key: cord-0324707-aqmy1w4u authors: Latif, Hasan; Starly, Binil title: A Simulation Algorithm of a Digital Twin for Manual Assembly Process date: 2020-12-31 journal: Procedia Manufacturing DOI: 10.1016/j.promfg.2020.05.132 sha: a3af0751b38b7ea6ccee6bd460aee1414d8ec285 doc_id: 324707 cord_uid: aqmy1w4u Abstract Digital twin (DT) is one of the key concepts for Industry 4.0 as it is a critical component in driving real-time simulation and decision making in complex systems. The existing scientific literature on Digital twin primarily refers to a product entity or a physical machine but the core concept can be applied to the entire product lifecycle, particularly the assembly process of a complex product system. In addition, majority of existing work focus on the Digital Twins of individual machines on a shop-floor. This paper focuses on aspects the process to build DTs of a production schedule for a complex defence weapon system. The process is inherently high variety and low quantity in a very manual assembly process. This paper consists of three elements. (1) It reviews the current state of art along with the research gap and discusses how DT can become a tool to the manual assembly process. (2) A data-driven simulation algorithm is proposed to model the complex and manual manufacturing process in a generic-reusable way. (3) Finally, an appropriate complex industrial case study is studied to exemplify the proposed framework. Results demonstrate that the production managers can make more informed early decisions that can help bring assembly schedules in check and limit wasteful efforts when disruptions in the supply chain of parts sourced for the assembly occur. The manufacturing industry is pushing for transformation in their approach to meet the demand for heavy mass customized products. Quick response to market demand necessitates a range of planning and control decisions at all levels. The ability to quickly react and respond to customer demands make smart manufacturing in factory floors necessary. According to Zheng et al. smart manufacturing systems are comprised of smart design, smart machining, smart monitoring, smart control, smart scheduling, and industrial applications [1] . The "smart" refers to the ability of a process to adapt to changing needs, perform the task automatically and to make intelligent decisions at the right time. Technologies such as statistical learning, artificial intelligence, recommendation-based system, deep learning, sensor-based automation, internet of things, big data analytics can be integrated in any of the decision-making hierarchy level to make it productive and smart. The widespread applications of smart manufacturing technologies infused into factory floors across the world and the digitization achieved in connecting various information and operational systems marks the advent of the fourth state of industrial production -Industry 4.0. Industry 4.0 does not have any systematic framework with various organizations implementing aspects of Industry 4.0 in steps that meet their current and future needs. It is still in a work-in-progress stage. But the Digital Twin technology is one of the key concepts to achieve Industry 4.0. The manufacturing industry has not fully embraced the DT concept because of the lack of execution knowledge. Especially, the manual assemblybased manufacturing line is far from reaching industry 4.0 and there is no guideline to cover the gap either. This research focuses on providing a guideline of DT for a Manual assemblybased manufacturing industry that caters to a high mix, low quantity complex defense system. The contributions of the paper include (1) identifying the research gap in the current The manufacturing industry is pushing for transformation in their approach to meet the demand for heavy mass customized products. Quick response to market demand necessitates a range of planning and control decisions at all levels. The ability to quickly react and respond to customer demands make smart manufacturing in factory floors necessary. According to Zheng et al. smart manufacturing systems are comprised of smart design, smart machining, smart monitoring, smart control, smart scheduling, and industrial applications [1] . The "smart" refers to the ability of a process to adapt to changing needs, perform the task automatically and to make intelligent decisions at the right time. Technologies such as statistical learning, artificial intelligence, recommendation-based system, deep learning, sensor-based automation, internet of things, big data analytics can be integrated in any of the decision-making hierarchy level to make it productive and smart. The widespread applications of smart manufacturing technologies infused into factory floors across the world and the digitization achieved in connecting various information and operational systems marks the advent of the fourth state of industrial production -Industry 4.0. Industry 4.0 does not have any systematic framework with various organizations implementing aspects of Industry 4.0 in steps that meet their current and future needs. It is still in a work-in-progress stage. But the Digital Twin technology is one of the key concepts to achieve Industry 4.0. The manufacturing industry has not fully embraced the DT concept because of the lack of execution knowledge. Especially, the manual assemblybased manufacturing line is far from reaching industry 4.0 and there is no guideline to cover the gap either. This research focuses on providing a guideline of DT for a Manual assemblybased manufacturing industry that caters to a high mix, low quantity complex defense system. The contributions of the paper include (1) identifying the research gap in the current The manufacturing industry is pushing for transformation in their approach to meet the demand for heavy mass customized products. Quick response to market demand necessitates a range of planning and control decisions at all levels. The ability to quickly react and respond to customer demands make smart manufacturing in factory floors necessary. According to Zheng et al. smart manufacturing systems are comprised of smart design, smart machining, smart monitoring, smart control, smart scheduling, and industrial applications [1] . The "smart" refers to the ability of a process to adapt to changing needs, perform the task automatically and to make intelligent decisions at the right time. Technologies such as statistical learning, artificial intelligence, recommendation-based system, deep learning, sensor-based automation, internet of things, big data analytics can be integrated in any of the decision-making hierarchy level to make it productive and smart. The widespread applications of smart manufacturing technologies infused into factory floors across the world and the digitization achieved in connecting various information and operational systems marks the advent of the fourth state of industrial production -Industry 4.0. Industry 4.0 does not have any systematic framework with various organizations implementing aspects of Industry 4.0 in steps that meet their current and future needs. It is still in a work-in-progress stage. But the Digital Twin technology is one of the key concepts to achieve Industry 4.0. The manufacturing industry has not fully embraced the DT concept because of the lack of execution knowledge. Especially, the manual assemblybased manufacturing line is far from reaching industry 4.0 and there is no guideline to cover the gap either. This research focuses on providing a guideline of DT for a Manual assemblybased manufacturing industry that caters to a high mix, low quantity complex defense system. The contributions of the paper include (1) identifying the research gap in the current The manufacturing industry is pushing for transformation in their approach to meet the demand for heavy mass customized products. Quick response to market demand necessitates a range of planning and control decisions at all levels. The ability to quickly react and respond to customer demands make smart manufacturing in factory floors necessary. According to Zheng et al. smart manufacturing systems are comprised of smart design, smart machining, smart monitoring, smart control, smart scheduling, and industrial applications [1] . The "smart" refers to the ability of a process to adapt to changing needs, perform the task automatically and to make intelligent decisions at the right time. Technologies such as statistical learning, artificial intelligence, recommendation-based system, deep learning, sensor-based automation, internet of things, big data analytics can be integrated in any of the decision-making hierarchy level to make it productive and smart. The widespread applications of smart manufacturing technologies infused into factory floors across the world and the digitization achieved in connecting various information and operational systems marks the advent of the fourth state of industrial production -Industry 4.0. Industry 4.0 does not have any systematic framework with various organizations implementing aspects of Industry 4.0 in steps that meet their current and future needs. It is still in a work-in-progress stage. But the Digital Twin technology is one of the key concepts to achieve Industry 4.0. The manufacturing industry has not fully embraced the DT concept because of the lack of execution knowledge. Especially, the manual assemblybased manufacturing line is far from reaching industry 4.0 and there is no guideline to cover the gap either. This research focuses on providing a guideline of DT for a Manual assemblybased manufacturing industry that caters to a high mix, low quantity complex defense system. The contributions of the paper include (1) identifying the research gap in the current 48th SME North American Manufacturing Research Conference, NAMRC 48 (Cancelled due to literature, summarizing digital twin definitions, and developing a data-driven simulation algorithm for digital twins of a manufacturing process; (2) providing a complex industrial case study from within the defense industry to demonstrate the digital twin algorithm. The rest of the paper is organized as follows: Section 2 provides the problem statement; Section 3 focuses on the current state of the literature and research gap analysis; Section 4 discusses the digital twin concept for manufacturing processes and proposed a simulation algorithm. Section 5 provides a real-life case study to exemplify the simulation algorithm. Finally, Section 6 provides a conclusion with future research opportunities. The aerospace, defense, and space industry are experiencing immense growth in recent years. The importance of having a successful production plan is also becoming more significant. The manufacturing process can get very complicated when it comes to a low volume, high value human assembled product. There are usually about 5,000 to 15,000 parts involved with different arrival times sourced from various internal and external partners from with the supply chain. The reliance on various partners adds uncertainty and uncontrollable breakdowns which inherently leads to production schedule delays and wasteful inventory stored at site locations. Even more critical is that often time production and factory managers are not informed of potential production delays until it is too late. In the event they are made aware of a schedule delay, they are not able to reliably tell how long the delay will last. This lack of information leads to customer dissatisfaction and relationships with supply chain partners both upstream and downstream to be negatively affected. This lack of an active decision support also limits the ability to optimize or strategize manufacturing operations so that actions taken can limit production schedule changes. A successful production plan largely depends on how well a production manager can handle production floor problems. The objective of the research is to find the best possible solutions of the manufacturing operations in the production floor considering uncertainty in part arrival, sudden machine breakdown, and part obsolescence. Methods and technologies including optimizing job shop scheduling, raw material arriving time, floor plan layout, material flow; inventory management; and simulation are commonly used to solve different manufacturing floor problems. Generally, it requires a lot of time from a production manager's standpoint to collect all the data, evaluate, and act accordingly. Although there is plenty of academic research been done on this specific topic, none of them are embraced completely by the industry. This is partially since most academic work does not fully capture the constraints and operational modalities specific to each organization that renders a plan useful. After analysing the extensive scientific literature for the current methods for manufacturing problems on the production floor, it can be divided into three major categories: traditional methods, advanced algorithms, and simulation with visualization. The traditional methods include optimizationbased problem formulation. Most of the literature in resource management use aggregated single objective algorithms. These algorithms are the traditional approach where a set of constraints are created based on factory resources. Advanced algorithms include all the latest hybrid methods and algorithms. Over the past decade, many research projects have been implemented using a new algorithm called Pareto based multi objective evolutionary algorithms (MOEAs) [2] . Genetic algorithm (GA) has been proved to be one of the best evolutionary algorithms over random search and particle swarm optimization. [3, 4] . Non-dominated sorting genetic algorithm (NSGA) provides better techniques and results in solving job shop scheduling problem [5] . The NSGA-II version came out to solve large multi-objective problems [1] . Another modified version of NSGA-II was proposed integrating random search, which was called non-dominated ranking genetic algorithm [6] . In a paper by Ahmadi et. al, uncertainty was integrated and tested among the discussed genetic algorithm techniques [7] . Apart from genetic algorithms, researchers tried several other techniques as well. Zhang et. al proposed a hybrid algorithm that combined particle swarm optimization and neighborhood function [8] . Xiong worked on a robust scheduling with machine breakdowns. The authors used surrogated model for approaching the multi objective problem [9] . Yuan and Zu created memetic algorithm that incorporates local search algorithm into NSGA-II [10] . He and Sun proposed an approach in which right shift scheduling and route changing scheduling are used [11] . Overall, it is obvious that no algorithm is adaptive enough to cover a broad area of manufacturing problems. Over time, researchers have proposed and created new algorithms to address different scenarios. However, all these algorithms are not proven in the pragmatic case studies. Most of the case studies are overly simplified. It is difficult to customize and adapted to other manufacturing problem contexts. There is robustness in the solution, but it is very specific to the field. Visualization is missing and too much computational effort is required to perform these algorithms. Therefore, all these algorithm-based techniques failed to make an impression in real industry where the situation changes, uncertainty looms, and intricacy increases often. The simulation tools reviewed in this paper are constantly evolving and they certainly lead towards more efficient manufacturing process solutions. But, in the current highly competitive business environment, which is constantly facing new challenges, there is always need for even more efficient and adaptive technologies [12] . At this point, simulation mostly considers the ideal situation where it generates the result without much consideration of building in detail necessary to replicate a running manufacturing floor. It does not have the critical portion where every decision requires a lot of effort and change. Hence, a simulation with machine learning capabilities aka predictive evaluating capabilities can be a game changer. That is where Digital Twin (DT) can play a big role. Digital twin can bridge the gap between current industry needs and available technologies. DT has a broad scope with a lot of integration opportunities. DT will support repeatable, adaptive, and convenient modeling of analysis of manufacturing operations. Further analysis and newer technologies are easier to incorporate. Table 1 provides the key issues for effective resource management (ERM) as well as whether they may be covered by different methods discussed. The "-" sign indicates "not covered" and the "+" indicates "covered". Apparently, digital twin will support more desired functionalities. However, there is no systematic approach that help guide the implementation of DT. In this paper, an implementation guideline along with an industrial case study from a defence system will be presented. A digital twin for a manual assembly process is an evolving profile of the manufacturing operations to provide insights based on real-time data from different business/process levels. For instance, a combination of a simulation of the manufacturing process, AI based recommendation, integration of different applications to exchange data, and process optimization will make it a true digital twin for a manufacturing process. The advancement of new generation technologies, such as internet of things (IOT), big data, cloud computing, artificial intelligence (AI), and their wide applications are driving manufacturing industry into an exciting era. New concepts are coming along to solve the manufacturing issues. Digital twin is one such concept that can help improve product performance, reliability, and maintenance. Digital twin is a virtual representation of physical process. Depending on the modelling objective and context, it could be a high-fidelity virtual model for a physical system in digital to simulate their behaviours. Gartner classified digital twin as a one of the top ten strategic technology trends for 2017 and 2018 [13] . It is critical to have a clear understanding of the digital twin concept, definition, and its implementation strategy. A digital twin can be defined as a digital representation of the historical and present outline of a physical entity or process to optimize performance. For a manufacturing production floor, the digital representation of the manufacturing process includes all the material path, machines, layout, part arrival rate, machine breakdown rate, etc. The University of Michigan used this conceptual model in their first executive PLM courses. It was referred as Mirrored Space Model and referenced in a journal article. In the seminal PLM book, Product Lifecycle Management: Driving the Next Generation of Lean Thinking, the conceptual model was referred to as the Information Mirroring Model [14] . The concept was greatly expanded in Virtually Perfect: Driving Innovative and Lean Products through Product Lifecycle Management [15] . The term "Digital Twin" was used to describe the model. Researchers from industry and academia have defined DT in various ways with many focused on the definition of a DT in the context of a physical machine. However, neither of the groups places emphasis on the ability of DT to replicate a physical assembly process carried out on a production floor in a production schedule. Especially in the manual assemblybased manufacturing domain, the digital twin concept is still being interpreted in different ways. For instance, Schluse and Rossmann defined the Digital Twin as virtual substitutes of real-world objects consisting of virtual representations and communication capabilities making up of smart objects acting as intelligent nodes inside the internet of things and services [16] . According to Kraft, an integrated multi-physics, multiscale, probabilistic simulation of an as-built system, enabled by Digital Thread, that uses the best available models, sensor information, and input data to mirror and predict activities/performance over the life of its corresponding physical twin [17] . Considering all the previous definitions, the authors believe digital twin definition for a manufacturing floor must be self-explanatory and simpler. The definition at the beginning of this section covers all the previous DT instances and captures the true essence of DT succinctly. Creating the digital twin of a manufacturing process requires a good understanding of the processes. First, a process map needs to be created with all the raw material input and operations. A bill of materials and work instructions can help prepare the process map. The core module of the DT is a simulation that is modelled based on the process map. The process map provides the requirements for the data points needs for the model. The key requirements for the implementation are (1) definition of the assembly scenario to be carried out, (2) operational data capture and analysis towards the identification of key parameters for the manual assembly process, (3) creation of the digital twin with the integration of key motion parameters and operational constraints, modelling the behaviour of the physical assets, and (4) simulation of the assembly process and its optimization according to a set of optimization constraints. For simulation algorithm, it has multiple stages. First, the time period needs to be defined and initial data needs to be provided. The simulation runs and stops after each time period and waits for "new data input". So, after each simulation run, the simulation stops and asks two questions: 1) is there new data input? 2) does user want to review the worst performed operations? Based on the answers, the simulation algorithm branched out to 4 different ways. The 4 ways are described below. a) "NO" new data input, and user does "NOT" want to review the performance: the simulation moves to the next time period using the previous data input. The performance review will not be executed either. The algorithm will continue afterwards. b) New data input "exists", and user does "NOT" want to review the performance: the simulation moves to the next time period using the new data input. The performance review will not be executed. The algorithm will continue afterwards. c) "NO" new data input, and user "wants" to review the performance: the simulation moves to the next time period using the previous data input. For performance review, the worst 5 operations will pop up based on user selected performance criteria. Then, a knowledge-based recommendation list will pop as well with respect to each of the worst performing operations. User will pick one or two from the recommendations list and implement in the actual manufacturing floor. At this point, the simulation will be divided into two sub-parts. First, it will assume a certain percentage of improvement will occur based on the recommended selection. The performance criteria parameter will be changed/improved and generate a new data input termed as "Estimated Data Input." Using the estimated data input, the simulation will go all the way to the end of the total time period without waiting for any more user selection. For the second sub-part, the simulation will continue stopping after each time period and will wait for a "new data input." Then, it will continue as per the 4 branch ways simulation algorithm. d) New data input "exists" and user "wants" to review the performance: the simulation moves to the next time period using the new data input. The performance review segment will be executed as per aforementioned (c) segment. . Fig. 1 . Digital Twin Simulation Algorithm Part 1 Another crucial part of the simulation algorithm is to rank and update the recommendation list. First, a generic knowledge-based recommendation list is created. Then, based on, user selection and actual improvement, the recommendation list will be ranked. As it is mentioned before, the user selects the recommendation/option to improve the worst performer. After the selection, a new estimated data input generates which is used to complete the simulation to the end of the total time period. Now, another sub-part simulation goes out which stops and seeks new data input. Once it receives the new data input, the simulation compares new data input with estimated data input with respect to the improved parameter aka recommendation impacted parameters. A threshold value will be assigned to assess whether the user selected recommendation really impacted the outcome. If it does, the recommendation list will remember the outcome, adjust its recommendation list, and rank accordingly. Next time, when user will be on similar stage and look for performance review, the recommendation list / option bank will be ranked and show historical significance. Henceforth, the recommendation list will be constantly updating and helping the user to make the right decision. Figure 1 and Figure 2 describe the whole simulation algorithm for an instantiated case for the simplicity. Figure 2 is a continuation of Figure 1 . It is representing simulation algorithm csubpart 2 One of the limitations from the literature review is the lack of strong case study. In this study, an actual complex part is selected from the defence industry. For reasons of product privacy, the exact system cannot be disclosed. Process plans are masked but are given code identities to each process step in the assembly. The product system in focus is a low volume, high variety and high value product. The objective of the study is to find the best sequence of manufacturing operations in the production floor considering uncertainty in part arrival, sudden machine breakdowns, data integration, and part obsolescence. The physical system is approximated in this study as a discrete event linear and time-invariant system. It consists of a production station/area for receiving the product to be assembled with expected production metrics and generates an output for presenting the evaluated outcome of the assembly. Figure 3 provides a schematic block diagram of the physical system. The schematic block diagram is divided into two parts for the ease of representation. The diagram represents one single part. 44 raw materials are coming at different stages from various suppliers. The single part in the assembly goes through approximately 85 stations including testing, assembling, soldering, torqueing, and inspecting. Finally, the model outputs Z, which is the final product. Each operation stations are denoted as "opxx" where xx is the operation number. Operations along with hour per unit (HPU), first pass yield (FPY), and failure data are collected for two-year time frame. For this project, the data covers a period of 24 months, starting from April 2017 to April 2019. First 12 months data are used as initial historical data. The last 12 months data are used to test and verify the algorithm outputs. The real data set comes from the factory floor after averaging the compiled factory data. This experimental data set includes operation numbers along with time requires to complete (HPU) and first pass yield (FPY), incoming raw materials and their estimated numbers, and assembly operations. There are several realworld issues that need to be considered before further analysis of the data, e.g., missing records of operations, missing operation-end timestamps, unintentional and deliberate errors, security issues related to privacy and nondisclosure of business secrets, etc. The block diagrams show the operations sequence. Important assumptions in data modelling are given as below. • Each day is 8 h and each month is 30 days. • Only one operation can be processed at a time for an individual operation. • There are no interruptions during the processing of an individual operation, which means that the work on an operation cannot be stopped in the middle and then continued later. • All work systems are considered as if they are each at their own location and the material is transported from one operation to another where the subsequent operation is performed. Due to missing information about transport, the time needed to transport material from one to another operation is set to zero. • Breaks at work, failures, troubleshooting, etc. are already included in the processing times. • Maintenance work is not considered. • Th number of workers in the systems is not considered. • The work orders are dependent. That is, first block (Z1) and second block (Z2) finishes before the final block starts. A python-based automation console is created to run the simulation. First, the console asks for data input. Once it gets the data input, it moves to next month and wait for new data input. Based on the user selection, it follows the simulation algorithm and provides a recommendation list. The user selected recommended Figure 4 (a) shows how simulation algorithm catches the trend better. And Figure 4 (b) shows the less error percentage than the traditional simulation methods. Figure 5 shows the initial recommendation list on the month of 2 and updated recommendation list on the month of 14 for OP120. The figure clearly shows how recommendation list gets updated and aids the user to make a better choice. Traditional simulation model is when the user provides all the data at the starting point, simulates the scenario for the whole time period, and provides the estimation of the end output. On the contrary, proposed simulation model provides all the data at the starting point, simulates to the fraction of the whole time period, feeds real data from the factory for the simulated period, provides the estimated output by adjusting the differences between starting data and actual data, and provides recommendations to capture initial production target. Thus real-time data input at incremental time points adds value to the simulation and provides a user with decision making availabilities in terms of forecasting production schedule delays. Therefore, it is clearly visible that the simulation algorithm of DT helps better predicting the output and aids the user to make a better choice. A limitation in our study is that finer grain real-time data from the manufacturing floor is lacking, and if available can further add to informing a DT simulation as to causes of delays. Much of data collection is still manual and therefore prone to errors and delays in digital entry into information systems. This can cause the simulation to not take into account true factory floor conditions. In its current implementation, the DT of the process is a valuable tool to aid production floor managers in making decision that are based on simulation and on a knowledge base of prior action taken when others have encountered the problem. The digital twin creates concrete value, generate revenue streams, and help the user to make key strategic decisions. DT has many applications across a product life cycle. It is easier to answer the critical what-if questions in real-time more accurately. It is expected that this kind of analysis can contribute to better management of operations and thus to more efficient execution of operations, better utilization of resources, shorter lead times, and higher due-date reliability. In this paper, a digital twin implementation framework has been proposed. A manufacturing process case study showed the proof of concept validation of the framework to implement a digital twin of the manufacturing process. Specifically, this paper helps human involve manufacturing process where Initial List Updated List automation is not practiced greatly. It can be speculated that this approach will have instructional significance when manufacturing industries build or revamp a plant, arrange the facilities or staff and polish up the process flow. Through the experiments, issues and challenges which need to be considered for further research towards the development and implementation of digital twins in real manufacturing environments is identified. Smart manufacturing systems for Industry 4.0: Conceptual framework, scenarios, and future perspectives Multi-objective optimization using evolutionary algorithms A hybrid of genetic algorithm and bottleneck shifting for multiobjective flexible job shop scheduling problems A simple and effective evolutionary algorithm for multiobjective flexible job shop scheduling Multi-objective optimization using evolutionary algorithms Non-dominated ranked genetic algorithm for Solving multiobjective optimization Problems Solving Stochastic Shortest Distance Path Problem by Using Genetic Algorithms An effective hybrid particle swarm optimization algorithm for multi-objective flexible job-shop scheduling problem Robust scheduling for multiobjective flexible job-shop problems with random machine breakdowns Multiobjective flexible job shop scheduling using memetic algorithms Scheduling flexible job shop problem subject to machine breakdown with route changing and right-shift strategies Simulation in manufacturing: Review and challenges Gartner's top 10 strategic technology trends for 2017 Digital twin: Mitigating unpredictable, undesirable emergent behavior in complex systems Virtually perfect: Driving innovative and lean products through product lifecycle management From simulation to experimentable digital twins: Simulation-based development and operation of complex technical systems The air force digital thread/digital twin-life cycle integration and use of computational and experimental knowledge The authors would like to acknowledge Dr. Guodong Shao from National Institute of Standards & Technology for his valuable comments and feedback. Author name / Procedia Manufacturing 00 (2019) 000-000