Digital Twins A Survey On Enabling Technologies Challenges Trends and Future Prospects

IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 24, NO.
4, FOURTH QUARTER 2022 2255
Digital Twins: A Survey on Enabling Technologies,

Challenges, Trends and Future Prospects
Stefan Mihai, Mahnoor Yaqoob , Dang V. Hung , William Davis , Praveer Towakel, Mohsin Raza,
Mehmet Karamanoglu , Member, IEEE, Balbir Barn, Member, IEEE, Dattaprasad Shetve,
Raja V. Prasad , Member, IEEE, Hrishikesh Venkataraman, Member, IEEE,
Ramona Trestian , and Huan X. Nguyen , Senior Member, IEEE
Abstract—Digital Twin (DT) is an emerging technology sur- paper provides a deep insight into the technology, lists design
rounded by many promises, and potentials to reshape the future goals and objectives, highlights design challenges and limitations
of industries and society overall. A DT is a system-of-systems across industries, discusses research and commercial develop-
which goes far beyond the traditional computer-based simula- ments, provides its applications and use cases, offers case studies
tions and analysis. It is a replication of all the elements, processes, in industry, infrastructure and healthcare, lists main service
dynamics, and firmware of a physical system into a digital providers and stakeholders, and covers developments to date,
counterpart. The two systems (physical and digital) exist side as well as viable research dimensions for future developments
by side, sharing all the inputs and operations using real-time in DTs.
data communications and information transfer. With the incor-
poration of Internet of Things (IoT), Artificial Intelligence (AI), Index Terms—Digital twin, digital transformation, smart man-
3D models, next generation mobile communications (5G/6G), ufacturing, industry 4.0, structural health monitoring, 5G.
Augmented Reality (AR), Virtual Reality (VR), distributed com-
puting, Transfer Learning (TL), and electronic sensors, the
digital/virtual counterpart of the real-world system is able to pro- I. I NTRODUCTION
vide seamless monitoring, analysis, evaluation and predictions. HE FOURTH Industrial Revolution is in full bloom,
The DT offers a platform for the testing and analysing of
complex systems, which would be impossible in traditional sim-
ulations and modular evaluations. However, the development of
T and the current global Covid-19 pandemic has even fur-
ther accelerated the digital transformation by several years.
this technology faces many challenges including the complex- The travel restrictions, lockdowns, and pending economic
ities in effective communication and data accumulation, data decline have forced industry executives to adapt their business
unavailability to train Machine Learning (ML) models, lack of prospects and shift their focus from saving costs to increasing
processing power to support high fidelity twins, the high need for
interdisciplinary collaboration, and the absence of standardized investments in digital development [1]. Additionally, the global
development methodologies and validation measures. Being in the viral outbreak has imposed a dynamic uncertainty upon the
early stages of development, DTs lack sufficient documentation. economic world, and companies found themselves compelled
In this context, this survey paper aims to cover the important to cope with and quickly adapt to ever-changing conditions
aspects in realization of the technology. The key enabling tech- and restrictions in order to survive or even rise above the cir-
nologies, challenges and prospects of DTs are highlighted. The
cumstances [2]. However, even before the sanitary crisis, the
efforts towards digitalization were considerable. Cisco pub-
Manuscript received 30 September 2021; revised 7 April 2022 and
12 August 2022; accepted 20 September 2022. Date of publication lished their annual Internet report in the first fiscal quarter
22 September 2022; date of current version 22 November 2022. This of 2020, and they predicted a significant growth in world-
work was supported in part by the U.K.–India Education and Research wide Internet users (66% of the population in 2023, compared
Initiative (UKIERI) under Grant DST UKIERI-2018-19-011; in part by the
British Council’s Institutional Links through the Newton Programme Vietnam to 51% in 2018), networked devices (3.6 devices per person
Partnership under Grant 429715093; and in part by the British Council in 2023, as opposed to 2.4 in 2018), and reduced commu-
Indonesia’s Going Global Partnerships Programme (2022–2024). The fund in nication latency that encourages the development of real-time
the programmes delivered by the British Council was provided by the U.K.
Department of Business, Energy and Industrial Strategy. The work of Huan interactive applications [3]. This forecast expansion of Internet
X. Nguyen was supported by the VinGroup Innovation Fund (VinIF) under coverage, speed, and connections is giving way to an increased
Grant VINIF.2021.DA00192. (Corresponding author: Ramona Trestian.) rate of information dissemination, availability, and accessi-
Stefan Mihai, Mahnoor Yaqoob, William Davis, Praveer Towakel,
Mehmet Karamanoglu, Balbir Barn, Ramona Trestian, and Huan X. Nguyen bility, as well as growing opportunities of development and
are with the London Digital Twin Research Centre, Middlesex University, innovation.
NW4 4BT London, U.K. (e-mail: sm3488@live.mdx.ac.uk; my365@live.mdx. The goals of Industry 4.0 (I4.0) align perfectly with this
ac.uk; wd085@live.mdx.ac.uk; pt445@live.mdx.ac.uk; m.karamanoglu@mdx.
ac.uk; b.barn@mdx.ac.uk; r.trestian@mdx.ac.uk; h.nguyen@mdx.ac.uk). fast-paced and continuously-evolving digital transformation.
Dang V. Hung is with the Department of Civil Engineering, National I4.0 aims to automate all the traditional, bare-metal indus-
University of Civil Engineering, Hanoi 0000, Vietnam (e-mail: hungdv@nuce. trial practices, and it hopes to do so by bringing as much of
edu.vn).
Mohsin Raza is with the Department of Computer Science, Edge Hill the of the equipment from the physical space into the vir-
University, L39 4QP Ormskirk, U.K. (e-mail: mohsin.raza@edgehill.ac.uk). tual domain. And this is where Digital Twins come into play.
Dattaprasad Shetve, Raja V. Prasad, and Hrishikesh Venkataraman are with DTs emerged as an experimental technology set to enable
the Indian Institute of Information Technology, Sri City 517646, India (e-mail:
dattaprasad.s@iiits.in; yrv.prasad@iiits.in; hvraman@iiits.in). replication of elements, functions, operations and dynamics
Digital Object Identifier 10.1109/COMST.2022.3208773 of physical systems into digital world, with better control at
1553-877X
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: J.R.D. Tata Memorial Library Indian Institute of Science Bengaluru. Downloaded on August 19,2023 at 02:07:46 UTC from IEEE Xplore. Restrictions apply.
2256 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 24, NO. 4, FOURTH QUARTER 2022
Fig. 2. Interest trend in Digital Twin as seen through Google searches.
frameworks, and case-studies that have been thoroughly doc-

umented, and finally discuss and share the learned lessons and
remaining challenges of this technological pacesetter.
Fig. 1. The Digital Twin’s central role in the Industry 4.0 era.
A. Background and Motivation

The Digital Twin gained traction in 2002, after Michael
testing, analysis, prediction and hazard prevention for sen- Grieves held a presentation at the University of Michigan,
sitive processes. However, the supporting technologies, until which was entitled “Conceptual Ideal for Product Lifecycle
recently, were not advanced enough to develop DTs for com- Management”. The original slide containing the proposed con-
plex systems or systems-of-systems. The recent developments cept was reproduced by Grieves and Vickers in [4] and it can
in Machine Learning, Artificial Intelligence, data integration be noted that the early architecture of what would later become
Virtual/Augmented Reality, sensing, security, cloud storage, the DT consisted of three main components:
Transfer Learning, data visualization and ultra-reliable low • the real space,
latency communications (uRLLC) have enabled the implemen- • the virtual space,
tation of the DT and its extended applications across several • the link serving as a communication medium between the
industries. A technology thought to be capable of dealing with two spaces.
isolated operations and processes, the DT can now offer poten- The implications of this idea were revolutionary for the
tial applications eventually replicating the processes, elements, manufacturing industry, and other economic domains would
dynamics, firmware, connections and operations of physical later pick up on this as well. The most important advantage
systems in digital world. Figure 1 illustrates the DT as a of the original DT was the conjoined lifetimes of the real and
supplier of various services across industries in I4.0. virtual entities, starting from the creation of the pair, and end-
The formation of a mirror image of a physical system in ing in their disposal. This feature suggests that the virtual asset
the digital world offers unlimited possibilities. Interlinking would, at all times, mirror the most recent representative char-
the physical and digital systems through seamless data trans- acteristics of the physical system, allowing remote monitoring
fer allows the virtual system to exist simultaneously with the throughout the whole lifetime of the physical object. As such,
physical system. Real-time data communications between the while it was initially intended as a tool for monitoring the life-
physical and digital systems enable a synchronized and coher- cycle of a manufactured product, academia and the industry
ent operation of the real and virtual counterparts. Once in soon realised that the DT concept can be fruitfully applied to
the digital domain, optimized learning, information transfer, other economic domains as well.
analysis, visualization, optimization and planning can easily As a result of this realisation, research output has surged
be incorporated to see the potential improvements with sug- and interest and representation of the digital twin has grown.
gested changes. Thus, the DT can be used effectively to assess, Figure 2 shows the global trend of interest into the term “dig-
observe and validate the physical system, suggest changes and ital twin” as expressed via the normalized number of Google
visualize the potential improvements. Previously deemed as an searches across the last 11 years, and it can be noted that the
impossibility, DTs have arisen as one of the key technologies last few years have drawn increased attention to the concept
with potential to reshape the future. that Michael Grieves introduced back in 2002. The “interest”
This survey paper aims to shine light on the recent advance- shown on the Y-axis takes values between 0 and 100, where
ments in the DT paradigm by reviewing the relevant literature a value of 100 denotes the highest popularity the search term
published over the last few years, in order to establish a has seen during one month of the given time window, and a
common understanding of this technology, explore its market value of 50 represents half of that maximum popularity.
potential and trends, enumerate its most prominent enabling While the Google Search trend is a representation of the
technologies, offer an in-depth look at several applications, DT’s increasing general popularity across the years, it is
MIHAI et al.: DTs: A SURVEY ON ENABLING TECHNOLOGIES, CHALLENGES, TRENDS AND FUTURE PROSPECTS 2257
its publishing. The authors set out to answer three research

questions, regarding the various definitions that the DT has
accumulated across the years, the paradigm’s main charac-
teristics (feature selection and extraction, that facilitate the
DT-characteristic Big Data analysis, pattern recognition and
ML, predictive and prescriptive analytics, etc.), as well as
the main domains where the DT had been applied at the
time of writing. In particular, the authors detail several DT
implementations across three different application domains:
manufacturing, healthcare, and aviation. The paper ends with
the design implications and challenges that developers should
take into account when considering such a system. In contrast,
this work’s goal is to expand Barricelli et al.’s contributions by
Fig. 3. Number of articles in the Scopus database whose abstracts contain additionally discussing the market potential of the DT, going
the phrase “digital twin” across the years.
into further details on the technologies that enable a DT’s
characteristics, and also providing a closer look at possible
use-cases in different industry domains.
important to assess and understand how this growing attention
Minerva et al. [7] proposed a comprehensive survey on
has been reflected in the research output throughout the same
the architectural models of a DT, and they also discussed
time period. In this regard, we analysed the number of pub-
the technical features of this concept (or, in other words, the
lications within the Scopus document database that included
technical “must-haves”) that consolidate the DT definition.
the phrase “digital twin”, for each year in the last decade.
The paper surveys different DT characteristics that were high-
Figure 3 shows the results of this analysis and, indeed, it can
lighted in literature pertaining to various technological and
be noted that increasing attention has been allocated to the
industrial domains, like manufacturing, AR/VR, multiagent
research and development of this revolutionary paradigm as
systems, virtualisation, and especially IoT. Additionally, the
well. The bar plot shown in Figure 3 is the result of leverag-
authors applied proactive thinking and evidentiated some other
ing the Scopus Search API to search for articles that contain
important characteristics of the DT that are often overlooked in
the specific phrase “digital twin” in their abstracts. While not
other research works, like data ownership, contextualisation,
all of the found articles might have been devoted entirely to
augmentation, servitisation, etc. The paper covers the value
the DT concept, it is still noteworthy that more and more aca-
of the DT concept, including its market potential, before div-
demics have taken some level of interest in the technology
ing into various detailed use-cases, like the digital city and
over the years. It is also important to mention that the search
the digital patient. The survey ends with a consolidated DT
was accomplished via the “pybliometrics” package of Python,
architectural model and illuminates the upcoming challenges.
developed by Rose and Kitchin [5].
While the article proposed by Minerva et al. contours the defi-
nitions, enabling technologies and applications of a DT, it does
B. Related Surveys not go into details about which technologies can be integrated
There is currently a considerable number of publications into the DT in order to build those applications. The survey
across literature that are dedicated to advance the concept of a at hand will attempt to fill that gap.
DT. In fact, there are so many articles that academics have also Löcklin et al. [8] published a survey paper that tackles the
put out a number of survey papers that were designed to review DT’s use for verification and validation (V&V) purposes. The
the state-of-the-art in DT development, illuminate fellow inno- authors mention that there are multiple DT definitions in the
vators on possible research gaps, questions and directions, literature, before settling on one definition that makes a dis-
as well as guide the industry towards possible DT use cases tinction between the Digital Twin, which is mainly referred to
that might generate significant business value in their specific as a tool for monitoring, verification and validation, and the
domain. This work aims to complement the other existing sur- Intelligent Digital Twin, that can provide meaningful feed-
vey efforts and facilitate a complete understanding of the DT, back to its corresponding asset based on the acquired data.
with its definitions, enabling technologies, applications, and Furthermore, the survey is conducted around three research
detailed use-cases and case-studies. questions that tackle how the DT can enable verification and
As such, Table I displays a comparison between a collection validation, the industrial domains where the DT is used for
of other survey articles covering comprehensive DT-related lit- V&V, and a classification of modalities in which the DT is
erature, and the survey paper at hand. Table I puts in contrast leveraged in the reviewed papers. As the title of the survey sug-
the contributions of this paper with other surveys’, in order gests, it is dedicated to studying the application of verification
to highlight the novelty brought by the work at hand. The and validation within multiple industrial domains.
findings of this study are analysed below in more detail. Biesinger et al. [9] conducted a survey on the necessity
Barricelli et al. [6] propose a study on the DT’s defini- of a DT in the automotive industry’s integration planning
tions, characteristics, applications, and design implications. processes. Unlike the previously presented articles, which
The paper provides full coverage of the DT’s evolution, from focused on reviewing scholarly works, this study has been con-
Michael Grieves’ concept to the state-of-the-art at the time of ducted by interviewing 22 production planners from various
TABLE I
C OMPARISON B ETWEEN T HIS S URVEY PAPER ’ S C ONTRIBUTIONS AND OTHER R ELATED W ORKS IN THE L ITERATURE
automotive manufacturing companies. The scope of the paper, collaborative robot. The manuscript at hand aims to comple-
which is a case-study by itself, covers the demand of an eas- ment this work with a wider array of references and case
ily accessible and configurable DT for the specific use-case studies.
of integration planning. As such, it is a great indicator of the Rasheed et al. [12] created a comprehensive DT sur-
market potential that the DT has in this industrial sector, but vey that analyses the paradigm from the perspectives of its
it does not delve into the possible solutions or enabling tech- value (expressed via services and software platforms), applica-
nologies that could help implement a DT that is compliant tions, enabling technologies, and challenges. The paper brings
with the definitions that the authors provide in their work. together works from various domains where the DT has been
He et al. [10] proposed a survey focused mostly on the tried and tested, as well as potential socio-economic impacts
monitoring and surveillance capabilities of the DT, and their that such a technology can have (i.e., loss of jobs, training
provided definition (“a dynamic digital replica of physical new workforce in DT specifics).
assets, processes, and systems, which comprehensively mon- In [13], Tao et al. introduce a thorough analysis of 50 papers
itors their whole lifecycle”) reflects this aspect of the study. and 8 patents related to the DT. The authors cover the concept
Although the authors do not emphasise the full extent of the of DT via a study of four perspectives: interaction and collab-
DT definition, they do present the technologies that enable oration between its constituents, data fusion, services, and DT
the DT to become an avant-garde methodology for the spe- modeling and simulation. Then, the authors describe the DT’s
cific application of surveillance. The survey concludes with applicability in three main areas in its most prominent industry,
an industrial use-case of the DT, dubbed Pavatar, and presents the manufacturing sector: DTs for product design, for produc-
the technical advantages and challenges of implementing this tion, and for prognostics and health management. As such, the
project for an ultra-high voltage converter station. paper draws its lessons from articles that are mostly focused
Pires et al. [11] present a compact review of the DT’s def- on the state of the art of DTs in the manufacturing industry.
initions, enabling technologies, and applications, ending the In [14], Huang et al. put together a detailed survey on
paper with the exemplification of a case-study of the DT and the AI-driven DTs in the context of smart manufacturing
the challenges the concept will face before the industry inte- and advanced robotics. The authors take into consideration
grates it into its businesses. The authors also provide a look the advantages of using a DT as a driver for sustainabil-
at on-going research efforts towards building a DT for a UR3 ity goals in the manufacturing and robotics, by facilitating
production planning and control, quality control, dynamics II. D EFINITIONS OF THE D IGITAL T WIN
control, predictive maintenance, and many other services. The Digital Twin is not a new paradigm. The premises that
Additionally, the authors focus on describing how AI tech- initiated its advancements have been introduced more than 50
niques specifically enable DTs across these two domains. years ago, amongst NASA’s many efforts of bringing man in
Finally, Fuller et al. in [15] propose an extensive survey on space. Indeed, the idea of virtually simulating real-life sce-
DT, with a focus on its integration with IoT and data analytics narios that would normally require extensive resources could
technologies. The paper also points out that there is a need for not have had a more appropriate origin than NASA’s early
a stable DT definition that consolidates all of the aspects that space programme, where high-fidelity (relative to the tech-
make up a true DT. Afterwards, the survey goes on to discuss nology standards at the time) simulators were used to train
the DT, its enabling technologies, and applications in three astronauts for their upcoming remote journeys in the outer
main use cases: smart city, healthcare, and manufacturing. The space. But, surely, there is more to DTs than just simula-
manuscript at hand aims to complement the aforementioned tion, so the true precursor of this paradigm actually came to
articles in providing a complete view of the DT. light during NASA’s Apollo 13 [16] mission, when an unex-
pected explosion caused a manned spacecraft to deviate from
C. Survey Contributions
its intended trajectory, endangering the astronauts on-board. In
To complement the previous works, this work provides a response, the ground-based Mission Control was then tasked
comprehensive survey on the DT concept, the enabling tech- to urgently simulate, in almost real-time, the erratic behaviour
nologies involved, the applications and use cases for deploying of the spacecraft, and make optimal decisions to ensure its
DTs across various industries. The main contributions of this safe return on Earth, in ever-critical conditions. The engi-
survey paper are summarized as follows: neers used the available spacecraft simulators, animated them
• Overview of the DT definitions from the literature; with real data coming from the space-bound physical ship
• Comprehensive discussions on the market potential of and its pilots, analysed possible scenarios, then communicated
DT; optimal instructions to the stranded pilots to maneuver their
• The enabling technologies for DT are surveyed, such as: ship back home safely. The mission was a success.
ML, cloud, fog and edge computing, IoT/IIoT, Cyber- Michael Grieves, in 2002, proposed a similar idea as a
Physical Systems, VR/AR, and modeling technologies; means to drive forward the Product Lifecycle Management
• Existing solutions of DT frameworks are reviewed across paradigm (PLM), although, back then, he had not dubbed it
three use cases examples, namely: smart factory, infras- as “digital twin”, but as “Mirrored Spaces Model”. This pre-
tructure, and future directions for 6th Generation Mobile cursor of the DT consisted of the same three main pillars that
Networks (6G). Then, we take a closer look at two DT lie at the base of this technology today: real space, virtual
services, irrespective of use case: anomaly detection and space, and the communication thread between them.
predictive maintenance; In the case of the Apollo incident, the real space was repre-
• Three real use cases of DTs as applied to tea indus- sented by the physical spacecraft stranded in space, the virtual
try in India, Festo Cyber-Physical Factory in the United space consisted of the ground-based simulators, and the link
Kingdom, and structural health monitoring for Vietnam between the two was characterised by the continuous com-
bridges are discussed in details; munication between the Mission Control, the spacecraft, the
• Lessons learned, remaining open challenges and future engineers, and the pilots. It was indeed the DT that saved the
directions of DTs are identified. day back in 1970.
The actual definition of the DT has always been at least
D. Survey Structure ambiguous, and the ever-growing number of publications in
This survey paper is organized as follows: Section II pro- the last few years has only added to the diversity of DT
vides a review and comparison of the existing DT definitions meanings. It is generally observed, however, that the DT
across the recent literature covered in this paper, as well as implementation attempts so far seem to have been accom-
our own comprehensive DT interpretation, Section III eval- plished with the support of a number of common enabling
uates the DT’s potential for market adoption and current technologies, such as: ML [17], TL [18], distributed comput-
trends in DT development, Section IV delves into some of the ing (i.e., cloud, fog, and edge computing) [19], the (Industrial)
most prominent DT enabling technologies and reviews how Internet of Things (I/IoT) [20], CPS [21], and VR/AR [22].
other researchers used them to build working DTs, Section V Consequently, to get a good grasp on the general under-
explores various DT frameworks and applications proposed in standing of the DT we analysed the recent literature on the
the literature, Section VI takes a closer look at three DT case DT definitions provided. The results indicate that the exist-
studies, Section VII draws the lessons we learned through- ing DT definitions seem to center around five approaches as
out this survey, establishes the current challenges that the summarised in Table III.
DT faces, and contours possible future directions, and finally Each of the five definitions capture the essence of the DT,
Section VIII summarizes the conclusions of this paper. but there is not complete overlapping between them:
An overview of the organization and structure of this • The first definition is a very popular one, and arguably
paper is illustrated in Figure 4. Table II provides a list of the oldest one, however it is very generic and it does not
abbreviations used in this article. offer insight into the constituent parts of the DT.
TABLE II
L IST OF ACRONYMS
•The second definition is more comprehensive, as it an autonomous car, where the virtual representation of the
demystifies the DT and provides a bit of understand- car is able to continuously communicate with the physical
ing regarding two of its components (the physical and asset and become aware at all times of its state and envi-
digital objects). However, it claims that the DT is just ronment in order to make appropriate control decisions. The
an intelligent digital model of a physical asset, with lit- use-case of autonomous driving imposes some strict require-
tle to no emphasis on the interaction between the two, ments on the DT, the most obvious of which are: ultra-low
its requirements, and limitations. Some works [23], [24], latency communication between the real and virtual twins,
however, do make observations about the twinning rate large data storage capacity, high processing power to reduce
(i.e., the rate of synchronization between the two objects), data-to-insight delays, and high-fidelity virtual rendering of
and mention that it is a requirement that depends on the the car and its environment.
DT’s use case. Now, seeing this example through the lenses of the five def-
• The third definition completely ignores the bi-directional initions, we would find that some of the above scenario’s very
communication requirement of a DT, essentially confus- important aspects are omitted. The first and third definitions
ing the DT with a digital shadow, which would more claim that the DT only mirrors the life of its twin, so they
accurately fit that description. do not envision the other half of the feedback loop between
• The fourth definition focuses entirely on the components the two entities, where the DT itself can control the car based
of the DT, but it does not hint towards the capabilities of on its real-time data. The second definition comes close to
a DT, making the definition too generic. describing our DT, except that it makes no mention of the DT’s
• The fifth definition shifts its attention to the services pro- requirements that might differ from use-case to use-case: a DT
vided by the DT, but not on its structure and technologies of an autonomous car will need a higher rate of synchroniza-
that enable said services. tion than the DT of an industrial water boiler, for instance. The
To put the five definitions into a better perspective, we will fourth definition makes no mention of the DT’s use-case at all,
work through a hypothetical example. Consider the DT of even though it is a deciding factor in the choice of enabling
Fig. 4. An overview of the survey paper’s contents.
TABLE III
T HE VARIOUS D EFINITIONS OF D IGITAL T WINS F OUND IN THE L ITERATURE
technologies and other requirements (a DT for lifetime moni- definition focuses entirely on the use-case and the services
toring of a car would not need a bi-directional communication the DT can provide, but the use-case’s requirements and
medium, unlike a DT for autonomous driving). Lastly, the fifth the DT components are only implied via the definitions
of the services themselves, and not explicitly stated as DT As such, a DT’s services should enable it to be self-adapting,
characteristics. self-regulating, self-monitoring, and self-diagnosing, or, in
It is thus clear that there is a need for a comprehensive fewer words, self-evolving.
definition of the DT. The definition needs to be specific enough
that it identifies the components of a DT, how they interact,
and what services it should provide, but also generic enough III. M ARKET P OTENTIALS AND T RENDS
that it can describe DTs across multiple industrial domains Academic research on the topic of DT has positioned it thus
and use cases. far as a central player in the race to I4.0. The widely-praised
As such, in this work, and many hereafter, we will refer potential of the paradigm presents attractive opportunities and
to the Digital Twin as a self-adapting, self-regulating, self- challenges that, once overcome, promise to bring benefits that
monitoring, and self-diagnosing system-of-systems with the far outweigh the costs. However, the issue of costs is more
following properties: (1) it is characterized by a symbiotic prominent in the industry than it is in the academia. While the
relationship between a physical entity and its virtual represen- DT is a versatile technology that can be successfully applied to
tation, (2) its fidelity, rate of synchronization, and choice of various domains businesses still remain reluctant to implement
enabling technologies are tailored to its envisioned use cases, the DT. That is because the Return On Investment (ROI) on a
and (3) it supports services that add operational and business DT is difficult to quantify, since it is not a product that directly
value to the physical entity. brings revenue, but rather a technology that aims to reduce and
We believe that this alternative provides a better understand- optimize costs. Nevertheless, this section will detail how the
ing of the DT concept, as it gives a precise indication of what leaders in the DT market have leveraged its aptitudes, and
the DT is (i.e., a system-of-systems), what its components are whether their results predict a good omen for the future of the
(i.e., physical and digital entities), how they interact (symbiotic DT.
relationship, i.e., a mutually beneficial two-way interaction), IBM [60], one of the top DT solutions providers, valued
how their interaction is leveraged (i.e., to offer services that the DT market at USD 3.1 billion in 2020 [61], and pre-
bring operational and business value), how accurate and syn- dicted that the technology would see significant adoption and
chronized the virtual asset should be, and what technologies economic growth in the following years. Although they do
should be used to build it (i.e., use case-dependent). Looking admit that the creation of a DT is not always a sound finan-
back on the DT example for autonomous driving, it is now cial investment, IBM’s case studies show encouraging returns
clear to see which are the DT’s components (the car and its and cost optimisations for DT implementations in manufac-
environment are the physical asset, while their virtual model turing and smart buildings. For example, ASTRI [62] used
is the digital asset), that they interact through a bi-directional DTs to validate software packages before their deployment
communication medium (one for collecting data from the real on the physical twin, reducing 30% of development costs and
space, and another for delivering insights and commands from expediting deployment by 40%. University of California San
the virtual twin), that the DT is supposed to, in our case, assist Francisco [63] implemented DTs for a branch of the Mission
the car in driving autonomously, and that this use-case implies Bay Hospital which helped engineers to reduce the diagnosis
some specific restrictions (low-latency, high security, etc.). and repair process of the building’s pipes from 2-3 days to
The fundamental characteristics of a DT, which are the just a few hours.
motors that actually bring operational and business value to Another big player on the DT market is Ansys [64], who
the physical entity, lie centrally in the DT philosophy. The praise the DT as the bridge between equipment development
“self-X” constructs distinguish a true DT from digital models and equipment operations, allowing manufacturers to monitor
and shadows, and emphasize the usefulness of a DT in I4.0. the behaviour of their product throughout its whole lifetime.
These traits are explained below: Mecuris [65] have used Ansys products to develop a DT for
• Self-adapting - a DT automatically reacts to changes tailor-made orthoses and prostheses development to reduce
in its real twin’s environment and configuration, but it product testing costs. Similarly, Jet Towers [66] implemented
should do so in a way that continually ensures operational DTs of modular wireless towers to reduce installation and
excellence (i.e., as measured via use case-appropriate design time by 80%.
performance measures). General Electric [67] lead the DT market in the power
• Self-regulating - the changes a DT undergoes while adapt- systems industry, where their solution claims to reduce start-
ing to its real twin’s environment should not exceed up time by 50% and maintenance costs by 10%, deliver up
the physical twin’s own limitations for the sake of to $5 million additional MWhr, and save costs on outages
maximising its performance measures (e.g., productivity, of up to $150 million per year. In the telecommunications
throughput, etc.). industry, Spirent [68] are taking the stage as the leading
• Self-monitoring - the DT is always aware of its real twin’s DT solutions provider for 5G networks. To optimise network
environment and configuration, by means of monitoring design, testing, and deployment, Spirent propose leveraging
the parameters that are relevant to its use cases. the DT for use cases such as: cellular Vehicle-to-Everything
• Self-diagnosing - the DT should be able to assess its own (V2X) virtual drive testing, private 5G networks for smart fac-
health and know, based on its current and historical con- tories, and testing and design for Communications Service
ditions, when and why it is no longer able to maintain Providers. Another Spirent report [69] praises the DT for
optimal operations. its Predictive Maintenance (PdM) aptitudes in I4.0’s smart
factories, claiming to reduce machine breakdown by up to or maximise a given process parameter. For example, in [39],
70% and save up to 25% of maintenance costs [70]. where the DT of a Mobile Edge Computing system had been
A Markets and Markets report [71] highlights attractive implemented, a Deep Neural Network (DNN) was used to
growth opportunities in the DT market and stresses on the maximize energy consumption efficiency across the network
increasing need for PdM and business optimisation. The report based on features like user association and resource alloca-
also states that the market in North America recorded as high tion. Similarly, the authors in [40] used Genetic Algorithms
as $1.32 billion, the largest share of the overall DT market (GA) to predict the circumstances that would favour most
in 2019, and it is projected to reach $16.94 billion by 2026. devastating forest fires, such that they could be proactively
The paper forecasts that the global DT economy is expected to prevented. In [55], four ML algorithms (Random Forest,
grow from $3.2 billion in 2020 to $48.27 billion by 2026 with AdaBoost, LightGBM, and XGBoost) were able to learn from
a Compound Annual Growth Rate of 58%. Another report by the equipment’s sensor data and optimise production yield in
the Institution of Engineering and Technology [72] echos the the petrochemical industry. The same approach is especially
idea that the DT market is on the rise, although it also men- common in literature focused on RL, where the algorithm
tions that industry-agnostic adoption at the time of publication learns by trying to maximise a reward function. The reward
amounted to only 5% of enterprises. Nevertheless, the authors that these models attempt to maximise could be product
are optimistic that the DT should pave the way to I4.0. quality outcome [23] or other arbitrary mathematical reward
functions [58].
IV. D IGITAL T WIN : E NABLING T ECHNOLOGIES Besides objective function optimisation, another applica-
tion of ML within DTs is to make predictions about the
Although it is often referred to as a piece of technology in
future behaviour of the physical asset. In this context, Artificial
and of itself, the DT can be more accurately thought of as
Neural Networks (ANNs) have been used in [30] to predict
a system-of-systems, a meld of several enabling technologies
future samples of the active power component based on histor-
that construct an intelligent virtual representation of a phys-
ical time series data. However, ML models, and in particular
ical entity and support a continuous two-way feedback loop
DL methods, are generally perceived as black-boxes [81]. This
between the twins. At the same time, the enabling technolo-
is because they do not offer sufficient transparency into what
gies themselves can take many forms depending on the DT’s
is motivating their predictions. On the other hand, in DT appli-
use case. For example, although it is known that a communi-
cations, transparency is desirable, and often required (e.g., in
cation medium is needed between the real and digital twins,
applications such as fault identification and urgency classifica-
the choice of the specific communication protocol is entirely
tion [41], or anomaly detection and root cause analysis [82]).
dependent on the communication requirements of the DT’s
Thus, researchers have looked for ways of integrating both
application. These being said, this section will explore the
physics-based and data-driven models into the DT. On this
most common enabling technologies of the DT and provide
note, the authors in [42] combined these two approaches to
insight into how researchers from various industries have cho-
enable a prognosis service that predicts the future parameters
sen algorithms and frameworks that were fitting to their use
of a physical asset, even though said parameters evolved at dif-
cases.
ferent time scales. As such, physics-based models were used to
preprocess data coming from multiple time-scale data streams,
A. Machine Learning while the ML models, Mixture of Experts and Gaussian
One of the advantages of DT is that it brings awareness (or Processes (ME-GP), combined the extracted information to
intelligence, or understanding) to a physical asset that would predict future bahaviours of each time series parameter.
otherwise lack it. Of course, we are not referring to the human Another use case of ML in DT implementations is appli-
understanding of “awareness” [73], but rather a new kind of cation security. Indeed, the authors in [56] proposed a DT
specialized intelligence that is able to understand significant for remote surgery services, and used neural networks to pro-
amounts of numerical data and draw domain-specific conclu- tect the crucial connection between the physical and virtual
sions from it faster than a human expert could. Thus, the DT worlds, which has stringent requirements for availability and
should be able to infer meaningful and actionable information latency, by detecting and preventing Denial of Service (DoS)
from the data that is generated by its physical twin and its attacks. Whereas in other works, ML is used to bring intel-
environment. In this scenario, ML techniques represent the ligence to the virtual representation of a physical asset, the
foundation, or the brain, of a DT. study conducted in [43] proposed Deep Convolutional Neural
Across the reviewed literature, researchers have employed Networks (DCNNs) to create the digital replicas of fibrous
the whole array of ML algorithms types in their DT implemen- materials from real and synthetic images, and this constitutes
tations: traditional ML [74], Deep Learning (DL) [75], super- an example of ML being used to build DTs.
vised ML [76], unsupervised ML [77], classification ML [78], On the note of synthetic data creation, DTs can also be
regression ML [79], Reinforcement Learning (RL) [80], etc. used to generate artificial training data for ML models [76].
The applications and choices of ML models encountered in Of course, for such models to be able to generalise well
this review are various and they depend on the use cases and on real data, the distribution of the synthetic data they were
services of the proposed DTs. However, since ML algorithms trained with has to closely resemble the distribution of the real
are ultimately used to solve optimization problems, one com- data from the test set [31]. This requires extremely accurate,
mon approach is to employ data-driven models to minimise and thus complex, simulation engines. A workaround for this
TABLE IV
OVERVIEW OF R EVIEWED L ITERATURE I NTEGRATING ML I NTO D IGITAL T WIN
challenge is offered by TL, where the ML model trained with analytics imposes some challenges: (1) traditional ML mod-
artificial data can be adapted to perform good predictions on a els are actually built upon a sequence of carefully engineered
real test set. TL requires that the distributions of artificial and functional blocks that are tailored to increase efficiency on
real data be somewhat similar, such that only a small amount the task at hand. Thus, designing such a pipeline can be a
of real data is needed to make the model generalise well once resource-expensive task; (2) on the other hand, end-to-end DL
deployed in production [44]. techniques remove some of that complexity, but they require
Another important use case of ML in DT implementa- significant amounts of data for training and tuning, while also
tions is remote control assistance. In scenarios such as remote offering no transparency into their predictions; (3) finally, on
surgery [83], [84] or space station maintenance [85], the DT the note of training data, generative ML models can be used
can help bridge the distance between the operators and the to create artificial data to compensate for the lack of real data.
physical twin. In the latter paper in particular, the authors However this could also induce bias in predictions if the two
proposed the Hierarchical Attention Single-Shot Detector distributions (i.e., real and synthetic) are not aligned.
Network (HA-SSD) for astronaut gesture recognition. The
system is based on the popular MobileNet architecture [86]
for fast and computationally inexpensive feature extractions, B. Cloud, Fog, and Edge Computing
which could be easily deployed on chips with low processing Depending on its use-case, DT can be used to mirror
power. Such a system is ideal for space station DTs where systems across the whole spectrum of complexity, from uni-
cameras and surveillance equipment can detect and monitor tary elements, such as the movement axis of the Computerized
faces, human postures, gestures and body language. Numerical Control (CNC) machine tool [87], to an entire fleet
Table IV summarizes the findings of integrating ML of aircraft [88]. The virtualisation of composite heterogeneous
into DT. It can be noted that most of the efforts towards machines or services always requires heavy computational
DT implementation come from the manufacturing industry. prowess. This demand, coupled with the DT’s characteristic of
Concurrently, the versatility of the technology makes it a real-virtual synchronisation, often calling for almost real-time
promising tool for other domains as well, such as civil engi- responsiveness, illuminates a need for distributed and paral-
neering or robotics. It is actually the vast array of existing and lel computing. For this purpose and many others, cloud, fog,
up-and-coming ML and DL algorithms that grants the DT its and edge computing are frequently encountered in DT-related
versatility. At the same time, the DT’s reliance on data-driven literature. As such, this section will review how researchers
TABLE V
OVERVIEW OF R EVIEWED L ITERATURE I NTEGRATING D ISTRIBUTED C OMPUTING I NTO D IGITAL T WIN
have integrated distributed computing into their DT imple- formed by three functional blocks: the physical asset, its vir-
mentations, with an emphasis on the reason why this enabling tual counterpart, and the communication medium that binds
technology was mandatory for the works’ use-cases. them together. Any DT proposal that misses or does not envi-
Table V highlights an overview of the papers reviewed sion the inclusion of any one of these three components is
in this section. It is noteworthy that our search indicated therefore not a real DT. In this section, the focus will fall on
that distributed computing and DT have been conjoined in the role of the two-way connection between the digital and
works mostly pertaining to the manufacturing sector. Even real twins, and the recent literature works that detail how this
papers focusing on DTs for logistics are mainly created to connection benefits the implementation of the DT. Table VI
help the logistics departments of the manufacturing industry, provides an overview of the articles discussed in this section,
again showing that there is an overwhelming focus on DT as well as their target domain and main contributions.
development in this area. In this context, the Internet of Things and the Industrial
The unison of cloud computing and DT creates a pros- Internet of Things are the main enabling technologies that
perous environment for complex simulations, multi-variable can perfectly fit in-between the digital and real twins to
analysis, DL-based analytics, and Big Data storage. In this converge virtuality and reality. These paradigms are at the
type of system, the cloud platform acts as the data ware- center of I4.0 due to their ability to aggregate data from
house and also provides heavy-processing capabilities, while multiple heterogeneous data sources via disparate communi-
the DT deals with synchronising the physical and virtual cation mediums to facilitate data mining and alaytics through
assets [25], [32], [45]. Additionally, a cloud platform allows distributed computing frameworks.
the harmonious connection and hosting of the virtual counter- The main appeal of these technologies is represented by
parts of the heterogeneous subsystems that form a complex the IIoT devices, like smart sensors, RFID tags, and smart
DT [33]. In the healthcare industry, the cloud represents wearables, that are useful and cheap data sources [35] which
a shared information platform between the medical service can paint meaningful virtual reflections of reality that the
provider and the patients [26], while in the manufacturing cloud-based DT can interpret and analyse in order to reduce
domain it can serve as a common medium where enterprises manufacturing uncertainty and complexity in fixed-position
can share data regarding the failure modes and maintenance assembly islands [57], optimise the functioning of power
needs of similar equipment to support DT-enabled Predictive equipment switchgear [24], provide PdM for automotive brake
Maintenance [89]. pads [46], and visualise in real time the stress endured by metal
In applications that imply a great amount of data consump- shelving brackets via Augmented Reality [47]. Other use-cases
tion and processing, even the cloud can become overwhelmed. benefit from these technologies as well, since smart sensors
To prevent this from happening, Hu et al. [90] reduce the and wearables can also be integrated in mobile equipment
cloud workload by using the MTConnect protocol and a that people carry, urban infrastructure, and interior appliances.
new Knowledge Resource Centre to manage all commu- The richness of heterogeneous data that the IoT sensors bring
nications with the cloud-hosted DT. Another approach to can be used to virtualise and visualise cities, allowing struc-
avoid over-dependence on the cloud is to use more forms tural simulations for hazards prevention [48], and to remotely
of distributed computing (i.e., cloud, fog, and edge) to man- manage safety issues in the workplace [28].
age different layers of complex logistics and manufacturing Concurrently, the IoT/IIoT brings more than just data to the
systems [27], [34]. DT system. In [91], the DT is built upon the traditional IoT
framework, and it is split into two parts: one at the edge, and
C. Internet of Things one in the cloud. Both essentially working also as the gate-
As previously mentioned, the research communities of ways that connect the two corresponding media of the IoT
academia and the industry recognise the DT as a system framework. Furthermore, while the IoT is a powerful enabling
TABLE VI
OVERVIEW OF R EVIEWED L ITERATURE I NTEGRATING I OT/II OT I NTO D IGITAL T WIN
technology of the DT, the DT itself can also act as a sup- like rarely-used manufacturing equipment, this requirement is
porting pillar for the IoT by providing a self-adaptive and not so stringent.
self-integrating digital abstraction of the IoT devices to make
the IoT framework resilient to dynamic changes [49], or by
allowing virtual simulations of large sensor networks [50]. In D. Cyber–Physical Systems
the context of IIoT, equipping edge devices with ML solutions With the advancement towards the digitisation of conven-
can be a challenging task due to the limited resources on the tional physical systems, the term Cyber-Physical Systems
devices as well as the concerns about communications with the has gained ample attention from the academia and industry.
cloud (low latencies, raw data privacy). As such, the authors However, in the literature works surveyed in this paper, we
in [93] proposed building a DT of the Edge Network that was have found that there are at least two definitions that fall under
able to leverage Federated Learning (FL) to re-train aggre- the span of the CPS abbreviation, and while they are quite
gated models locally, on edge devices (thus avoiding raw data similar in meaning, they still represent different concepts.
transmissions), as well as optimize communication efficiency In the original vision of CPS, they represent the ubiqui-
using the DT’s updated mirroring of the network. For addi- tous and holistic convergence between real complex systems
tionally improved communication security for the previously of heterogeneous systems and their virtual intelligent control
reviewed framework, the authors in [94] proposed using the instances. The physical space is represented by an ecosys-
blockchain technology to store the aggregated model parame- tem of physical equipment, sensors, actuators, and human
ters of the edge devices on the Base Station, making the FL operators that labor together towards the same goal. The
process even more robust to data privacy issues. cyber elements are the virtual representations of the physical
Another aspect that IoT contributes with into the DT devel- components and they offer a layer of intelligence that pro-
opment is that it provides a platform that can understand vides self-configuration, self-adaptation, and self-preservation
and translate data from multiple protocols. IoT devices are to each physical instance, to ensure that the ecosystem is
usually built with certain communication standards in mind, resilient to changes and failures that would affect its abil-
like MQTT, CoAP, MTConnect [90], OPC-UA [35], or 5G ity to reach its goals (this paradigm is sometimes referred to
uRRLC [92], and IoT is a bridge that connects those stan- as Cyber-Physical Production Systems (CPPS) [95]). In other
dards with the higher-level abstraction that is the DT. The words, this definition of CPS envisions them as a system of
choice of IoT devices, communication protocols, and IoT plat- inter-connected DTs so, in this sense, the DT is an enabling
forms can influence a very important characteristic of the technology for CPS. This interpretation of CPS leaves a blurry
DT: the synchronization rate between the real and virtual boundary between CPS and DTs, as they boast similar features
twins. We have established before that this synchronization and advantages and represent the smooth convergence between
rate will depend on the use-case of the DT. For time-sensitive reality and virtuality. To clear confusion, a study detailing a
applications, ranging from remote healthcare to traffic manage- comparison and correlation between the two paradigms has
ment in smart cities, the communication link between the two been conducted in [36].
entities should include secure uRLLC, while other use-cases The second meaning of CPS is more down-to-earth. In
where synchronization latency is not necessarily a problem, the literature, CPS sometimes refers to physical systems with
TABLE VII
OVERVIEW OF R EVIEWED L ITERATURE I NTEGRATING CPS I NTO D IGITAL T WIN OR V ICE -V ERSA
varying levels of complexity that are equipped with built-in strategy that allows DT services to interact independently and
sensors, actuators, networking and computation capabilities, efficiently.
and that are controlled digitally via computer-based algo-
rithms. It is clear that such CPS, compared to the ones in the
previous definition, require a lower level of intelligence and E. Virtual Reality and Augmented Reality (VR/AR)
digitalization. However, they do represent an enabling tech- The Digital Twin’s goal of virtuality and reality convergence
nology for the DT, given that they are I4.0-ready physical seems to perfectly align with the driver behind two developing
equipment that are proficient in reliable data acquisition, pro- technologies: Virtual Reality and Augmented Reality. Indeed,
cess optimization with feedback inputs, and improved built-in VR aims to improve Human-Machine Interactions (HMI) via
monitoring and control capabilities. Such CPS are an asset for 3D computer-generated simulations with which the user can
DTs to gather data securely from the physical processes and intuitively interact through wearable electronic devices. In
perform regulatory control operations at the edge. other words, VR can help immerge human operators into
In the works that reference the first definition of CPS, the a digital environment. On the other hand, AR technologies
authors rely on a generic architecture of CPS onto which make use of wearable devices render 3D digital images onto
they build the DT [51], [59]. As such, the DT makes use of a real-world background. In essence, AR helps bring vir-
the CPS-specific Service-Oriented Architecture (SOA) and act tual information in a physical environment. This section will
as Cyber-Physical System Nodes [29] in the virtual ecosys- explore how researchers have leveraged these two cutting-edge
tem, or it assures managerial independence of heterogeneous technologies to drive forward the DT paradigm.
interconnected systems [37]. In the healthcare industry, Laaki et al. [56] created a DT of
The articles that interpret CPS according to the sec- a remote surgery environment. The virtual representation of
ond definition provided above use the computer-controlled the medical equipment in a given location can be accessed via
system as the physical counterpart of a virtual twin. These VR by health professionals from a remote location. In turn,
works dive deeper into the implementation of adaptable DT wearable devices (head-mounted displays) allow the doctors
simulations via Functional Mock-up Units [52] or GA for to control a virtual robotic arm that operates on a dummy
scheduling optimization [53]. Others use the CPS-integrated patient. The DT then synchronizes the real and virtual twins of
Manufacturing Execution System (MES) software as a down- the robotic arm, such that the user can directly and intuitively
link between the virtual twin and the physical CPS [38], control the physical asset via its DT.
or leverage the DT to facilitate, control, and monitor highly Besides immersive and remote control of the real twin, VR
complex material flows [54]. also enables human operators to interact with the virtual twins
Table VII summarizes the contributions and advantages of industrial equipment as they normally would with the equip-
of the literature works revised in this section. Again, the ment itself, without interrupting the normal functioning of the
bulk of CPS-integrating DTs comes from the manufactur- real entities. As such, engineers can devise new deployable
ing industry, where significant efforts have been made to Circular Economy strategies to be implemented on the real
turn conventional factory equipment into CPS by populat- twin once thoroughly tested in the virtual world [96], create
ing them with sensors and connecting them to their vir- high-quality artificial training sets for safety training in sce-
tual twins that generate intelligent insight. The SOA, which narios where real data acquisition is risky or costly [97], or
is a founding principle of CPS, serves as a decoupling have students and trainees learn how to operate the physical
TABLE VIII
OVERVIEW OF R EVIEWED L ITERATURE I NTEGRATING VR/AR I NTO D IGITAL T WIN
twin by immersively interacting with and practising on their can impose significant alterations to that architecture. This is
DTs [98], [99], [100]. also part of the reason why the DT is also difficult to define.
AR technologies can facilitate quick access to the DT For example, it is not always the case that the physical twin
interfaces of real entities by superimposing their virtual data is an actual, real object, characterized by a physical geom-
and images onto the camera feed [101], when the camera is etry. As such, proposed modeling methodologies that imply
pointed at the physical twin itself [102]. This feature allows the existence of a physical object would not map well to
human operators to dynamically monitor DTs, without having use cases like the DT of processes. On this note, the authors
to go out of their way and connect to the computer that hosts in [106] proposed an architecture for the manufacturing indus-
them. try that breaks down the DT into three constituents, namely
According to one study [103], the VR/AR-enabled DT can product DT, process DT, and operation DT, with each com-
address three current challenges in HMI development: high- ponent having a seemingly different architecture. A different
fidelity virtual representations of physical assets, availability of modeling methodology, presented in [107], summarizes a man-
both real and simulated data, and intuitive interfaces for human ufacturing DT as the synchronization between 3D modeling
operators. However, for a complete merge between the real and and mechanism modeling.
virtual worlds, neither technology is enough by itself. Both Nevertheless, researchers in the field have come up with
technologies allow the user to interact with the virtual repre- DT frameworks that abstract away from the use case-specific
sentation of a physical entity, but they do not allow the real and details of implementing. Instead, they take a step back in
virtual surrounding environments to interact with each other. order to focus on the components one could reasonably expect
For such cases, Mixed Reality (MR) technologies combine to find in any DT, no matter the application scenario. The
the advantages of both VR and AR, to bring digital models in authors in [108] claim that a DT only requires two main
the physical world and simulate their processes under real cir- components to be whole, which are: a virtual representation
cumstances [104]. Table VIII summarizes the main findings of of the physical entity and an API. The authors also men-
this review, highlighting the main domains where VR/AR/MR tion that the virtual model does not require to include the
technologies have enabled the DT to provide immersive HMI, 3D geometry of the physical object, unless that is required
training and monitoring. by the DT’s use case. In addition, other dimensions can be
added to the DT, such as: data storage, access control, meth-
ods, events, and a human-machine interface. Similarly, other
F. Modeling Methodologies works [109] insist on the importance of creating an API-like
While not an enabling technology in-and-of-itself, the middleware that allows the DT to connect to external systems.
umbrella of modeling methodologies covers a great range of Riedelsheimer et al. [110] proposed a methodology for build-
frameworks and software meant to guide developers towards ing DTs for already-built, complex, inter-disciplinary physical
building a core component of the DT: the virtual representa- objects, with the aim to optimize the systems’ sustainability
tion of the physical entity. Similarly to the DT definitions, the throughout its lifetime. With the goal of creating the DT of a
modeling approaches also vary greatly across the literature, smart factory able to manufacture customizable products, the
that some researchers concluded that there is no consensus on authors introduced a planning and development framework that
the subject [13], [105]. The challenge to overcome here sits in integrates several design and management frameworks, such
the interdisciplinary and use case-specific nature of the DT. It as V-IoT, 8D-Model, Design Elements, SCRUM, etc.
is indeed difficult to create a one-size-fits-all architecture for A more common approach to modeling methodologies that
such a versatile technology, given that its application scenario can be found in the literature is the 5-dimensional DT initially
In [115], the authors took a closer look at the specific frame-

works and software that can be used for DT development.
V. D IGITAL T WIN : U SE C ASES AND S ERVICES

As we have stated previously, the choice of enabling tech-
nologies for the DT will be highly dependent on the DT’s
envisioned applications and services. In this section, we will
present how DT frameworks differ in their components from
one application scenario to another, and it will become appar-
ent how, even within the same use case, the DT structure can
vary greatly. As such, this section will be divided into two cen-
tral characteristics of the DT, as presented in the definition we
provided in Section II, namely DT use cases and DT services.
There are many potential DT services and use cases; however,
within this section, we limit our survey and discussion to some
typical examples in each category to demonstrate the concept
Fig. 5. Five-dimensional DT architecture. and potential of the DT. This choice of DT services and use
cases also align well with the specific case studies we present
in Section VI.
proposed by Tao et al. [111]. According to this work, the
five dimensions of the DT are: physical entity, virtual entity, A. Use Cases
communication, data, and services. This break-down of the 1) Smart Factory and Industry 4.0: The current vision of
paradigm allows for a decoupled architecture with orthogo- I4.0 aims to cut the costs of production, build efficiency and
nal elements, making it easier to understand. For this reason, give companies an increasingly versatile approach to produc-
many researchers have based their modeling methodology pro- tion. Factory stations now have the ability to communicate
posals on Tao’s work. Wang et al. [112] build on top of the directly with one another, eliminating the need to commu-
5D DT to introduce a System Design Digital Twin which aims nicate via a central processing controller. This decentralisa-
at reducing the complexity of model-based system engineering tion through modularisation and the IoT increases flexibility,
by closing the gap between the physical and theoretical design opportunity and efficiency. Rather than a centralised control
processes. The authors in [113] proposed a DT information unit delivering instructions to each machine to carry out linear
modeling method dubbed GHOST (Geometry, History, Object, sequential steps, individual machines now inter-communicate
Snapshot, Topology), representing an expansion of the data directly enabling the partly-finished product to be passed
element of Tao’s architecture. Its aim is to provide a flex- straight on to the next station. As everything is now processed
ible framework for combining multi-source heterogeneous locally, the production line is equipped to produce any num-
information in complex DT systems. Wu et al. [114] presented ber of unique products, which was not previously possible on
a methodology for building 5D DT models that is supported by single unit lines. By not having to communicate with a cen-
an improved version of the TRIZ function model. The TRIZ tralised unit, the production line can run more smoothly and
function model describes a complex system by breaking it efficiently. In addition to increased efficiency, the new secu-
down into various types of elements and relationships. In order rity sensors built into the autonomous modular systems create
to provide further nuance and fidelity to the model, the authors a safe working environment for human operatives, ensuring
enhance TRIZ with behavioural logic via conditional flow robots halt if they encounter an obstruction. This has the added
control, rules, and interactions with the external environment. benefit of workers being able to touch a robot to stop its
Finally, other existing architectures partially overlap with Tao’s motion without the need to activate an isolator. The decentral-
methodology. For example, Bazaz et al. [51] defined the DT as ization of manufacturing processes and the increasing demand
an interconnection of five layers: data store layer, primary pro- for customization leads to a need for adaptive and intelligent
cessing layer, model and algorithms layer, analysis layer, and production equipment. The DT aims to address this challenge.
the user interface component. Judging by the description of Makarov et al. in [116] investigated the design concept of a
each layer provided in the paper, Tao’s data dimension cor- DT, coining two new types of system and splitting the def-
responds to Bazaz’s data storage layer, the communication inition of a DT into four parts. A pre-DT is defined as a
element is similar to the primary processing layer, elements virtual prototype for a system to reduce technical risks and
of the virtual entity dimension can be found in the models root out design problems before development. Any issues
and algorithm layer, while Tao’s services element, it can be with the system found on the virtual twin can be solved and
argued, includes both the analysis and user interface layers. corrected on the physical system. An adaptive DT uses a
Abstract modeling methodologies have seemingly slowly user interface, linking the two systems, allowing the virtual
begun to converge in the DT-related literature around Tao’s twin to understand the preferences from the human oper-
5D model, depicted in Figure 5. However, the in-depth imple- ators in different scenarios. Finally, an intelligent DT has
mentation details and technologies remain use case-specific. all the characteristics mentioned before; however, it contains
Fig. 6. Digital Twin framework for smart manufacturing.
unsupervised ML for pattern detection in the physical system The results show that the addition of software circuit break-
environment. The results found to reduce repair costs and ers in DT systems can station errors cascading into higher-level
increase quality control for a lowered amount of product systems, and keeps the fault local to that physical station. The
defects. goal of this, however, is not directed at stopping the spread of
Lee et al. in [117] proposed a DT framework for smart errors, more to isolate them, as different errors hold different
manufacturing. More specifically, the authors introduced a sys- levels of severity.
tematic integration of the DT in various levels of shop floor In [119], Lee et al. investigated a DT-enabled predictive
design: unit level, system level, CPS level, and business level. maintenance framework for a CNC machine tool system using
Figure 6 displays the workflow of the proposed architecture: DL. The authors outlined vital design characteristics and
shop floor designs are initially tested against unit level Key requirements for integration of DTs in a CPS. The rapid
Performance Indicators (KPI) and are only selected for the growth and requirements for IoT and ML mean data trans-
subsequent level testing if they satisfy them. Consequently, a fer latency has to be as low as possible. The authors suggest
triage-based system filters out the designs that do not meet 5G will significantly contribute to the integration of ML and
KPI requirements at each level during the virtual implemen- be the backbone of DT technology. They also indicate that
tation of the shop floor. In the physical space, the selected all sensors should be developed into smart sensors for ease of
designs are implemented gradually across all levels, and the plug and play and scalable networking.
real performance of the designs help optimize the KPIs in the Gericke et al. [120] investigated the efficiency and latency
virtual domain. of communication and production rate of 500ml water bot-
In [118], Preuveneers et al. proposed the use of safeguard- tles with a cyber-physical bottling plant. Using Open Platform
ing systems throughout the software of the DT system. These Communication (OPC), they were able to give possible posi-
safeguarding systems, coined “software circuit breakers,” are tions of bottle-necks in production. Using this, the DT can
designed to handle local system errors to stop faults propagat- also detect drops in production rates. The authors conclude by
ing through the levels of the DT, as these have the potential pointing out the machines cannot rely on OPC as the only con-
to be catastrophic. Such local failures can include: nection to the physical system due to the significant latency,
• Missing sensor data ranging from 100ms to 500ms. They also found that as the
– Failed transmissions system scales, more OPC connections will be needed, further
– Disconnected sensors slowing the system.
• Broken Sensors Lin and Low [121] carried out research on the design and
– Hardware failures or disconnected actuators implementation of a DT on a CPS. The authors suggested there
– Denial of service attacks are three layers of modules required to produce a DT system:
• Operation Layer—used for tracking all physical assets on and OPC data server of SIEMENS have been used to process
the CPS this information transfer.
• Visualisation Layer—used for real-time simulation, tak- 2) Infrastructure: Civil infrastructures are highly valuable
ing the data gathered from the operation layer and assets, having vital societal roles and involving a large
provides a remote-monitoring function to allows com- number of people at every stage of its complex working
panies to be continuously updated on current production life from initial conceptual drawings, 3D numerical model,
status construction activities to operational service, as shown in
• Intelligence Layer—This takes all the information from Fig. 7. Thus, infrastructure management has been a sub-
the previous two layers to create a historical data bank. ject of intense research activity, aiming to maximize their
This layer will use this data bank, along with real-time safety and service life while minimizing the building and
data, to perform efficiency and health monitoring. maintenance costs. In the following section, recent works
The authors then took this design concept, implemented investigating DT-based paradigms unifying practical tools
it onto a prototype Surface Mounted Technology production and expert knowledge with novel advanced technologies are
system, built of seven stations. This system worked and acted reviewed. Furthermore, a DT application for Structural Health
as a validation to their prosed three-layer DT design concept. Monitoring developed by the authors will be described in
Hasan et al. [122] investigates the use of blockchain tech- Section VI.
nology as a way of optimising the DT format on a CPS. a) Smart Building: A building consists of a number of com-
The problem outlined in this paper is with the development ponents spanning different domains from energy, ventilation,
side of DTs. There is a need for monitoring the interaction heating, air-condition, plumbing, mechanics, and so on. Thus,
between design teams and workflows, so each change made having effective building management is challenging, espe-
to the software can be accredited to a specific person, who can cially with a high-rise building or a commercial complex.
be held accountable. This employs transparent history moni- Therefore, Lu et al. [127] have developed a smart O&M man-
toring, security and trust and ensures the trusted creation of agement tool using DT specialized in detecting anomalous
virtual models. The blockchain platform used was Solidity - behaviors. At first, a dynamic and distributed data integration
Ethereum smart contracts. component was built to integrate heterogeneous data from var-
In [123] Kanak et al. presents a blockchain-based model for ious daily-updated databases using corresponding object IDs.
distributed and collective DT environments which is becom- Secondly, intelligent anomaly detection functions were imple-
ing essential in new “Any 4.0” era. They proposed a novel mented using the BOCD to identify suspected change points,
approach to use security as a symmetric and asymmetric cryp- related time instants, locations, and even elaborate the causes
tographic tool to be implemented at a hardware level. The of the change points.
DT ecosystem proposed includes “X-by-design” and “X as-a- Lu et al. [128] have developed a dynamic DT in order to
service” principles where “X” is security, accountability and improve asset maintenance and asset failure prediction in a
integrity. campus of the University of Cambridge. The DT framework
Similarly, the authors in [124] developed a simulation-based consists of five layers: (1) first, acquisition layer collecting
CPS DT for blockchain enabled Industrial Hemp Supply Chain data from multiple sources including Building Information
(IHSC), which is utilized to improve the understanding of Modelling (BIM), real-time IoT sensor data, asset registry, and
a complete process pf supply chain, assist in quality con- asset tagging data and space management data; (2) second,
trol verification and fast track the development of secure and transmission layer transmitting data collected from physical
automated supply chain system. They present the two-layer device in the first layer to a central database using WiFi,
blockchain based data tracking, information sharing, and inter- 5G, low-power wide-area networks, etc.; (3) the third layer
operability framework for the end-to-end IHSC which can is digital modeling, where different types of digital models
greatly improve both security and efficiency. could be developed for various requests in DTs; (4) fourth, the
In [125] a DT for an experimental assembly system based data/model integration layer provides real-time analysis, then
on a belt conveyor system and an automatized line for quality assess up-to-date asset condition and maintenance status with
production check is proposed. They have created a DT for the help of AI-based functions; (5) and lastly, the application
Bowden holder from a 3D printer, which is composed of some layer with visual interface facilitates the interaction between
plastic components, fastener parts and a stepper motor. The DT and facility managers.
assembly was positioned in a fixture with ultra-high frequency Thyssenkrupp, in collaboration with Microsoft [129],
tags and IoT devices for identification of status and position. developed a DT framework for the elevator system in a
The inspection system included an industrial vision system high-rise building in Rottweil, Germany. The new advanced
for checking presence of parts, inspecting the dimensions and elevator, which could move both vertically and horizontally,
looking for errors before and after assembly operation. was equipped with IoT systems and deployed via the Azure
Vachálek et al. [126] presented a DT model of the produc- DT framework. This DT is able to reduce the downtime of the
tion line based on the simulation tool called Plant Simulation, elevator significantly, provides information related to elevator
made by SIEMENS. This model was a detailed virtual copy occupancy and usage to enhance the availability of the eleva-
of the physical process involved in assembling the hydraulic tor, which serves more than 10,000 people per day. Moreover,
pistons. The goal was to extract the information about the the DT is reinforced by AI learning to optimize travel times
number of times there has been movement, to the data storage of frequent users.
Fig. 7. Digital Twin in infrastructure.
b) Smart Infrastructure: Ganguli and Adhikar [130] thor- closely approach the real behavior of the structure over its
oughly presented a DT for a Single Degree Of Freedom working life. Another concrete example of DT application
(SDOF) dynamic system, in which a double time-scale system in SHM is the work of Kaewunruen and Xu [133] for rail-
was proposed the first time. Specifically, the fast time scale way turnout systems, which consists of many delicate and
reflected the dynamic responses of the real system and the complex details and has a critical role in ensuring the safety
slow one for the DT, and it was found that such a multiple of the railway system. The application is built by expanding
time-scales DT was able to capture effects of mass and stiff- the conventional 3D BIM models to 6D models, involving
ness evolution on the SDOF simultaneously. Ding et al. [131] three geometric dimensions, time dimension, cost dimen-
proposed a DT for a steel bridge construction using BIM and sion and its sub-categories, and carbon footprint dimension.
IoT data from embedded sensors, able to dynamically monitor Moreover, not only actual operation information are inves-
the construction processes and key related factors such as site tigated, but data from historical phases such as planning,
resources, business processes, field workers, as well as their design, pre-assembly to predicted future action, i.e., mainte-
live interaction, thus ensuring a lean construction. nance, demolition are also taken into account. A similar DT
In order to develop a proactive maintenance system for approach via 6D BIM is applied to the renovation manage-
bridge structures, Shim et al. [132] proposed twin models fus- ment of the King’s Cross station [134], aiming at a more
ing entire lifecycle information from design, construction to resource-economic and environment-friendly model than tra-
operation, and maintenance. The first model was built from ditional construction methods such as 2D paper drawings, or
as-built documents using BIM, while the second model was 3D static numerical models.
generated with the help of the 3D scanning technique using With regards to the offshore structures, Akselos [135] has
Unmanned Aerial Vehicles and laser scanner. A maintenance- developed a holistic DT framework coupling with parallel
oriented digital process connecting two models was also cloud computations, which can provide real-time risk-based
elaborated to update the structure’s status continuously. decisions in response to time-varying uncertainty encoun-
Structural Health Monitoring (SHM) is an important topic tered by offshore structural engineering involving wave, wind,
in civil engineering, thus Ritto and Rochinha [41] explored marine environment, and so on. As one of the leaders in the
a SHM-oriented DT. Essentially, data-augmented modeling power plant industry, General Electric has been devising a
is implemented to compensate for the discrepancy between sophisticated DT framework, a.k.a. Predix, [136] including
numerical models and physical counterparts, which involves a spectrum of aspects of the physical entity ranging from
two phases. First, the discrepancy is qualitatively and quan- thermal, mechanical, material, electrical, economic, statistical,
titatively measured from the physical entity. Second, bias- environmental. This DT is expected to perform a wide range
corrected models are calibrated based on the numerical model of applications such as optimizing profitability, maximiz-
and measured discrepancy, yielding augmented outputs which ing plant safety, accurately forecasting productivity, profiling
customers, synchronizing operations of many machines on help of BIM, energy simulation model, weather simulation
plants. For this purpose, various advanced technologies have model; data/model integration Layer powered by data analytics
been implemented in Predix: firstly, multimodal data analyt- and AI; and finally service layer providing multiple services
ics are required to automatically collect, update, and store a such as energy management, asset management, security man-
vast amount of heterogeneous data involving: parametric data agement, etc. at different levels for different stakeholders.
(temperature, pressure), graphical data (infra imaging, cam- Lin and Cheung [141] explored a DT application for Smart
era), spectral data (spectroscopy), time-series data (sensors, City’s underground parking garage using Wireless Sensor
accelerometers), text data (service records), historical data Networks (WSN) in conjunction with BIM technologies for
(maintenance database), and so on. Secondly, the company improving environmental management. The WSN was com-
deploys a number of digital models encompassing physics- posed of various sensors able to monitor gas, temperature, the
based models, i.e., thermodynamic model, combustion model, humidity of the garage, whose data were later transferred to a
transient dynamic model, and data-driven models such as sta- central host with the help of communication routers. On the
tistical process control, ML-based anomaly detection, and DL other hand, a BIM model of the garage is built by Autodesk
regression and classification. Finally, a knowledge network, Revit and Naviswork, providing a detailed digital representa-
called Expert Twin is explored to connect experts across tion. When coupling the BIM model with environmental data
the enterprise around the world for sharing data, knowledge, from sensors, the risk status, i.e., CO gas level, user comfort
solution, and best practice. level, are lively represented and can be visually noticed with
c) Smart City: In an attempt towards sustainable growth green/red color code.
of the city as well as a better quality of life for citizens, 3) Towards 5G/6G With Digital Twin: The future Industry
Francisco et al. [137] investigated a DT paradigm for Smart 5.0 paradigm envisages removing any physical limitations and
City using spatio-temporal data. At first, a digital replica of the building in virtual connectivity and capabilities that will enable
city is rebuilt in a virtual space using the Unity cross-platform; the seamless interaction between devices, humans and infras-
after that, the researcher can navigate across the virtual city tructure [142]. Even though this digital transformation across
via VR devices. In addition, an AR crowd-sourcing module various industries will enable applications that serve different
allows for integrating feedback of citizens about real infras- purposes, they all have something in common: dependency
tructures into the platform parallelly. By doing so, the triangle on reliable and strong connectivity enabled by the underlying
interaction human-infrastructure-technology is captured, ana- next generation network infrastructure (e.g., 5G/6G). The fifth
lyzed, and updated, serving to improve the sustainability and generation networks is already a key component in I4.0, since,
wellness of the city. even in the DT technology, the connection of components and
Ruohomäki et al. [138] explored a Smart city platform using devices is of utmost importance and communication latency
DT for the city of Helsinki to enhance city management in varis expected to be less than a few milliseconds. In this context,
ious aspects, including urban landscape, energy consumption, the relationship between 5G/6G and DT can be seen from two
environment. At the base of the platform is the 3D city models different point of views. The first one, sees the 5G/6G network
called CityGML integrating geographic information, geome- as an enabler for different DT applications, while the second
try, topology, and appearance. Next, sensor data are linked to one sees the DT as an enabler for 5G/6G by looking at the
city models via an IoT platform called SenSorThings, com- DT of the network itself. Both point of views are addressed
posed of two main parts: sensors for observation and thing, in this section.
i.e., API, for connection to the network. By doing so, the Communication technology is going to be the foundation of
initial 3D model is transformed into a semantic ecosystem industrial IoT, hence in [143] authors have presented a detailed
with high interoperability. Du et al. [139] presented a Proof overview of 5G wireless transmissions and their application
of Concept of DT for Smart City’s Information System at prospects according to cyber-physical-based manufacturing
an individual level, namely, Cog-DT, to reduce the cognitive systems. Furthermore, a novel 5G-based industrial IoT archi-
overload for residents and workers in the city. The first step tecture for smart manufacturing is proposed. In [144] authors
of Cog-DT involves using VR technology to gather personal have taken an industrial robotic arm as a use case and have
cognitive information such as neuroimaging, physiological, performed an analysis of simulated robot with the effects of
ergonomic. Then, the second step is to simulate human cog- simulated network for CPPS. The key contribution of the study
nition in response to various information stimuli. Finally, an is the comparison and analysis of effects of using different
adaptive information system is implemented to display engi- kinds of network types in a Gazebo simulated robot. The three
neering information adjusted in a real-time fashion. In an types of networks used between robot controller and robot
attempt to improve the long-term performance of the O&M are wired link, public LTE and the 5G uRLLC network. The
service of building and other infrastructure, Lu et al. [140] results showed that 5G outperforms LTE and wired network in
developed a 5-layer DT architecture and applied it to the West terms of productivity as well as the processing time increased
Cambridge site with more than 20 buildings and other facil- by 50%.
ities. The proposed DT architecture can be extended further Furthermore, 5G (and beyond) and the DT can revolution-
up to a city-sized application. Their five main layers are data ize the cooperative vehicle infrastructure. The authors in [145]
acquisition layer including sensor data, weather, energy, secu- have explored the implications of 5G based communication to
rity, culture, policy data and so forth; transmission layer via the intelligent V2X system called Providentia and proposed
the Internet, 4G/5G, HTTP; digital modeling layer with the scene detection, fusion and object detection strategies. The
vehicles in the Providentia system play two roles, first being

the source of information about the current state of vehicle and
its surroundings, second as a user of the DT system. All the
communications in this infrastructure is realized through on-
board 5G modems as well as 4G LTE. Similarly in [146] 5G
communication networks, more specifically edge devices, are
recommended for the real time data collection in DTs designed
for traffic congestion avoidance. The proposed solution for Fig. 8. The vision for DT-enabled next generation communications.
road connectivity infrastructure involves real time situational
information gathering, collection of driver history using ML
algorithms, data lake, intelligent transport system, DTs and yet completed at a large scale, many industry pioneers and
blockchain services for vehicles. The application domain of technology leaders are looking ahead at defining the next
the DT technology is very vast, as the authors in [39] have generation networks, 6G. 6G envisions interactions between
optimized the services of mobile edge computing using DT three worlds: the human world (e.g., senses, bodies, intelli-
and DL algorithms. Achieving ultra-reliable and low-latency gence, etc.), the digital world (information, communication,
in mobile edge computing can be challenging because of the computing, etc.); and the physical world (objects, organisms,
possibility of losing the packets in case of deep fading chan- processes, etc.) [152]. Furthermore, 6G envisions network
nels. Thus, in order to optimize the offloading probability, a speed of 100 to 1000 times faster than that of 5G for accom-
DT is developed in [39] which can evaluate the normalized modating new service classes like further enhanced mobile
energy consumption, reliability of user association schemes broadband (FeMBB), ultra-massive machine type communi-
and delays. The DT will save the optimal option and store it in cation (umMTC), and enhanced ultra-reliable and low latency
memory as a training data for DL algorithms. Tactile Internet communication (eURLLC) [153] and latency less than 1ms
or 5G will also revolutionize the traditional multimodal appli- for ensuring safety in mission critical communications and
cations. Research has been carried out to evolve the current IIoT applications [154]. DT has great potential to provide
state of the art media to multimodal media, where DTs for a digital environment where future generation networks
can facilitate the high quality interactions, like touching and like 6G can evolve. Integrating DT within mobile networks
smelling the objects of remote environment [147]. Moreover, is gaining popularity in the industry from major tech compa-
the DT and 5G/6G will also play a vital role in autonomous nies like Ericsson, Huawei and Nokia [155]. In this context,
navigation systems [148] where autonomous ships can easily DT has the capability to continuously monitor and analyze the
be commanded and decisions regarding navigations are made performance of the network, predict any unanticipated failures
easy. and optimize the network performance by triggering intelli-
Smart manufacturing is one of the most important verti- gent decisions accordingly. Figure 8 illustrates a vision of the
cal industries identified by 5GPPP and with the maturing of 6G DT that facilitates the live virtual replica of the whole or
network virtual functions and 5G, use of Virtual Network parts of the 6G network to perform continuous monitoring and
Functions (VNF) in smart manufacturing is gaining popu- assessment through a closed loop process between the physi-
larity in research community. To this context the authors cal entities and the digital counterparts. The 6G DT powered
of [149] have presented a use case in the manufacturing indus- by AI will enable design and performance improvements and
try using the experience of a manufacturing company named real time optimized operations enforced on the physical 6G
Weidmuller Group. The manufacturing network services in the network.
proposed use case are composed of different VNFs and it is
developed using SDK and 5GTANGO lightweight Network
Function Virtualization (NFV) prototyping platform. Similarly, B. Services
in [150], an efficient solution is proposed, called MIGRATE, 1) Anomaly Detection: The Digital Twin has gained its sig-
that implements virtual functions and virtual mobile devices nificance in I4.0 by virtue of its intuitive and insightful role in
to represent physical processing devices. MIGRATE ensures integrating data analytics into traditional manufacturing facil-
the successful and seamless transfer of software entities. ities[15]. Through the analysis of run-time data incurred from
Connectivity in the future is highly dependent on develop- the physical systems, DTs have enabled smart factories’ engi-
ment of DT environments [151] that are actual representation neers to know their facilities better. One of the many strategies
of their physical counter parts. With the advent of future gen- involved in analysing the data and discovering better insights
eration networks, it is expected from the DT technology to is Diagnostic Analytics (DA), i.e., having a profound look at
represent not only physical objects, but the biological world the data to observe and interpret the causes of events and
as well. behaviors. An effective and popular technique recognized for
With the rapid advancements in smart technologies and DA is anomaly detection [156]. Anomaly detection is a tech-
applications like holographic projection, VR, AR as well as nique of fault diagnosis that detects sections of data disobeying
mission critical applications like remote surgery, that have the normal, expected behaviour [157], [158]. The DT-enabled
strict Quality of Service (QoS) requirements, current networks anomaly detection has proven itself an asset for the Operation
including 5G will no longer be able to meet these expec- and Maintenance (O&M) phase of smart factories. As illus-
tations. Although the deployment of 5G networks is not trated in the subsequent literature, it is clear that incorporation
Fig. 9. DT application: Anomaly Detection Framework.
of anomaly detection in DT architectures has had a significant values of asset parameters that compared with the instanta-
impact on the popularity of DTs in the I4.0 ecosystems. neous IoT data from the physical systems to detect potential
The anomaly detection work flow followed by researchers anomalies.
in general is represented pictorially in Figure 9. Workflow of A DT-enabled anomaly detection mechanism was proposed
Anomaly detection application with most common key phases in [127] for built asset monitoring. The detection process
involved are shown in the Figure 9. It is observed that uni vari- asserted the need for cross-referencing the multiple data
ate or multi-variate models with single or multi step detection sources for building facilities information. Owing to the data
models have significant effect on Anomaly detection appli- inter-operability and re-usability aspects of the task, DT was
cations performance. The primary sources of data for such accepted as a comprehensive solution for the data integration
anomaly detection applications are derived from industry run- problem. Storing the data from heterogeneous sources into a
time data from the connected IoT devices or historical data single, integrated format eased the detection process, meant for
from database logs available. The DT models close to their every key asset in the building structure. The run-time data of
physical counterparts in an industry may also simulate the the assets, acquired in various data forms, was encapsulated
data required along with the run-time and historical data avail- in a single data format that invariably assisted the detection
able.Data set identification and creation based on the user process many-fold.
specific applications is also determining factor to improve the An extensive amount of research has been followed up
performance of the anomaly detection application. Anomalies in implementing anomaly detection with various ML algo-
can be visualized with a right choice of algorithms based on rithms. As a consequence, several research problems have
standard machine learning models. The choice among super- been encountered by the researchers. Castellani et al. [160]
vised, unsupervised and semi supervised learning models is implemented anomaly detection on industrial data with a
based on the availability of labelled data versus large data semi-supervised learning approach. The dataset under study
requirements for the training. The DT application can arrive at consisted of a major portion of unlabelled data (unsupervised
the next level PdM suggestions from the observations derived learning) and a smaller portion of labelled data (supervised
from the anomaly detection outcomes shown in Figure 9. learning). The choice of hybrid dataset for the said study
In [159], the authors proposed an IoT-enabled “Living was influenced from the following facts: i) for the supervised
Digital Twin” for additive manufacturing. The twin essen- approach, additional efforts need to be invested in labelling
tially rooted for assuring higher productivity by monitoring the data that is prone to be impractical for larger datasets,
the system with analog sensors such as acoustic, vibration, and ii) an unsupervised algorithm learns by the means of
magnetic, etc. The basic protocol might seem rudimentary bulk statistics of majority behaviour, thus a smaller unlabelled
due to the sensor-actuator action involved, but the twin had dataset might result in an ambiguous output. A DT frame-
a significant contribution in the aspect of anomaly detection. work, by virtue of its simulation capabilities, generated the
The digital counterpart introduced the concept of a fingerprint unlabelled portion of dataset with normal samples that simu-
library to detect anomalies. The fingerprints are the run-time lated normal industrial operation. The latter half of the dataset
comprised of anomalous samples derived from a real factory the device maintenance is completed. Through this article, the
data. The article claims to have achieved a higher score in eminence of machine failure prediction by anomaly detection,
its performance indices for semi-supervised approach than the reconfiguration and rescheduling has been emphasized.
fully unsupervised approach (AUC scores 0.872 and 0.756 From the literature discussed above, it is observed that DT
respectively). Bevilacqua et al. [161] proposed a reference proves its expertise in the field of anomaly detection owing
model using DTs for risk assessment analysis in industrial to its varied capabilities. The various problems or limitations
process plants. They put forth a framework for a smart fac- encountered in anomaly diagnostics such as the generation of
tory that would enhance its productivity while prioritizing the a reference model for comparison based on past data, lack of
safety of human operators. The risk identification and assess- simulation datasets, development of early prediction models
ment aspect of the model called for inclusion of an “Anomaly for analysis in real-time etc. have been addressed efficiently by
Detection and Prediction” tool in the twin. The tool was the DT. In conjunction with the enabling technologies, the DT
developed while being fairly inclusive of the twin-enabling showcases capabilities such as flexibility to include analytical
technologies: i) development of a communication and control tools in its core architecture [161], [162], [163], simulation of
system with sensors-actuators at the edge and Programmable training dataset [160] and anomalous dataset [88] for better
Logic Controllers (PLCs) at controller station, where the wire- model training, compatibility with heterogeneous data and its
less sensor networks helped to gather the ground-level data to pragmatic integration [127], real-time analytics [162], [163]
the twin, ii) development of ML models by employing the and generation, preservation of signature copies of every entity
received data to analyse, predict the risk factors and edu- in extensive sensor networks [159] to resolve the problems that
cate the DT to invoke precautionary actions. Notably in [88], are confronted with.
a framework was proposed for the monitoring and diagnos- 2) Predictive Maintenance: Another service that the
tics of a fleet of aero-engines. The framework implemented Digital Twin promises to provide is Predictive Maintenance.
anomaly detection for fault detection, isolation and identifica- The advantages in terms of cost, time, and resources that
tion. A physics-based model of a three-shaft turbofan engine PdM can demonstrably bring to the industry have been long
was developed and simulated in-house. The simulation data for sought-after. Thus, significant research effort has already been
the said model and the signatures of potential component faults invested into developing working architectures that can accu-
were generated by a DT. In order to simulate an entire fleet of rately predict a machine’s failure (i.e., self-diagnosing). For
engines, production scatter simulation was implemented by the example, Wang et al. [164] proposed a basic architecture that
twin that cloned multiple replicas of the engines with minimal outlines the components of a PdM system’s pipeline: data
random variation. acquisition, data analysis & state detection, health assessment
The capability of DTs to analyse temporal data in real time and prognosis, and maintenance actions & alerts. Although at
has proven to be a greater asset for targeting spurious events. a first glance these functional blocks might seem rudimentary,
This statement has been validated by Xie et al. [162] who they in fact represent the founding pillars that support PdM
have proposed a DT framework for crucial asset monitor- services, and they are predominant in many framework pro-
ing in a building facility. The multi-layered twin architecture posals to this day. For example, the authors in [165] present a
mainly featured: i) the Digital Modelling layer that acquired PdM architecture for nuclear infrastructure whose components
time series data in real-time and stored simulation data as greatly resemble the ones in the previously referenced work.
well as historical data, and ii) the Data/Integration layer that Another PdM architecture that is rooted in the principles of
analysed the data at hand and took informed decisions. The the one presented by Motaghare et al. is Cheng et al.’s pro-
said analysis was assisted by the Bayesian Online Change- posal in [166]. The paper introduces a solution for a bearing
point Detection (BOCD) algorithm that detected suspicious production line, and its framework is based on Edge-Cloud
instances upon investigating the sudden variations in the cooperation. In this architecture, the data pre-processing is
time-series (change-points). The framework has been demon- done at the edge-level, and the processing-heavy tasks, like
strated in a DT demonstrator laboratory at the University Remaining Useful Life (RUL) prediction, is delegated to the
of Cambridge. The experiment was conducted on two iden- cloud. This type of structure is preferable when the ML algo-
tical cooling pumps for vibration monitoring and the twin rithm used is actually a DL method. In that case, the training
successfully identified anomalous vibrations. On the similar can be too computationally-intensive to perform at the edge, so
lines, a group of researchers put forth a blueprint for a uni- instead it is deployed in the cloud. Indeed, for the estimation
fied DT for anomaly detection in Smart Manufacturing [163]. of the RUL parameter, the authors used a DL mixed algorithm,
A novel aspect of the said architecture was a DT platform named in the paper ARIMA-LSTM model (Autoregressive
consisting multiple twins for every crucial process/entity. A Integrated Moving Average - Long Short Term Memory). In
demonstration on a CNC facility proved the significance of this system, ARIMA handles the prediction of the linear part of
an anomaly detection scheme developed by framework. A the time-series data, while LSTM predicts the non-linear com-
range of limits is devised by the twin after thorough anal- ponents, which are then summed up to offer a final prediction.
ysis of historical data. The instantaneous values are referred ARIMA is also employed in [167] as a technique of extracting
against this range and dubious instances, if any, are reported the underlying trends in various data streams coming from sen-
by the twin prior to unfortunate tool damage. Further protocol sors. The trends identified in heterogeneous time series data
dictates switching the device state as “faulty”, requesting for are then fed as features to a Principal Component Analysis
maintenance and reconfiguration of the facility topology until (PCA) algorithm that extracts the most uncorrelated features
Fig. 10. DT-based PdM scheme using hybrid modeling.
to be fed into the RUL predictor. For remaining life estimation, would undoubtedly require large amounts of resources in terms
the authors propose the use of a regression technique called of storage space, computational power, and smart equipment.
the Support Vector Regression model. Cachada et al. [168] Liang et al. in [170] proposed a layered architecture for a
presented a complex and detailed architecture of an intelligent low latency deployment of a Convolutional Neural Network
and PdM system that explores the interdependent modules (CNN)—based prognosis system. The proposed scheme con-
that would constitute a functional block in a PdM frame- sists of three layers that share responsibilities effectively,
work. As such, the paper explains how different approaches keeping high-speed processing capabilities on the terminal and
are needed for the data acquisition block, depending on the fog layers, close to the manufacturing equipment, and leav-
type of input data: automated, semi-automated, or manually ing the training of the CNN to the cloud layer. Drawing the
introduced by an operator; it also presents an offline data anal- line, it seems like the focus has shifted from the development
ysis scheme that also relies on the LSTM DL algorithm for of extremely intricate mathematical health prognosis models,
prediction of machine state, a dynamic monitoring block that which were tailored to be specific to the equipment, towards
deals with visualisation and early detection of failures, and an data-driven models which predominantly rely on ML and, of
intelligent decision support system for maintenance interven- course, Big Data. The ability to reliably and quickly trans-
tion that guides the operator through simple instructions and port, store, and process huge amounts of data has opened new
visualisations, in a way that reduces the need for technical doors in the world of PdM, immensely facilitating the task.
knowledge, leaving room for focus on the maintenance task However, that is not to say PdM has become an easy job.
at hand. In this scheme, adjacent blocks communicate with The new approach presents other challenges in terms of time,
each other in a sequential manner, but as-needed communica- resources, and sets of skills that are required to deliver accurate
tion between non-adjacent blocks is also allowed for further predictions of the RUL parameter.
automated optimisation. A compromise between the complex, but transparent
As an overview of the proposed Predictive Maintenance physics-based models and efficient, but opaque data-driven
frameworks in the literature up to date, it seems that they models, are hybrid models, where researchers have used both
adopt the MIMOSA Open System Architecture for Condition- approaches simultaneously in order to leverage the advantages
Based Maintenance (OSA-CBM) [169], either completely, or from both of them. In this context, Luo et al. [171] have
only the main parts of it. proposed a DT-based PdM scheme that uses physics-based
For example, a variation of the OSA-CBM architecture is degradation and simulation models to generate theoretical
proposed in [119], where the integration of CPS, DT, and baselines for the machine state, as well as data-centric mod-
DL extensively rely on each other’s advantages to reduce els that consume real-time streams of data from the sensors
the need of human intervention usually required in a PdM installed in the machine. This work has been summarised in
scheme. In fact, the work suggests a great reliance on DL algo- Figure 10, and the results of the hybrid model outperformed
rithms to completely remove manual feature engineering from both physics-driven and data-driven models in predicting the
the PdM architect’s list of responsibilities. Instead, the intro- tool wear of a CNC Machine Tool.
duced framework claims DL algorithms can single-handedly In order to achieve satisfying accuracy, PdM schemes
manage state detection (through automatic feature extraction), that rely solely on Big Data more often than not require
health assessment, RUL estimation, and advisory generation massive amounts of historical failure data. For example,
through closed feedback loop connections between different Choi et al. [172] proposed a method for predicting the main-
functional layers of the architecture. Such an implementation tenance needs of an induction furnace with the help of neural
networks. Due to the nature of the ML algorithms used, as well wind turbine bearing by making use of analytical tools like
as the data-hungry aspect of PdM, the authors ended up using the wavelet transform (for de-noising and feature extraction),
a data set that consisted of measurements across 24 months. and Bayesian statistics for providing RUL predictions backed
Even then, the conclusion of the article admits that collecting by 90% credible intervals, whose accuracy would increase
data during faulty scenario was deemed nigh impossible, as over time. However the goal in I4.0 is to bring PdM to all
running experiments with a failed induction furnace could have equipment, including those that are too complex for statistical
been fatally dangerous. As such, the solution to this problem, modeling, which is why the data-driven approach has gained
according to the researchers, was to place focus in their future significant traction. And this approach, as the name states,
work on generating and simulating erroneous data, instead of requires large amounts of data that is not always available,
measuring it. In [164], Wang et al. introduced a two-part PdM especially in old machinery where maintenance records have
scheme for the China’s High-Speed Railway equipment using not been kept, or in equipment whose uptime is so crucial that
LSTM-RNN. As stated before, any DL approach requires no run-to-failure scenarios were allowed.
tremendous amounts of data for training. Of course, there is
plenty of data that can be generated by a nation-wide business VI. D IGITAL T WIN : C ASE S TUDIES
like China’s railway system, however it should be kept in mind
that the work’s target equipment, the Traction Power Supply This section will take a closer look at DT applications and
System, is engineered to be sturdy enough to have as few fail- services by detailing three DT case studies that represent main
ures as possible per year. As such, gathering historical failure research directions carried out at the London Digital Twin
data proved once again to be a prolonged challenge. In this Research Centre.1 The subsequent sections will thus sum-
direction, the authors proposed splitting the PdM framework marize the research goals, findings, challenges, and future
into two: a proactive maintenance system, and a predictive directions for each case study in part.
maintenance system. The proactive maintenance system deals
with analysing the failure modes of the physical asset to gen- A. A Look at the Tea Industry in India
erate new failure data through stochastic modeling. The PdM An important case study that we carried out is from a multi-
system is then trained using solely simulated data, and the national tea manufacturing company. It is a semi-automated
overall method is then validated using both simulated data manufacturing company involving machines and human beings
and field data. While it turns out that the model performs to control them, who bring in several inaccuracies in their
predictably better on artificial test data rather than on real processes. The tea bag manufacturing machine operated daily
data, the performance is still very good and it shows promise on an average of 20 hours and 2 hours of rest period. Fig. 11
in the direction of simulation-assisted PdM. Gugulothu et al. shows the snapshot of the different activities in the conveyor
in [173] invented an innovative and practical approach to RUL belt of the tea manufacturing company; starting with tea and
estimation. Their work proves to be robust to noise, sensor herbs, dosage, blending, etc. Notably, there is a separate place-
inter-dependencies across time, as well as data unavailabil- holder for filter paper along with its cutting, folding, etc. There
ity, which are all issues that are very present in various data are two transport wheels: upper and lower where the oper-
repositories in the industry. The proposed system uses “embed- ation of thread stitching and the cardboard using needle is
dings”, or rather, hidden features extracted by a RNN encoder being done. This different steps in the process can be clearly
after it was fed a fixed window-sized signal input from dif- understood from the different steps in the process of the tea
ferent sensors, to extract the Health Index which will then be manufacturing, as shown in Fig. 12. There are several sequen-
used to predict the RUL. The remaining life estimation is done tial steps and a relation between each step, right from adding
by comparing the extracted Health Index with its previously tea and herbs, or filter paper to packaging, checking the weight
seen values from the training data. The work delivers on its of each box and also investigating rejections and possible reuse
promises, providing a robust RUL estimator that can show of the materials.
good results, as long as enough training data is provided. There were 6 to 8 downtimes per day over a 24 hour
It can be noted that many works in the literature are period, each lasting around 15 minutes. The activities during
researching ways to compensate for the low amount of his- the downtime included: change of product, loading new feed
torical failure data made available by their target monitored of raw materials, equipment/mechanical failure, misalignments
physical system. It becomes apparent that the quality of the of material feed, effect of new product on machine, etc. Also,
data and, equally important, the quantity of data are extremely an accurate supervision of the facility during night shift is dif-
important boxes to check when developing failure prognosis ficult. The tea-bag manufacturing process was studied in detail
solutions. Over the years, the focus has shifted from model- to detect and keep record on their anomalies, observe any pat-
based PdM, where researchers came to the conclusion that terns, identify prospects for twin modelling and predicting the
developing a stochastic model for a complex system can maintenance needs of the machines. The seven major anoma-
be nigh impossible, towards data-driven approaches, where lies identified included: (1) Thread knot Anomaly; (2) Outer
an algorithm can learn from tremendous amounts of histor- envelope print not centered; (3) No filter bag in the outer enve-
ical failure data. And this shift in research direction can be lope; (4) Tag paper print not centered; (5) Outer envelope
justified. As an example of model-based PdM, where lim- paper missing; (6) Faulty filter paper tube; (7) Filter paper
ited amounts of data is not considered a hard constraint,
Wang et al. [174] proposed a fault prognosis system for 1 LDTRC website: https://dt.mdx.ac.uk/
Fig. 11. Conveyor Belt Snapshot of a Tea Manufacturing Plant.
Fig. 12. Steps in the Process of Tea Manufacturing.
slicing. Given the varying nature of the anomalies, they could 1) Analysis of individual data instances
be detected using multiple step dedicated anomaly detection 2) Evaluation of the relationship between neighboring
technique, typically, an N-step approach. points
One of the major concerns for a DT is the quality of data 3) Identification of the anomalies in dataset by the detection
being fed. The lack of noise-free data being administered algorithm
has severe consequences such as sub-standard performance 4) Evaluation of the performance of N-Step
of the twin [15]. This compels the DT to eliminate poten- The techniques selected for each step are as follows:
tial aberrations from the data in order to keep up with its Step 1: DBSCAN (Density-Based Spatial Clustering with
performance. The demand for noise-free data has paved a way Applications in Noise) [176]; Step 2: Isolation Forest; Step
for anomaly detection algorithms that would filter spurious 3: LOF (Local Outlier Factor) [177]; Step 4: KNN (K-Nearest
instances from datasets. The need for faster and more accurate Neighbour); Step 5: Hierarchical Classification based methods.
results obliges the algorithm to have an equilibrium between The reason for choosing the N-step approach in a particular
the two performance parameters: accuracy and execution time. way is as explained. DBSCAN would provide a very good
The service of anomaly detection integrated within a virtual separation of outliers from the overall data points, removing
twin demonstrates three of the DT’s main characteristic traits all false positives. Isolation Forest removes the few left over
enumerated in the definition we provided in Section II, namely: outliers that are located isolated from one another. Similarly,
self-adapting, self-monitoring, and self-diagnosing. Through LOF and other following techniques would remove only the
anomaly detection, the DT of the tea factory raises alarms boundary located nodes; thereby removing the True nega-
whenever its external environment changes in a way that the tives and false positives, if any and increasing the success
DT is not able to recognize, allowing the operator or other pre- ratio. A careful design of the N-step approach would result in
defined routines to handle the exceptional anomalies. At the higher accuracy/success ratio with a minimal increase in the
same time, implementing anomaly detection implies the exis- computation time.
tence of a monitoring mechanism that continuously ingests
data and checks for divergent behaviour within it. Lastly,
anomalies can be a sign of degradation in some of the physical B. Festo Cyber-Physical Factory
asset’s systems, so they could trigger pre-defined maintenance Aside from anomaly detection, another important aspect of
pipelines to address these system health issues. DTs in manufacturing, which is also the DT’s original pur-
In order to validate this, a two-step anomaly detection tech- pose, is continuous real-time monitoring of equipment and
nique was developed by Shetve et al. [175] and was evaluated processes. It is important that smart factory workers have
for its data pre-processing capability. access to always-available digital factory status reports that
The main drawback of the two-step approach was the dimin- are intuitive and remotely-accessible. This core feature of DTs
ishing accuracy with increasing outliers. Hence, the two-step enable equipment owners and factory executives to oversee
approach could be generalized to “N-step approach”. Each the good functioning of their products and processes. Besides
step would have its own technique that would be followed by actual operators, the monitoring service can be used by the DT
another technique in a sequential manner. The N-step approach itself in order to maintain awareness, at all times, of its physi-
has been developed based on following key aspects: cal asset’s current state and environment. This service enables
Fig. 13. Digital Twin in I4.0.
a core principle of the DT, detailed in Section II, namely the The kinematic model of this assembly line, which initially
ability to self-monitor. only mirrored one station of the CP-Lab, was later migrated to
On this premise, the work we conducted in [178] developed the Unity game engine, where the whole structure and motion
a DT framework of a real production line, in order to of the smart factory was modeled (right side of Figure 13).
establish a continuous monitoring mechanism for the kine- In addition, the virtual model captures streams of sensor data
matics of the factory. The physical twin in question is the flowing from the CP-Lab via the TCP protocol. This led to
Festo Cyber-Physical Factory for I4.0 (CP-Lab) located at the development of a DT-based PdM framework that includes
Middlesex University, a didactic model of an assembly line a monitoring dashboard for the machine’s sensors (temper-
for mock mobile phones. The smart factory is composed of ature and power data) [179]. The framework makes use of
six functional stations, each equipped with a Human-Machine real data coming from the CP-Lab, as well as configuration
Interface, and two transport (or bridge) stations. Figure 13 data stored on the DT, to better position the working regime,
depicts the physical twin on its left side, and it can be noted identify working stations, and assess the health of individual
that it is composed of two islands, each equipped with four stations via data pertaining to the whole island. More specif-
stations. On the first island, the first station is tasked with ically, the framework targets the health of the furnace station
placing the back plastic cover of the mock mobile phone of the second island. Being that it is equipped with a power-
onto the carrier, which is then carried to the next islands via ful heating element, the malfunctioning of this station could
a conveyor belt. The second station is the manual station, potentially lead to a fire hazard, so it stands to reason that
where a human operator will place a Printed Circuit Board guaranteeing its good health is a necessity. For this reason,
(PCB) onto the back cover. The third station visually inspects the furnace is also equipped with a “Safety Shutdown” mech-
the product automatically, to verify that the PCB that was anism, that will completely halt operations if the temperature
previously added corresponds to the correct order specifica- inside its chamber surpasses 80◦ C, however, this system is
tions. Lastly, a bridge station takes the carrier and passes it also not infallible. As such, the framework captures tempera-
on to an Automated Guided Vehicle (AGV) that transports ture data from inside the furnace, as well as the power data
it towards the second island with its next four stations. The pertaining to the whole second island, to predict if, or when,
second island is also equipped with a bridge station that inter- the Safety Shutdown mechanism will be triggered, in order
cepts the AGV and sends the carrier to the island’s second to proactively prevent it. The DT provides the configuration
station, which places another plastic cover on top of the prod- data of the heating station (i.e., its real-time state) to help
uct. The next station applies a pre-defined pressure onto the the framework extract the furnace’s power consumption from
product to seal the two plastic covers. Finally, the last com- the second island’s power measurements. As such, in case the
ponent of the assembly line is the furnace, where the product temperature sensor or element inside the heating chamber are
is heated up to a user-defined temperature to complete the faulty, the normal behaviour of the station can still be veri-
order. fied via its power data. The DT of the CP-Lab can increase
Fig. 14. Cloud Digital Twin Structural Health Monitoring Web application.
productivity by preventing the unnecessary or anomalous trig- uniformly distributed across the model. The SHM database
gering of the Safety Shutdown mechanism, as well as promote was empirically generated by hand-shaking the model; then
safety by predicting when the temperature inside the heating vibration data were collected through an array of accelerom-
chamber of the Tunnel Furnace Station reaches critical levels. eter sensors MPU-6050 and a microcontroller board Arduino
Uno. The damaged states of the model were introduced by
C. Structural Health Monitoring for Vietnam bridges randomly removing one or two truss rods. After that, two data
In this section, the Digital Twin framework for SHM analytic algorithms, including a lite mathematical model and a
developed at London Digital Twin Research Centre, dubbed ML-based model are developed to detect the structure status.
cDTSHM is presented, including its main components, case Furthermore, the latter could spatially localize the damage’s
studies, and its application to bridges, mainly in Vietnam. location and quantize the damage severity.
The cDTSHM consists of four components: the real structures The second case study used to validate the correctness of the
equipped with sensors along their body providing data related framework is a simplified laboratory model of a stayed-cable
to their operational services, a fog layer with local compu- bridge whose physical and numerical models are shown in the
tational servers preprocessing measured data, a cloud layer corner of Fig. 14. The SHM procedure is realized similarly to
involving data storage services, and data analytic components the first example, but the experiment data, including excita-
leveraging both mathematical models and machine learning- tion, vibration data, structure’s deformation are controlled and
based model, and a Web application visualizing the data and measured more rigorously. The model is excited by introduc-
computed results. The cDTSHM is developed with the help ing an impulse force of very short duration through an impact
of the AWS cloud services; most of the programs are written hammer; the structure’s status is then assessed based on the
in Python, the ML-based model is implemented using the DL loss level of prestressing strands which can be manually mod-
library Pytorch, and the main input data for the framework are ified by alternating the anchor bolts tightness. As a result, the
the vibration data from accelerometer sensors. Fig. 14 depicts sDTSHM can provide highly accurate SHM results with much
the interface of the cDTSHM, including case studies carried less CPU time while bypassing the cumbersome preprocessing
out. modal analysis as in the conventional structural identification
The first case study to demonstrate the applicability of the methods.
framework is a toy model of the Sydney Harbour bridge built For the third case study, the performance of the SHM frame-
by using K’nex interlocking plastic rods. Next, one manu- work is tested with real data collected from the Z24 bridge in
ally excites the model by hand-shaking, where its vibration is Switzerland [180]. For such a real structure, using a lite math-
recorded by using a set of accelerometer sensors MPU-6050 ematical model or shallow ML cannot provide reliable SHM
results. Thus, a highly modular architecture has been devised, One important aspect that must be factored into the costs
allowing switching different DL algorithms and combining when planning for DTs, is the ongoing maintenance require-
data from different sensors in a straightforward way [181]. The ments, such as: software updates that affect the DT, changes
results demonstrate that the data analytic component of the to the physical asset, etc. Considering the total life cycle of
cDTSHM outperforms competing methods with a structural the DTs, one could anticipate that the maintenance and man-
damage detection result of 90.1% with low time complexity agement cost might represent the largest proportion of the
and budget memory storage. investment costs.
Afterwards, the cDTSHM framework is applied to the Nam
O railway bridge located in central Vietnam. The bridge is
60 years old and constantly undergoes unfavorable factors B. Social and Ethical Challenges
involving the corrosive maritime environment, dynamic and Digital Twin technology and applications are experienc-
heavy train loadings, etc. From the structural perspective, ing a shift from engineering/physics based domains, where
the structure’s mode shapes and their high-order derivatives closed equations are an appropriate modelling abstraction,
are sensitive to damages; therefore, a knowledge-enhanced to one where the problem domain is socio-technical, lead-
deep 1D-CNN has been developed for automatically extracting to the notion of Socio-Technical Digital Twins (STDT).
ing modal characteristics from raw vibration data to accurately Such a problem domain utilises systems that comprise com-
detect and quantify connection stiffness reductions. The real- plex interaction between humans, machines and the work
ization steps and implementation details can be found in [182]. environment [183]. This class of system is characterised by
The results show that the framework could achieve accuracy heterogeneous networked agents, adaptive and goal oriented
up to 95% even with minor damage (5% of stiffness reduc- with respect to the environment and joint optimisation and
tion) with faster convergence speed and more stable results evolution of both technical and social systems [184]. These
than counterparts, including the Multi-Layer Perceptron and properties are also those that are characterised by agent based
other DL architectures. systems and make such technologies ideal for representation of
The four case studies illustrated in this section demon- STDTs despite computational cost [184], [185]. Hence, STDTs
strate that the enabling technologies supporting a DT will vary generate new and different research challenges.
greatly, depending on the DT’s use-case, as also mentioned in DTs of phenomena that include human interactions and
our definition, provided in Section II. Additionally, the cDT- behaviours acquire complexity simply due to the involvement
SHM framework has been developed to exhibit self-diagnosing of multiple disciplines. For example, DT models of cities for
and self-monitoring capabilities within several case studies of monitoring pandemic behaviour have included social geogra-
bridges. phers, economists, medical practitioners as well as computer
scientists. Arriving at a shared understanding, common lan-
VII. L ESSONS L EARNED , R ESEARCH C HALLENGES guage and a way of working demands new methodological
AND F UTURE D IRECTIONS approaches as well as intuitive access to underpinning theory
from different disciplines [186].
The previous sections provided a comprehensive view
Moving away from closed equations to systems that model
of the DT, commencing with its definition, market poten-
emergent behaviour presents expected validation challenges.
tial, enabling technologies, frameworks and applications, and,
Recognising that a STDT has purpose beyond prediction such
finally, three case studies. Throughout the literature surveyed
as explanation is the first response [187]. STDTs can be
in this manuscript, as well as our own experience in develop-
used for discovering new questions, demonstrating trade-offs
ing DTs, we have learned important lessons and encountered
or experimentation with prevailing theories that lack empiri-
significant challenges that will contour future directions for us
cal understanding. Hence they represent a move away from
and the research community. This section will delve into the
closed form analytical models. Models are a form of the-
most significant of these concepts, detailing the obstacles that
ory building [188] and as such they can only be invalidated
the DT needs to overcome in order to realize its potential.
so a more useful target is a form of accredited or accepted
model based on standardised criteria and metrics [189]. As
A. Investment Costs STDTs become more established, policy oriented domain-
As mentioned in Section III, businesses still remain reluctant specific practice could lead to libraries of accepted models,
to implement the DT because of its envisioned development encoding existing knowledge, that do not need to change
costs and difficult-to-quantify ROI. As a matter of fact, it is and are much less volatile. Given the emergent properties
rather challenging to put a price on the DT because of its of STDTs, such models need to be defined at both micro,
multi-disciplinary nature and use-case-specific particularities. meso and macro levels. Building libraries of STDT models is
Additionally, the DT is rarely a product that generates direct reminiscent of component based development practice and its
profit, since its core philosophy focuses primarily on saving inherent challenges [190].
costs. With the exception of DT solution providers and the Validation of a model for STDT is closely related to abstrac-
healthcare industry, where the DT can indeed be a source tion concerns and in particular the challenges that arise from
of revenue, other entrepreneurs will need detailed and long- establishing an appropriate framing structure. The complex-
term plans of investment that emphasize the merits of DT ity, multi-level and range of modelling required to represent a
development before diving into such expenditures. socio-technical problem domain within a DT require choices
to be made in determining the scope and detail of the envi- multiple industries. These complexities could be eliminated
ronment to be modelled. This choice we can refer to it as through standardisation.
a conceptual problem frame. For example, Barat et al. [186] Recently, one of the subcommittees (SC 41)2 of the
in their city DT of Pune, for modelling the COVID-19 pan- Joint International Organization for Standardization (ISO) and
demic, observe that existing agent based systems for pandemic International Electrotechnical Commission (IEC) Technical
modelling do not show sufficient granularity of types of peo- Committee (JTC 1)3 has widened its scope and terms of refer-
ple and their movements within the city, raising concerns of ences to include DT, looking now into standardization in the
model completeness. area of IoT and DTs, including their related technologies.
Perhaps the most striking challenge that needs to be The ISO 23247-14 standard for DT framework for man-
addressed is that arising from STDTs that include ML or ufacturing is currently under development. The first part
other algorithmic decision making. Ethical concerns come of the standard provides general principles and defines the
to the fore when (1) conclusions drawn from inference are requirements for developing DTs in manufacturing.
probable and therefore an epistemic limitation; (2) traceabil- Similarly, the National Institute of Standards and
ity between the input data and conclusion is not accessible Technology (NIST) in an attempt to standardize the DT
and open to critique; (2) conclusions are dependent upon the technology, have released a draft NISTIR 8356 [194] cov-
quality of data; or (4) the actions based on conclusions are dis- ering the definition, common low-level operations, usage
criminatory even if well-founded [191]. Underpinnning these scenarios, and use-cases examples.
epistemically-based ethical issues is the encoding of value Another initiative that has as a primary objective to influ-
systems such as privacy, transparency, security and so on. ence the requirements for DT standards is the Digital Twin
Understanding value sensitive concerns and related approaches Consortium.5 The consortium consists of members from indus-
that explore more fundamentally the nature of social require- try, government, and academia that form a global ecosystem
ments and (unintended social impacts) of software remains aiming to accelerate the development, adoption, interoperabil-
an ongoing project in software engineering [192] and requires ity, and security of DTs.
study in the context of STDTs. One proposed solution by Harper et al. [195] could be to
define a set of standardised Application Process Interfaces
(APIs) that could evolve over time. The advantage of this
C. Fidelity and Rate of Synchronization approach is that different DTs could be developed using differ-
A common misconception about the DT is that the vir- ent software and processes as long as they support the defined
tual twin should reflect the physical twin in its entirety, and set of APIs.
that it should gather and process all of its data in almost Microsoft developed the Digital Twin Definition Language6
real-time. However, these feats are not currently feasible, and (DTDL) that is used in their commercial services, such as IoT
certainly not always necessary. As specified in our definition Hub, IoT Central and Azure Digital Twins. However, DTDL
of the DT, provided in Section II, the virtual representation’s does not resource discovery and access and deals with resource
fidelity and rate of synchronization are specific to the DT’s description only.
use-cases. For instance, for ambitious, nation-wide project like Consequently, the current lack of standardised approaches
the U.S. Air Force’s DT for weapon system development, the when modelling digital twins open up new challenges when
required fidelity and responsiveness might impose prohibitive dealing with their interoperability in order to maximize the
costs [193]. On the other hand, for the purposes of traffic relief, interconnectivity.
a DT that stores the coordinates and synchronization rates of
traffic lights, as well as the real-time traffic density, might
E. Data Ownership and Governance
arguably perform almost as well as a DT that completely 3D
models the city’s infrastructure. As such, the DT’s granularity Apart from standardisation, another closely related chal-
and twinning rate requirements can be more, or less, lenient, lenge is the data ownership and governance, that brings a new
depending on its applications. The more stringent demands question, specifically: Who owns the DT information? One
for real-time, granular mirroring could be encountered in the can anticipate that with the advancements in technologies, the
healthcare industry, for scenarios where the DT is used to industry is moving towards a connected data ecosystem of DTs
facilitate remote surgery. with potential different owners of the physical assets as well
as the DTs. Additionally, the potential existence of an ecosys-
tem of DTs implies the need for a shared communication
D. Standardisation Efforts framework for standardized data exchange between multiple
One of the most important features that could acceler- heterogeneous data sources. This scenario brings technical,
ate the adoption of DTs within various industries is their financial and legal aspects challenges that need to be clarified.
modularity. This could enable the rapid reproduction of DT
2 SC 41: https://bit.ly/3tf0hSM.
processes and their components. However, this dynamic envi-
3 JTC 1: https://jtc1info.org/.
ronment could become very complex, with different digital 4 ISO 23247-1 Digital Twin Framework for Manufacturing: https://www.
twins custom built for different purposes, specific equipment iso.org/standard/75066.html.
type, specific manufacturers, etc., these represent factors that 5 Digital Twin Consortium: https://www.digitaltwinconsortium.org.
could inhibit the adoption and implementation of DTs across 6 Digital Twin Definition Language: https://bit.ly/3jJ3EOV.
One solution would be to adopt the Industrial Data Space physical twin would be carried out exclusively via the virtual
(IDS)7 concept that has been introduced in [196] to cater for model, through a synchronization gateway that filters all traf-
all these issues within the I4.0 framework. The IDS concept fic to the real twin, effectively isolating it from bad agents.
represents a virtual data space that enforces data ownership A similar approach was taken in [202], where the authors
within a distributed environment, based on open standards and proposed using the DT to filter out incoming commands to
existing technologies as well as common governance models for a smart inverter and ensure that only the non-malicious ones
data economy. As DTs are seen as part of I4.0, the IDS model are carried out.
could be the answer for DTs data ownership and governance. In
parallel, an effort for constructing an Information Management G. Artificial General Intelligence, Beyond Human
Framework (IMF) for the National Digital Twin is carried Performance
out in [197], where the goal is to create a common national In Section II of this manuscript, we have described the
information resource that can sustain a country-wide DT. DT as a self-adapting, self-regulating, self-monitoring, and
self-diagnosing system-of-systems, a definition which places
F. Data Security it under the span of another, broader category of human ambi-
There are two ways to approach the discussion on security tions: Artificial Intelligence. In fact, the idea of self-improving
issues in DTs. The first one addresses the security of the DT artificial systems was part of the original proposal made by
itself, starting from the physical servers that host the DTs up McCarthy et al. in 1955 for the Darthmouth Summer Research
until the safety and integrity of the software and data com- Project on AI, a project which is also sometimes dubbed as the
munication links that animate the DT. The second approach is birthplace of AI [203]. Since then, significant progress has been
about how the DT itself can provide security to its real twin, carried out towards realizing the objectives laid out in said pro-
as an additional valuable offering. This section will touch on posal, and the term “AI” has now become so popular that some
both of these aspects of security within DTs. voices have raised concerns about the implications of integrating
One of the central components of a DT is the communica- AI into our society [204]. However, as Shevlin et al. emphasize
tion medium that enables the symbiotic relationship between in [205], there is an important distinction to be made between
the physical and virtual twins. This link effectively transports the original meaning assigned to AI by McCarthy et al. in 1955,
all the data between the two entities, so it stands to reason and AI as it is understood and marketed today. McCarthy’s
that it needs to guarantee impeccable data security. Every proposal identified AI as a machine that behaves “in ways
time data flows to, and from, the real twin, or in-between that would be called intelligent if a human were so behaving”,
the servers hosting the DT itself, the risk of losing impor- whereas nowadays, AI is sometimes used to refer to systems
tant information is high, which calls for increased attention to that reach or surpass human performance in specific tasks [206],
preserving data integrity [198]. This communication medium [207]. The former interpretation corresponds to what is today
becomes a potential area of weakness in front of data corrup- understood as Artificial General Intelligence (AGI), while the
tion and theft, and it can create disturbances for businesses. latter is an appropriate example of Artificial Narrow Intelligence
As such, data security principles, like privacy, authentication, (ANI) [205].
integrity, and traceability, need to be taken into account during However, although there is a great overlap between the
DT development. Some important measures that provide secu- DT and AGI, there are some differences that stand out and
rity features are data encryption, access privileges, source code can make the DT an even more challenging task than AGI.
automated scanning, penetration testing, and routine check- First, the versatility of the DT paradigm across industries
ups [199]. Emerging approaches to deal with these issues implies that it will, by definition, have a specialized kind
include using blockchain technologies to ensure data privacy of intelligence that cannot be generalized to other domains,
in the communication between DT systems-of-systems [94]. but trains and excels in every possible task that pertains to
Yaqoob et al. [200] conducted an extensive research on how the physical twin’s use cases, including hypothetical scenar-
blockchain has been integrated in the DTs across literature to ios. In order for the DT to become truly self-evolving, just
also ensure trust and transparency in various use cases. like humans and animals are, it needs to be able to imple-
Given that the DT is still an emerging technology, the ment a level of creativity that can make maximal use of
related literature still lacks a consensus on its security require- its physical twin’s unique features (i.e., learn to become as
ments. As we have discovered throughout this research work, resourceful with its physical structure as an animal is with its
the use case and services envisioned for the DT can dictate body). This implies that, just like there is a need for stan-
the architecture of the virtual twin. According to Gehrmann dardized benchmarks for validating human AGI, or animal
and Gunnarsson in [201], security measures should also be at intelligence-mimicking AGI [205], there is also a requirement
the forefront of the DT architects’ minds since they can have for performance measures and benchmarks that can accurately
a significant impact on the final DT model’s structure as well. evaluate the DT for each industry and application. Validation
The authors also introduced the idea of using the DT as an metrics aside, while current AI systems can definitely learn to
enabler of security in the communication between the physi- perform tasks even beyond human-level performance, they also
cal twin and other cloud-based services, of which DTs might lack comprehension, and therefore cannot offer transparency
make use. In their work, all external communication with the into their “reasoning” [208]. This lack of transparency invokes
skepticism, and can even impede development of DTs or AIs
7 Industrial Data Space: https://internationaldataspaces.org. due to lack of trust.
VIII. C ONCLUSION [5] M. E. Rose and J. R. Kitchin, “Pybliometrics: Scriptable bibliometrics

using a python interface to Scopus,” SoftwareX, vol. 10, Jul.–Dec. 2019,
This paper took an in-depth look at the existing and expand- Art. no. 100263.
ing Digital Twin-related literature, and we drew some lessons [6] B. R. Barricelli, E. Casiraghi, and D. Fogli, “A survey on digital twin:
that will help the researchers in this field consolidate their Definitions, characteristics, applications, and design implications,”
IEEE Access, vol. 7, pp. 167653–167671, 2019.
understanding of the DT and choose future directions that
[7] R. Minerva, G. M. Lee, and N. Crespi, “Digital twin in the IoT context:
need further development. In Section II, we took a look at A survey on technical features, scenarios, and architectural models,”
the variety of DT definitions that the papers reviewed in this Proc. IEEE, vol. 108, no. 10, pp. 1785–1824, Oct. 2020.
manuscript provided and we clustered them together according [8] A. Löcklin, M. Müller, T. Jung, N. Jazdi, D. White, and M. Weyrich,
“Digital twin for verification and validation of industrial automation
to their similarity; then, we introduced a new, comprehensive systems—A survey,” in Proc. 25th IEEE Int. Conf. Emerg. Technol.
definition for the DT that covers important aspects that the Factory Autom. (ETFA), vol. 1, 2020, pp. 851–858.
others might have missed. In Section III, we evaluated the [9] F. Biesinger, B. Kraß, and M. Weyrich, “A survey on the necessity for
a digital twin of production in the automotive industry,” in Proc. IEEE
current level of adoption of the DT in the industry, and how 23rd Int. Conf. Mechatronics Technol. (ICMT), 2019, pp. 1–8.
this paradigm is currently valued by various leading busi- [10] Y. He, J. Guo, and X. Zheng, “From surveillance to digital twin:
nesses in this field. Section IV then dove into the technical Challenges and recent advances of signal processing for Industrial
aspects of the DT, namely its enabling technologies, and how Internet of Things,” IEEE Signal Process. Mag., vol. 35, no. 5,
pp. 120–129, Sep. 2018.
they are being used to support the DT in delivering business [11] F. Pires, A. Cachada, J. Barbosa, A. P. Moreira, and P. Leitão, “Digital
value across various industries. Consequently, in Section V, twin in industry 4.0: Technologies, applications and challenges,” in
we described how the previously presented technologies help Proc. IEEE 17th Int. Conf. Ind. Inf. (INDIN), vol. 1, 2019, pp. 721–726.
the DT in providing a handful of services, ranging from smart [12] A. Rasheed, O. San, and T. Kvamsdal, “Digital twin: Values, chal-
lenges and enablers from a modeling perspective,” IEEE Access, vol. 8,
manufacturing to new generation communication technologies. pp. 21980–22012, 2020.
Section VI offered a closer look at three DT case studies in [13] F. Tao, H. Zhang, A. Liu, and A. Y. C. Nee, “Digital twin in indus-
order to present how the technology can be leveraged and what try: State-of-the-art,” IEEE Trans. Ind. Informat., vol. 15, no. 4,
pp. 2405–2415, Apr. 2019.
requirements and limitations appear, based on the use case.
[14] Z. Huang, Y. Shen, J. Li, M. Fey, and C. Brecher, “A survey on AI-
Finally, Section VII summarized the lessons we have learned driven digital twins in industry 4.0: Smart manufacturing and advanced
throughout this review, in order to point out interesting new robotics,” Sensors, vol. 21, no. 19, p. 6340, 2021. [Online]. Available:
challenges to be addressed in the research community. https://www.mdpi.com/1424-8220/21/19/6340
[15] A. Fuller, Z. Fan, C. Day, and C. Barlow, “Digital twin: Enabling
Overall, the DT finds itself accelerating in full force towards technologies, challenges and open research,” IEEE Access, vol. 8,
I4.0, and its endless perceived potential makes it a central and pp. 108952–108971, 2020.
evermore popular player in the race. Its enabling technolo- [16] NASA. “Apollo expeditions to the moon: Chapter 13.” Accessed:
Jun. 22, 2021. [Online]. Available: https://www.hq.nasa.gov/office/pao/
gies are continuously evolving, and each step towards their History/SP-350/ch-13-1.html
improvement brings us closer to making true DTs a reality. [17] K. Dröder, P. Bobka, T. Germann, F. Gabriel, and F. Dietrich, “A
As the DT attracts research interest, the number of attempts at machine learning-enhanced digital twin approach for human–robot-
developing it contour some common and persisting obstacles. collaboration,” Procedia CIRP, vol. 76, pp. 187–192, Jan. 2018,
doi: 10.1016/j.procir.2018.02.010.
In the era of AI, the focus falls on the data, and the DT finds [18] M. Xia, H. Shao, D. Williams, S. Lu, L. Shu, and C. W. de Silva,
itself in the middle of an information loop: it needs to be fed “Intelligent fault diagnosis of machinery using digital twin-assisted
carefully considered data to power its complex ML algorithms, deep transfer learning,” Rel. Eng. Syst. Safety, vol. 215, Nov. 2021,
Art. no. 107938.
and then it further allows a better understanding of that data
[19] Q. Qi, D. Zhao, T. W. Liao, and F. Tao, “Modeling of cyber-physical
via its interactive and predictive feats. With some challenging systems and digital twin based on edge computing, fog computing and
puzzles in the way, the DT is steadily heading towards the cloud computing towards smart manufacturing,” in Proc. Int. Manuf.
automation of industrial processes. Sci. Eng. Conf., vol. 51357, Jun. 2018, Art. no. V001T05A018.
[20] V. Souza, R. Cruz, W. Silva, S. Lins, and V. Lucena, “A digital twin
architecture based on the Industrial Internet of Things technologies,”
in Proc. IEEE Int. Conf. Consum. Electron. (ICCE), 2019, pp. 1–2.
[21] J. Leng, H. Zhang, D. Yan, Q. Liu, X. Chen, and D. Zhang, “Digital
R EFERENCES twin-driven manufacturing cyber-physical system for parallel con-
[1] L. Laura, O. Clayton, S. Jeremy, and S. Kate, How COVID-19 trolling of smart workshop,” J. Ambient Intell. Humanized Comput.,
Has Pushed Companies Over the Technology Tipping Point and vol. 10, no. 3, pp. 1155–1166, 2019.
Transformed Business Forever, McKinsey Company, New York, NY, [22] S. Aheleroff, X. Xu, R. Y. Zhong, and Y. Lu, “Digital twin as a service
USA, 2020. (DTaaS) in industry 4.0: An architecture reference model,” Adv. Eng.
[2] S. Blackburn, L. LaBerge, C. O’Toole, and J. Schneider. “Digital Inf., vol. 47, Jan. 2021, Art. no. 101225.
Strategy in a Time of Crisis.” Apr. 2020. [Online]. Available: [23] K. Xia et al., “A digital twin to train deep reinforcement learning
https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/ agent for smart manufacturing plants: Environment, interfaces and
digital-strategy-in-a-time-of-crisis intelligence,” J. Manuf. Syst., vol. 58, pp. 210–230, Jan. 2021.
[3] Cisco. “Cisco Annual Internet Report (2018–2023) White Paper.” 2020. [24] P. Wang and M. Luo, “A digital twin-based big data virtual and real
Accessed: Sep. 12, 2021. [Online]. Available: https://www.cisco.com/ fusion learning reference framework supported by industrial Internet
c/en/us/solutions/collateral/executive-perspectives/Annu.-internet-Rep./ towards smart manufacturing,” J. Manuf. Syst., vol. 58, pp. 16–32,
whitepaper-c11-741490.html Jan. 2021.
[4] M. Grieves and J. Vickers, “Digital twin: Mitigating unpre- [25] W. Xu, J. Cui, L. Li, B. Yao, S. Tian, and Z. Zhou, “Digital twin-
dictable, undesirable emergent behavior in complex systems,” in based industrial cloud robotics: Framework, control approach and
Transdisciplinary Perspectives on Complex Systems: New Findings and implementation,” J. Manuf. Syst., vol. 58, pp. 196–209, Jan. 2021.
Approaches, F.-J. Kahlen, S. Flumerfelt, and A. Alves, Eds. Cham, [26] Y. Liu et al., “A novel cloud-based framework for the elderly
Switzerland: Springer Int., 2017, pp. 85–113. [Online]. Available: Healthcare services using digital twin,” IEEE Access, vol. 7,
https://doi.org/10.1007/978-3-319-38756-7_4 pp. 49088–49101, 2019.
[27] C. Liu, L. Le Roux, C. Körner, O. Tabaste, F. Lacan, and S. Bigot, [50] E. Y. Song, M. Burns, A. Pandey, and T. Roth, “IEEE 1451 smart
“Digital twin-enabled collaborative data management for metal addi- sensor digital twin federation for IoT/CPS research,” in Proc. IEEE
tive manufacturing systems,” J. Manuf. Syst., vol. 62, pp. 857–874, Sensors Appl. Symp. (SAS), 2019, pp. 1–6.
Jan. 2022. [51] S. M. Bazaz, M. Lohtander, and J. Varis, “5-Dimensional defini-
[28] Z. Zhao, L. Shen, C. Yang, W. Wu, M. Zhang, and G. Q. Huang, tion for a manufacturing digital twin,” Procedia Manuf., vol. 38,
“IoT and digital twin enabled smart tracking for safety management,” pp. 1705–1712, Jan. 2019.
Comput. Oper. Res., vol. 128, Apr. 2021, Art. no. 105183. [52] E. Negri, L. Fumagalli, C. Cimino, and M. Macchi, “FMU-
[29] C. Liu, P. Jiang, and W. Jiang, “Web-based digital twin modeling and supported simulation for CPS digital twin,” Procedia Manuf., vol. 28,
remote control of cyber-physical production systems,” Robot. Comput. pp. 201–206, Jan. 2019.
Integr. Manuf., vol. 64, Aug. 2020, Art. no. 101956. [53] E. Negri, H. D. Ardakani, L. Cattaneo, J. Singh, M. Macchi, and J. Lee,
[30] X. He, Q. Ai, R. C. Qiu, and D. Zhang, “Preliminary exploration on “A digital twin-based scheduling framework including equipment
digital twin for power systems: Challenges, framework, and applica- health index and genetic algorithms,” IFAC PapersOnLine, vol. 52,
tions,” 2019, arXiv:1909.06977. no. 10, pp. 43–48, 2019.
[31] Y. Xu, Y. Sun, X. Liu, and Y. Zheng, “A digital-twin-assisted [54] M. Glatt, C. Sinnwell, L. Yi, S. Donohoe, B. Ravani, and J. C. Aurich,
fault diagnosis using deep transfer learning,” IEEE Access, vol. 7, “Modeling and implementation of a digital twin of material flows
pp. 19990–19999, 2019. based on physics simulation,” J. Manuf. Syst., vol. 58, pp. 231–245,
[32] W. Hofmann and F. Branding, “Implementation of an IoT-and cloud- Jan. 2021.
based digital twin for real-time decision support in port operations,” [55] Q. Min, Y. Lu, Z. Liu, C. Su, and B. Wang, “Machine learning based
IFAC PapersOnLine, vol. 52, no. 13, pp. 2104–2109, 2019. digital twin framework for production optimization in petrochemical
[33] L. López-Estrada, M. Fajardo-Pruna, S. Gualoto-Condor, J. Ríos, and industry,” Int. J. Inf. Manag., vol. 49, pp. 502–519, Dec. 2019.
A. Vizán, “Creation of a micro cutting machine tool digital-twin using [56] H. Laaki, Y. Miche, and K. Tammi, “Prototyping a digital twin for
a cloud-based model-based PLM platform: First results,” Procedia real time remote control over mobile networks: Application of remote
Manuf., vol. 41, pp. 137–144, Jan. 2019. surgery,” IEEE Access, vol. 7, pp. 20325–20336, 2019.
[34] Y. Pan, T. Qu, N. Wu, M. Khalgui, and G. Huang, “Digital twin [57] D. Guo, R. Y. Zhong, P. Lin, Z. Lyu, Y. Rong, and G. Q. Huang,
based real-time production logistics synchronization system in a multi- “Digital twin-enabled graduation intelligent manufacturing system
level computing architecture,” J. Manuf. Syst., vol. 58, pp. 246–260, for fixed-position assembly islands,” Robot. Comput. Integr. Manuf.,
Jan. 2021. vol. 63, Jun. 2020, Art. no. 101917.
[35] E. P. Hinchy, N. P. O’Dowd, and C. T. McCarthy, “Using open- [58] M. Matulis and C. Harvey, “A robot arm digital twin utilising reinforce-
source microcontrollers to enable digital twin communication for smart ment learning,” Comput. Graph., vol. 95, pp. 106–114, Apr. 2021.
manufacturing,” Procedia Manuf., vol. 38, pp. 1213–1219, Jan. 2019.
[59] P. Zheng and A. S. Sivabalan, “A generic tri-model-based approach
[36] F. Tao, Q. Qi, L. Wang, and A. Nee, “Digital twins and cyber–physical
for product-level digital twin development in a smart manufacturing
systems toward smart manufacturing and industry 4.0: Correlation and
environment,” Robot. Comput. Integr. Manuf., vol. 64, Aug. 2020,
comparison,” Engineering, vol. 5, no. 4, pp. 653–661, 2019.
Art. no. 101958.
[37] M. Borth, J. Verriet, and G. Muller, “Digital twin strategies for SoS 4
[60] “What is a digital twin?” IBM. Accessed: Jan. 26, 2022. [Online].
challenges and 4 architecture setups for digital twins of SoS,” in Proc.
Available: https://www.ibm.com/uk-en/topics/what-is-a-digital-twin
IEEE 14th Annu. Conf. Syst. Syst. Eng. (SoSE), 2019, pp. 164–169.
[61] A. Mehra. “Digital Twin Market Worth $48.2 Billion by 2026.”
[38] E. Negri, S. Berardi, L. Fumagalli, and M. Macchi, “MES-integrated
Accessed: Jun. 22, 2021. [Online]. Available: https://www.
digital twin frameworks,” J. Manuf. Syst., vol. 56, pp. 58–71, Jul. 2020.
marketsandmarkets.com/PressReleases/digital-twin.asp
[39] R. Dong, C. She, W. Hardjawana, Y. Li, and B. Vucetic, “Deep learn-
ing for hybrid 5G services in mobile edge computing systems: Learn [62] IBM. “Hong Kong Applied Science and Technology Research Institute
from a digital twin,” IEEE Trans. Wireless Commun., vol. 18, no. 10, Company Limited.” Accessed: Jun. 22, 2021. [Online]. Available:
pp. 4692–4707, Oct. 2019. https://www.ibm.com/case-studies/hong-kong-astri-engineering-
rhapsody
[40] T. Zohdi, “A machine-learning framework for rapid adaptive
digital-twin based fire-propagation simulation in complex environ- [63] IBM. “Profile: UCSF Health and Maximo Lead the Way on Smart
ments,” Comput. Methods Appl. Mech. Eng., vol. 363, May 2020, Medical Buildings.” Accessed: Jun. 22, 2021. [Online]. Available:
Art. no. 112907. https://www.ibm.com/blogs/internet-of-things/iot-ucsf-health-and-
[41] T. Ritto and F. Rochinha, “Digital twin, physics-based model, and maximo-smart-medical-buildings/
machine learning applied to damage detection in structures,” Mech. [64] Ansys. “Digital Twins: Making the Vision Achievable—Ansys White
Syst. Signal Process., vol. 155, Jun. 2021, Art. no. 107614. Paper.” Accessed: Jun. 22, 2021. [Online]. Available: https://www.
[42] S. Chakraborty and S. Adhikari, “Machine learning based digital twin ansys.com/resource-center/white-paper/digital-twins-making-vision-
for dynamical systems with multiple time-scales,” Comput. Struct., achievable
vol. 243, Jan. 2021, Art. no. 106410. [65] F. Glas. “Virtual Quality Assurance in the Digital Orthopaedic
[43] M. A. Ali, Q. Guan, R. Umer, W. J. Cantwell, and T. Zhang, “Deep Workshop of the Future—Ansys Case Study.” Accessed: Jun. 22, 2021.
learning based semantic segmentation of µCT images for creating dig- [Online]. Available: https://www.ansys.com/resource-center/case-study/
ital material twins of fibrous reinforcements,” Composites A Appl. Sci. mecuris-virtual-quality-assurance-digital-orthopaedic-workshop-future
Manuf., vol. 139, Mar. 2020, Art. no. 106131. [66] R. Damian. “ANSYS + JET Towers.” Accessed: Jun. 22, 2021.
[44] B. Maschler, D. Braun, N. Jazdi, and M. Weyrich, “Transfer learning [Online]. Available: https://www.ansys.com/content/dam/product/3d-
as an enabler of the intelligent digital twin,” Procedia CIRP, vol. 100, design/aim/case-study-jet-towers.pdf
pp. 127–132, Jan. 2021. [67] G. P. D. Solutions. “GE Digital Twin.” Accessed: Jun. 22,
[45] Z. Wang et al., “A digital twin paradigm: Vehicle-to-cloud based 2021. [Online]. Available: https://www.ge.com/digital/sites/default/
advanced driver assistance systems,” in Proc. IEEE 91st Veh. Technol. files/download_assets/Digital-Twin-for-the-digital-power-plant-.pdf
Conf. (VTC-Spring), 2020, pp. 1–6. [68] Spirent. “5G Network Digital Twin—Spirent.” Accessed: Jun. 22, 2021.
[46] P. K. Rajesh, N. Manikandan, C. S. Ramshankar, T. Vishwanathan, [Online]. Available: https://www.spirent.com/assets/wp_simplifying-
and C. Sathishkumar, “Digital twin of an automotive brake pad for 5g-with-the-network-digital-twin
predictive maintenance,” Procedia Comput. Sci., vol. 165, pp. 18–24, [69] R. Jamnadass. “STL Partners: The Relationship Between 5G and
Jan. 2019. Digital Twins in Industry 4.0—Spirent.” Accessed: Jun. 22, 2021.
[47] R. Revetria, F. Tonelli, L. Damiani, M. Demartini, F. Bisio, and [Online]. Available: https://www.spirent.com/assets/the-relationship-
N. Peruzzo, “A real-time mechanical structures monitoring system between-5g-and-digital-twins-in-industry-4-0
based on digital twin, IoT and augmented reality,” in Proc. IEEE Spring [70] D. A. Institute. “Predictive Maintenance.” Accessed: Jun. 22, 2021.
Simulat. Conf. (SpringSim), 2019, pp. 1–10. [Online]. Available: https://www2.deloitte.com/content/dam/Deloitte/
[48] G. White, A. Zink, L. Codecá, and S. Clarke, “A digital twin smart city de/Documents/deloitte-analytics/Deloitte_Predictive-Maintenance_
for citizen feedback,” Cities, vol. 110, Mar. 2021, Art. no. 103064. PositionPaper.pdf
[49] A. Burger, C. Cichiwskyj, S. Schmeißer, and G. Schiele, “The elastic [71] Markets and Markets. “Digital Twin Market by Technology,
Internet of Things—A platform for self-integrating and self-adaptive Type, Application, Industry—COVID-19 Impact Analysis.” Accessed:
IoT-systems with support for embedded adaptive hardware,” Future Jun. 22, 2021. [Online]. Available: https://www.marketsandmarkets.
Gener. Comput. Syst., vol. 113, pp. 607–619, Dec. 2020. com/Market-Reports/digital-twin-market-225269522.html
[72] A. B. S. Evans, C. Savian, and C. Cooper. “Digital Twins for the Built [95] O. Cardin, “Classification of cyber-physical production systems appli-
Environment.” Accessed: Jun. 22, 2021. [Online]. Available: https:// cations: Proposition of an analysis framework,” Comput. Ind., vol. 104,
www.theiet.org/media/4719/digital-twins-for-the-built-environment.pdf pp. 11–21, Jan. 2019.
[73] M. Mitchell, “Why AI is harder than we think,” 2021, [96] R. Rocca, P. Rosa, C. Sassanelli, L. Fumagalli, and S. Terzi,
arXiv:2104.12871. “Integrating virtual reality and digital twin in circular economy prac-
[74] M. S. Ibrahim et al., “Machine learning and digital twin driven diagnos- tices: A laboratory application case,” Sustainability, vol. 12, no. 6,
tics and Prognostics of light-emitting diodes,” Laser Photon. Reviews, p. 2286, 2020.
vol. 14, no. 12, 2020, Art. no. 2000254. [97] M. di Benedetto, E. Meloni, G. Amato, F. Falchi, and C. Gennaro,
[75] P. Franciosa, M. Sokolov, S. Sinha, T. Sun, and D. Ceglarek, “Deep “Learning safety equipment detection using virtual worlds,” in Proc.
learning enhanced digital twin for closed-loop in-process quality Int. Conf. Content Based Multimedia Indexing (CBMI), 2019, pp. 1–6.
improvement,” CIRP Ann., vol. 69, no. 1, pp. 369–372, 2020. [98] V. Kuts, T. Otto, T. Tähemaa, and Y. Bondarenko, “Digital twin based
[76] K. Alexopoulos, N. Nikolakis, and G. Chryssolouris, “Digital twin- synchronised control and simulation of the industrial robotic cell using
driven supervised machine learning for the development of artificial virtual reality,” J. Mach. Eng., vol. 19, no. 1, pp. 128–144, 2019.
intelligence applications in manufacturing,” Int. J. Comput. Integr. [99] A. M. Karadeniz, İ. Arif, A. Kanak, and S. Ergün, “Digital twin of
Manuf., vol. 33, no. 5, pp. 429–439, 2020. eGastronomic things: A case study for ice cream machines,” in Proc.
[77] F. Xue, W. Lu, Z. Chen, and C. J. Webster, “From LiDAR point cloud IEEE Int. Symp. Circuits Syst. (ISCAS), 2019, pp. 1–4.
towards digital twin city: Clustering city objects based on gestalt princi- [100] T. Kaarlela, S. Pieskä, and T. Pitkäaho, “Digital twin and virtual reality
ples,” ISPRS J. Photogrammetry Remote Sens., vol. 167, pp. 418–431, for safety training,” in Proc. 11th IEEE Int. Conf. Cogn. Infocommun.
Sep. 2020. (CogInfoCom), 2020, pp. 115–120.
[78] M. H. Farhat, X. Chiementin, F. Chaari, F. Bolaers, and M. Haddar, [101] R. Williams et al., “Augmented reality assisted calibration of dig-
“Digital twin-driven machine learning: Ball bearings fault severity clas- ital twins of mobile robots,” IFAC PapersOnLine, vol. 53, no. 3,
sification,” Meas. Sci. Technol., vol. 32, no. 4, 2021, Art. no. 044006. pp. 203–208, 2020.
[79] M. Samnejad, M. G. Shirangi, and R. Ettehadi, “A digital twin of [102] G. Schroeder et al., “Visualising the digital twin using Web services
drilling fluids rheology for real-time rig operations,” in Proc. Offshore and augmented reality,” in Proc. IEEE 14th Int. Conf. Ind. Inf. (INDIN),
Technol. Conf., 2020, p. 18. 2016, pp. 522–527.
[80] C. Cronrath, A. R. Aderiani, and B. Lennartson, “Enhancing digital [103] X. Ma, F. Tao, M. Zhang, T. Wang, and Y. Zuo, “Digital twin
twins through reinforcement learning,” in Proc. IEEE 15th Int. Conf. enhanced human–machine interaction in product lifecycle,” Procedia
Autom. Sci. Eng. (CASE), 2019, pp. 293–298. CIRP, vol. 83, pp. 789–793, Jan. 2019.
[81] S. E. Ahmed, O. San, K. Kara, R. Younis, and A. Rasheed, [104] S. Ke, F. Xiang, Z. Zhang, and Y. Zuo, “A enhanced interaction frame-
“Multifidelity computing for coupling full and reduced order models,” work based on VR, AR and MR in digital twin,” Procedia CIRP,
PLoS ONE, vol. 16, no. 2, 2021, Art. no. e0246092. vol. 83, pp. 753–758, Jan. 2019.
[82] J. Gammerman. “Predictive Maintenance With Conformal and [105] M. Van Den Brand, L. Cleophas, R. Gunasekaran, B. Haverkort,
Probabilistic Prediction: A Commercial Case Study.” Accessed: May 8, D. A. M. Negrin, and H. M. Muctadir, “Models meet data: Challenges
2021. [Online]. Available: https://www.jamesgammerman.com/files/ to create virtual entities for digital twins,” in Proc. ACM/IEEE Int.
Thesis.pdf Conf. Model Driven Eng. Lang. Syst. Companion (MODELS-C), 2021,
[83] A. Jayakumar, B. Mathew, N. Uma, and P. Nedungadi, “Interactive pp. 225–228.
gesture based cataract surgery simulation,” in Proc. IEEE 5th Int. Conf. [106] J. Bao, D. Guo, J. Li, and J. Zhang, “The modelling and operations for
Adv. Comput. Commun. (ICACC), 2015, pp. 350–353. the digital twin in the context of manufacturing,” Enterprise Inf. Syst.,
[84] H. F. Vargas and O. A. Vivas, “Gesture recognition system for surgical vol. 13, no. 4, pp. 534–556, 2019.
robot’s manipulation,” in Proc. IEEE XIX Symp. Image Signal Process. [107] H. Qin, H. Wang, Y. Zhang, and L. Lin, “Constructing digital twin for
Artif. Vis., 2014, pp. 1–5. smart manufacturing,” in Proc. IEEE 24th Int. Conf. Comput. Supported
[85] G. Lingyun, Z. Lin, and W. Zhaokui, “Hierarchical attention-based Cooper. Work Design (CSCWD), 2021, pp. 638–642.
astronaut gesture recognition: A dataset and CNN model,” IEEE [108] G. N. Schroeder, C. Steinmetz, R. N. Rodrigues, R. V. B. Henriques,
Access, vol. 8, pp. 68787–68798, 2020. A. Rettberg, and C. E. Pereira, “A methodology for digital twin
[86] A. G. Howard et al., “MobileNets: Efficient convolutional neural modeling and deployment for industry 4.0,” Proc. IEEE, vol. 109, no. 4,
networks for mobile vision applications,” 2017, arXiv:1704.04861. pp. 556–567, Apr. 2021.
[87] K. Liu, L. Song, W. Han, Y. Cui, and Y. Wang, “Time-varying error [109] G. N. Schroeder, C. Steinmetz, C. E. Pereira, and D. B. Espindola,
prediction and compensation for movement axis of CNC machine tool “Digital twin data modeling with automationml and a communication
based on digital twin,” IEEE Trans. Ind. Informat., vol. 18, no. 1, methodology for data exchange,” IFAC PapersOnLine, vol. 49, no. 30,
pp. 109–118, Jan. 2022. pp. 12–17, 2016.
[88] V. Zaccaria, M. Stenfelt, I. Aslanidou, and K. G. Kyprianidis, “Fleet [110] T. Riedelsheimer, S. Gogineni, and R. Stark, “Methodology to develop
monitoring and diagnostics framework based on digital twin of aero- digital twins for energy efficient customizable IoT-products,” Procedia
engines,” in Proc. Turbo Expo Power Land Sea Air, vol. 51128, CIRP, vol. 98, pp. 258–263, Jan. 2021.
Jun. 2018, Art. no. V006T05A021. [111] F. Tao, M. Zhang, and A. Y. C. Nee, Digital Twin Driven Smart
[89] S. Mi, Y. Feng, H. Zheng, Y. Wang, Y. Gao, and J. Tan, “Prediction Manufacturing. London, U.K.: Academic, 2019.
maintenance integrated decision-making approach supported by digi- [112] H. Wang, H. Li, X. Wen, and G. Luo, “Unified modeling for digital twin
tal twin-driven cooperative awareness and interconnection framework,” of a knowledge-based system design,” Robot. Comput. Integr. Manuf.,
J. Manuf. Syst., vol. 58, pp. 329–345, Jan. 2021. [Online]. Available: vol. 68, Apr. 2021, Art. no. 102074.
https://www.sciencedirect.com/science/article/pii/S0278612520301345 [113] Y. Fan et al., “A digital-twin visualized architecture for flexible
[90] L. Hu et al., “Modeling of cloud-based digital twins for smart manu- manufacturing system,” J. Manuf. Syst., vol. 60, pp. 176–201, Jul. 2021.
facturing with MT connect,” Procedia Manuf., vol. 26, pp. 1193–1203, [114] C. Wu, Y. Zhou, M. V. P. Pessôa, Q. Peng, and R. Tan, “Conceptual
Jan. 2018. digital twin modeling based on an integrated five-dimensional frame-
[91] Z. Jiang, Y. Guo, and Z. Wang, “Digital twin to improve the virtual- work and TRIZ function model,” J. Manuf. Syst., vol. 58, pp. 79–93,
real integration of industrial IoT,” J. Ind. Inf. Integr., vol. 22, Jun. 2021, Jan. 2021.
Art. no. 100196. [115] Q. Qi et al., “Enabling technologies and tools for digital twin,” J.
[92] J. Tan, X. Sha, B. Dai, and T. Lu, “Wireless technology and protocol Manuf. Syst., vol. 58, pp. 3–21, Jan. 2021.
for IIot and digital twins,” in Proc. IEEE ITU Kaleidoscope Ind. Driven [116] V. Makarov, Y. B. Frolov, I. Parshina, and M. Ushakova, “The design
Digit. Transf. (ITU K), 2020, pp. 1–8. concept of digital twin,” in Proc. IEEE 12th Int. Conf. Manag. Large
[93] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, Scale Syst. Develop. (MLSD), 2019, pp. 1–4.
“Communication-efficient federated learning for digital twin edge [117] J. Lee, M. Azamfar, and B. Bagheri, “A unified digital twin framework
networks in industrial IoT,” IEEE Trans. Ind. Informat., vol. 17, no. 8, for shop floor design in industry 4.0 manufacturing systems,” Manuf.
pp. 5709–5718, Aug. 2021. Lett., vol. 27, pp. 87–91, Jan. 2021.
[94] Y. Lu, X. Huang, K. Zhang, S. Maharjan, and Y. Zhang, [118] D. Preuveneers, W. Joosen, and E. Ilie-Zudor, “Robust digital twin
“Communication-efficient federated learning and permissioned compositions for industry 4.0 smart manufacturing systems,” in
blockchain for digital twin edge networks,” IEEE Internet Things J., Proc. IEEE 22nd Int. Enterprise Distrib. Object Comput. Workshop
vol. 8, no. 4, pp. 2276–2288, Feb. 2021. (EDOCW), 2018, pp. 69–78.
[119] J. Lee, M. Azamfar, J. Singh, and S. Siahpour, “Integration of digital [142] V. V. Martynov, D. N. Shavaleeva, and A. A. Zaytseva, “Information
twin and deep learning in cyber-physical systems: Towards smart man- technology as the basis for transformation into a digital society and
ufacturing,” IET Collaborative Intell. Manuf., vol. 2, no. 1, pp. 34–36, industry 5.0,” in Proc. Int. Conf. Qual. Manag. Transp. Inf. Security
2020. Inf. Technol. (IT QM IS), 2019, pp. 539–543.
[120] G. A. Gericke, R. B. Kuriakose, H. J. Vermaak, and O. Mardsen, [143] J. Cheng, W. Chen, F. Tao, and C.-L. Lin, “Industrial IoT in 5G envi-
“Design of digital twins for optimization of a water bottling plant,” ronment towards smart manufacturing,” J. Ind. Inf. Integr., vol. 10,
in Proc. IEEE 45th Annu. Conf. Ind. Electron. Soc. (IECON), vol. 1, pp. 10–19, Jun. 2018.
2019, pp. 5204–5210. [144] G. Szabó, S. Rácz, N. Reider, H. A. Munz, and J. Pető, “Digital twin:
[121] W. Lin and M. Y. Low, “concept and implementation of a cyber- Network provisioning of mission critical communication in cyber phys-
physical digital twin for a SMT line,” in Proc. IEEE Int. Conf. Ind. ical production systems,” in Proc. IEEE Int. Conf. Ind. 4.0 Artif. Intell.
Eng. Eng. Manag. (IEEM), 2019, pp. 1455–1459. Commun. Technol. (IAICT), 2019, pp. 37–43.
[122] H. R. Hasan et al., “A blockchain-based approach for the creation of [145] G. Hinz, M. Büchel, F. Diehl, M. Schellmann, and A. Knoll,
digital twins,” IEEE Access, vol. 8, pp. 34113–34126, 2020. “Designing a far-reaching view for highway traffic scenarios with
[123] A. Kanak, N. Ugur, and S. Ergun, “A visionary model on blockchain- 5G-based intelligent infrastructure,” in 8. Tagung Fahrerassistenz.
based accountability for secure and collaborative digital twin environ- Heidelberg, Germany: Springer, 2017.
ments,” in Proc. IEEE Int. Conf. Syst. Man Cybern. (SMC), 2019, [146] S. A. Kumar, R. Madhumathi, P. R. Chelliah, L. Tao, and S. Wang,
pp. 3512–3517. “A novel digital twin-centric approach for driver intention prediction
[124] K. Wang et al., “Simulation-based digital twin development for and traffic congestion avoidance,” J. Rel. Intell. Environ., vol. 4, no. 4,
blockchain enabled end-to-end industrial hemp supply chain risk pp. 199–209, 2018.
management,” in Proc. IEEE Winter Simulat. Conf. (WSC), 2020, [147] A. El Saddik, “Multimedia and the tactile Internet,” IEEE Multimedia,
pp. 3200–3211. vol. 27, no. 1, pp. 5–7, Mar. 2020.
[125] K. Židek, J. Pitel’, M. Adámek, P. Lazorík, and A. Hošovskỳ, “Digital [148] M. Höyhtyä, “Connectivity manager: Ensuring robust connections for
twin of experimental smart manufacturing assembly system for industry autonomous ships,” in Proc. IEEE 2nd Int. Conf. Intell. Auton. Syst.
4.0 concept,” Sustainability, vol. 12, no. 9, p. 3658, 2020. (ICoIAS), 2019, pp. 86–90.
[126] J. Vachálek, L. Bartalskỳ, O. Rovnỳ, D. Šišmišová, M. Morháč, and [149] S. Schneider, M. Peuster, D. Behnke, M. Müller, P.-B. Bök, and
M. Lokšík, “The digital twin of an industrial production line within the H. Karl, “Putting 5G into production: Realizing a smart manufacturing
industry 4.0 concept,” in Proc. IEEE 21st Int. Conf. Process Control vertical scenario,” in Proc. IEEE Eur. Conf. Netw. Commun. (EuCNC),
(PC), 2017, pp. 258–262. 2019, pp. 305–309.
[127] Q. Lu, X. Xie, A. K. Parlikad, and J. M. Schooling, “Digital [150] J. Santa et al., “MIGRATE: Mobile device virtualisation through state
twin-enabled anomaly detection for built asset monitoring in oper- transfer,” IEEE Access, vol. 8, pp. 25848–25862, 2020.
ation and maintenance,” Autom. Construction, vol. 118, Oct. 2020, [151] H. Viswanathan and P. E. Mogensen, “Communications in the
Art. no. 103277. 6G era,” IEEE Access, vol. 8, pp. 57063–57074, Mar. 2020,
[128] V. Q. Lu, A. K. Parlikad, P. Woodall, G. D. Ranasinghe, and J. Heaton, doi: 10.1109/ACCESS.2020.2981745.
“Developing a dynamic digital twin at a building level: Using cam- [152] T. G. I. Association. “European Vision for the 6G Network Ecosystem.”
bridge campus as case study,” in Proc. Int. Conf. Smart Infrast. 2021. Accessed: Jun. 12, 2021. [Online]. Available: https://5g-ppp.eu/
Construction (ICSIC) Driving Data Informed Decis. Making, 2019, wp-content/uploads/2021/06/WhitePaper-6G-Europe.pdf
pp. 67–75.
[153] Z. Zhang et al., “6G wireless networks: Vision, requirements, architec-
[129] Á. Segura et al., “Visual computing technologies to support the operator
ture, and key technologies,” IEEE Veh. Technol. Mag., vol. 14, no. 3,
4.0,” Comput. Ind. Eng., vol. 139, Jan. 2020, Art. no. 105550.
pp. 28–41, Sep. 2019.
[130] R. Ganguli and S. Adhikari, “The digital twin of discrete dynamic
[154] P. Yang, Y. Xiao, M. Xiao, and S. Li, “6G wireless communications:
systems: Initial approaches and future challenges,” Appl. Math. Modell.,
Vision and potential techniques,” IEEE Netw., vol. 33, no. 4, pp. 70–75,
vol. 77, pp. 1110–1128, Jan. 2020.
Jul./Aug. 2019.
[131] K. Ding et al., “Smart steel bridge construction enabled by BIM and
Internet of Things in industry 4.0: A framework,” in Proc. IEEE 15th [155] H. X. Nguyen, R. Trestian, D. To, and M. Tatipamula, “Digital twin
Int. Conf. Netw. Sens. Control (ICNSC), 2018, pp. 1–5. for 5G and beyond,” IEEE Commun. Mag., vol. 59, no. 2, pp. 10–15,
[132] C.-S. Shim, N.-S. Dang, S. Lon, and C.-H. Jeon, “Development of Feb. 2021.
a bridge maintenance system for prestressed concrete bridges using [156] H.-N. Dai, H. Wang, G. Xu, J. Wan, and M. Imran, “Big data ana-
3D digital twin model,” Struct. Infrastruct. Eng., vol. 15, no. 10, lytics for manufacturing Internet of Things: Opportunities, challenges
pp. 1319–1332, 2019. and enabling technologies,” Enterprise Inf. Syst., vol. 14, nos. 9–10,
[133] S. Kaewunruen and Q. Lian, “Digital twin aided sustainability-based pp. 1279–1303, 2020.
lifecycle management for railway turnout systems,” J. Clean. Prod., [157] J. Zhang, “Advancements of outlier detection: A survey,” ICST Trans.
vol. 228, pp. 1537–1551, Aug. 2019. Scalable Inf. Syst., vol. 13, no. 1, pp. 1–26, 2013.
[134] S. Kaewunruen and N. Xu, “Digital twin for sustainability evalua- [158] V. Kotu and B. Deshpande, Data Science: Concepts and Practice,
tion of railway station buildings,” Front. Built Environ., vol. 4, p. 77, 2nd ed. Amsterdam, Netherlands: Elsevier, 2019.
Dec. 2018. [159] S. R. Chhetri, S. Faezi, A. Canedo, and M. A. A. Faruque, “QUILT:
[135] M. Podskarbi and D. J. Knezevic, “Digital twin for operations-present Quality inference from living digital twins in IoT-enabled manufactur-
applications and future digital thread,” in Proc. Offshore Technol. Conf., ing systems,” in Proc. Int. Conf. Internet Things Design Implement.,
2020, p. 18. 2019, pp. 237–248.
[136] S.-K. Jo, D.-H. Park, H. Park, and S.-H. Kim, “Smart livestock farms [160] A. Castellani, S. Schmitt, and S. Squartini, “Real-world anomaly detec-
using digital twin: Feasibility study,” in Proc. IEEE Int. Conf. Inf. tion by using digital twin systems and weakly supervised learning,”
Commun. Technol. Converg. (ICTC), 2018, pp. 1461–1463. IEEE Trans. Ind. Informat., vol. 17, no. 7, pp. 4733–4742, Jul. 2021.
[137] A. Francisco, N. Mohammadi, and J. E. Taylor, “Smart city digital [161] M. Bevilacqua et al., “Digital twin reference model development to
twin–enabled energy management: Toward real-time urban build- prevent operators’ risk in process plants,” Sustainability, vol. 12, no. 3,
ing energy benchmarking,” J. Manag. Eng., vol. 36, no. 2, 2020, p. 1088, 2020.
Art. no. 04019045. [162] X. Xie, Q. Lu, A. K. Parlikad, and J. M. Schooling, “Digital twin
[138] T. Ruohomäki, E. Airaksinen, P. Huuska, O. Kesäniemi, M. Martikka, enabled asset anomaly detection for building facility management,”
and J. Suomisto, “Smart city platform enabling digital twin,” in Proc. IFAC PapersOnLine, vol. 53, no. 3, pp. 380–385, 2020.
IEEE Int. Conf. Intell. Syst. (IS), 2018, pp. 155–161. [163] Y. Qamsane et al., “A unified digital twin framework for real-time
[139] J. Du, Q. Zhu, Y. Shi, Q. Wang, Y. Lin, and D. Zhao, “Cognition digital monitoring and evaluation of smart manufacturing systems,” in Proc.
twins for personalized information systems of smart cities: Proof of IEEE 15th Int. Conf. Autom. Sci. Eng. (CASE), 2019, pp. 1394–1401.
concept,” J. Manag. Eng., vol. 36, no. 2, 2020, Art. no. 04019052. [164] Q. Wang, S. Bu, and Z. He, “Achieving predictive and proactive main-
[140] Q. Lu et al., “Developing a digital twin at building and city levels: Case tenance for high-speed railway power equipment with LSTM-RNN,”
study of west cambridge campus,” J. Manag. Eng., vol. 36, no. 3, 2020, IEEE Trans. Ind. Informat., vol. 16, no. 10, pp. 6509–6517, Oct. 2020.
Art. no. 05020004. [165] H. A. Gohel, H. Upadhyay, L. Lagos, K. Cooper, and A. Sanzetenea,
[141] Y.-C. Lin and W.-F. Cheung, “Developing WSN/BIM-based environ- “Predictive maintenance architecture development for nuclear infras-
mental monitoring management system for parking garages in smart tructure using machine learning,” Nucl. Eng. Technol., vol. 52, no. 7,
cities,” J. Manag. Eng., vol. 36, no. 3, 2020, Art. no. 04020012. pp. 1436–1442, 2020.
[166] C. Cheng, B.-K. Zhang, and D. Gao, “A predictive maintenance solu- [190] B. Barn, A. W. Brown, and J. Cheesman, “Methods and tools for com-
tion for bearing production line based on edge-cloud cooperation,” in ponent based development,” in Proc. IEEE Technol. Object Oriented
Proc. IEEE Chin. Autom. Congr. (CAC), 2019, pp. 5885–5889. Lang. (TOOLS), 1998, pp. 385–395.
[167] F. Francis and M. Mohan, “Arima model based real time trend analysis [191] B. D. Mittelstadt, P. Allo, M. Taddeo, S. Wachter, and L. Floridi, “The
for predictive maintenance,” in Proc. IEEE 3rd Int. Conf. Electron. ethics of algorithms: Mapping the debate,” Big Data Soc., vol. 3, no. 2,
Commun. Aerosp. Technol. (ICECA), 2019, pp. 735–739. p. 19, 2016.
[168] A. Cachada et al., “Maintenance 4.0: Intelligent and predictive main- [192] B. Barn, R. Barn, and F. Raimondi, “On the role of value sensitive
tenance system architecture,” in Proc. IEEE 23rd Int. Conf. Emerg. concerns in software engineering practice,” in Proc. IEEE/ACM 37th
Technol. Factory Autom. (ETFA), vol. 1, 2018, pp. 139–146. Int. Conf. Softw. Eng., vol. 2, 2015, pp. 497–500.
[169] MIMOSA. “MIMOSA OSA-CBM—MIMOSA.” Accessed: Aug. 10, [193] T. D. West and M. Blackburn, “Is digital thread/digital twin affordable?
2020. [Online]. Available: https://www.mimosa.org/mimosa-osa-cbm/ A systemic assessment of the cost of DoD’s latest manhattan project,”
[170] Y. Liang, W. Li, X. Lu, and S. Wang, “Fog computing and con- Procedia Comput. Sci., vol. 114, pp. 47–56, Jan. 2017.
volutional neural network enabled prognosis for machining process [194] J. Voas, P. Mell, and V. Piroumian, “Considerations for digital twin
optimization,” J. Manuf. Syst., vol. 52, pp. 32–42, Jul. 2019. technology and emerging standards,” NIST Internal Interagency Rep.,
[171] W. Luo, T. Hu, Y. Ye, C. Zhang, and Y. Wei, “A hybrid predictive Gaithersburg, MD, USA, Rep. NISTIR 8356, 2021.
maintenance approach for CNC machine tool driven by digital twin,” [195] E. Harper, C. Ganz, and S. Malakuti, Digital Twin Architecture and
Robot. Comput. Integr. Manuf., vol. 65, Oct. 2020, Art. no. 101974. Standards, Ind. Internet Consortium, Boston, MA, USA, 2021.
[172] Y. Choi, H. Kwun, D. Kim, E. Lee, and H. Bae, “Method of [196] C. Lange, Industrial Data Space—Digital Sovereignty Over Data,
predictive maintenance for induction furnace based on neural network,” Vienna Data Sci. Group, Vienna, Austria, 2016.
in Proc. IEEE Int. Conf. Big Data Smart Comput. (BigComp), 2020, [197] J. Hetherington and M. West, The Pathway Towards an Information
pp. 609–612. Management Framework—A Commons for Digital Built Britain, Univ.
[173] N. Gugulothu, V. Tv, P. Malhotra, L. Vig, P. Agarwal, and G. Shroff, Cambridge, Cambridge, U.K., 2020.
“Predicting remaining useful life using time series embeddings based [198] M. Dietz, B. Putz, and G. Pernul, “A distributed ledger approach to
on recurrent neural networks,” 2017, arXiv:1709.01073. digital twin secure data sharing,” in Proc. IFIP Annu. Conf. Data Appl.
[174] J. Wang, Y. Liang, Y. Zheng, R. X. Gao, and F. Zhang, “An integrated Security Privacy, 2019, pp. 281–300.
fault diagnosis and prognosis approach for predictive maintenance of [199] M. Hearn and S. Rix, Cybersecurity Considerations for Digital Twin
wind turbine bearing with limited samples,” Renew. Energy, vol. 145, Implementations,
pp. 642–650, Jan. 2020. Ind. Internet Consortium, Boston, MA, USA, pp. 107–113, 2019.
[175] D. Shetve, R. VaraPrasad, R. Trestian, H. X. Nguyen, and [200] I. Yaqoob, K. Salah, M. Uddin, R. Jayaraman, M. Omar, and M. Imran,
H. Venkataraman, “CATS: Cluster-aided two-step approach for “Blockchain for digital twins: Recent advances and future research
anomaly detection in smart manufacturing,” in Advances in challenges,” IEEE Netw., vol. 34, no. 5, pp. 290–298, Jan. 2020.
Computing and Network Communications. Singapore: Springer, 2021, [201] C. Gehrmann and M. Gunnarsson, “A digital twin based industrial
pp. 103–115, doi: 10.1007/978-981-33-6987-0_9. automation and control system security architecture,” IEEE Trans. Ind.
[176] A. Nowak-Brzezińska and T. Xieski, “Outlier mining using the Informat., vol. 16, no. 1, pp. 669–680, Jan. 2020.
DBSCAN algorithm,” J. Appl. Comput. Sci., vol. 25, no. 2, pp. 53–68, [202] T. Hossen, M. Gursoy, and B. Mirafzal, “Digital twin for self-security
2017. of smart inverters,” in Proc. IEEE Energy Convers. Congr. Expo.
[177] Z. Cheng, C. Zou, and J. Dong, “Outlier detection using isolation forest (ECCE), 2021, pp. 713–718.
and local outlier factor,” in Proc. Conf. Res. Adapt. Converg. Syst., [203] J. McCarthy, M. L. Minsky, N. Rochester, and C. E. Shannon, “A
2019, pp. 161–168. proposal for the dartmouth summer research project on artificial
[178] M. Raza, P. M. Kumar, D. V. Hung, W. Davis, H. Nguyen, and intelligence, August 31, 1955,” AI Mag., vol. 27, no. 4, p. 12, 2006.
R. Trestian, “A digital twin framework for industry 4.0 enabling next- [204] P. K. McClure, “‘you’re fired,’ says the robot: The rise of automation
gen manufacturing,” in Proc. IEEE 9th Int. Conf. Ind. Technol. Manag. in the workplace, technophobes, and fears of unemployment,” Soc. Sci.
(ICITM), 2020, pp. 73–77. Comput. Rev., vol. 36, no. 2, pp. 139–156, 2018.
[179] S. Mihai, W. Davis, D. V. Hung, and R. Trestian, “A digital twin frame- [205] H. Shevlin, K. Vold, M. Crosby, and M. Halina, “The limits of machine
work for predictive maintenance in industry 4.0,” in Proc. Int. Conf. intelligence: Despite progress in machine intelligence, artificial general
High Perform. Comput. Simulat., 2021, pp. 1–9. intelligence is still a major challenge,” EMBO Rep., vol. 20, no. 10,
[180] H. V. Dang, M. Raza, T. V. Nguyen, T. Bui-Tien, and H. X. Nguyen, 2019, Art. no. e49177.
“Deep learning-based detection of structural damage using time-series [206] D. Silver et al., “Mastering the game of go with deep neural networks
data,” Struct. Infrastruct. Eng., vol. 17, no. 11, pp. 1474–1493, 2021. and tree search,” Nature, vol. 529, no. 7587, pp. 484–489, 2016.
[181] H. V. Dang, H. Tran-Ngoc, T. V. Nguyen, T. Bui-Tien, G. De Roeck, [207] M. Campbell, A. J. Hoane, Jr., and F.-H. Hsu, “Deep blue,” Artif. Intell.,
and H. X. Nguyen, “Data-driven structural health monitoring using vol. 134, nos. 1–2, pp. 57–83, 2002.
feature fusion and hybrid deep learning,” IEEE Trans. Autom. Sci. Eng., [208] G. Marcus and E. Davis, Rebooting AI: Building Artificial Intelligence
vol. 18, no. 4, pp. 2087–2103, Oct. 2021. We Can Trust. New York, NY, USA: Vintage, 2019.
[182] H. V. Dang, M. Raza, H. Tran-Ngoc, T. Bui-Tien, and H. X. Nguyen,
“Connection stiffness reduction analysis in steel bridge via deep CNN Stefan Mihai received the B.Sc. degree in telecom-
and modal experimental data,” Struct. Eng. Mech., vol. 77, no. 4, munications from the Politehnica University of
pp. 495–508, 2021. Bucharest and the M.Sc. degree in telecommunica-
[183] E. Trist and F. Emery, Socio-Technical Systems Theory. Oxford, U.K.: tions engineering from Middlesex University, where
Pergamon, 1960. he is currently pursuing the Ph.D. degree work-
[184] B. Heydari and M. J. Pennock, “Guiding the behavior of sociotechnical ing on the Digital Twin Modelling for Automation,
systems: The role of agent-based modeling,” Syst. Eng., vol. 21, no. 3, Maintenance and Monitoring in Industry 4.0 Smart
pp. 210–226, 2018. Factory Project. His research interests include
[185] E. Bonabeau, “Agent-based modeling: Methods and techniques for sim- machine learning, digital twin, and predictive
ulating human systems,” Proc. Nat. Acad. Sci. USA, vol. 99, no. s3, maintenance.
pp. 7280–7287, 2002.
[186] S. Barat et al., “An agent-based digital twin for exploring Localized Mahnoor Yaqoob received the B.Sc. degree in
non-pharmaceutical interventions to control COVID-19 pandemic,” software engineering from Fatima Jinnah Women
Trans. Indian Nat. Acad. Eng., vol. 6, no. 2, pp. 323–353, 2021. University Pakistan in 2016 and the M.Sc. degree in
[187] J. M. Epstein, “Why model?” J. Artif. Soc. Soc. Simulat., vol. 11, no. 4, computer engineering from Middle East Technical
p. 12, 2008. University Northern Cyprus Campus in 2020. She
[188] B. S. Barn and T. Clark, “Revisiting Naur’s programming as theory is currently pursuing the Ph.D. degree within the
building for enterprise architecture modelling,” in Proc. Int. Conf. Adv. discipline of Design Engineering with Middlesex
Inf. Syst. Eng., 2011, pp. 229–236. University London. Her current research interests
[189] J. Oden et al., Simulation-Based Engineering Science: Revolutionizing include digital twin, blockchain, machine learning
Engineering Science Through Simulation. Report of the National and artificial intelligence, wireless cellular commu-
Science Foundation Blue Ribbon Panel on Simulation-Based nications, computer networks, analytical modelling,
Engineering Science, Nat. Sci. Found., Alexandria, VA, USA, 2006. and queueing theory.
Dang V. Hung received the M.Sc. and Ph.D. degrees Mehmet Karamanoglu (Member, IEEE) is cur-
in structural dynamics from the University of Lyon, rently serving as the Head of Department of Design
France, in 2009 and 2013, respectively. He was a Engineering and Mathematics with the Faculty of
Postdoctoral Fellow with the Faculty of Science and Science and Technology, Middlesex University, and
Technology, Middlesex University London. He is a Professor of Design Engineering. He is a mem-
currently a Lecturer with the National University ber of several professional bodies and societies,
of Civil Engineering, Hanoi, Vietnam. He has pub- including a Fellowship at Institution of Mechanical
lished his research works internationally in France, Engineers and the Royal Society of Arts. He has
Italy, USA, HongKong, and Morocco. His research spent significant length of time working in collabo-
interests include structural dynamic, numerical sim- ration with industry in a variety of sectors and has
ulation, structural health monitoring, data analysis, managed numerous Knowledge Transfer Partnership
machine learning, and digital twin. Projects in the field of Manufacturing Engineering and Automation. His
research interest includes engineering education, interplay between art, design
and engineering, advanced manufacturing including optimisation, mecha-
tronics, and robotics. His recent work included mass customisation, and
developing autonomous systems. He is currently working on mathematical
optimisation techniques and cognitive manufacturing. In his wider area of
work, he is the U.K. National Expert for Mechatronics and Automation
competitions for WorldSkills U.K.
William Davis received the B.Eng. degree in
mechatronics engineering. He is currently pursu-
ing the master’s degree working on the Digital
Twin Modelling for Automation, Maintenance and
Monitoring in Industry 4.0 Smart Factory Project.
His current research interests include blockchain,
digital twin, and mechatronics systems.
Balbir Barn (Member, IEEE) is a Professor of
Software Engineering with the Computer Science
Department, Middlesex University. He has exten-
sive commercial research experience working in
research centres with Texas Instruments and JP
Morgan Chase as well as leading on academic
funded research (Over 2.5 million). In collaboration
with TCS Research Labs, he is working on model
driven approaches for supporting Manufacturing 4.0
contexts through the design and implementation of
Praveer Towakel received the B.Sc. degree in a simulation environment for Digital Twins that
physics from the University of Mauritius in 2016. accommodates value sensitive design principles. He has published over 120
He is currently pursuing the Ph.D. degree within peer-reviewed papers in leading international conferences and journals and
the discipline of Design Engineering with Middlesex is currently editing a book on the “Digital Enterprise” with IGI-Global. His
University. His current research interests include research is focused on model driven software engineering where the goal is
gesture recognition, radar systems, and machine to use models as abstractions and execution environments to support complex
learning. decision making.
Dattaprasad Shetve received the B.E. degree in

electronics engineering from the Goa College of
Engineering and the M.Tech. degree in indus-
trial automation and robotics from the Manipal
Mohsin Raza received the B.S. (Hons.) and M.S. Institute of Technology, Manipal, in 2019. He
degrees in electronic engineering from Mohammad worked as a Junior Research Fellow on the Digital
Ali Jinnah University, Pakistan, and the Ph.D. degree Twin Modelling for Automation, Maintenance and
from Math, Physics and Electrical Engineering Monitoring in Industry 4.0 Smart Factory Project.
Department, Northumbria University, U.K. He is a His area of interest includes embedded systems, real-
Senior Lecturer with the Department of Computer time operating systems, bare metal programming,
Science, Edge Hill University, U.K. Prior to this, he embedded linux, and device driver development.
worked as a Lecturer with Northumbria University,
U.K., from 2019 to 2020, a Postdoctoral Fellow
with Middlesex University, U.K., from 2018 to 2019,
as a Demonstrator/Associate-Lecturer and Doctoral
Fellow with Northumbria University, U.K., from 2015 to 2017, a Junior
Lecturer from 2010 to 2012 and later as a Lecturer from 2012 to 2015
from Engineering Department, Mohammad Ali Jinnah University, Pakistan, Raja V. Prasad (Member, IEEE) received the Ph.D.
and a Hardware Support Engineer with Unified Secure Services, Pakistan, degree from the Indian Institute of Technology
from 2009 to 2010. He served as a Technical Committee Member for ICET Hyderabad in 2016, under the supervision of
2012, SKIMA 2015, SKIMA 2017, WSGT 2017, CSNDSP 2018, SKIMA Dr. P. Rajalakshmi. He is currently working as
2018, ICT 2019, and CSoNet 2019. He has also been a Guest Editor to an Assistant Professor with the Indian Institute of
special issue on Heterogenous Internet of Medical Things in International Information Technology, Sri City. His research areas
Journal of Distributed Sensor Networks and a Reviewer of several journals, are centered on wireless sensor networks, wireless
including IEEE ACCESS, IEEE Communications Letters, Sensors (MDPI), sensor and actuator networks, smart buildings, net
Vehicular Communications (Elsevier), and Arabian Journal for Science and zero energy buildings, wireless protocols for Internet
Engineering (Springer). His research interests include IoT, 5G and wireless of Things applications, smart cities automated wire-
networks, autonomous transportation systems, machine learning, Industry 4.0, less sensor networks, green networks, and Internet
and digital twins. of Things.
Hrishikesh Venkataraman (Member, IEEE) Huan X. Nguyen (Senior Member, IEEE) received
received the M.Tech. degree from the Indian the B.Sc. degree from the Hanoi University of
Institute of Technology Kanpur in 2004, and the Science and Technology, Vietnam, in 2000, and
Ph.D. degree from Jacobs University Bremen, the Ph.D. degree from the University of New
Germany, where he was awarded the best Graduate South Wales, Australia, in 2007. He is currently
Student in September 2007. His M.Tech. thesis a Professor of Digital Communication Engineering
from Vodafone Chair for Mobile Communications, with Middlesex University London, U.K., where he
Dresden, TU, Germany, from 2003 to 2004. is also the Director of the London Digital Twin
He has more than 13 years of industry and Research Centre and the Head of the 5G/6G &
research experience, having worked with Irish IoT Research Group. He leads research activities
national centre–Research Institute for Networks and in digital twin modelling, 5G/6G systems, machine-
Communication Engineering, CTO Office of Tech Mahindra and Microverse type communication, digital transformation and machine learning within
Automation Private Ltd. He is currently an Associate Professor and the his university with focus on industry 4.0 and critical applications (disaster
Faculty-in-Charge for Research and Development activities with the Indian recovery, intelligent transportation, e-health, and smart manufacturing). He
Institute of Information Technology, Sri City, India. He has three Ph.D. has been leading many council/industry funded projects, publishing 130+
students and several research Honours students working under him. He has peer-reviewed research papers, and serving as the Chairs for international
more than 60 publications in different international conferences and journals, conferences (ICT’21, ICEM2021, ICT’20, ICT’19, IWNPD’17, PIMRC’20,
including in ACM, Elsevier, IEEE, IET, and Springer. He has edited three FoNeS-IoT’20, and ATC’15).
books, has one granted U.S. patent, one contribution in European Telecom
Standards Institute, and has been an Editor of the European Transactions
of Telecommunications for five years. His area of interest is in wireless
communication, connected cars, and device-to-device communication.
Ramona Trestian received the Ph.D. degree from

Dublin City University, Ireland, in 2012. She is
a Senior Lecturer with the Design Engineering
and Mathematics Department, Middlesex University,
London, U.K. She published in prestigious inter-
national conferences and journals and has five
edited books. Her research interests include mobile
and wireless communications, quality of experience,
multimedia streaming, handover and network selec-
tion strategies, and digital twin modelling. She is an
Associate Editor of the IEEE C OMMUNICATIONS
S URVEYS AND T UTORIALS.

Digital Twins A Survey On Enabling Technologies Challenges Trends and Future Prospects

Uploaded by

Copyright:

Available Formats

Digital Twins A Survey On Enabling Technologies Challenges Trends and Future Prospects

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Digital Twins A Survey On Enabling Technologies Challenges Trends and Future Prospects

Uploaded by

Copyright:

Available Formats

IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 24, NO.

4, FOURTH QUARTER 2022 2255

Digital Twins: A Survey on Enabling Technologies,

Fig. 2. Interest trend in Digital Twin as seen through Google searches.

frameworks, and case-studies that have been thoroughly doc-

A. Background and Motivation

its publishing. The authors set out to answer three research

Fig. 4. An overview of the survey paper’s contents.

In [115], the authors took a closer look at the specific frame-

V. D IGITAL T WIN : U SE C ASES AND S ERVICES

Fig. 6. Digital Twin framework for smart manufacturing.

Fig. 7. Digital Twin in infrastructure.

vehicles in the Providentia system play two roles, first being

Fig. 9. DT application: Anomaly Detection Framework.

Fig. 10. DT-based PdM scheme using hybrid modeling.

Fig. 11. Conveyor Belt Snapshot of a Tea Manufacturing Plant.

Fig. 12. Steps in the Process of Tea Manufacturing.

Fig. 13. Digital Twin in I4.0.

VIII. C ONCLUSION [5] M. E. Rose and J. R. Kitchin, “Pybliometrics: Scriptable bibliometrics

Dattaprasad Shetve received the B.E. degree in

Ramona Trestian received the Ph.D. degree from

You might also like