
A new weighted distance-based approximation methodology for ow shop
scheduling group decisions under the interval-valued fuzzy processing time. Applied
Soft Computing, 91, Article 106248. https://doi.org/10.1016/j.asoc.2020.106248
Fang, Y., Peng, C., Lou, P., Zhou, Z., Hu, J., & Yan, J. (2019). Digital-twin-based job shop
scheduling toward smart manufacturing. IEEE Transactions on Industrial Informatics,
15(12), 6425–6435. https://doi.org/10.1109/TII.2019.2938572
Feng, S., Li, L., Cen, L., & Huang, J. (2003). Using MLP networks to design a production
scheduling system. Computers & Operations Research, 30(6), 821–832. https://doi.
org/10.1016/S0305-0548(02)00044-8
Fera, M., Greco, A., Caterino, M., Gerbino, S., Caputo, F., Macchiaroli, R., et al. (2020).
Towards digital twin implementation for assessing production line performance and
balancing †. Sensors, 20(97), 1–18.
Frantz´
en, M., Ng, A. H. C., & Moore, P. (2011). A simulation-based scheduling system for
real-time optimization and decision making support. Robotics and Computer-
Integrated Manufacturing, 27(4), 696–705. https://doi.org/10.1016/j.
rcim.2010.12.006
Frazzon, E. M., Agostino, ´
I. R. S., Broda, E., & Freitag, M. (2020). Manufacturing
networks in the era of digital production and operations: A socio-cyber-physical
perspective. Annual Reviews in Control, 49, 288–294. https://doi.org/10.1016/j.
arcontrol.2020.04.008
Fumagalli, L., Polenghi, A., Negri, E., & Roda, I. (2019). Framework for simulation
software selection. Journal of Simulation, 13(4), 286–303.
Gajate, A., Haber, R. E., Vega, P. I., & Alique, J. R. (2010). A transductive neuro-fuzzy
controller: Application to a drilling process. IEEE Transactions on Neural Networks, 21
(7), 1158–1167. https://doi.org/10.1109/TNN.2010.2050602
Güçdemir, H., & Selim, H. (2018). Integrating simulation modelling and multi criteria
decision making for customer focused scheduling in job shops. Simulation Modelling
Practice and Theory, 88, 17–31. https://doi.org/10.1016/j.simpat.2018.08.001.
Guerra, R. H., Quiza, R., Villalonga, A., Arenas, J., & Castano, F. (2019). Digital twin-
based optimization for ultraprecision motion systems with backlash and friction.
IEEE access : practical innovations, open solutions, 7, 93462–93472. https://doi.org/
10.1109/ACCESS.2019.2928141
Gurjanov, A. V., Zakoldaev, D. A., Shukalov, A. V., & Zharinov, I. O. (2019). Formation
principles of digital twins of cyber-physical systems in the smart factories of Industry
4.0. IOP Conference Series: Materials Science and Engineering, 483, 12070. https://doi.
org/10.1088/1757-899x/483/1/012070
Haber, R. E., & Alique, J. R. (2007). Fuzzy logic-based torque control system for milling
process optimization. IEEE Transactions on Systems, Man and Cybernetics Part C:
Applications and Reviews, 37(5), 941–950. https://doi.org/10.1109/
TSMCC.2007.900654
Haber, R. E., Alique, J. R., Alique, A., Hern´
andez, J., & Uribe-Etxebarria, R. (2003).
Embedded fuzzy-control system for machining processes: Results of a case study.
Computers in Industry, 50(3). https://doi.org/10.1016/S0166-3615(03)00022-8
Haber, R. E., Beruvides, G., Quiza, R., & Hernandez, A. (2017). A simple multi-objective
optimization based on the cross-entropy method. IEEE access : practical innovations,
open solutions, 5, 22272–22281. https://doi.org/10.1109/ACCESS.2017.2764047
Haber-Haber, R., Haber, R., Schmittdiel, M., & del Toro, R. M. (2007). A classic solution
for the control of a high-performance drilling process. International Journal of
Machine Tools and Manufacture, 47(15), 2290–2297. https://doi.org/10.1016/j.
ijmachtools.2007.06.007
Hamid, M., Nasiri, M. M., Werner, F., Sheikhahmadi, F., & Zhalechian, M. (2019).
Operating room scheduling by considering the decision-making styles of surgical
team members: A comprehensive approach. Computers & Operations Research, 108,
166–181. https://doi.org/10.1016/j.cor.2019.04.010
Han, J., Liu, Y., Luo, L., & Mao, M. (2020). Integrated production planning and
scheduling under uncertainty: A fuzzy bi-level decision-making approach.
Knowledge-Based Systems, 201-202, Article 106056. https://doi.org/10.1016/j.
knosys.2020.106056
Havard, V., Jeanne, B., Lacomblez, M., & Baudry, D. (2019). Digital twin and virtual
reality: A co-simulation environment for design and assessment of industrial
workstations. Production & Manufacturing Research, 7(1), 472–489. https://doi.org/
10.1080/21693277.2019.1660283
Iarovyi, S., Martinez Lastra, J. L., Haber, R., & Del Toro, R. (2015). From articial
cognitive systems and open architectures to cognitive manufacturing systems. In
Industrial Informatics (INDIN), 2015 IEEE 13th International Conference On (pp.
1225–1232).
Isermann, R. (2005). Model-based fault-detection and diagnosis - Status and applications.
Annual Reviews in Control, 29(1), 71–85. https://doi.org/10.1016/j.
arcontrol.2004.12.002
Ivanov, D., & Sokolov, B. (2012). The inter-disciplinary modelling of supply chains in the
context of collaborative multi-structural cyber-physical networks. Journal of
Manufacturing Technology Management, 23(8), 976–997. https://doi.org/10.1108/
17410381211276835
Jardine, A. K. S., Lin, D., & Banjevic, D. (2006). A review on machinery diagnostics and
prognostics implementing condition-based maintenance. Mechanical Systems and
Signal Processing, 20, 1483–1510. https://doi.org/10.1016/j.ymssp.2005.09.012
Jiang, Z., Jin, Y., Mingcheng, E., & Li, Q. (2017). Distributed dynamic scheduling for
cyber-physical production systems based on a multi-agent system. IEEE access :
practical innovations, open solutions, 6, 1855–1869. https://doi.org/10.1109/
ACCESS.2017.2780321
Jung, T., Shah, P., & Weyrich, M. (2018). Dynamic co-simulation of internet-of-things-
components using a multi-agent-system. Procedia CIRP, 72, 874–879. https://doi.
org/10.1016/j.procir.2018.03.084
Khan, S., Farnsworth, M., McWilliam, R., & Erkoyuncu, J. (2020). On the requirements of
digital twin-driven autonomous maintenance. Annual Reviews in Control, 50(June),
13–28. https://doi.org/10.1016/j.arcontrol.2020.08.003
Kritzinger, W., Karner, M., Traar, G., Henjes, J., & Sihn, W. (2018). Digital Twin in
manufacturing: A categorical literature review and classication. IFAC-PapersOnLine,
51(11), 1016–1022. https://doi.org/10.1016/j.ifacol.2018.08.474
La F´
e-Perdomo, I., Beruvides, . G.., Quiza, R., Haber, R., & Rivas, M. (2019). Automatic
selection of optimal parameters based on simple soft-computing methods : A case
study of micromilling processes. IEEE Transactions on Industrial Informatics, 15(2),
800–811. https://doi.org/10.1109/TII.2018.2816971
Leit˜
ao, P., Colombo, A. W., & Karnouskos, S. (2016). Industrial automation based on
cyber-physical systems technologies: Prototype implementations and challenges.
Computers in Industry, 81, 11–25. https://doi.org/10.1016/j.compind.2015.08.004
A. Villalonga et al.
Annual Reviews in Control xxx (xxxx) xxx
17
Li, K., Xiao, W., & Yang, S.lin. (2019). Scheduling uniform manufacturing resources via
the Internet: A review. Journal of Manufacturing Systems, 50(December 2018),
247–262. https://doi.org/10.1016/j.jmsy.2019.01.006
Long, J., Zheng, Z., & Gao, X. (2017). Dynamic scheduling in steelmaking-continuous
casting production for continuous caster breakdown. International Journal of
Production Research, 55(11), 3197–3216. https://doi.org/10.1080/
00207543.2016.1268277
Ma, Z., Yang, Z., Liu, S., & Wu, S. (2018). Optimized rescheduling of multiple production
lines for owshop production of reinforced precast concrete components. Automation
in Construction, 95, 86–97. https://doi.org/10.1016/j.autcon.2018.08.002
Morel, G., Pereira, C. E., & Nof, S. Y. (2019). Historical survey and emerging challenges
of manufacturing automation modeling and control: A systems architecting
perspective. Annual Reviews in Control, 47, 21–34. https://doi.org/10.1016/j.
arcontrol.2019.01.002
Napoleone, A., Macchi, M., & Pozzetti, A. (2020). A review on the characteristics of
cyber-physical systems for the future smart factories. Journal of Manufacturing
Systems, 54(December 2019), 305–335. https://doi.org/10.1016/j.
jmsy.2020.01.007
Nazarenko, A.A., .& Camarinha-Matos, L.M. (.2017). Towards collaborative Cyber-Physical
Systems. 12–17. https://doi.org/10.1109/YEF-ECE.2017.7935633.
Negri, E., Berardi, S., Fumagalli, L., & Macchi, M. (2020a). MES-integrated digital twin
frameworks. Journal of Manufacturing Systems, 56, 58–71. https://doi.org/10.1016/j.
jmsy.2020.05.007
Negri, E., Pandhare, V., Cattaneo, L., Singh, J., Macchi, M., & Lee, J. (2020b). Field-
synchronized Digital Twin framework for production scheduling with uncertainty.
Journal of Intelligent Manufacturing. https://doi.org/10.1007/s10845-020-01685-9
Panetto, H., Iung, B., Ivanov, D., Weichhart, G., & Wang, X. (2019). Challenges for the
cyber-physical manufacturing enterprises of the future. Annual Reviews in Control,
47, 200–213. https://doi.org/10.1016/j.arcontrol.2019.02.002
Pedrycz, W., & Kwak, K. (2007). The Development of Incremental Models. IEEE
Transactions on Fuzzy Systems, 15(3), 507–518. https://doi.org/10.1109/
TFUZZ.2006.889967
Poongothai, V., Kannan, M., & Godhandaraman, P. (2021). Performance analysis of a
single scheduling machine with cluster supply system, retardation, makespan and
deterrent protection using genetic algorithm. Materials Today: Proceedings. https://
doi.org/10.1016/j.matpr.2021.01.530
Qamsane, Y., Chen, C., Balta, E. C., Kao, B., Mohan, S., Moyne, J., et al. (2019). A unied
digital twin framework for real-time monitoring and evaluation of smart
manufacturing systems. In 2019 IEEE 15th international conference on automation
science and engineering (CASE) (pp. 1394–1401). https://doi.org/10.1109/
COASE.2019.8843269
Qi, Q., Zhao, D., Liao, T.W., .& Tao, F. (2018). Modeling of cyber-physical systems and
digital twin based on edge computing, fog computing and cloud computing towards smart
manufacturing. https://doi.org/10.1115/MSEC2018-6435.
Ramírez, M., Haber, R., Pe˜
na, V., & Rodríguez, I. (2004). Fuzzy control of a multiple
hearth furnace. Computers in Industry, 54(1). https://doi.org/10.1016/j.
compind.2003.05.001
Redelinghuys, A.J.H., Kruger, K., & Basson, A. (2020). A six-layer architecture for digital
twins with aggregation bt - Service Oriented, holonic and multi-agent manufacturing
systems for industry of the future (T., Borangiu, D., Trentesaux, & P. Leit˜
ao et al.,
(eds.); pp. 171–182). Springer International Publishing.
Redelinghuys, A., Basson, A., & Kruger, K. (2019). A six-layer digital twin architecture for a
manufacturing cell bt - Service Orientation in holonic and multi-agent manufacturing (T.
Borangiu, D. Trentesaux, A. Thomas, & S. Cavalieri (eds.); pp. 412–423). Springer
International Publishing.
Rosen, R., Von Wichert, G., Lo, G., & Bettenhausen, K. D. (2015). About the importance
of autonomy and digital twins for the future of manufacturing. IFAC-PapersOnLine,
48(3), 567–572. https://doi.org/10.1016/j.ifacol.2015.06.141
Shafto, M., Conroy, M., Doyle, R., Glaessgen, E., Kemp, C., LeMoigne, J., et al. (2012).
Modeling, simulation, information technology & processing roadmap. National
Aeronautics and Space Administration.
Stecken, J., Lenkenhoff, K., & Kuhlenk¨
otter, B. (2019). Classication method for an
automated linking of models in the co-simulation of production systems. Procedia
CIRP, 81, 104–109. https://doi.org/10.1016/j.procir.2019.03.019
Tabbussum, R., & Dar, A. Q. (2021). Comparison of fuzzy inference algorithms for stream
ow prediction. Neural Computing and Applications, 33(5), 1643–1653. https://doi.
org/10.1007/s00521-020-05098-w
Tao, F., & Zhang, M. (2017). Digital twin shop-oor: A new shop-oor paradigm towards
smart manufacturing. IEEE access : practical innovations, open solutions, 5,
20418–20427. https://doi.org/10.1109/ACCESS.2017.2756069
Tao, Fei, Cheng, J., Qi, Q., Zhang, M., Zhang, H., & Sui, F. (2018). Digital twin-driven
product design, manufacturing and service with big data. The International Journal of
Advanced Manufacturing Technology, 94(9), 3563–3576. https://doi.org/10.1007/
s00170-017-0233-1
Tao, Fei, Qi, Q., Wang, L., & Nee, A. Y. C. (2019). Digital twins and cyber–physical
systems toward smart manufacturing and industry 4.0: Correlation and comparison.
Engineering, 5(4), 653–661. https://doi.org/10.1016/j.eng.2019.01.014
Uhlemann, T. H.-J., Lehmann, C., & Steinhilper, R. (2017). The digital twin: realizing the
cyber-physical production system for industry 4.0. Procedia CIRP, 61, 335–340.
https://doi.org/10.1016/j.procir.2016.11.152
Valckenaers, P. (2019). ARTI reference architecture – PROSA revisited. Studies in
Computational Intelligence, 803, 1–19. https://doi.org/10.1007/978-3-030-03003-2_1
Villalonga, A., Beruvides, G., Castano, F., & Haber, R. E. (2020). Cloud-based industrial
cyber-physical system for data-driven reasoning: A review and use case on an
industry 4.0 pilot line. IEEE Transactions on Industrial Informatics, 16(9), 5975–5984.
https://doi.org/10.1109/TII.2020.2971057
Wang, X. V., & Wang, L. (2018). Digital twin-based WEEE recycling, recovery and
remanufacturing in the background of Industry 4.0. International Journal of
Production Research, 1–11. https://doi.org/10.1080/00207543.2018.1497819. 0(0).
Wolf, W. (2009). Cyber-physical Systems. Computer, 42(3), 88–89.
Zhang, H., Liu, Q., Chen, X., Zhang, D., & Leng, J. (2017). A digital twin-based approach
for designing and multi-objective optimization of hollow glass production line. IEEE
access : practical innovations, open solutions, 5, 26901–26911. https://doi.org/
10.1109/ACCESS.2017.2766453
A. Villalonga et al.
