REPLICA: A Solution for Next Generation IoT and Digital Twin Based Fault Diagnosis and Predictive Maintenance Rosaria Rossini, Davide Conzon, Gianluca Prato, João Reis, Gil Gonçalves Claudio Pastrone SYSTEC, Research Center for Systems and Technologies IoT and Pervasive Technology Area Faculty of Engineering, University of Porto LINKS Foundation Porto, Portugal Turin, Italy {jpcreis , gil}@fe.up.pt {name.surname}@linksfoundation.com Abstract—Nowadays competitiveness goes through several as- monitoring a replica instead of the real object, a perfect virtual pects: digitalization, productivity and environmental impact. replica that interacts with both humans and machines: a Digital Technology is advancing fast and helping industries to obtain Twin. The concept of “twin” is originally derived from Na- more and more detailed data about their processes and equip- ment. In fact, the possibility to monitor and control each part tional Aeronautics and Space Administration (NASA)’s Apollo of the process is a strong base on which a more intelligent Project when the aircraft’s twin body was a real physical and focused control can be built. Technology advance brings system [1]. Twin models help astronauts and staffs make innovation and the possibility to manage the production in decisions under emergency situations. Digital twin integrates terms of ”near future” through AI prediction and decision- the life cycle of a machine [2], and achieves a closed loop making support. Forecasting demands and planning production, optimizing process by reducing costs and improving efficiency and optimisation of the machine design, production, operation, without corrupting the quality of the product is a big challenge and maintenance, etc. In Magargle et al. [3], a multi-physical at the plant level. In this paper, a flexible, scalable architecture twin model is built to monitor the status of the brake system for intelligent digital twin realization called REPLICA has been through multiple angles. NASA hopes to realise the health proposed to cope with such problem and help industries to management and residual life prediction of the aircraft by advance and discover possible optimizations. This architecture sits on top of two European projects, namely CPSwarm and building a multi-physical, multi-scale Digital Twin model [4], RECLAIM, where their contribution focus on distributed sim- furthermore serveral roles are envisioned for Digital Twin in ulation and optimization, and Adaptive Sensorial Networks, the industry 4.0 scenario [5]. correspondingly. As a validation process, a hypothetical use case The present paper focuses on the proposition of an intel- is presented, detailing the key differentiating points and benefits ligent digital twin architecture called REclaim oPtimization of the proposed architecture. Index Terms—IoT, Digital Twin, AI, Fault Diagnosis, Predictive and simuLatIon Cooperation in digitAl twin (REPLICA), that Maintenance focuses on two important aspects: 1) plug’n’play of models on demand and 2) Workflow design to orchestrate the models I. I NTRODUCTION used in the Digital Twin itself. Aspect (1) aims to ease the integration and removal of models into digital twins, In the era of Industrial Internet of Things (IIoT) and Industry whenever a new version of the software is available or it 4.0, complex electromechanical systems can be equipped with performs any necessary correction in the used models. The a variety of sensors providing new opportunities for the goal of the latter aspect is dedicated to the creation of a development of Health Monitoring and Management Systems. pipeline that can manage the flow of data among all the These new opportunities target an optimum exploitation of available models. These models might be from pure data available information in order to maximize the performance processing and decision making, from sensor and actuator of the machinery and optimize the process. Focusing on the integration, to third party synchronization with information increase of production reliability and safety, as well as on the systems such as Manufacturing Execution Systems (MES) reduction of costs, there is an ever increasing industrial need or Enterprise Resource Planning (ERP). Most digital twins, not only for accurate and on time online diagnostics, but also and in particular intelligent digital twin architectures focus for a robust and early estimation of the Remaining Useful Life on providing the best set of models that one should have (RUL) of the defected components, within a high confidence to accomplish, e.g. predictive maintenance, from the type of interval, independent of the operating conditions. simulation required to machine learning models that should be For this reason, and many others, a new concept of ’inter- refined based on newly acquired data. Furthermore, REPLICA action’ with the process arose; a concept of controlling and can allow a flexible and distributed deployment in such a Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). 55 way that both completely cloud or mixed edge/cloud solutions In [11], [12] and [13], the authors show how it is possible deployment are accepted, with respect to the need of the to build a DT of machines and systems of systems to allow specific application. autonomous smart manufacturing, but these works, while However, these architectures are normally rigid and do not interesting, are not specifically presenting a solution for fault support changing software models or even easily set up the diagnosis and predictive maintenance. orchestration of those resources. Fixed data flows are generally The authors of [14] introduce a solution for predictive main- hardcoded, meaning that if an implementation needs to be tenance of computer numerical controlled machines, based modified, changes in code are required. This is often not on DT. They demonstrate how the exploitation of a DT for recommended since these changes might negatively influence predictive maintenance can provide better results compared to the stability of the system. To this intent, an architecture that more traditional approaches. Even if this work provides a good addresses these challenges is presented. example of application of DT for predictive maintenance, it The paper is organized as follows: in Section II, the authors doesn’t aim to present a solution that can be leveraged in other introduce a literature review about predictive maintenance and scenarios. fault diagnosis based on digital twin. Complementary, Section In this paper, the authors intend to propose a novel ar- III presents the core technologies of the solution presented chitecture that supports several features missing in the other in this paper. Section IV describes the architecture of the solutions presented above. Specifically, the proposed solution solution proposed and Section V presents a first prototype is not based on a set of fixed components but it can integrate implementation followed by Section VI in which a possible heterogeneous modules, in terms of Internet of Things (IoT) application is described. Finally, Section VII concludes the sensors, AI algorithms and simulation tools, easing its cus- paper by summarizing and discussing the work. tomization in different use-cases. Furthermore, thanks to the flexibility guaranteed by the distributed nature of the system, II. L ITERATURE R EVIEW the setting-up of the platform can easily be adapted to the each With the rapid advancement of Cyber-Physical Production specific industrial infrastructure, selecting the most suitable Systems, Artificial Intelligence (AI) and IIoT, Digital Twin mix of edge/cloud deployed components. (DT) has gained increasing attention due to its capability to Moreover, the proposed architecture can support the creation adapt and replicate the industry processes. Accordingly to at runtime of workflows both among the AI modules as well these changes, many different DT architectures have been pro- as between the IoT sensors and the models. This drastically posed to realize several use cases in an intelligent and complex reduces the time needed to run and collect results from the production system. Industrial AI [6] brings to the processes AI algorithms. In these terms, a possible process optimization self-aware, self-adapt, and self-configure functionalities and can be quickly evaluated and eventually discarded if not facilitates the integration of the DT. appropriate. Finally, all the entities (sensors, AI modules and In [7], the authors propose to insert an intelligent DT in simulators) can be substituted following a plug&play approach the Cyber Layer architecture. The concept has been partially that ease the adaption of the system to the changes in the realized with two industrial use cases, namely a modular pro- physical world. duction system as well as a metal forming industrial process III. C ORE TECHNOLOGIES to show its potential and gains over the challenges in Cyber- Physical System (CPS), i.e., synchronization throughout the A. RECLAIM platform lifecycle of a cyber-physical production system; development Following the industry 4.0 paradigm, the business models of the DT, which can contain different models; the interac- of manufacturing companies need to be transformed, resetting tion between DT, both for the purpose of co-simulation and their strategies to improve productivity and quality. The current operation data exchange; and the active data acquisition. maintenance strategies often require the user to manually In [8], the authors present a methodology for enabling analyse data collected to extract useful information from them DT using advanced physics-based modelling in predictive and, furthermore, periodic human inspection is required to maintenance. This methodology for advanced physics-based assess the real condition of the assets monitored. modeling aims to enable the DT concept in predictive main- Currently, the lack of continuous operation and health status tenance application and consists of two main points: digital monitoring tools and predictive maintenance solutions lead model creation and DT enabling. Then, the user is able to to unpredictable situations in industry like sudden machine define, create and utilize the digital model of a resource, as operation failures. In this case, the current common procedure well as its DT. The integration of DT and deep learning in CPS is to ask the intervention of technicians, which then try to environment has been also proposed in [9] for the development repair and solve the problem. This causes several problems: and realization of smart manufacturing. it is time-consuming; it leads to production delays since the In [10], the authors present solutions for fault diagnosis machine is stopped until it is not repaired; it doesn’t support based on DT. The paper includes an experiment and interesting resources distribution. The industry 4.0 paradigm goes in the results obtained with the software proposed. Compared to direction to address such problems through different actions: 1) the solution presented in this paper, this work has a limited re-manufacturing systems for material and resource efficiency, flexibility since it is only suitable for fault diagnosis. 2) increased flexibility in changing machine operation purpose, 56 3) application of big data analytics techniques, and 4) pre- number of CPSs used, leveraging evolutionary design method- dictive analytics and model-based forecasts and optimization ologies. In the latter case, candidate parameter sets are ranked procedures, based on completely data-driven processes. based on a fitness score computed after the controller was These four suggestions have been the funding principles executed with those parameters in a predefined environment. of the RE-manufaCturing and Refurbishment LArge Industrial Successful parameter sets are then adapted to produce a new equipMent (RECLAIM) concept definition. The main objec- generation of candidates to be tested. This is a high time- and tive of the project is to increase productivity, extending the resource-consuming process, which requires a high number of lifetime of the machines and reducing the time and cost of simulation runs. To address this, the CPSwarm solution allows machinery refurbishment and/or re-manufacturing. This objec- to parallelize the execution of these simulations, reducing the tive will be achieved designing and developing a set of tools times required to complete an optimization. For this objective, supporting several activities: from the monitor of machines’ the Simulation and Optimization Environment has a network- health status, to the implementation of adequate recovery based architecture, allowing to parallely use a set of STs strategy (e.g., refurbishment, re-manufacturing, upgrade, main- distributed on different machines [16]. This architecture has tenance, repair, recycle, etc.). To achieve this, the RECLAIM been implemented leveraging the eXtensible Messaging and outcomes will include two main components: an Adaptive Presence Protocol (XMPP) protocol, already tested executing Sensorial Network used to collect data and a Decision Support multiple simulations on Robot Operating System (ROS)-based Framework (DSF) for optimization based on different criteria. STs, i.e., Stage, Gazebo and Virtual Robot Experimentation Specifically one of the technologies supporting the DSF is the Platform (V-REP). In the last release of the software (available proposed REPLICA where simulation and optimization is used as open-source on github1 ), a set of technologies have been for fault diagnosis. The Adaptive Sensorial Network is one of integrated to improve its scalability and easy-to-use, i.e., the key elements to be used in the proposed architecture and docker and Kubernetes. Such final release has been tested, is seen as an entry point for the essential data to be used. showing that it is able to scale till 128 SMs and that the time required to complete one optimization is inversely proportional B. CPSwarm Simulation and Optimization Environment to the number of STs used. Finally, a proof of concept has demonstrated the ability to deploy the controller with the As indicated in [15], the CPSwarm Workbench - the set optimized parameters onto CPSs. of tools released by the project for the development of CPS The concept of distributed simulation and optimization swarms applications - includes also a Simulation and Opti- is brought to the proposed REPLICA architecture by the mization Environment, used to evaluate the performance of CPSwarm results and the whole orchestration process and a swarm solution. Such solution is composed mainly by: the main building blocks are inspired by this project. Simulation and Optimization Orchestrator (SOO), which over- sees the simulation and optimization tasks; a set of Simulation IV. A RCHITECTURE Managers (SMs), which provide common Application Pro- gramming Interface (API) to control heterogeneous Simulation This section introduces the REclaim oPtimization and sim- Tools (STs); and an Optimization Tool (OT) used to perform uLatIon Cooperation in digitAl twin (REPLICA) architecture the optimization processes. The network-based architecture is that has been designed to provide an infrastructure and be depicted in Fig. 1. used for Digital Twin-based fault diagnostics and predictive maintenance solutions, which can be easily deployed and customize in different Industrial IoT environments. REPLICA is composed by several modules (shown in Figure 2), mainly subdivided in two blocks: Backend and Frontend. The first one contains three main components: Artificial Intelligence (AI) Environment that hosts the AI modules, Digital Twin Orchestrator (DTO) that is used to orchestrate the operations done by the REPLICA and the Sim- ulation Environment that is a distributed environment includ- ing several heterogeneous simulators deployed into different machines. The latter one instead contains two applications: one devoted to show the results obtained and another one for Fig. 1. Network-based Architecture from CPSwarm for Distributed Optimiza- the configuration of the component. These modules will be tion and Simulation. described in the remainder of this section. As explained in Section I, the DT concept concerns the Such environment is useful both to simulate the behaviour integration of three main components: the data collected by of a designed swarm solution in a ST, leveraging the ST’s IoT sensors; the realistic models of the real devices and the Graphical User Interface (GUI) to evaluate its behaviour; and, on the other side, to optimize the controller parameters of 1 https://github.com/cpswarm/SimulationOrchestrator/wiki/Simulation-and- algorithm/module, and possible aspects of the problem, i.e., the Optimization-Environment 57 simulators data produced by the AI algorithms (for example to simulate failures); 4) receive from the simulators the produced results. Finally, the Simulation Environment supports for each integrated simulator one advanced Simulator GUI that allows to monitor the simulated device. This GUI is the one integrated in the simulator, which provides a graphical representation (also 3D) of the simulated device. Finally, in REPLICA the AI algorithms are hosted and executed in the AI Environment. This environment allows to host and run heterogeneous algorithms for fault diagnosis and predictive maintenance. Besides the algorithms, the environ- ment also hosts a module called AI engine. This module has the objective to orchestrate the algorithms creating the needed workflows among them. Furthermore, the AI Engine uses the API provided by the DTO to interconnect the AI modules with the machine models and the data coming from the shop-floor. The architecture is completed by the two interfaces in the Frontend: the OutputMonitor GUI is used to monitor in real time the results produced by the running solutions in a user friendly interface. Instead, the Configuration GUI is leveraged by the users of the system to configure the AI engine for the needed tasks. The proposed architecture aims to provide the following key Fig. 2. REPLICA Architecture. features: 1) Allowing the integration of heterogeneous compo- nents in terms of sensors data collected from the field, AI algo- rithms and simulators running accurate machine models in the synchronization with those using the data collected; and a set shop-floor; 2) Supporting the creation at runtime of workflows of AI modules connected to these models. REPLICA fully not only among both AI modules and the IoT sensors, but also supports this concept, providing the infrastructure to integrate among themselves; 3) Supporting the plug&play at runtime these technologies. of the IoT sensors, the AI modules and machine models, In REPLICA, the Digital Twin Orchestrator (DTO) is the without the need to restart the system; 4) Easing the adaptation module in charge to manage all the IoT data flow coming from of the digital twin to the changes in the physical world; 5) the field: machinery data, historical data and other data from Enabling a flexible and distributed deployment: supporting legacy systems already present in the shop-floor. As described both completely cloud or mixed edge/cloud solutions, based in Section III, these data are flowing through a component on the need of the specific application. of the RECLAIM platform, the Adaptive Sensorial Network. A first partial implementation of the proposed architecture Furthermore, the DTO is in charge to create the correct flow and a set of possible future works for the part not yet among the AI modules running in the AI environment and implemented is presented in the next section. the machine models running in the Simulation Environment. Finally, the DTO oversees the storing and organization of V. P RELIMINARY IMPLEMENTATION AND FUTURE the processed and simulated data, which are saved in a local PROSPECTS database. In REPLICA every machinery of interest has a corre- This section presents the first prototype of the proposed sponding realistic model running in one of the simulators architecture. The solution is a combination of newly developed integrated in the Simulation Environment. Specifically, the components and the evolved version of components already Simulation Environment is a distributed environment based on developed in previous European Union (EU) projects. For the one presented in Section III-B. Similar to what has been the new components, this section will introduce only some presented for the original solution about swarm intelligence, possible technologies that the authors are evaluating and the environment supports a set of heterogeneous simulators testing so far to leverage and implement the architecture, while distributed in different machines. Each of these simulators for the existing components a more concrete implementation is is wrapped by a Simulation Manager, and the role of this presented. More specifically, the software already implemented component is to abstract the functionalities provided by the is one algorithm for predictive maintenance, one for fault simulators using the standard API exported by the DTO. In diagnosis and a distributed simulation environment already this way, the DTO can: 1) control the simulators to run the developed in CPSwarm, while the components yet to be required simulations; 2) inject the data needed to keep the implemented are the AI environment, the AI engine, the DTO, models synchronized with the real machines; 3) inject in the the Configuration GUI and the OutputMonitor GUI. 58 The authors have chosen to base the AI environment on a case, the deployment of the containers, the interconnection of docker container based solution. Each predictive maintenance the components and the workflow will be handled by tools and fault diagnosis algorithm will be wrapped in one container. included in the framework. At the moment of writing, the In this way the AI environment will support the integration of final solution to be used is still under evaluation and the AI modules based on different technologies. authors are investigating if Acumos AI satisfies all the needed Two examples of solutions currently supported in the AI requirements of the AI environment, particularly focusing on environment are the fault diagnosis and predictive maintenance the possibility to add and remove AI modules at run-time and modules presented in Fig. 3 and 4. The fault diagnosis module the dynamic change of their interconnections to create new is composed by techniques to find abnormal behaviors that workflows. deviate from normal process conditions to raise warnings For the implementation of the DTO and the Simulation En- and find root causes for the problem. This algorithm will be vironment, the Simulation and Optimization Environment so- fed directly with sensor data (when possible and pertinent) lutions provided by CPSwarm will be leveraged and extended. or transformed data from the field in order to be more Specifically, REPLICA will incorporate the communication interpretable. Based on the analysis of data streaming, the API based on XMPP and the deployment system, based on algorithm should indicate if a warning should be sent to the docker and Kubernetes [21]. The use of these technologies will key personnel to check the system. This algorithm is the first allow to integrate heterogeneous simulators, to simply deploy front-line of analysis from shop-floor components in order to and run the simulations needed on distributed machines, add understand machine’s health. and remove at run-time the simulators running different mod- Additionally, the predictive maintenance module is com- els. More specifically, in the Simulation Environment, consid- posed of 1) a component failure prediction in the future ering that the solutions proposed in CPSwarm was integrating (e.g. 48h and which maintenance action should take place); only ROS based simulators, new types of simulators, e.g., 2) Optimization module for scheduling future maintenance java based simulators, will be included during the RECLAIM actions based on the existing scheduling; 3) Simulation module project. For this scope, a specific SM will be developed for the that aims at assessing the impact of changes in the machine required simulators and the API will be refactored to support and shop-floor [17]. The main idea of this method is to predict also these new simulators. Beside the SM, also docker contain- what kind of maintenance and when it will be required based ers to easily deploy such simulators will be created. Instead, on the failing component in the machine. With this, it will be for the implementation of the DTO, the authors have defined possible to understand what changes need to be done in order that the SOO implemented in CPSwarm will be completely to compensate the downtime of the failing machine. refactored and extended to support the functionalities required As can be seen from Figure 3, the implementation already by REPLICA. Additionally, only some of the functionalities follows a block based approach which allows a better flexibil- of the SOO will be leveraged, extending them to support ity once building the required data workflows among models. data storage and data analysis features. Finally the DTO will For this particular case, the Dynamic INtelligent Architecture provide a set of API based on some standard technologies, for Software MOdular REconfiguration (DINASORE) [18] e.g., Message Queue Telemetry Transport (MQTT), which will platform was used, which is a run-time environment devel- allow to collect data to be used by the algorithms and to keep oped in python language for the International Electrotech- the models updated and synchronized with the physical world. nical Commission (IEC) 61499 standard [19] and integrated Thanks to these API, the DTO will be able to collect data with the Eclipse based Framework for Distrubeted Industrial from heterogeneous devices that, in the RECLAIM platform, Automation and Control (4DIAC) Integrated Development are integrated through the IoT Gateway (see Section II). Also Environment (IDE) [20]. Moreover, this implementation does in this case, it will be possible to add and remove devices not only allow for the orchestration of models, but also the at run-time, without the need to restart the system. The new plug&play of such models in a distributed system, where devices can be immediately used by the solution developed, software can be reconfigured on-demand. This supports both just after they have registered themselves. completely cloud or mixed edge/cloud systems, depending on Finally, for the implementation of the GUI included in the the required application and the number of machines available architecture, the presented modules are in different phases for execution. Finally, since DINASORE is implemented in of development. Specifically, for the Configuration GUI, the python, the state of the art implementations of AI can be authors have not yet chosen how to implement it and different promptly used. solutions will be evaluated, keeping in consideration a thor- For the implementation of the AI Engine that interconnects ough integration with the rest of the platform. Instead, for the the AI modules, the authors have already evaluated several Output/Monitor GUI, for the monitoring and assessment of the solutions. One is the possibility to run the modules in a results of the algorithms, a simple implementation based on docker environment, making them read and write from text the work done in CPSwarm is already available. Specifically, files located in specific folders and then interconnect them this solution is based on Thingsboard, which has been used through a software that allows to handle the workflow, e.g., to develop two different GUI: one for process monitoring and Node-red or NiFi. Another evaluated solution is the possibility another one for the assessment of results. The first one shows to use Acumos AI to implement the AI environment; in this live data in a chart and allows the monitoring of the process; 59 Fig. 3. Fault Diagnosis algorithm. exploitation of fault diagnosis and predictive maintenance techniques based on the use of digital twin will increase the efficiency of the maintenance activity with respect to the performance obtained with the traditional methods based on a fixed schedule and a simple telemetry analysis. The flexibility and adaptability of REPLICA can be well demonstrated both during the system setup in the industrial site and in case of replacement of one machine in the Wood- working Production Line. In the first case, when the platform is going to be deployed, the data collected by the Adaptive Sensorial Network and a set of AI algorithms for fault diagnosis and predictive main- tenance developed by different analysts have to be integrated. Furthermore, to allow the simulation of different operative Fig. 4. Predictive Maintenance algorithm [17]. scenarios, the realistic models of different machines have to be imported and executed in one simulator. Using REPLICA, the effort to make all these components the latter one, instead, shows the data in a table, where it can be provided by heterogeneous vendors working together is signif- sorted by column (one column for each data). These GUIs have icantly reduced, allowing their integration by simply exposing been used in this first implementation, but the possibility to their inputs/outputs through the REPLICA defined interfaces. enhance or completely replace them with something different, In particular, for the machine simulation, if the simulator used based on the requirements of the solution proposed, will be to execute it is already supported by REPLICA, no further evaluated in the future. developments are needed; otherwise only the SM for that simulator needs to be developed to allow its integration with VI. RECLAIM USE CASE the rest of the solution. The same advantage can be considered The aim of this section is to present a possible application also for the AI algorithms: if the ones already integrated of the solution presented in this paper, focusing on the predic- in the AI environment are suitable for the specific case, no tive maintenance and refurbishment of a large Woodworking developments are necessary and the platform should be just Production Line. The main objective of such use case will be configured to enable the correct flow of data among different to show the benefits of the adoption of advanced maintenance components. Otherwise, to integrate a new algorithms, the only strategies in a large scale industrial scenario. requirement is to implement the inputs/outputs API defined in The selected scenario presents different challenges: firstly, REPLICA. the need to integrate in a single environment both heteroge- Once all the components are connected, the AI modules can neous data collected by installed sensors at the shop-floor be trained using the data coming from the Adaptive Sensorial and AI modules; together with the realistic models of the Network (as it is usually done), but also with data produced machines, enabling and easing the construction of a shop- by the simulated machines. The integration of this secondary floor’s digital twin. Moreover, the proposed solution will have source of data not only allows to speed up the training process to optimize the use of a large industrial equipment providing (more data available means less time to learn) but also to add novel machine learning solutions able to monitor the current the possibility of using data that are generally more difficult to system status and predict possible failures. In particular, the collect, such as the one associated with specific failures that, 60 for obvious reasons, are not so common in a real industrial of the workflows needed for fault diagnosis and predictive plant. maintenance, adding/removing/replacing entities to reflect the The advantages of the REPLICA solution can be further situation of the components available in the system. demonstrated taking into consideration the scenario of a In this work, the authors introduced the details of the machines part replacement in the production line which is first implementation of the proposed architecture; since the already monitored by the system. In this case, the administrator development currently is only in a preliminary phase, Section of the platform needs to update the components used for V presents the implementations of the modules that were the fault diagnosis and predictive maintenance to reflect the evolved from previous EU projects’ outcomes. Instead, for new situation on the field. In a traditional system, this is a the new components only some initial design choices are process that requires a complete shutdown of the system in presented. In particular, the AI environment and the DTO have order to setup and reconfigure it; instead, by using REPLICA been just designed and will be developed in next phases of the the process is fast and mainly automatic. Indeed, for the project, while the Simulation Environment is already available replacement of the simulation models, this can be done just and only some SMs for new simulators will be developed. removing the old simulator and instantiating a new one with In the same way, a dashboard for monitoring and results the updated model, taking advantage of the integration of assessment is ready, while the GUI for configuration has still the solution with Kubernetes. Once the updated simulator to be implemented. is instantiated, the same process can be applied to the AI Finally, the authors have provided in Section VI two use algorithms, which can replace the previous ones. All these cases based on one realistic industrial scenario, which show updates can be done without the need to interrupt the execution the advantages of using the proposed solution to apply fault of the system. Obviously, if this change requires the integration diagnosis and predictive maintenance techniques based on of some new AI modules or the development of a new SM, digital twin. these have to be implemented in advance. Also the workflow of the components need to be updated to interconnect the new ACKNOWLEDGMENT ones, and can be done simply by using the tools provided The work presented here was part of the project by AI engine and DTO. These will automatically update the ”RECLAIM- RE-manufaCturing and Refurbishment LArge workflow to reflect the status of the components available in Industrial equipMent” and received funding from the European the system and that allow the administrator to easily create the Union’s Horizon 2020 research and innovation programme new workflows. Finally as for the previous use-case, REPLICA under grant agreement No 869884. can also be used to speed up the training time needed to have the new algorithms ready to be used, supporting the use of R EFERENCES simulated data, instead of using the actual devices. [1] R. Rosen, G. [von Wichert], G. Lo, and K. D. Bettenhausen, “About the importance of autonomy and digital twins for the future of manufac- VII. CONCLUSION turing,” IFAC-PapersOnLine, vol. 48, no. 3, pp. 567 – 572, 2015, 15th IFAC Symposium onInformation Control Problems inManufacturing. The paper has presented REPLICA, a solution that enables [2] F. Tao, J. Cheng, Q. Qi, M. Zhang, H. Zhang, and F. Sui, “Digital twin- the application of innovative fault diagnosis and predictive driven product design, manufacturing and service with big data,” The maintenance techniques based on DT. The software proposed International Journal of Advanced Manufacturing Technology, vol. 94, 02 2018. is part of a more complex platform developed by the RE- [3] R. Magargle, L. Johnson, P. Mandloi, P. Davoudabadi, O. Kesarkar, CLAIM project that provides a complete software solution S. Krishnaswamy, J. Batteh, and A. Pitchaikani, “A simulation-based for both extending machinery lifetime while also improving digital twin for model-driven health monitoring and predictive main- tenance of an automotive braking system,” in Proceedings of the 12th productivity and performance. International Modelica Conference, Prague, Czech Republic, May 15- The paper shows the main key innovations introduced by 17, 2017, 07 2017, pp. 35–46. REPLICA in terms of fault diagnosis and predictive mainte- [4] E. Negri, L. Fumagalli, and M. Macchi, “A review of the roles of digital twin in cps-based production systems,” Procedia Manufacturing, vol. 11, nance techniques based on DT. Specifically, one of the basic pp. 939–948, 12 2017. concepts of REPLICA is to build the solution not as unique [5] ——, “A review of the roles of digital twin in cps-based production suite of technologies, but as a run-time environment with APIs systems,” Procedia Manufacturing, vol. 11, pp. 939–948, 12 2017. [6] J. Lee, H. Davari, J. Singh, and V. Pandhare, “Industrial that allow the integration of heterogeneous AI modules and artificial intelligence for industry 4.0-based manufacturing systems,” simulators. In this way, the solution can be easily customized Manufacturing Letters, vol. 18, pp. 20 – 23, 2018. [Online]. Available: to be used in different industrial sites, integrating components http://www.sciencedirect.com/science/article/pii/S2213846318301081 [7] B. A. Talkhestani, T. Jung, B. Lindemann, N. Sahlab, N. Jazdi, provided by different vendors, without requiring the devel- W. Schloegl, and M. Weyrich, “An architecture of an intel- opment of new components, but just adapting the existing ligent digital twin in a cyber-physical production system,” at- ones. Furthermore, using REPLICA the user can integrate Automatisierungstechnik, vol. 67, no. 9, pp. 762–782, 2019. [8] P. Aivaliotis, K. Georgoulias, Z. Arkouli, and S. Makris, “Methodology and replace AI modules and simulation models at run-time, for enabling digital twin using advanced physics-based modelling in without the necessity to stop the system and reconfigure it. predictive maintenance,” Procedia CIRP, vol. 81, pp. 417–422, 2019. REPLICA provides tools to interconnect among each other [9] J. Lee, M. Azamfar, J. Singh, and S. Siahpour, “Integration of digital twin and deep learning in cyber-physical systems: towards smart man- the different modules with the data collected from the field. ufacturing,” IET Collaborative Intelligent Manufacturing, vol. 2, no. 1, This allows the creation and the modification at run-time pp. 34–36, 2020. 61 [10] Y. Xu, Y. Sun, X. Liu, and Y. Zheng, “A digital-twin-assisted fault [16] M. Rappaport, D. Conzon, M. Jdeed, M. Schranz, E. Ferrera, and W. El- diagnosis using deep transfer learning,” IEEE Access, vol. PP, pp. 1– menreich, “Distributed simulation for evolutionary design of swarms 1, 01 2019. of cyber-physical systems,” in Proc. Int. Conf. on Adaptive and Self- [11] K. Ding, F. T. Chan, X. Zhang, G. Zhou, and F. Zhang, “Defining Adaptive Systems and Applications (ADAPTIVE). IARIA, Feb. 2018, a digital twin-based cyber-physical production system for autonomous pp. 60–65, ISBN: 978-1-61208-610-1. manufacturing in smart shop floors,” International Journal of Production [17] L. Antão, J. Reis, and G. Gonçalves, “Continuous maintenance system Research, vol. 57, no. 20, pp. 6315–6334, 2019. [Online]. Available: for optimal scheduling based on real-time machine monitoring,” in 2018 https://doi.org/10.1080/00207543.2019.1566661 IEEE 23rd International Conference on Emerging Technologies and [12] G. Zhou, Z. Chao, L. Zi, K. Ding, and C. Wang, “Knowledge-driven Factory Automation (ETFA), vol. 1. IEEE, 2018, pp. 410–417. digital twin manufacturing cell towards intelligent manufacturing,” In- [18] Digital and I. I. Lab. Dinasore - dynamic intelligent architecture ternational Journal of Production Research, 04 2019. for software and modular reconfiguration. [Online]. Available: https: [13] M. Borth, J. Verriet, and G. Muller, “Digital twin strategies for sos 4 //digi2-feup.github.io/dinasore/ challenges and 4 architecture setups for digital twins of sos,” 05 2019, [19] G. Cengic, O. Ljungkrantz, and K. Akesson, “Formal modeling of pp. 164–169. function block applications running in iec 61499 execution runtime,” [14] W. Luo, T. Hu, Y. Ye, C. Zhang, and Y. Wei, “A hybrid predictive in 2006 IEEE Conference on Emerging Technologies and Factory maintenance approach for cnc machine tool driven by digital twin,” Automation, Sep. 2006, pp. 1269–1276. Robotics and Computer-Integrated Manufacturing, vol. 65, p. 101974, [20] T. Strasser, M. Rooker, G. Ebenhofer, A. Zoitl, C. Sunder, A. Valentini, 10 2020. and A. Martel, “Framework for distributed industrial automation and [15] M. Jdeed, M. Schranz, A. Bagnato, S. Suleri, G. Prato, D. Conzon, control (4diac),” in 2008 6th IEEE International Conference on Indus- M. Sende, E. Brosse, C. Pastrone, and W. Elmenreich, “The cpswarm trial Informatics, July 2008, pp. 283–288. technology for designing swarms of cyber-physical systems,” in Proc. [21] C. Consortium, “Finalintegration of external simulators, d6.7,” Deliver- Int. Conf. on The Research Project Showcase of Software Technologies: able of the CPSwarm Project, 2020. Applications and Foundations (STAF), Jul. 2019. 62