Skip to Content
MDPIMDPI
Applied SciencesApplied Sciences
PDF
  • Article
  • Open Access

6 January 2022

A Novel Embedding Model Based on a Transition System for Building Industry-Collaborative Digital Twin

,
,
,
and
1
Department of Smart Factory Convergence, Sungkyunkwan University, 2066 Seobu-ro, Jangan-gu, Suwon 16419, Korea
2
Department of Systems Management Engineering, Sungkyunkwan University, 2066 Seobu-ro, Jangan-gu, Suwon 16419, Korea
*
Author to whom correspondence should be addressed.
Appl. Sci.2022, 12(2), 553;https://doi.org/10.3390/app12020553
This article belongs to the Special Issue Artificial Intelligence in Machine Learning Approaches for Smart Manufacturing
Version Notes
Order Reprints

Abstract

Recently, the production environment has been rapidly changing, and accordingly, correct mid term and short term decision-making for production is considered more important. Reliable indicators are required for correct decision-making, and the manufacturing cycle time plays an important role in manufacturing. A method using digital twin technology is being studied to implement accurate prediction, and an approach utilizing process discovery was recently proposed. This paper proposes a digital twin discovery framework using process transition technology. The generated digital twin will unearth its characteristics in the event log. The proposed method was applied to actual manufacturing data, and the experimental results demonstrate that the proposed method is effective at discovering digital twins.

1. Introduction

As the complexity of manufacturing systems increases alongside rapid changes in recent consumption trends, forecasting to prevent future losses is becoming more important. A lot of research is being done on this related technology, and among them, digital twins technology is considered as an important core component of Industry 4.0. In particular, this technology makes process predictions possible through digital data, not just physical implementations [1]. The output of these digital twins can provide important indicators for long term and short term decision-making in the manufacturing process.
A lot of research on digital twins is being conducted, and recently, approaches using process mining have become popular. However, there are more specialized studies for business processes than manufacturing technology, and research using process mining technology to create a digital twin of the manufacturing industry is insufficient. It is necessary to be able to create digital twins using event log data to make appropriate decisions for manufacturing operations using the vast amounts of data generated in the manufacturing process. In this paper, we propose a digital twin discovery model that can predict the remaining cycle time, which is an important indicator in manufacturing environments, using event log data.
Process characteristics are extracted through embeddings based on transition technology of event log data with activity and Class Id. Extracted data are used to uncover digital twin models that predict the manufacturing remaining cycle times. In this way, this study solves the problem of determining the time required to terminate the remaining processes based on the execution log data of the previously completed processes. We intend to solve this problem using the GRU(Gated Recurrent Unit) artificial intelligence model.
In this paper, we devised a digital twin model that predicts the remaining cycle time of a manufacturing system using event log data, and propose a cloud type architecture using this setup. There are three main contributions of our study.
(1)
Transition system based embedding technology to easily extract process characteristics from event logs is proposed, and this technology can help interpret numerous log data accumulated by manufacturing companies.
(2)
Creating a digital twin model requires a lot of technology and is costly, and in most cases additional professional manpower and costs are required to reflect real data in the created twin. In this paper, we tried to implement this process through artificial intelligence learning, and it can satisfy the need for cost reduction.
(3)
By providing a cloud computing style digital twin architecture, our setup can help small companies with insufficient technical skills to realize digital twin methods.
The rest of this thesis is structured as follows. Section 2 describes digital twins and introduces research that links process mining to digital twins. Section 3 provides a description of the components of the proposed approach and the main idea of this study, as well as describes the transition embedding digital twin model proposed in this paper. Section 4 describes the experimental environment, the structure of the data set, and the experimental results. Section 5 concludes the paper with a summary of the paper, evaluation results, and future research directions.

3. TED: Transition System Based Embedding for Digital Twin

3.1. System Architecture

This paper uses a cloud computing based digital twin framework to predict the remaining manufacturing cycle time, and solves the prediction problem through machine learning using a transition based embedding method. In addition, we propose a framework that can solve insufficient technical problems through collaboration with other companies.
Herein, we propose a twin collaboration framework. Figure 2 illustrates a digital twin framework based on cloud computing. In this paper, after preprocessing data by tracking activities in the event log, machine learning is performed with a GRU based artificial intelligence model. We consider cloud based digital twins created through learning through the cloud, augmenting the artificial intelligence technology based on data provided by each company, allowing for the construction of precise digital twins through improved models.
Figure 2. Collaborative digital twin framework for smart manufacturing.
We provide an integrated solution to implement digital twins in the form of pictures in smart factory manufacturing processes. A service provider receives event log data from a customer, creates a digital twin model suitable for the process, and when the customer wants to conduct experiments in the manufacturing process, creates a customer in the virtual environment of the digital twin model. The results of the experiment are delivered to the smart factory so that the customer can make a more correct mid to long term plan and operate the manufacturing process. The generated digital twin can create a more precise model while learning another customer’s dataset, and the customer can receive more reliable experimental results. In fact, most SMEs are unable to implement even a basic digital twin due to a huge lack of costs such as data analysis costs, server construction costs, and consulting costs. Most small and medium sized enterprises (SMEs) are unable to reap the many benefits of implementing a digital twin, and are often stuck in a low-level technology environment.Through this collaborative architecture that we propose, both service providers and customers can benefit, and it can provide more growth opportunities for small businesses in low-skill environments due to lack of skills to implement digital twins.

3.2. Modeling Procedures

We use event logs to model digital twins. First, the activity of the manufacturing process is tracked in the event log, and the tracked activity is classified as a transition system. To proceed with machine learning using the classified data, the activity is newly mapped according to the sequence and embedding is carried out.
The name of the embedding model is TED (Transition System Based Embedding For Digital Twin). Figure 3 is an overview of the TED model. The manufacturing event log of the smart factory is injected into the model, and the output value is learned using a GRU neural network, an artificial intelligence model. Through the trained model, a digital twin that predicts the remaining cycle time in the process is created. By using the generated model, small factories of small and medium sized enterprises can also implement digital twins of the manufacturing process, helping bring factories closer to becoming a smart factories.
Figure 3. Transition system based embedding for digital twin modeling.
The data generated in the industrial process are collected through IoT sensors attached to the machine in the form of an event log including the activities and case IDs. The collected data are classified through the transition system (TS), and the TS layer describes all situations that can occur during the process. This transition relationship describes the transition from one state to another. It is expressed as a tree shaped graph explaining the direction. All transitions have start and end points, and the flow of these states can be expressed as a diagram as shown in Figure 4. The main purpose of this classification system is to understand the state of the process in detail and extract meaningful results that can improve operation of the plant based on the event log data.
Figure 4. Petri net of event log.
As described in Figure TED, this paper collects all the names of all machines operating during product production based on the event log in the manufacturing environment and the transition system, and then maps them in order. At this time, the mapped states are collected and expressed as Process = <1,2,3,4>. Based on the state information collected in the process cases, a relationship based on a Petri net can be constructed as shown in Figure 4.
Each node in Figure 4 represents the product creation process and status in the manufacturing process. Each node is a mapping of the machine name that the product was produced from. For example, 1,2 represents a state that went through Machine 1 and Machine 2, while 1,3,2 represents a state for Machine 1 and Machine 2 such that Machine 3 has been passed.
To determine the remaining cycle time of the manufacturing process through the transition system, the time remaining until the corresponding instance is terminated is estimated based on the process information for each tracked process. For example, if one process takes a total of 4 h, 3 h are left in the case of Machine 1, and 2 h are left in the case of Machine 2. When Machine 3 starts, the remaining processing time is 2 h. The remaining processing time can be calculated using the following method. The remaining process cycle time is first mapped, then the remaining time until the final process according to each state is predicted based on the record.
Using Table 1 as log data, we will explain this method by showing an example of estimating the remaining cycle time in the manufacturing process. For each case ID, the start time of the product is set to 0, and the manufacturing process time is normalized as shown in Figure TED. After calculating the end time based on the start time, the remaining time until the end of the process is tracked for each state.
Table 2 shows the normalized remaining cycle time in the manufacturing process using the data tracked in Table 1. The number at the top of the trace column represents the progress of the manufacturing process time according to the state. By designating the start time of the first machine in the tracked state as “0” as in the following table, the start time of all machines is normalized, and the remaining process time can be expressed. In Table 1, the time expressed as 3 July 2021: 13:00 is considered as 0, i.e., the start time of the process, so the start time of Machine 2, 3 July 2021: 13:20 is 0 minus the normalized start time. Here, 20 is the processing time of Machine 1. That is, the second machine started 20 min after the first machine was operated, and the process ends at 3 July 2021: 14:30, where the end time indicates the total processing time of the product. The end time of the first case ID in Table 2 is 90. That is, the difference between the end time of the last machine and the start time of the first machine is expressed in minutes (3 July 2021: 14:30–3 July 2021: 13:00) and calculated as follows.
Table 2. Normalization of the event log.
This paper creates a digital twin via artificial intelligence based on a data set classified using the transition system shown in Figure 3. For machine learning, the normalized data are tracked in the database according to the machine operation process situation, and the newly tracked data are output in a vector format. Each element of the vector is an integer indicating the operation of the machine, and the O1 Case ID in Table 2 is coded as in Table 3 below.
Table 3. Transformation of Database.
The columns in Table 3 represent the number of each machine for production, and each machine number is assigned 0 or 1 based on whether the machine is operated or not. To calculate the remaining time for each product to be produced, the progress time for each state calculated in the previous table is subtracted from the end time and divided by the number of products. The formula for this calculation is as follows:
Remaining Time S , P product = S End time − S Machine start time P product
Remaining Time ( < 1 > , 10 ) = 90 − 0 10 = 9.0 Remaining Time ( < 1 , 2 > , 10 ) = 90 − 20 10 = 7.0 Remaining Time ( < 1 , 2 , 3 > , 10 ) = 90 − 55 10 = 3.5 Remaining Time ( < 1 , 2 , 3 , 4 > , 10 ) = 90 − 80 10 = 1.0
Using the method explained through the following execution example, vectorization of all states of the event log is carried out to make a data set that can train an artificial intelligence neural network, embedding and retaining time as a Y vector in deep learning learning. Using the vectorized matrix according to tracking as an X vector, learning is carried out with the GRU artificial intelligence layer as shown in the TED picture. The next section describes the AI model used for learning in this paper.

3.3. Training Procedures

Since the model proposed in this paper is a digital twin based on regression analysis, we used the MAE (mean absolute error) as a performance indicator of the learning model and used a GRU model suitable for training ordered data. The TED model presented in this paper uses event log data, as shown in Figure TED, which is input into a transition system based embedding process and is trained on the GRU neural network layer. The residual cycle value of the manufacturing process according to the subsequent state is predicted using the GRU neural network training model proposed in [31].
The GRU model in Figure 5 is a type of RNN(Recurrent Neural Network). It is very similar to LSTM (long short term memory), a model that can effectively learn sequential data, but has a relatively simpler structure than LSTM.
Figure 5. Reset gate.
As Figure 6 shows, the purpose of the Reset Gate is to properly remove past information. Multiplying the previous hidden layer by a value between 0 and 1, the output value of the sigmoid function is used to create the next layer:
r ( t ) = σ W r h ( t − 1 ) + U r x ( t )
Figure 6. GRU model flow chart.
As Figure 7 shows, the update gate is a combination of a forget gate and input gate of LSTM. In the update gate, the resulting output of the sigmoid function determines the amount of information at the present time, and the value subtracted from 1 is multiplied by this information in the hidden layer of the previous layer. Each layer is similar to the input gate and forget gate of LSTM.
u ( t ) = σ W u h ( t − 1 ) + U u x ( t )
Figure 7. Update gate.
The candidate step shown in Figure 8 calculates the information candidates at the present time. The key is not to use information from previous hidden layers as it is, but to first multiply and use the results of the reset gate:
h ¯ ( t ) = τ W h ( t − 1 ) ∗ r ( t ) + U x ( t )
Figure 8. Candidate step.
Here, τ is a tangent hyperbolic and ∗ is a pointwise operation.
Finally, by synthesizing the update gate result and the candidate results, the hidden layer at the current time is calculated as described in Figure 9. As mentioned earlier, the result of the sigmoid function determines the amount of information at the present time, and the result of unity minus this sigmoid function determines the amount of information in the past:
h ( t ) = 1 − u ( t ) ∗ h ( t − 1 ) + u ( t ) ∗ h ¯ ( t )
Figure 9. Hidden layer calculation step.
The clear advantage of GRU compared to LSTM is that it uses an update gate that combines the forget gate and input gate of the LSTM, so that the learning weight is relatively small compared to the LSTM [31].

4. Experiments and Results

4.1. Experiment Environment

For configuration of the model proposed in this paper, an experiment was conducted using an Intel core i7, 64 GB RAM, and GeForce 1660TI environment. In addition, to create a deep learning network model, the Keras Python library was used for implementation.
Table 4 expresses the configuration parameters of the GRU neural network, and the experiment was conducted with epochs of 1000 and a batch size 40. AdaGrad was used as an optimization function for optimization, and a GRU neural network of 128 units was trained on the real world dataset.
Table 4. Description of Parameters.

4.2. Real World Dataset

This paper used event log data from real world process data. The datasets of the factory that the paper is analyzing can be accessed at [32]. This real world factory produces metal parts such as ball nuts and wheel shafts. With this production process, we have 28 machines for milling, wire cutting, laser marking, round grinding, flat grinding, turning, turning and milling, and lapping. Table 5 shows some of the event logs.
Table 5. Industrial event log of Case 106.
Table 5 shows 106 sets of cases from the data sets. The activity name in the table simply expresses the name of the actual process included in the log data [32]. Each case includes the type of machine used, the start and end times of activities, and the number of manufactured products. The quantity of products to be manufactured is also used as an important factor, and this variable distinguishes cases where the same product is produced by the same set of machines but with a different number of productions. In the absence of these data, bias may occur. These data describe the process of producing 31 products. We use the transition based embedding method proposed earlier to map the event log data into a vector format, and then proceed with modeling through artificial intelligence. After manufacturing residual cycle prediction modeling is completed, the model of the training dataset and the model of the testing dataset are compared to demonstrate the model being appropriate, where we implemented the MAE:
MAE = 1 n ∑ t = 1 n A t − F t
Figure 10 visually expresses the flow of the process at the manufacturing site, as well as how the activities in Table 5 proceed. This figure visually expresses the trace of case 106 described in the previous table, where this process starts from Machine 8 and proceeds along the blue dotted line. Each process is finally terminated with QC2. The following process is a relatively simple process, but since a smart factory process is very complex, it may be difficult to train an artificial intelligence neural network [19]. This is because it can interfere with the creation of a valid digital twin with repeated activities that interfere, making it difficult to understand. In this paper, to create a more meaningful digital twin model, in the case of a repeated process, the process was abstracted and considered as one activity.
Figure 10. Manufacturing process overview.

4.3. Model EvaluationResults

This section shows that the approach proposed in this paper can produce meaningful results. This paper used three AI algorithms for AI learning. The most basic models of MLP (multi layer perceptron), LSTM, and GRU were used, and learning was carried out with a hyper parameter with a batch size of 40, number of epochs of 1000, and learning rate of 0.001. As an evaluation index, an evaluation was conducted based on the aforementioned MAE. As a result, when comparing the training data model and the testing data model for all three models, we found that the MAE change in value according to the increase in epochs is similar, which shows that it is an accurate model that it is not overfitted to one side. In addition, all three models showed convergence close to 0.1 (hours).
Figure 11 shows the progression of training data used to predict the remaining cycle time of the manufacturing process using the MLP algorithm. The left side is a graph of the results after training using the training data, and the right side is a graph where the MAE is extracted by inputting the test data into the MLP model trained using the testing data. Both the left and right graphs show a relatively similar downward trend, and the fact that the model is not overfitted is confirmed by the shape of both graphs, indicating that the data are not too biased to one side. This graph using the MLP algorithm shows mild convergence after 400 epochs. As a result, it can be confirmed that the best MAE value is close to 0.14 (hours).
Figure 11. Comparison between train data and test data of the model using the MLP algorithm.
Figure 12 shows the progression of training data used to predict the remaining cycle time of the manufacturing process using the LSTM algorithm. As in Figure 11, the left side is after training using the training data, and the right side considers the MAE extracted by inputting the test data into the LSTM model trained using the testing data. These models also have a downward trend to some extent, similar for both the left and right graphs. Thus, it can be seen that it is a valid model by noting it is not overfitted to one side. In the case of the LSTM learning model, it can be seen that the MAE rapidly decelerates after 100 epochs. It can be seen from the testing graph on the right that the best MAE value converges to 0.12 (hours).
Figure 12. Comparison between train data and test data of the model using the LSTM algorithm.
Figure 13 shows the results of training to predict the remaining cycle time of the manufacturing process using the GRU algorithm. As in Figure 11 and Figure 12, the left side is after training using the training data, and the right side is obtained by extracting the MAE by inputting the test data into the GRU model trained using the testing data. The model also has a downward trend in both left and right graphs, demonstrating that the learning results are valid by noting there is no overfitting to one side too much. In the case of the GRU learning model, it can be seen that the MAE rapidly decelerates after 50 epochs, showing the best results among the three models. It can be seen from Figure 13 that the MAE value of the GRU learning model converges to 0.1 (hours).
Figure 13. Comparison between train data and test data of the model using the GRU algorithm.

4.4. Model Comparison

In this experiment using real world data, a comparative experiment was conducted considering three different artificial intelligence algorithms, MLP, LSTM, and GRU. The training was conducted in a GeForce 1660TI environment, and Figure 14 shows the results of comparing the models according to these three algorithms. With a batch size of 40, the execution time for each epoch is less than 9 s on average. For accurate comparison of the three models, the early stop function was not used, and all models were run with 1000 epochs.
Figure 14. Comparing training models.
As shown in Figure 14, as a result of model evaluation with the test data, the model with the fastest converging MAE values among the three AI algorithms was the GRU, which also had the lowest error value. The learning results of GRU are comparatively better starting from 100 epochs. This value is six times faster than for the MLP model, which produces the best results only after reaching 600 epochs, and four times faster than the LSTM, which exhibits meaningful results only after reaching 400 epochs. The difference in MAE between the GRU model and the LSTM model shows that the difference between the GRU model and the LSTM model is small, while the experimental results demonstrate that the GRU model is superior to LSTM in terms of the learning speed.
Although the difference in MAE between the three models is not large, it can be seen from the following graph that GRU is the most effective in predicting the number of remaining cycles in the manufacturing industry.

5. Conclusions

In this paper, to create a digital twin model to predict the remaining cycle time in manufacturing processes, the GRU model using transition based embedding was validated. Research on predicting processes using event log data are a common research topic recently, but as far as we know, this is the first time that event log data have been used for digital twin modeling alongside transition based embedding. The proposed method shows good results, and we conducted experiments using various AI approaches for comparison. The results of the experiments performed in this paper show the possibility of using event log data and transition based embedding methods for constructing digital twin models.
In keeping with the rapidly changing environment of the manufacturing industry, the need for predicting industrial processes is increasing, and interest in digital twins is also increasing accordingly. However, in the case of small and medium sized enterprises (SMEs), which typically exhibit a shortage of data sets as well as insufficient manpower and consulting costs, there are many factors to consider. The digital twin model using the transition based embedding method proposed in this paper can help more users cooperate via digital twin technology. In addition, partners using the technology proposed in this paper can test manufacturing environments at a lower cost, use more data, and test various situations, leading to more meaningful results.
In this paper, the GRU algorithm is used for digital twin modeling. As a future study, we plan to upgrade this model to produce more precise results, increase their reliability, and apply the model to other industrial domains.

Author Contributions

Conceptualization, M.Y. and J.J.; methodology, J.M.; software, M.Y. and J.M.; validation, M.Y. and J.J.; formal analysis, M.Y.; investigation, M.Y. and J.M.; S.S. and J.K.; resources, J.J.; data curation, M.Y.; writing—original draft preparation, M.Y. and J.M.; writing—review and editing, J.J.; visualization, M.Y.; supervision, J.J. project administration, J.J. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2021R1F1A1060054).

Institutional Review Board Statement

Not applicable.

Data Availability Statement

Not applicable.

Acknowledgments

This research was supported by the MSIT(Ministry of Science and ICT), Korea, under the ICT Creative Consilience Program(IITP-2021-2020-0-01821) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation).

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Wang, X.; Wang, Y.; Tao, F.; Liu, A. New paradigm of data-driven smart customisation through digital twin. J. Manuf. Syst. 2021, 58, 270–280. [Google Scholar] [CrossRef]
  2. Tao, F.; Qi, Q.; Liu, A.; Kusiak, A. Data-driven smart manufacturing. J. Manuf. Syst. 2018, 48, 157–169. [Google Scholar] [CrossRef]
  3. Söderberg, R.; Wärmefjord, K.; Carlson, J.S.; Lindkvist, L. Toward a Digital Twin for real-time geometry assurance in individualized production. CIRP Ann. 2017, 66, 137–140. [Google Scholar] [CrossRef]
  4. Zhang, C.; Chen, D.; Tao, F.; Liu, A. Data driven smart customization. Procedia CIRP 2019, 81, 564–569. [Google Scholar] [CrossRef]
  5. Saldivar, A.A.F.; Li, Y.; Chen, W.-n.; Zhan, Z.-h.; Zhang, J.; Chen, L.Y. Industry 4.0 with cyber-physical integration: A design and manufacture perspective. In Proceedings of the 2015 21st International Conference on Automation and Computing (ICAC), Glasgow, UK, 11–12 September 2015; pp. 1–6. [Google Scholar]
  6. Singh, V.; Willcox, K.E. Engineering design with digital thread. AIAA J. 2018, 56, 4515–4528. [Google Scholar] [CrossRef]
  7. Tuegel, E.J.; Ingraffea, A.R.; Eason, T.G.; Spottswood, S.M. Reengineering aircraft structural life prediction using a digital twin. Int. J. Aerosp. Eng. 2011, 2011, 154798. [Google Scholar] [CrossRef] [Green Version]
  8. Bevilacqua, M.; Bottani, E.; Ciarapica, F.E.; Costantino, F.; Di Donato, L.; Ferraro, A.; Mazzuto, G.; Monteriù, A.; Nardini, G.; Ortenzi, M.; et al. Digital twin reference model development to prevent operators’ risk in process plants. Sustainability 2020, 12, 1088. [Google Scholar] [CrossRef] [Green Version]
  9. Luo, W.; Hu, T.; Zhang, C.; Wei, Y. Digital twin for CNC machine tool: Modeling and using strategy. J. Ambient. Intell. Humaniz. Comput. 2019, 10, 1129–1140. [Google Scholar] [CrossRef]
  10. Ghosh, A.K.; Ullah, A.; Kubo, A.; Akamatsu, T.; D’Addona, D.M. Machining phenomenon twin construction for industry 4.0: A case of surface roughness. J. Manuf. Mater. Process. 2020, 4, 11. [Google Scholar] [CrossRef] [Green Version]
  11. Aheleroff, S.; Xu, X.; Zhong, R.Y.; Lu, Y. Digital twin as a service (DTaaS) in industry 4.0: An architecture reference model. Adv. Eng. Inform. 2021, 47, 101225. [Google Scholar] [CrossRef]
  12. Zhou, H.; Yang, C.; Sun, Y. Intelligent Ironmaking Optimization Service on a Cloud Computing Platform by Digital Twin. Engineering 2021, 7, 1274–1281. [Google Scholar] [CrossRef]
  13. Chen, J.; Hu, K.; Wang, Q.; Sun, Y.; Shi, Z.; He, S. Narrowband internet of things: Implementations and applications. IEEE Internet Things J. 2017, 4, 2309–2314. [Google Scholar] [CrossRef]
  14. Lugaresi, G.; Matta, A. Real-time simulation in manufacturing systems: Challenges and research directions. In Proceedings of the 2018 Winter Simulation Conference (WSC), Gothenburg, Sweden, 9–12 December 2018; pp. 3319–3330. [Google Scholar]
  15. van der Aalst, W.M. Process mining and simulation: A match made in heaven! In Proceedings of the SummerSim, Bordeaux, France, 9–12 July 2018; pp. 1–4. [Google Scholar]
  16. Bergmann, S.; Feldkamp, N.; Strassburger, S. Approximation of dispatching rules for manufacturing simulation using data mining methods. In Proceedings of the 2015 Winter Simulation Conference (WSC), Huntington Beach, CA, USA, 6–9 December 2015; pp. 2329–2340. [Google Scholar]
  17. Ferreira, D.R.; Vasilyev, E. Using logical decision trees to discover the cause of process delays from event logs. Comput. Ind. 2015, 70, 194–207. [Google Scholar] [CrossRef]
  18. Moon, J.; Park, G.; Jeong, J. POP-ON: Prediction of Process Using One-Way Language Model Based on NLP Approach. Appl. Sci. 2021, 11, 864. [Google Scholar] [CrossRef]
  19. Lugaresi, G.; Matta, A. Automated manufacturing system discovery and digital twin generation. J. Manuf. Syst. 2021, 59, 51–66. [Google Scholar] [CrossRef]
  20. Pan, Y.; Zhang, L. A BIM-data mining integrated digital twin framework for advanced project management. Autom. Constr. 2021, 124, 103564. [Google Scholar] [CrossRef]
  21. Tran, T.a.; Ruppert, T.; Eigner, G.; Abonyi, J. Real-time locating system and digital twin in Lean 4.0. In Proceedings of the 2021 IEEE 15th International Symposium on Applied Computational Intelligence and Informatics (SACI), Timisoara, Romania, 19–21 May 2021; pp. 000369–000374. [Google Scholar]
  22. Lugaresi, G.; Zanotti, M.; Tarasconi, D.; Matta, A. Manufacturing Systems Mining: Generation of Real-Time Discrete Event Simulation Models. In Proceedings of the 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC), Bari, Italy, 6–9 October 2019; pp. 415–420. [Google Scholar]
  23. Ghahramani, M.; Qiao, Y.; Zhou, M.; Hagan, A.O.; Sweeney, J. AI-based modeling and data-driven evaluation for smart manufacturing processes. IEEE/CAA J. Autom. Sin. 2020, 7, 1026–1037. [Google Scholar] [CrossRef]
  24. Lee, J.; Singh, J.; Azamfar, M.; Pandhare, V. Industrial AI and predictive analytics for smart manufacturing systems. In Smart Manufacturing; Elsevier: Amsterdam, The Netherlands, 2020; pp. 213–244. [Google Scholar]
  25. Ding, H.; Gao, R.X.; Isaksson, A.J.; Landers, R.G.; Parisini, T.; Yuan, Y. State of AI-based monitoring in smart manufacturing and introduction to focused section. IEEE/ASME Trans. Mechatron. 2020, 25, 2143–2154. [Google Scholar] [CrossRef]
  26. Kotsiopoulos, T.; Sarigiannidis, P.; Ioannidis, D.; Tzovaras, D. Machine Learning and Deep Learning in Smart Manufacturing: The Smart Grid Paradigm. Comput. Sci. Rev. 2021, 40, 100341. [Google Scholar] [CrossRef]
  27. Huang, Z.; Shen, Y.; Li, J.; Fey, M.; Brecher, C. A Survey on AI-Driven Digital Twins in Industry 4.0: Smart Manufacturing and Advanced Robotics. Sensors 2021, 21, 6340. [Google Scholar] [CrossRef]
  28. Mostafa, F.; Tao, L.; Yu, W. An effective architecture of digital twin system to support human decision-making and AI-driven autonomy. Concurr. Comput. Pract. Exp. 2021, 33, e6111. [Google Scholar] [CrossRef]
  29. Örs, E.; Schmidt, R.; Mighani, M.; Shalaby, M. A Conceptual Framework for AI-based Operational Digital Twin in Chemical Process Engineering. In Proceedings of the 2020 IEEE International Conference on Engineering, Technology and Innovation (ICE/ITMC), Cardiff, UK, 15–17 June 2020; pp. 1–8. [Google Scholar]
  30. Rojek, I.; Mikołajewski, D.; Dostatni, E. Digital Twins in Product Lifecycle for Sustainability in Manufacturing and Maintenance. Appl. Sci. 2021, 11, 31. [Google Scholar] [CrossRef]
  31. Cho, K.; Van Merriënboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; Bengio, Y. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv 2014, arXiv:1406.1078. [Google Scholar]
  32. Tello-Leal, E.; Roa, J.; Rubiolo, M.; Ramirez-Alcocer, U.M. Predicting activities in business processes with LSTM recurrent neural networks. In Proceedings of the 2018 ITU Kaleidoscope: Machine Learning for a 5G Future (ITU K), Santa Fe, Argentina, 26–28 November 2018; pp. 1–7. [Google Scholar]
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Article Metrics

Citations

Article Access Statistics

Journal Statistics
Multiple requests from the same IP address are counted as one view.
Properties of CAD/CAM 3D Printing Dental Materials and Their Clinical Applications in Orthodontics: Where Are We Now?
Previous
Evidence for Human-Centric In-Vehicle Lighting: Part 1
Next