Challenges in Digital Twin Development for Cyber-Physical Production Systems
CChallenges in Digital Twin Development forCyber-Physical Production Systems ∗ Heejong Park − − − , Arvind Easwaran − − − , andSidharta Andalam − − − Nanyang Technological University, 50 Nanyang Avenue 639798, Singapore { hj.park,arvinde } @ntu.edu.sg Delta Electronics, 50 Nanyang Avenue 639798, 50 Nanyang Avenue 639798,Singapore [email protected]
Abstract.
The recent advancement of information and communicationtechnology makes digitalisation of an entire manufacturing shop-floorpossible where physical processes are tightly intertwined with their cy-ber counterparts. This led to an emergence of a concept of digital twin,which is a realistic virtual copy of a physical object. Digital twin willbe the key technology in Cyber-Physical Production Systems (CPPS)and its market is expected to grow significantly in the coming years.Nevertheless, digital twin is still relatively a new concept that peoplehave different perspectives on its requirements, capabilities, and limita-tions. To better understand an effect of digital twin’s operations, mitigatecomplexity of capturing dynamics of physical phenomena, and improveanalysis and predictability, it is important to have a development toolwith a strong semantic foundation that can accurately model, simulate,and synthesise the digital twin. This paper reviews current state-of-arton tools and developments of digital twin in manufacturing and discussespotential design challenges.
Keywords:
Digital twin · Cyber-physical system · Industry 4.0 · Smartmanufacturing · Model of computation · Modelling tool.
The advancement of today’s information and communication technologies (ICT)has enabled a collection and effective use of big data which give useful insightsabout various industrial assets and their operations. High availability of low-cost, low-powered sensors and Internet-of-Thing (IoT) devices together withtheir communication networks are the key enablers of cyber-physical systems.Cyber-phycial system is the main technological concept of the fourth industrialrevolution, so called Industry 4.0 [15], characterised by a tight integration of ∗ This work was supported by Delta-NTU Corporate Lab for Cyber-Physical Sys-tems with funding support from Delta Electronics Inc. and the National ResearchFoundation (NRF) Singapore under the Corp Lab@University Scheme. a r X i v : . [ c s . S E ] F e b Park et al.
Cyber-PhysicalProduction SystemDigital Twin Factory
Real-Time Computing – Closed-loop control – WCET analysis – Latency – Communication network – Protocols – Execution platform
Tools – Model-Driven Engineering – APIs – Development framework – Runtime environment
Modelling – Finite State Machine – ODE – Hybrid/Timed automata – Petri-Nets – SynchronousReactive/GALS – Finite Element Method
Analytics – Big data – Machine learning – Data fusion, cleaning, visualisation – Simulation, optimisation
Fig. 1.
Components of Cyber-Physical Production Systemcomputations in a cyber world with physical processes in a real world. In cyber-physical system, changes in a physical process affect computations in a cyberworld or vice-versa [17] where ICT enables a feedback between these two. A con-cept of digital twin, which is a ultra-realistic virtual counterpart of a real-worldobject, was introduced firstly by Grieves in 2003 [11]. Since the introduction ofIndustry 4.0 at the Hannover Fair in Germany in 2013, the capability the digitaltwin in cyber-physical system has been received a great attention along with therecent advancement information technologies. The digital twin market is likelyto be worth USD 15.66 billion by 2023 at a compound annual growth rate of37.87% [23].Implementation of cyber-physical system in industrial manufacturing is calledCyber-Physical Production System (CPPS). In this domain, the use of digitaltwin has been mainly studied to improve time-to-market and MRO (Mainte-nance, Repair and Overhaul) costs, predict potential failures, and estimate re-maining life of individual components, through high-fidelity simulation as well asreal-time monitoring and control of manufacturing process. Since digital twinsare built using the best available ICTs to mirror the physics of target objects ina virtual world, they can perform various simulations as if physical systems aretested in a real-life situation. Furthermore, digital twins can pause, resume, save,and restore their states to validate various corner-cases, which would rather betime-consuming and costly, if not impossible, to accomplish with physical sys-tems. It is particularly valuable for the organisations who cannot afford veryexpensive resources to conduct exhaustive testing that may ruin their physicalprototype or result in catastrophic events. For example, NASA and U.S. pro-posed digital twin concept to accelerate development of their future vehicles [10].An overview of technologies that create digital twin in CPPS is shown inFig. 1. CPPS is a combination of both logical and physical components that canbe characterised by continuous and discrete dynamics. In addition, modellingand implementing a digital twin may require skills from multiple disciplines,such as electromagnetism, fluid dynamics, and kinematics etc., to capture phys- hallenges in Digital Twin Development for CPPS 3 ical properties of the manufacturing process. Therefore, a modelling techniquewith varying levels of abstraction would be needed for both flexibility and ex-pressiveness. The tight integration with the physical system often put real-timeconstraints on the operations of digital twin. As a result, designers would alsoneed to consider real-time aspects of the twin such as worst-case execution timeand communication latency in the time-analysable networks. Data measuredfrom sensors and simulation are useful for predictive maintenance and optimis-ing production process. A platform for data analytics that enriches digital twincapabilities is also a significant part of the CPPS. Lastly, a proper softwareframework will be required that incorporates APIs, runtime environment, andmodel-driven engineering.Although digital twin is employed for tackling various problems [25], it is stillrelatively a new research area where there exist several open research questionsthat have not yet been thoroughly explored such as:1. How to build a highly-accurate, yet scalable, digital twin for both simulationand real-time closed-loop control.2. How to mitigate issues related to uncertainties and discrepancies betweenthe twin and the physical plant.3. How digital twin merges with big data. Are big data related technologiessuch as machine learning and statistical approaches part of digital twin orseparate services? What are use-case scenarios of digital twin with big data?4. How to evaluate a digital twin that it faithfully mirrors its twinned system,how to quantitatively measure and compare two or more different digitaltwins of the same factory?5. What are the requirements of a digital twin development tool that addressesthe aforementioned questions.The purpose of this paper is to review current state-of-art on tools and de-velopments of digital twin in CPPS, discuss open research problems and suggestpotential directions to address the aforementioned gaps.The rest of this paper is organised as follows. Section 2 presents literaturereview on digital twin architectures and modelling strategies. Section 3 discussesan overview of digital twin in the context of smart manufacturing and CPPS.Section 4 introduces different models of computation. General-purpose as wellas field-specific modelling tools are presented in Section 5. An industrial man-ufacturing case study with open research questions are presented in Section 6.Finally, conclusions are given in Section 7.
Digital twin is the cornerstone of the Industry 4.0 wave which has been an activeresearch area in the past several years. Table 1 shows some of the recent literaturewhere digital twin is introduced for addressing various problems in a number ofdomains. The use of different modelling tools and whether the author highlightsthe use of big data analytics in their work are also indicated in the table.
Park et al.
Table 1.
Digital twin found in literature
Ref Year Domain Purpose Big data Tools used [16] 2016 General CPS Model-based design tool for multidisciplinary,collaborating modelling of CPS No Modelio, Overture,Modelica, 20-sim, FMI[34] 2017 Manufacturing Predictive maintenance, simulation, resiliency Yes –[1] 2017 Vehicletelematics Optimising communication cost No Qfsm[19] 2017 Manufacturing Predictive maintenance, applicationdevelopment, simulation. Yes –[20] 2017 Process plants Predictive maintenance, monitoring, 3Dvisualisation, simulation Yes ANSYS Simplorer,SCADE, PTCThingWorx[4] 2017 General CPS CPS implementation No –[18] 2017 Structuralhealthmonitoring Predictive maintenance No –[31] 2018 Manufacturing Simulation-based systems engineering No SysML[33] 2018 Manufacturing Product lifecylcle management Yes –[27] 2018 Manufacturing Modelling, analysis, simulation No –[30] 2018 Manufacturing Modelling, simulation No Modelica, FEM
An architecture of digital twin shop-floor (DTS) is presented in [34]. Thereare four main components in this architecture: physical shop-floor (PS), virtualshop-floor (VS), shop-floor service system (SSS), and shop-floor digital twin data(SDTD). VS is a digital twin of the PS and data generated from both PS and VSare merged into the SDTD database. SSS consists of many sub-services, whichare transformed into composite services based on demands from PS and VS. VSis used to simulate, predict, and perform calibration using the real-time datagenerated from PS. While the work gives a good overview of digital twin basedphysical and virtual space interconnection, authors do not tackle modelling VSdirectly although they suggest a number of tools that can be used to model VSin different levels of hierarchy: geometry, physics, behaviour, and rule. The sameresearch group also showed employing digital twin for product development andfor managing entire product life-cycle [33,28]. In particular, [28] presented simi-larities and differences between digital twin and big data technologies and howthey can be complementary with each other to enhance an overall manufacturingprocess.A reference model for digital twin architecture for the cloud-based cyber-physical systems is presented in [1]. The architecture consists of three interme-diary layers, namely cyber-things layer, peer-to-peer communication layer, andintelligent service layer. In this work, closely related physical and cyber things can create communication groups and peer-to-peer communication channels be-tween those things are formed based on networking or communication criteriausing a Bayesian belief network.In [31], authors propose a concept of experimentable digital twin (EDT) thatcombines the ideas of digital twins with model-based systems engineering andsimulation technology. In addition to digital twin itself, EDT also comprises of amodel of external environment that interacts with the twin via simulated sensorsand actuators. The authors showed modelling of automotive headlight housingassembly using the EDT approach.A case study for modelling an industrial machine tool is presented in [30]. Inthis work, authors use Finite Element Method (FEM) based preprocessing ap-proach for modelling structural flexibility of machine’s components. The models hallenges in Digital Twin Development for CPPS 5 of cutting process, kinematic chains, and control systems are developed usingModelica [8].An integrated tool-chain for model-based design of cyber-physical systemsis introduced in [16]. The tool enables co-simulation of multi-domain modelsby providing an integrated framework that combines multiple tools. They useUnifying Theories of Programming (UTP) [13] as a foundation to give semanticsto their heterogeneous approach.ANSYS and PTC worked together to demonstrate how a digital twin of apump can help diagnose and solve operating problems faster. The pump modelis developed using ANSYS’s Simplorer and SCADE whereas PTC’s ThingWorxplatform, which provides data collection and analysis services, is used to createan IoT ecosystem for devices and sensors [20].A concept of software-defined control (SDC) has emerged recently [19,27]which is inspired from the traditional software-defined networking (SDN), forprogrammatically configuring a communication network. A basic idea of SDC isto provide a central controller that has a global view of a system and separatedecision making logic from the operations management solutions. In this frame-work, digital twin is suggested as a core simulation engine to improve decisionmaking and detect faults in the manufacturing systems.A concept of dynamic Bayesian network based digital twin is introduced inin [18] for monitoring the health of an aircraft. The authors propose a mod-ification to the traditional probabilistic network that significantly reduces thecomputational cost for Bayesian interference. The approach integrates physicsmodels with sources of uncertainty to predict crack growth on the airframe.Authors in [4] discuss modelling, discretisation, executability, simulation,and implementation of cyber-physical systems. The paper highlights the need ofmethods and tools with appropriate design languages underpinned by a solid se-mantic foundation which can model complex electromechanical systems. Variousmodelling techniques and respective challenges are discussed including detectionof Zeno behaviours, difficulties in simulation of hybrid models in the presence ofdifferential algebraic equations (DAE), physical systems modelling using linearand bond graphs. Although the authors do not particularly relate their workwith digital twin development, the techniques and the design flow presented inthe paper should definitely be considered in any digital twin development tools.In literature, it is shown that digital twin is appeared the most in the man-ufacturing domain for predictive maintenance and simulation purposes. Severalworks [34,33,19] propose the digital twin based software architectures and fore-see the importance of big data analysis and its linkage with the digital twin.The works in [4,18,30,27,16] focus more on the modelling aspects. Nevertheless,implementation details of the software architectures and their use case scenar-ios are somewhat abstracted. In addition, most of works lack a comprehensivemodelling framework except [27], whose target is yet limited to discrete eventsystems, and [16], which focuses more on simulation of cyber-physical systemsin general rather than digital twin itself. The next section discusses digital twinin the context of CPPS.
Park et al.
Cyber-Physical Production System
Digital Twin P l an t M ode l A b s t r a c t i on M anage r Network Component Models
Factory
Applications A pp li c a t i on s I n t e r f a c e ( A pp I ) W i - F i , E t he r ne t, B l ue t oo t h Network topologyScheduling optimisation
Predictive maintenance
Resilience Manager
Machine learning
Simulation
Real-time control
Shop-floor planner ???? F a c t o r y I n t e r f a c e ( F I ) Physical SpaceCyber Space
Conveyor Belt
PlantController
Conveyor Belt
PlantController
SMT
PlantController
SMT
PlantController
Drilling Station
PlantController
Drilling Station
PlantController
Pressing StationPressing Station
Plant
Controller
Pressing Station
Plant
Controller
Fig. 2.
An overview of the proposed Cyber-Physical Production System (CCPS)architecture
In literature and industries, the terms “Smart Manufacturing” and “Industry4.0” are being used interchangeably and they are now almost synonymous witheach other. The main objective is to leverage the recent advancement of in-formation and communication technologies, such as cloud computing, IoT, andbig data, to achieve autonomous, self-optimising, and self-diagnosing capabilitiesthat can mitigate various problems in complex manufacturing scenarios. Digitaltwin is typically used in the context of cyber-physical systems to mirror thelife of its real object via the best available physical models and sensor data. Asa result, the digital twin enables simulation of real-world scenarios in a cyberworld that otherwise would cost considerable amount of resource and time. Fig. 2shows our proposed digital twin architecture for CPPS.CPPS is a mechanism used in smart manufacturing and Industry 4.0 designprinciples. It is comprised of five main components: (1) a factory, (2) a digitaltwin and its runtime environment, (3) a factory interface to extract sensor/ac-tuator data from the physical space, (4) an application interface that providesApplication Programming Interfaces (APIs) to applications that wish to utilisethe digital twin, and (5) the application themselves.Most often a factory, also called a physical plant , is a hybrid system, whichis characterised with both continuous and discrete dynamics, modelled with asystem of ODEs and DAEs, and state transitions, respectively. Examples ofcontinuous dynamics of the plant are a movement of a workpiece on a conveyorbelt and a movement of mechatronic arms during an assembly operation. On theother hand, discrete dynamics such as switching the operation mode from activeto idle when no workpiece is detected for a certain amount of time.The Plant Model Abstraction Manager and the Network Component Modelsare the parts of the digital twin runtime environment which manage lifecycles ofindividual twins and communication channels between them. Each twin consistsof models of a plant and a controller that form a closed-loop control system viafeedback and control signals. Interactions between these two models in the cyber hallenges in Digital Twin Development for CPPS 7
Table 2.
Digital twin use case scenarios
Features References
Root-cause failure analysis andpredictive maintenance – Detection of a faulty valve [20]. – Data-oriented analysis and prediction for wind turbines [9].High-fidelity simulation – Water pump simulation [20]. – Financial and risk simulation [9]. – Testing machineries for filling and packaging medications [32]. – Simulation of a sheet metal punching machine [24].Closed-loop real-time control – Real-time control of the water pump [20]. – Turbine control [9]. – Human-robot collaborative assembly system [37].3D visualisation –
3D simulation model that shows cavitation inside a waterpump [20]. –
3D visualisation of pharmaceutical machines [32]. –
3D model of a brake pad wear [21]. –
3D visualisation of the punching machine process [24]. space will also be accurately reflected in the physical space through the FactoryInterface (FI).The framework also provides a set of APIs that allow user applications tointeract with the digital twin via the Application Interface (AppI). In this paper,we do not focus on the design of applications that utilise the digital twin, butrather on the digital twin development itself. However, it is worth to note herethat technologies such as big data analysis can be employed in the application todeduce useful insights about the physical plant. The digital twin can communi-cate with this application to enhance its functionalities, for example, predictingpotential failures and optimising throughput by adjusting its parameters. Sim-ilarly, the quality of data analysis can also be improved through data fusion ofthe physical plant and digital twin simulation.Digital twin can be modelled using various levels of abstraction. Dependingon requirements and resource availability, designers may choose a low-fidelitymodel such as finite-state machine or a higher-fidelity hybrid automaton or fi-nite element method (FEM). Specification of plant components at different lev-els fidelity allows generation of mixed-fidelity digital twin , which is a trade-offbetween accuracy and scalability. This trade-off makes implementing CPPS inbigger scale more practical since it would be computationally expensive to realiseevery aspect of physical plants using high-fidelity modelling approaches.In manufacturing, employing a digital twin based cyber-physical system isfavourable in a variety of scenarios. Table 2 shows examples of digital twin usecase scenarios, which are grouped in four different categories. Since digital twinsare often developed to handle multiple problems, most of the works overlap witheach other in those categories:1.
Root-cause failure analysis and predictive maintenance : When there is a faultin the system, operators can utilise digital twin to pinpoint a root-causethanks to its rich structural and behavioural information about the twinnedsystem. In this case, choosing an appropriate formal model of computation (MoC) of the plant would play a pivotal role that facilitates semantic preser-vation between different design phases. From this, the digital twin will be
Park et al. : Supported : Supported via extension : Notsupported
MoCs Physicaltime Hierarchy Concurrency Data ODEsSync AsyncFSMPNCPNTPNSRGALSTAHA (a) A feature comparison among differentMoCs
FSMTA,TPNHA
AccurateUncertainOpen-loop Highly-certainClosed-loop
PN,CPN,SR,GALSPN,CPN,SR,GALS
Observability F i d e l i t y (b) Relationship between model fidelityand observability of a physical plant Fig. 3.
Models of computation and observability of a production systemable to back-trace the failure from an executable code all the way back tothe modelling phase.2.
High-fidelity simulation : Data analytics, storage, and together with high-fidelity modelling techniques will enable the digital twin to assess differentmanufacturing as well as fault scenarios which provides highly accurate re-sults that would be difficult to obtain via traditional simulation techniques.3.
Closed-loop real-time control : Digital twin augments the quality of the clas-sical plant and controller closed-loop control where it can add additionalvalues such as dynamicity, reconfigurability, connectivity, global intelligence,and predictability. This is possible because the digital twin can directly in-fluence a behaviour of the physical plant and also acts as an intermediaryentity between the plant and the user applications which consume the twinas a service. In this context, we foresee that the digital twin has a greatchance to become a technological gateway to a wide range of cyber-physicalsystem applications.4.
3D visualisation : The multi-domain and multi-physics design approachesmake it natural for the digital twin to be a good candidate for 3D visualisa-tion for interactive validation and inspection.In the next section, a number of well-known modelling strategies is presentedthat could be a back-end of the digital twin development framework.
Model of computation (MoC) deals with a set of theoretical choices that buildan execution model of the design language. MoCs define how computations arecarried out, for example sequentially or concurrently, interaction/synchronisa-tion between computational units, and a notion of time, etc., without bindingthem to specific implementation details. Programming language equipped with hallenges in Digital Twin Development for CPPS 9 formal MoC ensures that the resulting program follows the semantics of the cor-responding MoC. On the other hand, a programming model for a certain MoCcan also be built on a top of host programming language, for example througha library or a framework.Many formal MoCs available for modelling a digital twin [6]. Some of themare shown in Fig. 3, which can be basis of the digital twin development frame-work. Each MoC supports different features:
Physical time determines an abilityfor a MoC to relate its computations with a continuously evolving quantity, typ-ically in real or integer domains. A simple example is when a Timed Automata(TA) [2] generate an output event after a clock value reaches 5.
Hierarchicalcomposition describes an ability to compose two or more basic design entities toachieve a more advanced and complex functionality. This can be easily found inmany digital systems designs, for example an arithmetic logic unit composed ofadders and subtractors, etc.
Concurrency describes an ability of basic entitiesin a model that can be executed in overlapping time. Typically, it refers to ei-ther synchronous or asynchronous concurrency.
Data indicates whether a MoCsupports for variables, expressions, constructs, type systems, and etc., to per-form data-oriented algorithmic computations. Lastly,
ODEs refer to an abilityto capture continuous dynamics of a system via a system of ordinary differentialequations. There is no MoC that supports all the features. An FSM can be usedto capture various control-dominated behaviours. However, it alone is not wellsuitable for accurately capturing a complex nature of physical systems that areinherently concurrent, time-dependent, and data-rich.Petri-Nets [26] are good models for describing concurrent and distributed sys-tems. However, it lacks expressiveness in data computations and timing prop-erties. Coloured Petri-Nets (CPN) [14] and Timed Petri-Net (TPN) [29] areextensions of the PN with data and timing features, respectively. Nevertheless,they lack in capturing continuous dynamics of a system. Although the originalformal definition of the PN do not include hierarchy, there is an extension of PNthat supports hierarchical composition [7].Synchronous Reactive (SR) MoC [3] provides a set of constructs for captur-ing reactive behaviours of a system. The execution semantics of SR constructsunderpinned with rigorous mathematical foundation enables bug avoidance inthe early design phase using correct-by-construction compilation and verifica-tion techniques [3], which increase confidence in the correctness of final designs.While SR MoC has a notion of time, they are only logical. Moreover, the SR MoCis not amenable to a large distributed systems since every design component issynchronised with a single global clock.Globally Asynchronous Locally Synchronous (GALS) [22] is a superset ofthe SR MoC where a system is modelled using several synchronous subsystems,which run asynchronously and communicate with each other via a message pass-ing mechanism. Still, the most of GALS-based modelling languages often lack anability to model real-time behaviours and continuous dynamics of many physicalsystems.
Hybrid Automata (HA) [12] supports modelling of both discrete and contin-uous dynamics of a system. It uses a state machine with a finite set of real-valuedvariables whose values evolve according to a set of ODEs. A notion of referencetime can also be defined in HA using a simple formula such as ˙ t = 1. However, itis a flat structure similar to an FSM, and does not directly support concurrencyand hierarchical composition.Generating a highly accurate, time-synchronised, and scalable digital twin,therefore requires a technique that combines various MoCs to complement eachother’s weaknesses. Generally, fidelity of a model is based on how well it cancapture dynamics of physical phenomena. On the other hand, effectiveness of amodelling technique varies depending on observability of the physical plant. Asshown in Fig. 3b, the relationship between fidelity of a model and observabilityof a physical plant categorises the digital twin based CPPS into largely fourgroups:1. Accurate but uncertain model : High-fidelity models can accurately capturea plant’s behaviour according to the specification. However, since a less ob-servable plant only provides feedback at certain discrete instants, there is alow confidence on dynamics of the plant between two feedback events. Forexample, a digital twin can only speculate the current position of an itemthat moves along a conveyor belt based on the motor speed. Any unexpectedevents that occur between two photoelectric sensors cannot be captured bythe digital twin unless the item is enabled with RFID tracking.2.
Less accurate and uncertain model : In this case, a model does not capture theplant’s behaviour accurately and the plant is also less observable. An exampleof this scenario is using an untimed model such as FSM for modelling achange of the temperature of the boiling water and the plant only providesfeedback when the temperature reaches at T max threshold.3. Less accurate and highly-certain model : In this scenario, the plant is fullyobservable where it provides changes its states in a frequent manner. Low-fidelity models, however may not fully utilise such information. For example,for a multi-axis arm movements, discrete event models may only react toevents when the arm only reaches its final destination of the movement,ignoring how it reaches there.4.
Accurate and highly-certain model : This is when both the model and theplant fully synchronises and the model can accurately trace dynamics of theplant at any instants of time.In the next section, an overview of existing modelling tools are presented.These tools are used for general cyber-physical system development, but alsocan target digital twin.
Table 3 summarises various types of modelling tools available which targetgeneral-purpose discrete and continuous systems as well as more application hallenges in Digital Twin Development for CPPS 11
Table 3.
A summary of modelling tools
Tools Licence Formal MoC Codegeneration
SCADE Simplorer Commercial Multi-paradigm –SCADE Suite Commercial Synchronous DataflowANSYS Twin Builder Commercial Multi-paradigm –SCILAB GPLv2 – via thrid-party plug-ins: X2C, Project-PModelica families Open source and commercial – (some implemen-tations)MATLAB/Simulink Commercial –BCVTB Modified BSD Actor-based –INTO-CPS Open source VDM, SysMLCIF MIT HAFlow* GPL HA –Flexsim Commercial – –HyST LGPL HA (To other HAmodels)SL2SX GPLv3 – (Simulink toSpaceEx)SpaceEx GPLv3 HA –IOPT Free (Web-based) Petri-NetCPN Tools GPLv2 Coloured Petri-Net –Ptolemy II Mixed Actor-oriented on topof heterogenous MoCEsterel Mixed and commercial Synchronous ReactiveLustre Free and commercial Synchronous DataflowC´eu MIT Synchronous Reactive specific fields. It is not possible to cover all of them here, but we selected someof the notable tools currently available in the market and academia.Tools like Scilab, Modelica, and Matlab/Simulink provide general-purpose,numerical computing environment for modelling a wide range of systems such asmechanical and electrical systems, fluid dynamics, and etc. The SCADE Suite isa model-based development environment for mission critical embedded software https://simulationresearch.lbl.gov/bcvtb https://into-cps.github.io/ http://cif.se.wtb.tue.nl/index.html https://flowstar.org https://github.com/nikos-kekatos/SL2SX http://spaceex.imag.fr/ http://gres.uninova.pt/IOPT-Tools/login.php http://cpntools.org/ https://ptolemy.berkeley.edu with Lustre as its core language. ANSYS Simplorer and Digital Twin Builderprovide multi-domain, co-simulation environment with support for VHDL-AMS,Modelica, C/C++, and SPICE languages along with MIL (Model-in-the-Loop),and SIL (Software-in-the-Loop) capabilities. The model can be connected tovarious industrial IoT platforms such as PTC ThingWorx, GE Predix, and SAPLeonardo.There are also tools equipped with formal models of computation. For exam-ple, synchronous reactive languages such as Esterel, Lustre, and C´eu and Petri-Net based IOPT and CPN Tools. Building Controls Virtual Test Bed (BCVTB)is a software environment based on Ptolemy II project [6], to couple differentsimulation programs for co-simulation.Modelling tools which target hybrid systems have direct support for capturingboth discrete and continuous dynamics and transition between these modes. Forexample, Compositional Interchange Format (CIF) is a automata-based languagefor the specification of discrete event, timed, and hybrid systems. HyST andSL2SX do source-to-source transformation to enable evaluation of HA modelsusing different tools. Model checking tools for HA models also exist, for instanceSpaceEx and Flow*.Flexsim is a simulation software for manufacturing factories including 3Dvisualisation and statistical reporting and analysis features. The Integrated ToolChain for Model-based Design of Cyber-Physical Systems (INTO-CPS) is anintegrated tool chain for comprehensive model-based design of cyber-physicalsystems. It aims to provide a framework for model-based design and analysis bycombining multiple models generated from different tools using the FunctionalMock-up Interface (FMI) standard [5].In the next section, a case study called the IMPACT manufacturing line isintroduced where an initial concept of the digital twin modelling is presented. To illustrate the use of heterogeneous models and discuss open issues, we proposedeveloping a cyber-physical production system case study called the IMPACT
Fig. 4.
The IMPACT line – A testbed for future manufacturing hallenges in Digital Twin Development for CPPS 13 line shown in Fig. 4. It consists of four linear modules with parallel conveyorsand seven processing stations for manufacturing smart phones. The followingoutlines the operation of each station:1.
High-Bay Rack – When there is a request from an external source for man-ufacturing a product, a cartesian robot is triggered to pick-up a workpiecepallet. The pallet is then placed on the conveyor belt. A motor for the con-veyor belt is turned on to move the pallet to the drilling module2.
Drilling Module – Two drilling spindles are advanced in the Z and X direc-tions to make two pairs of holes into the workpiece.3.
Robot Assembly Module – The 6-axis arm picks up the rear panel from thepallet and place it on the processing bay. The arm places a PCB on the paneland switches to a smaller gripper so that it can pick up and install a fuseinto the PCB. The arm switches to the original gripper to place the rearpanel with the finished PCB back to the pallet.4.
Mobile Station – This module delivers boxes of PCBs to the IMPACT line.5.
Robotino – This automated guided vehicle transfers a box of PCBs from theMobile Station to the Robot Assembly Module.6.
Magazine Module – This module places a front panel on the PCB.7.
Pressing Module – This module applies pressing force to seal the product.This CPPS has a number of characteristics for demonstrating the need fordigital twin development: (1) modelling continuous dynamics such as the 6-axisarm and cartesian robot movements, (2) choosing the right modelling strategiesfor the machines with different observabilities, (3) detecting faults such as unsat-isfactory drilling operations due to wear and tear of a drill bit and misplacementof fuses on a PCB, (4) and real-time closed-loop control.
This section focuses modelling the motion of the conveyor belt on the linearmodules and the cartesian robot arm in the High-Bay Rack for the illustrationpurpose. The first model of the conveyor belt is shown in Fig. 5. This is thesimplest possible case where the conveyor belt operates in either one of two macrostates:
Idle or On mode. When T urnOn signal is set high by the controller, themachine makes transition to state On after setting v = 0 .
03, which indicates thespeed of the conveyor belt. The conveyor belt stays on as long as an incomingworkpiece
W P resets a timer via
Reset . It goes
Idle mode when
T urnOn signal
Idle start On TurnOnv =0 . ¬ TurnOn ∨ TimeOutv =0 WPReset
Fig. 5.
An FSM model of the conveyor belt q (cid:48) ( x, (cid:48) ( y, (cid:48) ( z, U T d = x ∧| x d − i | ≥ . d,i ) q UinM x x d T ( d,i ) q U T ( d,i ) T d = y ∧| y d − i | ≥ . d,i ) inM y y d q UT d = z ∧| z d − i | ≥ . d,i ) inM z Z d q U T ( d,i ) T ( d,i ) inT x einT y einT z e ( d,i )( d,i )( d,i ) ( d,i )( d,i ) ( d,i )( d,i ) (a) A Petri-Net model of the High-Bay Rack ˙ x = 0start q ˙ x = 0 . | x d − x | ≤ . q ˙ x = − . | x d − x | ≤ . q x d − x ≥ . x d − x ≤ − . ee (b) A hybrid automaton for cap-turing movement of the carte-sian robot arm Fig. 6.
Modelling the High-Bay Rackis unset by the controller or there were no incoming workpieces for the last x time period indicated by T imeOut .The cartesian robot arm in the High-Bay Rack (HBR) is able to move inx,y,z directions simultaneously. Let us assume a designer has chosen ColouredPetri-Net (CPN) to model the robot since it features both concurrency and datamanipulation for capturing the position of the arm in the cartesian space. Thecorresponding CPN model is shown in Fig. 6a. This CPN has four colours (datatypes) and two variables defined as follows.type d i r = | x | y | ztype pos = i n ttype U = d i r ∗ postype E = evar d : d i rvar i : posThe three outgoing arcs from q indicate the robot arm can move in x,y,zdirections simultaneously when it is requested via input signals inM x , inM y ,and inM z . The initial marking at q shows there are three tokens of type U . Theplus sign indicates the tokens are combined into a multiset. Consider a tokentravelling from q to q via transitions T and T . The guard on T ensures thatit is only enabled when there is a token of type U in q whose first element is x .When there is a request to move the arm in x direction, i.e. when | x d − i | ≥ . q indicating the model is now in the “moving”state. The colour set U on q specifies type of tokens which may reside on theplace.One of the requirements of the digital twin framework is to provide a real-time view of the current state of the modelling plant. However at this point, thedesigner realises that he/she cannot easily capture continuous dynamics of the hallenges in Digital Twin Development for CPPS 15 ˙ x = v ˙ v = 0 . x < Acc ˙ x = 1 Const ˙ x = v ˙ v = − . x > Dec ˙ x = 0˙ v = 0 Idle ˙ x = 1 P ˙ x = 0 P P P WPReset
Propositions: P = TurnOnP = ¬ TurnOn ∨ TimeOut (a) The conveyor belt in HA (b) The speed of the conveyor belt
Fig. 7.
The conveyor belt refinedarm movement using CPN since it does not support the flow of continuous vari-ables. Assume that the tool allows to mix different models so that the designerdoes not have to redesign the robot arm from scratch. In this case, HA can beadded that runs together with the previously designed CPN as shown in Fig. 6b.Here, the HA makes a transition from q to q or q when the final position x is greater or smaller than the current coordinate, respectively. States q and q show continuous evolution of variable x using its derivative ˙ x , which is equivalentto i of the token ( x, i ) in q in Fig. 6a. When the arm reaches its destination, i.e. | x d − x | ≤ .
03, the HA makes transition back to q while generating an event e .This event enables transition T in Fig. 6a and results in returning of the tokenback to q for serving the next request for x direction movement.By supporting different levels of abstraction, a model can be enhanced furtherto incorporate finer details of operating modes. For example, after turning on aconveyor from the idle state, it may take some time until the motor reaches fullrotational speed. Thus, the original conveyor belt modelled in an FSM can befurther refined using HA, which is shown in Fig. 7a. Here, the system starts withthe state Acc where the speed increases according to the flow variable ˙ v = 0 . Const . The system stays in this state as long as the next workpiecearrives before
T imeOut or a controller resets signal
T urnOn . The state
Dec decelerates the speed of the conveyor belt until it reaches zero and the systemstays in
Idle until
T urnOn signal is set by the controller. The change of speedof the conveyor belt over the period of 30 seconds is shown in Fig. 7b, which ismodelled in Modelica. Similar approaches can be applied to model rest of thestations of the IMPACT line.
Discrepancy between a model and aphysical system may arise due to several reasons. For example, increase in com- plexity, choosing an inadequate modelling strategy, accumulation of errors overtime, lack of documentation about the system, and lack of observable states asexplained in Section 4. Unlike traditional software development process, wherea set of test cases can be executed on a single development computer, func-tionalities of digital twins cannot be easily validated unless they are tested andcompared with the real system, which is often difficult and time consuming.Moreover, it is unrealistic to validate every aspect of physical phenomenon ofthe twinned system – should a designer consider all the circuitries, propertiesof the materials used, geometry, multibody, and etc. of the IMPACT line? Howmuch of the physical system can be abstracted without causing significant dis-crepancy? For example, after deploying the digital twin for the IMPACT line,discrepancies between the physical plant and the twin might be found based onfeedback data collected from the two. In this case, there should be a facilityto minimise the error by automatically adjusting the parameters of the twin oreven modify the model during runtime. If the error cannot be minimised, thecorresponding issue can be reported and possible solutions may be suggested tothe user. To realise a faithful digital twin, therefore, the development frameworkshould utilise both modelling (offline) and post-calibration (online) techniquesto manage and reduce discrepancy and uncertainty in the digital twin model.
Interplay with Big Data.
Arguably, big data plays an important role inIndustry 4.0. However, there are still many questions need to be answered onhow digital twin models interact with big data: what are the gains and losseswhen employing digital twin based simulations or data-driven approaches forpredictive maintenance, decision making, fault detection, etc. For example, dataanalytics might detect anomalies in a physical system but may not have a formalmodel of the plant to draw possible solutions and the root-cause of the problem.On the other hand, a formal model of digital twin itself may not have an abilityto deduce useful insights from data generated from the physical plant or high-fidelity simulation. Undoubtedly, there is a potential synergy between the bigdata and formal modelling of digital twin. Recently, big data topic in digitaltwin modelling has been recognised in several literature [28,11,38]. However, aconcrete use-case scenarios need to be further investigated.A concept of Data-driven and Model-based Design (DMD) is discussed in[35] which is an extension of the Model-Based Design (MBD) with elementsof data-driven approaches such as machine learning and model learning. Themain focus of DMD is to leverage the advances of AI while preserving meritsof MBD approaches such as mathematical formalism and analysability. In par-ticular, model learning [36] can be useful for building formal models of legacy(black-box) models or for analysing complex systems when there is a lack oftools. Nevertheless, scalability issue when increasing number of states in thesystem and its applicability for building hybrid system are still in question.
Integration of Heterogeneous Models.
There should be a consistent way tocombine heterogeneous models to support design of the “mixed-fidelity” digital hallenges in Digital Twin Development for CPPS 17 twin. This requires the formal modelling framework that can capture physicalbehaviour in different levels of abstraction. In case of designing the IMPACT line,it would be easier for a designer to first model a behaviour of the system in coarserlevel, e.g. flow of workpieces between modules using Petri-Nets or abstractingcontinuous dynamics using macro states in FSM, and further refine individualmodules using high-fidelity models as required. In this case, a consistent methodfor exchanging information between different models should be developed. Forexample, when combining two models in Fig. 6, it is crucial that the event e generated from the HA model is always captured in the Petri-Net model. Thisimplies that the models should synchronise whenever the invariant condition | x d − x | ≥ .
03 becomes unsatisfied. Implementation-wise, the FMI standard [5]can be employed for interconnecting the heterogeneous models and implementinga deterministic mechanism for synchronisation and data exchange.
Time-Critical Systems.
Some of the cyber-physical production systems aretime-critical systems, for example Surface-Mount Technology (SMT) placementequipment. In this case, the worst-case response time for communication betweenthe twin and the physical system must be guaranteed and within a boundedtime. Furthermore, the worst-case execution time of the twin and its controllershould also be bounded and analysable. Many digital twin architectures foundin literature rely on information flow through the cloud and the non time-criticalnetwork interfaces, which is not suitable to build time-critical applications. Asoftware architecture that comprises time-critical components are required forcertain types of application.
Security.
Since digital twins are tightly coupled with physical environment, anattack on a such cyber-physical system may endanger people’s safety and resultin significant economic loss. Applications interact with digital twins to accesssensor data and actuate the physical system when necessary (see Section 3).Since user applications can be a major source of security threat, a secure accesscontrol mechanism needs to be implemented for those that access digital twins.In summary, security aspects of digital twin should be studied in the areas ofmalicious activities detection, cryptography, resiliency to cyber attacks.
To address the aforementioned issues, the desired capabilities of digital twinneed to be clearly defined first which will be the basis for the development ofthe underlying software/hardware architecture (see Fig. 2). More specifically, weanticipate that through digital twin, users should be able to (1) control physicalobject, (2) monitor and analyse data for minimising discrepancy and uncer-tainties in a model, and (3) simulate the physical counterparts by instantiatingdigital twins in a cyber space. To achieve this, a basic design block will be in-troduced namely a digital twin component (DTC). It consists of a twin itself,a virtual controller, database for storing cyber and physical data, and a digital twin runtime that manages data collection and synchronisation between the twinand the physical object. Furthermore, DTC should support incremental updateof the twin as well as the controller to handle discrepancies and uncertaintiesas much as possible. Multiple DTCs can be combined to create a bigger systemand we plan to adopt Globally Asynchronous Locally Synchronous paradigm fortheir integration. On the highest level of the architecture, we plan to provide aset of services, for example DTC repository, DTC management (i.e. creation anddisposal), analytics, and data access etc., with which more intelligent applica-tions such as prognostic and health management and virtual commissioning canbe developed. The digital twin framework therefore will consist of a tool and thearchitecture that enables the aforementioned capabilities for CPPS.
In this paper, we presented an overview and open research questions in digitaltwin development for cyber-physical production systems (CPPS). Digital twinis still a relatively new concept that requires more researches in the fields ofmodelling and integration with other technologies. We propose development ofa tool that supports modelling CPPS in various levels of abstraction under-pinned by formal mathematical models. The result is a mixed-fidelity digitaltwin which is a trade-off between accuracy and scalability. In future work, weplan to develop formal semantics for translating different models into an inter-mediate representation for analysis and code generation and implement a digitaltwin architecture described in Section 6.3 that enables control, monitoring, andsimulation of physical plant in the cyber space.
References
1. Alam, K.M., El Saddik, A.: C2PS: A Digital Twin Architecture Reference Modelfor the Cloud-based Cyber-Physical Systems. IEEE Access , 2050–2062 (2017)2. Alur, R., Dill, D.: A Theory of Timed Automata. In: Workshop/School/Symposiumof the REX Project (Research and Education in Concurrent Systems). pp. 45–73.Springer (1991)3. Berry, G., Gonthier, G.: The ESTEREL Synchronous Programming Language:Design, Semantics, Implementation. Science of computer programming (6), 61–70 (2015)15. Kagermann, H., Helbig, J., Hellinger, A., Wahlster, W.: Recommendations forImplementing the Strategic Initiative INDUSTRIE 4.0: Securing the Future ofGerman Manufacturing Industry; Final Report of the Industrie 4.0 Working Group.Forschungsunion (2013)16. Larsen, P.G., Fitzgerald, J., Woodcock, J., Fritzson, P., Brauer, J., Kleijn, C.,Lecomte, T., Pfeil, M., Green, O., Basagiannis, S., et al.: Integrated Tool Chainfor Model-based Design of Cyber-Physical Systems: The INTO-CPS Project. In:Modelling, Analysis, and Control of Complex CPS (CPS Data), 2016 2nd Interna-tional Workshop on. pp. 1–6. IEEE (2016)17. Lee, E.A.: Cyber Physical Systems: Design Challenges. In: Object Oriented Real-Time Distributed Computing (ISORC), 2008 11th IEEE international symposiumon. pp. 363–369. IEEE (2008)18. Li, C., Mahadevan, S., Ling, Y., Wang, L., Choze, S.: A Dynamic Bayesian Net-work Approach for Digital Twin. In: 19th AIAA Non-Deterministic ApproachesConference. p. 1566 (2017)19. Lopez, F., Shao, Y., Mao, Z.M., Moyne, J., Barton, K., Tilbury, D.: A Software-Defined Framework for the Integrated Management of Smart Manufacturing Sys-tems. Manufacturing Letters (2017)20. MacDonald, C., Dion, B., Davoudabadi, M.: Creating a Digital Twin for a Pump.ANSYS Advantage Issue 1 p. 8 (2017)21. Magargle, R., Johnson, L., Mandloi, P., Davoudabadi, P., Kesarkar, O., Krish-naswamy, S., Batteh, J., Pitchaikani, A.: A Simulation-Based Digital Twin forModel-Driven Health Monitoring and Predictive Maintenance of an AutomotiveBraking System. In: Proceedings of the 12th International Modelica Conference,Prague, Czech Republic, May 15-17, 2017. pp. 35–46. No. 132, Link¨oping Univer-sity Electronic Press (2017)0 Park et al.22. Malik, A., Salcic, Z., Roop, P.S., Girault, A.: SystemJ: A GALS Language forSystem Level Design. Computer Languages, Systems & Structures (2), 365–373 (2017)25. Negri, E., Fumagalli, L., Macchi, M.: A Review of the Roles of Digital Twin inCPS-based Production Systems. Procedia Manufacturing , 939–948 (2017)26. Peterson, J.L.: Petri Net Theory and the Modeling of Systems. Prentice Hall PTR,Upper Saddle River, NJ, USA (1981)27. Potok, M., Chen, C.Y., Mitra, S., Mohan, S.: SDCWorks: A Formal Framework forSoftware Defined Control of Smart Manufacturing Systems. In: Proceedings of the9th ACM/IEEE International Conference on Cyber-Physical Systems. pp. 88–97.IEEE Press (2018)28. Qi, Q., Tao, F.: Digital Twin and Big Data Towards Smart Manufacturing andIndustry 4.0: 360 Degree Comparison. IEEE Access , 3585–3593 (2018)29. Ramchandani, C.: Analysis of Asynchronous Concurrent Systems by Petri Nets.Tech. rep., MASSACHUSETTS INST OF TECH CAMBRIDGE PROJECT MAC(1974)30. Scaglioni, B., Ferretti, G.: Towards Digital Twins Through Object-Oriented Mod-elling: A Machine Tool Case Study. IFAC-PAPERSONLINE pp. 613–618 (2018)31. Schluse, M., Priggemeyer, M., Atorf, L., Rossmann, J.: Experimentable Digi-tal Twins—Streamlining Simulation-Based Systems Engineering for Industry 4.0.IEEE Transactions on Industrial Informatics (9-12), 3563–3576 (2018)34. Tao, F., Zhang, M.: Digital Twin Shop-Floor: A New Shop-Floor Paradigm To-wards Smart Manufacturing. IEEE Access , 20418–20427 (2017)35. Tripakis, S.: Data-Driven and Model-based Design. In: 2018 IEEE IndustrialCyber-Physical Systems (ICPS). pp. 103–108. IEEE (2018)36. Vaandrager, F.: Model Learning. Communications of the ACM (2), 86–95 (2017)37. Wang, X.V., Kem´eny, Z., V´ancza, J., Wang, L.: Human–Robot Collaborative As-sembly in Cyber-Physical Production: Classification Framework and Implementa-tion. CIRP annals (1), 5–8 (2017)38. Zhuang, C., Liu, J., Xiong, H.: Digital Twin-based Smart Production Managementand Control Framework for the Complex Product Assembly Shop-Floor. The In-ternational Journal of Advanced Manufacturing Technology96