Abstract
Due to the rapid increase in constantly changing client requirements, companies are undergoing rapid transformation and product life cycles are becoming shorter, making typical methods in modeling and simulation unsuitable for dealing with this situation. Meanwhile, utilization of digital twins in intelligent production domain can improve making accuracy, reduce production costs and control energy use by simulating the realistic states of production platforms in real time. Therefore, a universal data-guided architecture is proposed for automatically generating virtual models as the foundation for intelligent production digital twins. The novelty of the presented architecture lies in its data-guided method, which leverages advances in techniques of process mining and machine learning, as well as progressive model enhancement and verification. The objective of this architecture is to reduce and completely specify, or even remove, the demand for specialist knowledge when eliciting the appropriate virtual models. Finally, the proposed architecture is described using application example.
Keywords
Introduction
Although the expression “digital twin (DT)” is defined in multiple ways in the recent years,1,2 these definitions share several characteristics. For example, a “physical substance” is “digitized” to reach definite objectives, and the “instantaneous” essence of digitized model, that is, executable virtual model, can be used to assess different hypothetical contexts. The term “instantaneous (or real-time)” means the fact that updates to the virtual model are the result of alterations in the physical substances. It allows the digitally captured virtual model to become a high-fidelity representation of its real-world equivalents and operate within realistic simulation.1,3,4 The main difference in definition is that digital twin needs to communicate with physical substance in some way. Some researches depict this communication as a “bidirectional flow of messages,”5,6 leading to alterations in both real and virtual substances. Technologies such as the IoTs (Internet of Things) that enable such connections can also be regarded as a critical portion of a DT. 4 In order to clearly express what is meant by DT in this study, the following components are defined: (1) the data that are being or have been created by the physical substance. As depicted in Section 2.2, the object is a critical portion of the DT that affects the extent to which the digital modeling targets are reached; (2) data-guided virtual model, including (i) algorithms depicting the virtual models, data mining and machine learning for eliciting the model from data, as well as their applications, and (ii) connected objects, that is, IoT platforms; and (3) physical substance, for example, an appliance, an equipment, or an entire shop floor.
The notion of DT can be regarded as an expansion of conventional virtual modeling, combined with enhanced data usability, connectedness, and the progression of user requirements. The capability to comprehend, control and validate complicated physical platforms is the common aspect of both simulation modeling 7 and DT. Another function of the DT is to offer feedback to the real-world platform using the information elicited from the simulations. Afterward, the feedback can be utilized to conduct certain facets of it to allow optimization of certain parameters given by the user.
DT has a variety of purposes, from the automotive industry 7 to manufacturing.8–10 Table 1 shows its applications in intelligent manufacturing. In this study, the scenario of data-guided DT modeling is considered for intelligent manufacturing and its scope is associated with the notion of automatic virtual modeling. Table 2 reveals the previous researches that tackle automatic virtual modeling. Summarizing available researches on semi-automatic or automatic virtual modeling, the modeling methods encompass knowledge/data and incident-guided. Hence, in terms of model accuracy and reliance on expert knowledge, adjustments must still be made on the basis of the specific demands of the scenario. The shift from conventional manual virtual modeling to data-guided methods - particularly in reconfigurable/reorganizable production platforms—essentially involves reallocating the focus of effort and expertise. The summary of Trade-offs is listed in Table 3.
List of digital twins in intelligent manufacturing.
A brief review of automated virtual modeling.
means rapid reconfiguration of manufacturing system.
means manufacturing operations management.
Summary of Trade-offs.
Furthermore, these researches have used various types of data and fabrication contexts. The emphasis of this study is on the product life management (PLM) in the reorganizable production platforms. This results in challenges in the following aspects: (1) reorganizable production—This production model is driven by changing product demands, 21 which require workshop to quickly alter its architecture, and digital twin is particularly well adapted to this context.22,23 Although reorganizable production has many advantages, such as high flexibility to meet different market needs, it also faces many challenges associated with production reliability, cost and efficiency. However, these challenges can be overcome by utilizing powerful DT that reflect real production scenario; (2) re-production—this facet of production involves converting used and/or old products into new items. DT has been utilized in recycling solutions for discarded electronic and electrical devices24,25; (3) whole PLM—all facets of the product lifecycle, from initial concept to eventual disposal, can be tackled by dedicated digital twins.26,27 PLM is attractive due to its inclusion of all facets of the manufacturing platform, and this is particularly critical for the creation of virtual models of a whole intelligent shop floor.
Relevant work
Industry 5.0 is characterized by a people-oriented, flexible, and sustaining transformation, and DT is one of the most promising technological drivers for implementing these features. 28 Basically, DT is an evolutionary virtual models of the past and present activities of real components or processes, which can be utilized to help optimize related performance indicators. Its implementation is facilitated by several new technologies, for example, XR (extended reality), IoT sensors, and artificial intelligence. 29 The concept of DT helps address several obstacles currently facing the manufacturing industry, which are as follows: (1) complexity of systems—because of the change in system mode (from general to reconfigurable), the deployment and manipulation of entire production systems has become extremely complex. DT can tackle and optimize these platforms in an automated and data-guided manner to satisfy client needs, assisting to overcome this obstacle4,30; (2) fluctuation of client needs—the quantities and styles of products are constantly changing in response to the needs of customer and markets. The same is true for the production lines. Thus, the production model need to be transformed into customizable, flexible, and reconfigurable ones, resulting in customized virtual models becoming obsolete shortly after development31,32; (3) efficiency of production—with intense competition in the manufacturing industry, more innovative ways are required to enable production more cost-effective. Without promptly and precise virtual models of various facets of production and other relevant processes, it is impossible to improve cost-effectiveness. On the contrary, more efficient and valuable analysis can be conducted, including description, prediction, and prescription, thereby reaching cost-effectiveness33; (4) instant monitoring—as mentioned earlier, production platforms undergo frequent alterations to satisfy the requirements, greatly increasing the requisite for instantaneous and continuous monitoring. This capability enables better and smarter decisions, which in turn influence the achievement of relevant performance targets.34,35
In short, DT helps address the above-mentioned obstacles by offering a way of successively evaluating, controlling, and optimizing the performance of a given production platform.
The challenges of typical virtual models
Technique of virtual modeling is often used to assess the layout, manipulation, and efficiency of complex platforms. Its approaches and modes are used to model platform activities down to the lowest layer of detail and use them to forecast subsequent activities for the purpose of making decisions. Discrete event simulation is a computational approach that can be applied to express the activities of complex platforms. 36 These models are created with expertise in choosing proper abstractions.
However, this approach of manual virtual modeling cannot exactly reveal the variations occurring in the reorganizable production platforms, where the digital and real structure are constantly modified to meet changing market requirements. 37 As a result, these models quickly become obsolete after development. To address this issue, it is essential to continuously create new virtual models or manually update current ones, an extremely challenging and costly process. By contrast, data-guided virtual models (i.e. digital twin) can seize these changes and their data in real time. 38
Digital twin for intelligent shop floors
In production installations with dynamic and quickly shifting marginal situations, some critical issues can be addressed by adaptive processes as portion of an intelligent workshop. DT can be used in an intelligent shop floor to effectively dominate different facets in complex production platforms. There are several benefits included by using DT 29 : (1) tailored services to clients; (2) real-time monitoring of machine state; (3) the possibility of predicting maintenance; (4) enhanced reliability. The above-noted advantages can be realized in intelligent workshop through data continuously gathered and transmitted by IoT appliances. This data can then be used for critical tasks, for example, decision-making, scheduling, logistics, and maintenance. The eventual target of Industry 5.0 is to achieve smart and data-guided decision-making in intelligent production context to promote profitability and sustainability.
Virtual model with data-guided mode
A virtual model that retrieves and sets parameters from data is termed a data-guided virtual model. 16 There are several advantages highlighted regarding data-guided virtual modeling. They are depicted as follow: (1) involvement of machining learning (ML) and artificial intelligence (AI)—data can be modeled using AI and ML approaches to model situations that are difficult to reflect in typical virtual modeling. In addition, the two methods can be combined to simulate and integrate, in order to further comprehend the actions of certain platforms.39,40 Meanwhile, several researches describe algorithms (e.g. Reliability-based Multidisciplinary Design Optimization, RBMDO; surrogate-model-based reliability evaluation) for virtual modeling, data mining and machine learning for extracting models from data, and their applications41,42; (2) close to real-world context— compared with conventional approaches, data-guided models more closely resemble the real platforms being modeled. Physical substance, for example, an appliance, an equipment, or an entire shop floor.43,44 In other words, physical platforms may display activities that are difficult to predict and model beforehand, and they can only be caught by data.33,45
The above description, coupled with several successful cases of data-guided methods, highlights the need to shift from typical virtual modeling to data-guided methods. Advances in various areas, such as data management and algorithms, as well as the ever-altering demands of the product market, have made this transformation possible. There are several shortcomings in previous researches (as presented in Table 2) in data-guided virtual modeling: (1) reorganizable platforms—reorganizable production platforms have been used to support the market demand for large-scale customization. 46 The concepts of virtual modeling depicted in above-noted studies are not applicable to frequently changing contexts, for example reorganizable production platforms; (2) Data verification—data gathered from intelligent workshops has many issues because it usually includes missed values and noise (which is very common for sensors). The frequence and quantity of these data are usually subject to noise interference, which stems from the workshop conditions and operator negligence. Therefore, it is crucial to have a mode for automatically verifying streaming data, as the elicited model’s quality is strongly based on the efficiency and availability of the data. 47 Most of the past researches are not concentrate on data verification, a critical step in assuring the accuracy of the models; (3) model produced primarily by data-guided method—except for the approaches depicted in previous researches,48,49 most of the existing model generated methods cannot be categorized as automatic virtual modeling. Furthermore, these studies do not tackle the entire PLM of a typical reorganizable intelligent shop floor. This topic is receiving attention because it focuses on all aspects of the production platform that are related to product life; (4) data diversity—intelligent workshops have a lot of data sources. Furthermore, data transmitted by IoT appliances are available in many various formats, such as diaries, videos, graphics, and so on. Deploying integration approaches to leverage this diverse data. 50 Previous studies mostly concentrate on homogenized data sources.
An architecture is developed for data-guided DT for intelligent production in this study. This architecture offers many benefits to tackle the existing obstacles of intelligent production. In this study, there are several issues discussed, including the primary objects of proposed architecture, its advantages, obstacles and possible directions. By using an application example to further depict the presented method. This study is outlined below. Section two outlines the data-guided DT architecture proposed for intelligent production, covering detailed main factors. Section three provides an application example for the purpose of analyzing the reliability of DT. The obstacles and potential chances of the presented architecture are explored in Section four. Eventually, the conclusion is depicted in the last section.
Digital twin architecture for data-guided intelligent production
Due to the fact that intelligent workshops continuously generate massive volumes of data, the notion of data-guided DT has emerged. Data has become a critical factor in evaluating and monitoring the status of complex platforms in the contexts of shop floor. Thus, effective utilization of this data helps to create high-fidelity models of production platforms. On the other hand, two critical approaches (process mining and ML) can be introduced to detect information from this data.51,52 Their function within the DT structure is to enhance the model performance by leveraging elicited data. This data is updated online in real time to reflect any alterations that happen in the physical production context. This allows many functions to constantly enhance various facets of the production process, even if only minor alterations happen. The innovative facet of the presented architecture is that it utilizes process mining and ML to address issues with past methods of virtual model development and establish a data-guided model that can be constantly renewed and verified.
Critical phases
The critical phases for achieving a data-guided DT in an intelligent workshop are described below:
Phase of data gathering
The scenario of intelligent workshop, a massive quantity of data acquired from multiple sectors. Before gathering data, the activity that must be executed first is identification of substance. In this phase, the real-world data related to the DT target in the modeled physical platform are identified and gathered. After this phase is accomplished, the data constantly created by the identified substances in the production platform are kept in a repository. There are several data sources included: (1) intranet—the appliances and interconnection between various portions of the shop floor can gather overall data from the shop floor; (2) sensor—different sensors are attached to the fabricated parts, and the data recorded by these sensors is extremely beneficial to execute specific functions, for example, trouble checking and preventive monitoring; (3) terminal input—members in production platforms (e.g. clients) usually offer data that shape their priorities, restrictions, and demands.
As for the data generated, the types include the below: (1) feedback from clients—the comments, restrictions, and requirements from end clients are typically used as a basis for modifying appliances operations to achieve predefined objectives; (2) diaries and tracks—recording the operational history of parts or appliances as a work log is very valuable for comprehending current processes in the workshop scenario; (3) inspection—this data is usually gathered by sensors attached to the appliances, which help monitor the operating status of the production platforms.
Meanwhile, data integration is another essential portion of this phase, which involves converting the gathered data in different formats (including diaries, videos, graphs, etc.) into a unified format appropriate for task execution.
Phase of data confirmation
Because the body of the model is data, it must be thoroughly inspected to confirm its correctness. The checking process includes transformation and validity tests. Some issues related to data are explained below: (1) deficient data—possible causes for this problem include malfunction of appliance, network issues, and others; (2) vast amounts of data—it is common to create vast amounts of data in the shop floors, and the high dimensionality and size of this data can make it difficult and inaccurate to create models; (3) random data—such data may be caused by operator negligence, the measurement instrument and the environment; (4) Multiple modes—sensors within the architecture of intelligent shop floor can capture various types of data, such as audio, video, graph and text.
If the fundamental data involves undiscovered errors, it can severely impact the exactness and quality of extraction models. 53 Thus, the availability of data is extremely important. The preprocessing activities contain supplementation and transformation. The former means the activity of patching in the deficient data utilizing ML or statistical methods. The latter denotes converting data into a form suitable for use in the retrieval procedure. To address the data verification issue, an anomaly discovery method is used, which includes the concept of “normal” and marks deviations from this normal value as anomalies. Currently, there are several catalog-based methods that use categorizers (e.g. supervised ML algorithm) to identify the limits of normal to abnormal data. However, the problem with utilizing supervised categorizers for anomaly discovery is that the uneven distribution of examples in the dataset often significantly affects discovery efficiency. Thus, non-supervised approaches are also applied successfully to discover abnormal data. Although most methods in the aforementioned anomaly detection domains are already verified against reference datasets, their capabilities on physical data usually depends heavily on the application. These different methods have been discussed in previous literature, and none of them have demonstrated a clear advantage. Hence, an integrated approach is used for discovering anomalies to assure data availability in the proposed architecture. Specifically, the most proper way for discovering anomalies is selected according to the kind of data received by the sensors.
Phase of information extraction
The activity of retrieving information from the useful data acquired in the preceding phase. The procedure is as follows: (1) discovery and marking of incident—certain related incidents serve as inputs for virtual models, so discovering these related incidents involving virtual targets occurring in an intelligent shop floor is a crucial portion of data-guided virtual modeling. Incidents are critical factors in guiding the establishment of discrete incident virtual models. The capability to auto-discover incidents occurring in the workshop decides the efficacy of virtual modeling. There are different ways to discover incidents from the data of intelligent shop floor, like bunching54,55 and catalog-based methods.56,57 Meanwhile, in the reliability analysis of production platforms, the discovery of particular incidents, for example, defects, is considered a critical portion. 58 The reactivity and exactness of the intelligent workshop model is influenced by its capability to discover when defects occur. From a ML aspect, this activity of defect discovery falls under the classification of anomaly discovery activity, with numerous algorithms already available in the past researches59,60; (2) discovery of process—another source of information for recognizing workshop status is the process, which helps to comprehend several parameters, for example, time, steps, reliability, and so on. Incident diaries are one of the critical data sources for process recording. By employing mathematical modeling language, such as place/transition net, 51 a formal description of the workflow within the platform is extracted from the primitive incident diaries. Apart from that, the relevant compliance test is executed through the formal specification.
Phase of model establishment
The incidents and processes obtained in the previous phase (Section 2.1.3) are used as inputs for establishing the virtual model in this phase, though a certain degree of human involvement is still required initially. Afterward, this preliminary model is automatically updated to respond to alterations in the intelligent shop floor. To promote this update, it is essential to clearly recognize the correlations of the model with the data flows within the intelligent production environment. Additionally, approaches for extracting and updating model have to be established to assist the processes of virtual modeling.
Phase of model confirmation
The target of this phase is to develop high-fidelity models to support timely and accurate decisions. Because data is continuously collected, the virtual models must be constantly validated. The usability of such data provides a chance to construct novel methods of model validation that can handle the validation process incrementally, enabling virtual models to be verified immediately once “enough” data is available. Since model validation relies on data, the data itself must undergo strict verification. The aim of this phase is to assure that the model not only performs well on seen data, but demonstrates strong predictive capabilities on unseen data. Existing research indicates that out-of-sample bootstrapping methods are inclined to produce estimates with minimal bias. 61 Within the proposed architecture, the out-of-sample bootstrapping approach operates as depicted below. Data instances are randomly sampled from the raw dataset using replacement sampling. Model drilling is performed with this sample, while benchmarking is conducted with the source dataset. Repeat this procedure n times and record the mean performance index obtained for each iteration of the train-test cycle. As for the procedure of model verification, it is executed at preset intervals to assure the model remains synchronized with the data generated to date. In the proposed architecture, when platform alterations result in new data becoming available, the situation where models become outdated can be addressed through a cyclical process involving data validation, information retrieval, model building and verification. More specifically, data gained over time t is used to extract information and build models based on it. This model is validated to assess its accuracy regarding the observed data and the model’s goal. If the model meets the accuracy, store its parameters. When time reaches t + n, the platform undergoes a change and creates new data. At this point, the identical process is performed repeatedly to assure consistency between the existing model and the data.
Data-guided digital twin architecture
Figure 1 illustrates the architecture for data-guided DT for intelligent workshops. As a constantly modeled substance, the intelligent shop floor generates data by means of its sensors and IoT appliances and sensors. This data serves as the base for data-guided modeling method. The process of data accumulation involves recognizing related substances and storing data in repositories. Substance recognition involves clearly indicating related substances, for example, control techniques, fabrication platform, and other objects in the intelligent workshop. The Project Haystack 62 is one of the methods used for collecting and interpreting production data. It is an open-source technology suite for modeling IoT data, designed to offer a standardized semantic data model. Its objective is to retrieve value from the data produced by different IoT appliances, involving those within production facilities. The next step involves data verification through common data conversion or integration and preprocessing. The data records the incidents taking place in the workshop. To clarify these incidents and assist the model creation, incident marking is executed semi-automatically, meaning it is completed with human involvement.
Architecture of data-guided digital twin in intelligent shop floor.
The process of incident discovery and marking can be performed through an incident-based unmonitored learning additional procedure, such as bunching. In this procedure, all discovered incidents undergo feature description and bunching analysis on the basis of their similarity. In summary, a specialist observes the established groups and their feature descriptions, and intervene when necessary to offer tags for the groups. Through this process of event discovery and marking, several related incidents are recognized and marked. ML models can utilize prior marked data to automatically and constantly discover incidents. Hence, incident diaries are generated, which are further utilized for process detection through process mining approaches. 63 The identified incidents and processes are subsequently utilized to establish a virtual model for a specific intelligent shop floor. During the phase of model establishment, ceaseless verification of the model is required. This process is conducted incrementally, with verification tasks completed concurrently during the model extraction process.
The associated model data are saved for subsequent utilization when an extracted model passes verification. This model can be further applied to simulate and analyze different hypothetical contexts, and assessed based on predefined performance metrics. These experimental outcomes can assist business in making beneficial decisions about the optimal operating mode for a specific intelligent workshop.
Implementation of the proposed architecture
This study demonstrates the potential and practical value of applying data-guided virtual modeling technology to DTs in the manufacturing sector through a case study of impeller production. The objective of the virtual scenario research is to study the reliability of the production route. Additionally, relevant background information on the production process is provided, along with a depiction of a Place/Transition net (PT net) virtual model obtained from the production appliance. This model serves to illustrate the process and explore potential expansions and influences. Eventually, by applying this architecture to data produced through the illustrated production route, its effectiveness is demonstrated.
Depiction of route and appliance in intelligent production
The rotating parts of the turbomachinery in an automotive engine are primarily located in the turbine and compressor stages (as shown in Figure 2(a)). Their geometric structure is quite complex, consisting of a series of blade disks arranged across the single rotor. 9 Over the past few decades, a new trend has emerged in which blades and rotors are integrated into a single, monolithic part. This involves machining from a single initial disk, using five-axis milling to remove excess material, ultimately producing a part that combines maximum strength and the lowest weight. Such parts are generally defined as Integrally Bladed Rotors (IBR), or “Blisk” (derived from “Blade on Disk”). The categorization of fabrication procedures depends on part features, application areas, material demands, and geometric definitions. In the IBR specification, geometric shapes are classified by the relationship between the disk diameter and the length or area of blade. 10
A rotating part in an: (a) is automotive turbocharger system and (b) is impeller. 64
The IBR geometries with small blades size, such as high-compressor phases. Thanks to the high flexibility of this process and a deep understanding of this traditional method, high-speed machining (HSM) technology is employed to machine the part directly from a solid block of material. Hence, the production process of impellers (as shown in Figure 2(b)) involves machining, inspection, and heat treatment. There are several properties involved in the production route65,66 and its conceptual layout is illustrated in Figure 3: (1) a graphical user interface (GUI) for monitoring and operating fabrication procedure; (2) a storage device equipped with automated picking and stocking functions; (3) an automated guided vehicle (AGV) with a robotic hand can perform part loading and unloading activities; (4) two production blocks are provided with collaborative robotic arms to execute certain activities; (5) three high-performance equipment consisting of a five-axis machining center, an optical inspection instrument, and a laboratory box furnace (for heat treatment). Meanwhile, the production steps are depicted briefly below:
Step I: an operator (such as an engineer or technician) initiates the production process by sending a demand through the GUI;
Step II: upon receiving demand messages, inventory prepares workpiece for processing;
Step III: the workpiece is delivered by the AGV equipped with a robotic arm to the five-axis machining center, then load the workpiece into the machine’s chuck and fasten it securely;
Step IV: after machining (includes rough and fine) is complete, Unit 1 unloads the workpiece and transfers it to the inspection instrument;
Step V: after the workpiece undergoes inspection (including dimensional, geometric, and surface roughness testing), it is unloaded via Unit 2 and subsequently transferred to the heat treatment equipment;
Step VI: the finished product is unloaded by AGV equipped with robotic arm after heat treatment;
Step VII: the AGV delivers product to the stocking area for storage;
Step VIII: the stored message is returned and displayed on the GUI to notify the operator.
Conceptual layout of production route.
The aforementioned appliances and their associated resources have facilitated a series of technological developments associated with the Industry 5.0 Award. The target is to build a dynamic and imaginative workplace environment for manufacturing enterprises to collaborate closely with academics in exploring new and efficient technologies in production.
Reliability of virtual model
To display the model retrieved from the production route through the proposed architecture, a PT net is created, as shown in Figure 4. The PT net clearly illustrates the strict sequential character of the production process. Because the existing production route configuration lacks redundancy or buffer mechanisms, a failure in any single piece of properties would cause the whole production to shut down completely. The potential faults that may occur during property operation are modeled as special arcs to prevent specific conversions triggers. When a property fails, its fault conversion triggers and generates a mark at a specific “fault” location, thereby activating the special arcs. Any conversion dependent on the failed property cannot be triggered. The mark at the location of the property failure is erased once it is repaired, and the special arc is disabled.
Diagram of reliability—focused P/T net model in the present state of production route.
Common defects on the impeller production route that may lead to platform malfunction are investigated and depicted as follows: (1) when a workpiece is tilted on the staging table, the collaborative robot arm can fail to grasp the workpiece; (2) the AGV loads the workpiece from the machining center to the heat treatment equipment; (3) the AGV cannot identify the base plate containing the workpiece offered by the storage; (4) workpieces occasionally fall off the staging table due to accidental bumps.
A similar production route with PT net mode is mentioned 67 to show and verify the efficiency of a timed PT net virtual system. A model of maintenance circuit is established similar to the property failure circuit to reduce intermission of production route. During maintenance on a definite property, a special arc is utilized to prevent the associated timed conversion. In this study, the proposed architecture and approach allow extracting models, for example, Figure 4 shows the reliability PT model is extracted from the data that is collected through sensors and require restricted specialist knowledge assistance.
Virtual model with data-guided reliability
As revealed in Figure 4, it is manually constructed to act as a standard for the proposed framework and method to extract key messages from data. The automatic extraction of such a reliability-centric PT net utilizing process and data mining techniques are detailed next. By investigating previous studies,68,69 the data used to support data-guided reliability modeling in the workshop are divided into three categories, including status, condition tracking and incident. They all have a sequence arranged according to a time pattern, so each record is composed of a timestamp (Tp) and a certain type of data. The recognized data sorts are illustrated in Figure 5: (1) status data—it records the working statuses of the respective properties of a platform. The table reveals the example statuses for each property; (2) condition tracking data—it records the related health data of a platform. The table indicates the example data acquired through sensors; (3) incident data—it records the individual incident produced by the properties of a platform. There are three items listed in the table: instance label, property and incident. The instance label categorizes all the tasks that must be executed to develop a product. The starting of each task is marked by incidents. 70
Data types with schematic diagram and example dataset: (a) is status data, (b) is condition tracking data, and (c) is incident data.
On the basis of incident data, there are several process discovery approaches utilized to extract process models depicting the production of impeller.71–73 For status data, extraction and modeling can be utilized to depict alterations in the working status of the properties. These modeling facets are reflected in PT net. Proper norms are adopted to assess its quality after deriving the reliability-centered PT net. This can be finished on the basis of two methods: the data diaries from which the PT net is elicited (consistency check) or the PT net with base fact. The past research 74 has mentioned suitability and appropriateness indicators to deal with the former. When the base fact process model is recognized (as in this study), the quality of the mined model can be verified and assessed utilizing the general indicators, including accuracy (Ac), retract (Re), and harmonic mean (Hm). 75 Ac means the ratio between the amount of exactly allocated brims and the total amount of allocated brims. Re represents the ratio between the amount of exactly allocated brims and the amount of all brims exist in the primitive model. The corresponding formulas are listed below:
Where Bap denotes the group of brims that present in both elicited model and primitive model at the same time. Bp is the group of brims that reveal only with the primitive model. Ba represents the group of brims that appear only with the elicited model.
This study uses reliability analysis to evaluate the system’s availability and accessibility. To perform reliability analysis using elicited PT net, it is crucial to obtain the fault probability distributions and mend times for each production property (“Mend” and “Fault” conversions in Figure 4). The status data offers this messages by estimating the time to each fault and mend on the basis of the related working status alterations. To calculate the probability distribution followed by empirically gathered data, MLE, a statistical method, is utilized.76,77 It computes the parametric values of a probability distribution using maximum likelihood estimation, ensuring that the hypothesized theory distribution closely fits the empirical messages. Meanwhile, the reliability function (R(t)) can be sketched when the correct dispersion and its relative variables are calculated. The R(t) is a cumulative function of the dispersion, representing the probability that a certain manufacturing property appears within a definite time period. 78
Another method for evaluating a platform reliability is fault tree analysis (FTA). 79 A data-guided FTA method is depicted on the basis of time series data. The incidents that lead to a platform failure have to be recognized to create FT in a production environment. If the incident data is not fine-grained enough to seize defect incidents, the approach of defect discovery and categorization on the basis of condition tracking data of a production workshop can be utilized. 80
Explication of instance
To demonstrate the data-guided DT architecture presented in this study, the data of production route depicted in Section 3.1 is utilized, and techniques of process mining are used to elicit the corresponding PT net model and the reliability perceptions associated with the fault and mend time of production properties. Next, the steps performed are depicted in detail as follows.
In addition to the 1000 entities built on the impeller production route, the relevant incidents and working status alterations of production properties are recorded to generate data. A portion of the status diary and incident diary that are created in the period of this procedure (as listed in Tables 4 and 5). To draw the PT described in Figure 4 and evaluate the distributions for the fault and mend conversions, the two libraries, reliability 81 and PM4PY, 82 are used. The former is python library for survival analysis. The latter is a process mining library specifically developed for Python, designed to bridge the chasm from process to data science. Also, it provides an extended code library for PT nets in modeling and simulation. By applying the above-mentioned libraries, the whole elicited PT net is illustrated in Figure 6.
List of the partial status diary.
List of the partial incident diary.
Diagram of the elicited PT net.
On the basis of the created incident diary, the α-miner approach
72
is utilized to elicit the model of material flow in the production route (middle portion of PT net in Figure 4). Afterward, the fault and mend rings for each property are elicited on the basis of the status diary (upper right and lower left portion of PT in Figure 4) and connected to the correlated property operating conversions, such as “HTF operation,”“Deliver to MC.” Firstly, establish two positions (
Histogram of each fitted distributions superimposed for the failure of machine center.
Fault time and survival function of each production property: (a) machine center, (b) inspection instrument, (c) heat treatment furnace, (d) AGV, (e) unit 1, and (f) unit 2.
Discussion
There are several issues with the proposed architecture for data-driven DT in smart workshops, shown below: (1) decision making with real-time mode—practitioners can make beneficial decisions on various performance indexes with the help of DT. This decision-making assistance is possible because data-guided DT can derive related information from all gathered data, which may prove valuable to the practitioners. Regarding the promptness and importance of decision-making, the demand for promptness should be aligned with the importance and significance of the related decisions. This must be evident in the level of detail of the data gathered. Safety-related decisions, particularly in safety-critical platforms, demand higher degree of time-sensitivity than production-associated decisions; (2) growing objectives—the goal of the intelligent production platform is to maintain a steady and affordable productivity. This target is tightly intertwined into other features, including reducing material waste and energy consumption, as well as enhancing the reliability of the production platform, and so on. Additionally, environmental objectives such as reducing greenhouse gas emissions are equally important considerations that must be incorporated. Developing virtual models to update, represent, and feed into these changing and expanding objectives is an extremely complex and urgent task; (3) proper human interference—to facilitate the data-guided method, it is necessary to identify and define the scope of human participation. This issue denotes the utilization of primarily data-guided method that requires and relies on human participation to achieve automation. Clearly defined manual input spots assure the correct configuration of data-guided virtual models, and their explanability and effectiveness. The former means the data-driven/intelligent method, offering practitioners a principled mode for understanding the decision-making processes within DT. The latter represents the facet of data-guided DT that involves exactness and promptness.
Conversely, the proposed architecture also reveals new chances and depicts below: (1) improvement of reliability—the proposed architecture demonstrates reliability due to the consistent verification of model and data. It focuses on continuous validation, assuring the platform can properly evolve at the pace of new data arrivals. The verification platform assures that the critical data within the proposed architecture is accurate and that the model truly presents the data it models; (2) evolution of high-accuracy DT—because the proposed architecture is data-guided, the model can accurately represent the corresponding production platforms to display the latest behavior of processes in intelligent shop floor. Additionally, the continuous validation of models and data assures the sustained accuracy; (3) nimble production approach—a well-established system can be provided by data-guided DT for implementing nimble production method, enabling manufacturing companies to rapidly adapt to market demands and changes as well as keep quality high and costs under control; (4) enhance comprehension of procedures and decision making—process finding is a key portion of the proposed method. Apart from other data-guided parts, this study achieve process finding by mining gathered incident diaries. This enhances comprehension of processes within the platform and aids in improving the quality of decisions. Through the use of processes within the platform, practitioners are able to make informed choices regarding the potential of the platform and the DT.
In short, the utilization of DT in the intelligent workshop presents numerous chances, but also involves different obstacles. The above-mentioned chances can be seen as the driving force behind the employment of data-guided DT in production platforms.
Summary
In order to successfully implement Industry 5.0, data-guided DT is essential. Because modern production platforms undergo many quick reconfigurations throughout their lifecycle, manual virtual modeling is not a good selection. On the basis of this reason, there are several characteristics introduced in this research: (1) a data-guided approach is utilized to establish demands for the virtual model of an intelligent production platform, serving as the foundation for its DT; (2) Data-guided DT with real-time mode is established based on data collected from IoT devices in intelligent shop floors to meet critical requirements, such as maintaining up-to-date virtual models and showing alterations in workshop configurations; (3) data-guided DT enables integrated continuous model verification, allowing immediate initiation after extracting individual model parts, including the reliability of each property (e.g. MC_R(t = 20) = 74.08%; II_R(t = 20) = 70.22%; HTF_R(t = 8) = 63.21%; AVG_R(t = 8) = 71.19%; Unit1_R(t = 8) = 47.17%; Unit2_R(t = 5) = 10.90%) and the corresponding repair distribution (MC_M (median) = 15.2 h; II_M = 5.19 h; HTF_M = 10 h; AGV_M = 1.24 h; Unit1_M = 4.94 h; Unit2_M = 4.5 h). The data usability in real platforms allows this to be possible.
The chances also bring numerous threats and demands that must be addressed: (1) Some procedures still require human participation even after full automation. Nonetheless, clear definitions and differentiations can help it gain mainstream acceptance. For example, defining objectives and identifying associated incidents requires specialists to execute; (2) The appropriate ratio of data to human expertise is largely dependent on the existing system architecture and the objectives set for the data-guided DT; (3) To meet the demands of data-guided DT development, an architecture of high-level concept has been established. As modes for assuring data quality become increasingly refined, this data-guided virtual modeling is expected to gain greater attention; (4) A high-level example is also provided to explain the proposed architecture and explore related chances and difficulties.
The production route will be expanded to include additional production stations and assembly units. These units, for instance, may include types where robots collaborate with technicians, as well as types that rely entirely on technicians for operation. Meanwhile, a hierarchical or distributed digital twin architecture for complex production routes that operate hundreds of machines and engage in mixed-model production will be explored, including developing explainable artificial intelligence (XAI) modules to enhance decision-makers’ trust in the future. In addition, this study will introduce goodness-of-fit tests such as the Kolmogorov-Smirnov or Anderson-Darling tests in the future to verify the validity of the selected distributions. Eventually, more robust alternatives like Inductive Miner and Heuristics Miner are generally better suited for such scenarios, particularly when model fidelity matters alongside real-time performance for information extraction tasks.
Footnotes
Handling Editor: Sharmili Pandian
Funding
The author disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study is supported in part by the university-industry-Local Collaboration project under contract number KH 24157. Also, it is supported by the research start-up funding for introducing high-level talents to Sanming University under contract number 25YG04.
Declaration of conflicting interests
The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data availability statement
The data are not publicly available due to privacy restrictions.