臺灣博碩士論文加值系統

程序中的質量變量，是在一個運行批次中需要調節和提高的關鍵因素。然而，質量變量是一個難以線上監控的變數。軟測量提供了即時洞察過程之質量的直接替代方案，但仍然存在的問題包括過程與質量測量之間的不平衡、測量雜訊以及複雜的三維動態批次資料結構。為了應對這些挑戰，第2章開發了魯棒半監督雙注意力潛在動態條件狀態空間模型 (Robust Semi-Supervised Dual Attention Latent Dynamic Conditional State Space Model, RS2DA-LDCSSM) 用於批次內的質量預測。為了防止資料展開過程中資訊的失真，採用注意力序列到序列RNN編碼器-解碼器 (Attentional Sequence-to-Sequence RNN Encoder-Decoder, AS2S-RNNED) 處理三維批次數據。RS2DA-LDCSSM將AS2S-RNNED與概率狀態空間模型結合，以過濾雜訊過程資料並穩定質量資料的概率預測，從過去的過程和質量數據中提取空間和時間潛在資料，同時儘量減少多步預測的損失。
為了確保模型的可靠性，將把不完全的一階原理模型（First-principle Models, FPMs）的物理見解與概率資料驅動序列到序列（Sequence-to-sequence, S2S）模型集成。由於工業複雜性， FPMs本質上是不完整的，收集數據的過程中還會受到過程不確定性和缺失資料的影響。因此，第3章提出網路物理半監督雙注意力潛在動態條件狀態空間模型（Cyber-Physics Semi-Supervised Dual Attention Latent Conditional State Space Model, CPS2DA-LDCSSM）。通過協調不完整的FPM與S2S模型，以及採用網路物理模型（Cyber-physics model, CPM）處理缺失的質量數據。CPS2DA-LDCSSM利用條件狀態空間和時間注意力序列到序列網路，從三維批次數據視窗中提取動態潛在特徵並減少過程不確定性。
儘管數據驅動的機器學習模型構建容易，但在處理稀疏質量數據時，其可靠性和可解釋性會受到影響。FPM提供了可解釋性，但其在應用於預測質量問題時面臨挑戰，因為其性能高度依賴於參數，導致其對條件變換的適應性較差。為了克服這些模型特徵造成的局限性，第4章提出了一種三階段強化網路物理集成學習方法（Reinforced Cyber-Physics Ensemble Learning, RCPEL），可溝通不同操作條件下基於FPM的網路物理模型，來增強線上質量預測適應性。
在工業製造過程中，完整的FPM結構或準確的FPM參數由於過程的複雜性難以獲得。因為FPM之性能嚴重依賴於結構，導致對變化條件的適應性較差，為了解決FPM在準確地預測質量時表現欠佳的問題，第5章提出了一種三階段強化不完全網路物理集成加誤差補償學習（Reinforced Incomplete Cyber-Physics Ensemble Plus Error Compensation Learning, RICPE-P-ECL）方法。該方法提高了基於不完全網路物理模型（Incomplete Cyber Physics Model, IncompCPM）在變化操作條件下的線上質量預測適應性。RICPE-P-ECL的主要創新在於其集成設計和誤差補償策略。
為了評估模型的可靠性和預測準確性，第2章至第5章分別通過半批次青黴素發酵案例 (Semi-batch Penicillin fermentation process) 研究展示了RS2DA-LDCSSM、CPS2DA-LDCSSM、RCPEL和RICPE-P-ECL的性能，並在每章末尾與傳統數據驅動模型進行比較，來呈現本論文提出的方法優勢。

The pivotal factor to regulate and enhance within an operating batch is the process quality, a challenging variable to monitor online. Soft sensors offer an immediate alternative for providing real-time insights into process quality, yet persisting issues include the imbalance between process and quality measurements, noisy measurements, and the intricate 3-dimensional dynamic batch data structure. The Robust Semi-Supervised Dual-Attention Latent Dynamic Conditional State-Space Model (RS2DA-LDCSSM) is developed in Chapter 2 to address these challenges for within-batch quality prediction. Given the frequent absence of quality data due to measurement inconveniences, an imputation network is embedded within the RS2DA-LDCSSM to facilitate the prediction of future quality. To prevent information distortion during data unfolding, the Attentional Sequence-to-Sequence RNN Encoder-Decoder (AS2S-RNNED) is employed to process the 3-dimensional batch data. The proposed method integrates AS2S-RNNED with a probability state-space model to filter out noisy process data and stabilize the probability prediction of quality data, extracting spatial and temporal latent data from past process and quality data while minimizing the loss in multi-step predictions.
To ensure model reliability, integration of physical insights from incomplete first-principle models (FPMs) with a probabilistic data-driven sequence-to-sequence (S2S) model is considered. Industrial FPMs are inherently incomplete due to complexity, compounded by process uncertainties and missing data during collection. Our proposed solution, the cyber-physics semi-supervised dual attention latent dynamic conditional state-space model (CPS2DA-LDCSSM), addresses these challenges in Chapter 3. By coordinating the incomplete FPM with the S2S model, a cyber-physics model (CPM) is employed to handle missing quality data. The CPS2DA-LDCSSM utilizes quality data from CPM to apply spatial and temporal attention sequence-to-sequence networks for dynamic latent feature extraction and process uncertainty reduction using 3-dimensional batch data windows.
While data-driven machine learning models are easily constructed, their reliability and interpretability are compromised when dealing with sparse quality data. First-principle models (FPM) offer interpretability but face challenges in solving quality solutions accurately, as their performance is highly dependent on parameters, resulting in poor adaptability to changing conditions. To overcome these limitations, a three-phase reinforced cyber-physics ensemble learning method (RCPEL) is proposed in Chapter 4. It can enhance the adaptability of the cyber-physics model (CPM) built upon FPM for online quality prediction under different operating conditions.
In practical application on industrial manufacturing processes, the complete FPM structure or the accurate FPM parameter are hard to be obtained due to the complexity of the process. To address the problem where the FPMs struggle with accurately solving quality solutions, as their performances heavily depend on structure, leading to poor adaptability to changing conditions, a three-phase reinforced incomplete cyber-physics ensemble plus error compensation learning (RICPE-P-ECL) method is proposed in Chapter 5. This method enhances the adaptability of the incomplete cyber-physics model (IncompCPM), which is based on partially-available FPMs, for online quality prediction under varying operating conditions. The main innovation in RICPE-P-ECL lies in its ensemble design and error compensation strategy.
To evaluate the reliability and prediction accuracy, the proposed RS2DA-LDCSSM, CPS2DA-LDCSSM, RCPEL and RICPE-P-ECL are each presented by a semi-batch penicillin fermentation case study. Their performances are also compared to traditional data-driven models at the end of each chapter.

摘要 I
Abstract III
Acknowledgement V
Table of contents VI
List of tables IX
List of figures X

Chapter 1. Introduction 1
1 Significance of the research 1
2 Literature review of data-driven soft-sensors 2
3 Literature review of physical knowledge models 4
4 Literature review of ensemble learning 5
5 Literature review of optimization methods 6
6 Research motivations 8
7 Dissertation outline 9

Chapter 2. A robust semi-supervised learning scheme for development of within-batch quality prediction soft-sensors 14
1 Research motivations and problem formulations 14
1.1 Research motivations 14
1.2 Problem formulations 15
2 Methodology of RS2DA-LDCSSM 22
2.1 PQ-networks 22
2.2 Satt-networks 23
2.3 Tatt-EDs 26
2.4 Emission network 32
3 Results and discussion 36
3.1 Numerical case 36
3.2 Industrial penicillin fermentation process 50
4 Conclusions of this chapter 55

Chapter 3. Enhancing within batch quality prediction by cyber-physical latent state models and incomplete first-principles 56
1 Research motivations, background knowledge and problem formulations 56
1.1 Research motivations 56
1.2 Background knowledge of physical informed neural network (PINN) 57
1.3 Problem formulation 61
2 Methodology of CPS2DA-LDCSSM 63
2.1 PQ-CPMs 63
2.2 Satt-networks 64
2.3 Tatt-EDs 66
2.4 Emission network 67
3 Penicillin fermentation case study 71
4 Conclusions of this chapter 77


Chapter 4. Boosting quality prediction with reinforced cyber-physical ensemble learning: A case study of fed-batch bioprocess 78
1 Research motivations 78
2 Process description and data collection strategy 79
2.1 Fed-batch bioprocess 79
2.2 Data collection and organization 81
3 Methodology of RCPEL 81
3.1 Phase 1: Constructing FPM parameter supply agent with actor-critic (AC) 85
3.2 Phase 2: Learning the CPM to immediately get quality from the process data 90
3.3 Phase 3: Constructing ensemble weight supply agent with RL 92
4 Case demonstration 97
5 Conclusions of this chapter 107




Chapter 5. A novel reinforced incomplete cyber-physics ensemble plus error compensation learning for within-batch quality prediction 109
1 Research motivations 109
2 Data collection strategy 110
3 Methodology of RICPE-PECL 112
3.1 Phase 1: Constructing CPMs with incomplete FPMs 113
3.2 Phase 2: Constructing ensemble weight supply agent with RL 114
3.3 Phase 3: Constructing error compensation agent with RL 123
4 Results and discussions 133
4.1 Case study 1 137
4.2 Case study 2 139
4.3 Case study 3 142
5 Conclusions of this chapter 144


Chapter 6. Conclusions and future directions 146
1 Concluding remarks 146
2 Recommendations for future work 147


References 148

Appendices 154

Table 1 1. Summary of the problems in past researches and their corresponding model performances 8
Table 2 1. R2 indices of the comparison models corresponding to the noisy and real data in Scenario 1, Scenario 2, and Scenario 3 44
Table 2-2. RMSE indices for the comparison models of the noisy and real data in Scenario 1, Scenario 2, and Scenario 3 44
Table 2 3. R2 and RMSE indices of the comparison models in Scenario 1, which include both noisy and real data with high and low SNRs 47
Table 2-4. The detail of the measured process variables in IndPenSim 50
Table 2-5. RMSE and R2 values of all the comparison models. 54
Table 3 1. The prediction indices of the comparative methods corresponding to the measured data and true data 76
Table 4-1. Network configurations and training hyperparameters for Phase 1 RCPEL 100
Table 4 2. Network configurations and training hyperparameters of the trained RNN, DNN, and CPM 102
Table 4 3. Prediction performance of the RNN, DNN, and CPM under different scenarios 103
Table 4 4. Network configurations and training hyperparameters for Phase 3 RCPEL 104

Fig. 1 1. Innovation framework of this thesis 10
Fig. 2 1. Rearrangement of 3-dimensional batch data into 2-dimensional overlapping window data with batch independency. 11
Fig. 2 2. Process and quality inferencing mechanism in CPS2DA-LDCSSM 17
Fig. 2 3. The pathway to learning RS2DA-LDCSSM. 20
Fig. 2 4. The illustration of (a) the PQ-Post-Network and (b) the PQ-Prior-Network. 22
Fig. 2 5. Illustration of the network structure of (a) PV-Satt-Post-Network, (b) PV-Satt-Prior-Network, (c) QV-Satt-Post-Network, and (d) QV-Satt-Prior-Network. The blue rectangle indicates the spatial attention mechanism; pink nodes represent the forward-backward RNN; blue nodes represent the DNNs, which are used to map the RNN states to the latent spaces. 24
Fig. 2 6. The network structure of (a) PV-Tatt-Post-ED, (b) PV-Tatt-Prior-ED, (c) QV-Tatt-Post-ED, and (d) QV-Tatt-Prior-ED. The green dashed rectangle indicates the temporal attention mechanism; pink nodes represent the encoder-decoder RNNs; blue nodes represent the DNNs, which map the RNN states of the decoder to obtain the transition 7posterior and prior distributions. 28
Fig. 2 7. Illustration of the emission network. 34
Fig. 2 8. Prediction results of the comparison methods: (a) RS2DA-LDCSSM, (b) DA-LDCSSM, (c) DNNCSS, (d) ED, (e) RNN, (f) DNN, and (g) SAE, under Scenario 1. 40
Fig. 2 9. Prediction results of the comparison methods (a) RS2DA-LDCSSM, (b) DA-LDCSSM, (c) DNNCSS, (d) ED, (e) RNN, (f) DNN, and (g) SAE, under Scenario 2. 41
Fig. 2 10. Prediction results of the comparison methods (a) RS2DA-LDCSSM, (b) DA-LDCSSM, (c) DNNCSS, (d) ED, (e) RNN, (f) DNN, and (g) SAE, under Scenario 3. 43
Fig. 2 11. The degree of degradation of the comparison models in terms of (a) R2 and (b) MSE in the three scenarios. 46
Fig. 2 12. Prediction results of the comparison methods (a) RS2DA-LDCSSM, (b) DA-LDCSSM, (c) DNNCSS, (d) ED, (e) RNN, (f) DNN, (g) SAE, using low SNR. 48
Fig. 2 13. The schematic graph of the penicillin fermentation process (IndPenSim). 49
Fig. 2 14. The data trend of the 5 over the 100 online available batches of the penicillin fermentation process, (a) the aeration rate, (b) the sugar feed rate, (c) the pH value, (d) the temperature and the quality variable, and (e) offline penicillin concentration. 51
Fig. 2 15. The measured quality and the quality predicted by soft sensors in (a) RS2DA-LDCSSM, (b) DA-LDCSSM, (c) ED, (d) DNNCSS, (e) RNN, (f) SAE and (g) DNN, respectively. 54
Fig. 3 1. The model structure of PINN is equivalent to the integration of the first principle model and the ODE solver. 58
Fig. 3 2. Process and quality inferencing mechanism in CPS2DA-LDCSSM 60
Fig. 3 3. The pathway to learning the CPS2DA-LDCSSM. 62
Fig. 3 4. Illustration of the network structure of (a) PV-Satt-Post-Network, (b) PV-Satt-Prior-Network. The blue dotted dash box indicates the spatial attention mechanism; pink nodes represent the forward-backward RNN; blue nodes represent the DNNs, which are used to map the RNN states to the latent spaces. 65
Fig. 3 5. The network structure of (a) PV-Tatt-Post-ED and (b) PV-Tatt-Prior-ED. The green dashed rectangle indicates the temporal attention mechanism; pink nodes represent the encoder-decoder RNNs; blue nodes represent the DNNs, which map the RNN states of the decoder to obtain the transition posterior and prior distributions. 67
Fig. 3 6. Illustration of the emission network. 68
Fig. 3 7. (a) Training batches and (b) test batches collected from the simulation. 73
Fig. 3 8. Comparison between the real (green) and disturbance-contaminated measurements (blue) in the 1st training batch. 74
Fig. 3 9. Prediction results of the comparison methods for the 1st testing batch: (a) PINN (b) DA-LDCSSM (c) S2DA-LDCSSM, and (d) CPS2DA-LDCSSM. 75
Fig. 4 1. Illustration of the fed-batch bioprocess. 79
Fig. 4 2. The framework of the proposed RCPEL method. 83
Fig. 4 3. (a) The CPM models constructed based on the operating conditions; (b) Process and quality data trends during online batch operation. 84
Fig. 4 4. Graphical learning strategy of the RL problem for the FPM parameter supply agent in Phase 1. 87
Fig. 4 5. The model structure of CPM for condition . The neural networks (in the upper right side) is equivalent to the integration of the first principle model and the ODE solver (in the bottom). 91
Fig. 4 6. Graphical learning strategy of the RL problem for ensemble weight parameter supply agent in Phase 3. 93
Fig. 4 7. The (a) biomass (b) product and (c) substrate concentration difference between the 10 operating conditions (conditions 1~10 with solid lines) and the 5 conditions (conditions 11 ~15 with dashed lines) for testing the RNN, S2-LDVAE, and RCPEL. 98
Fig. 4 8. The RMSE (left) and R2 (right) values of the 5 testing conditions with the FPM parameter supply agent in Phase 1. 101
Fig. 4 9. The RMSE (left) and R2 (right) values of the 5 testing conditions with the FPM parameter supply agent in Phase 2. 103
Fig. 4 10. The RMSE (left) and R2 (right) values of (a) the proposed Phase 3 RCPEL, (b) RNN, and (c) S2-LDVAE. 106
Fig. 5 1. Entire framework of the proposed RICPE-P-ECL method 112
Fig. 5 2. (a) The IncompCPMs constructed based on the historical operating conditions; (b) Process and quality data trends during online batch operation. 113
Fig. 5 3. The model structure of CPM for condition . The neural network (in the upper right side) is equivalent to the integration of the first principle model and the ODE solver (in the bottom). 114
Fig. 5 4. Method of ensemble the historical operated CPMs (upper part) and the graphical learning strategy of the RL problem for the EWSA in Phase 2 (lower part). 117
Fig. 5 5. Batch data rearrangement to retain dynamic information while constructing the EWSA. 118
Fig. 5 6. The model structure of the (a) actors and (b) critics in the ESWA. 121
Fig. 5 7. Graphical learning strategy of the RL problem for error compensation agent (ECA) in Phase 3. 124
Fig. 5 8. The model structure of the (a) actors and (b) critics in the ECA. 128
Fig. 5 9. Illustration of the fed-batch bioprocess. 133
Fig. 5 10. The quality trends of (a) biomass, (b) product, and (c) substrate concentrations for different operating conditions. 136
Fig. 5 11. The RMSE (left) and R2 (right) of the (a) CPM based on complete FPM, (b) CPM based on incomplete FPM, (c) DNN and (d) RNN corresponding to the quality variables in the 5 operating conditions 139
Fig. 5 12. The RMSE (left) and R2 (right) value of the biomass (X), product (P) and substrate (S) concentration for each constructed CPMs with respect to the testing conditions. 139
Fig. 5 13. Learning curve of the reward for the EWSA. 140
Fig. 5 14. The RMSE (left) and R2 (right) of RICPEL of the 3 quality variables with respect to the new testing conditions for (a) S2-LDVAE, (b) S2DA-LDCSSM, (c) RICPEL and (d) RICPE-P-ECL. 142
Fig. 5 15. The learning curve of the reward for the error compensation agent. 143
Fig. 6 1. Concluding flow of this thesis. 146

[1]P. N. Sharratt, “Chemicals manufacture by batch processes,” in Handbook of Batch Process Design, 1997, pp. 1–23. doi: 10.1007/978-94-009-1455-1_1.
[2]M. A. Henson, “Biochemical reactor modeling and control,” IEEE Control Syst., vol. 26, no. 4, pp. 54–62, 2006, doi: 10.1109/mcs.2006.1657876.
[3]Z. Ge, “Process data analytics via probabilistic latent variable models: A tutorial review,” Ind. Eng. Chem. Res., vol. 57, no. 38, pp. 12646–12661, 2018, doi: 10.1021/acs.iecr.8b02913.
[4]T. Kourti, P. Nomikos, and J. F. MacGregor, “Analysis, monitoring and fault diagnosis of batch processes using multiblock and multiway PLS,” J. Process Control, vol. 5, no. 4, pp. 277–284, 1995, doi: 10.1016/0959-1524(95)00019-M.
[5]C. Undey and A. Cinar, “Statistical monitoring of multistage, multiphase batch processes,” IEEE Control Syst., vol. 22, no. 5, pp. 40–52, 2002, doi: 10.1109/MCS.2002.1035216.
[6]G. Tomasi, F. Van Den Berg, and C. Andersson, “Correlation optimized warping and dynamic time warping as preprocessing methods for chromatographic data,” J. Chemom., vol. 18, no. 5, pp. 231–241, 2004, doi: 10.1002/cem.859.
[7]M. Jia, F. Chu, F. Wang, and W. Wang, “On-line batch process monitoring using batch dynamic kernel principal component analysis,” Chemom. Intell. Lab. Syst., vol. 101, no. 2, pp. 110–122, 2010, doi: 10.1016/j.chemolab.2010.02.004.
[8]J. M. Lee, C. K. Yoo, and I. B. Lee, “Fault detection of batch processes using multiway kernel principal component analysis,” Comput. Chem. Eng., vol. 28, no. 9, pp. 1837–1847, 2004, doi: 10.1016/j.compchemeng.2004.02.036.
[9]X. Xiao, G.Q. Liu, B.F. Hu, X. Zheng, L.N. Wang, S.J. Chen, A. Ullah, “A comparative study on Arrhenius-type constitutive equations and artificial neural network model to predict high-temperature deformation behaviour in 12Cr3WV steel,” Comput. Mater. Sci., vol. 62, pp. 227–234, 2012, doi: 10.1016/j.commatsci.2012.05.053.
[10]X.Y. Huang, H. Wang, W.H. Xue, A. Ullah, S. Xiang, H.L. Huang, L. Meng e, G. Ma, G.Z. Zhang, “A combined machine learning model for the prediction of time-temperature-transformation diagrams of high-alloy steels,” J. Alloys Compd., vol. 823, p. 153694, 2020, doi: 10.1016/j.jallcom.2020.153694.
[11]J. Wang, A. S. Mohammed, E. Macioszek, M. Ali, D. V. Ulrikh, and Q. Fang, “A novel combination of PCA and machine learning techniques to select the most important factors for predicting tunnel construction performance,” Buildings, vol. 12, no. 7, 2022, doi: 10.3390/buildings12070919.
[12]M. Ali and S. Hin Lai, “Artificial intelligent techniques for prediction of rock strength and deformation properties – A review,” Structures, vol. 55, no. January, pp. 1542–1555, 2023, doi: 10.1016/j.istruc.2023.06.131.
[13]K. Wang, Z. Guo, Y. Wang, X. Yuan, and C. Yang, “Common and specific deep feature representation for multimode process monitoring using a novel variable-wise weighted parallel network,” Eng. Appl. Artif. Intell., vol. 104, no. 104381, pp. 1–10, 2021, doi: 10.1016/j.engappai.2021.104381.
[14]S. R. Vijaya Raghavan, T. K. Radhakrishnan, and K. Srinivasan, “Soft sensor based composition estimation and controller design for an ideal reactive distillation column,” ISA Trans., vol. 50, no. 1, pp. 61–70, 2011, doi: 10.1016/j.isatra.2010.09.001.
[15]Y. Tang, Y. Wang, C. Liu, X. Yuan, K. Wang, and C. Yang, “Semi-supervised LSTM with historical feature fusion attention for temporal sequence dynamic modeling in industrial processes,” Eng. Appl. Artif. Intell., vol. 117, no. 105547, pp. 1–11, 2023, doi: 10.1016/j.engappai.2022.105547.
[16]X. Yuan, L. Li, and Y. Wang, “Nonlinear dynamic soft sensor modeling with supervised long short-term memory network,” IEEE Trans. Ind. Informatics, vol. 16, no. 5, pp. 3168–3176, 2020, doi: 10.1109/TII.2019.2902129.
[17]K. Wang, X. Yuan, J. Chen, and Y. Wang, “Supervised and semi-supervised probabilistic learning with deep neural networks for concurrent process-quality monitoring,” Neural Networks, vol. 136, pp. 54–62, 2021, doi: 10.1016/j.neunet.2020.11.006.
[18]X. Zhang and Z. Ge, “Latent adversarial regularized autoencoder for high-dimensional probabilistic time series prediction,” Neural Networks, vol. 155, pp. 383–397, 2022, doi: 10.1016/j.neunet.2022.08.025.
[19]S. Lin, R. Clark, R. Birke, and S. Sch, “Anomaly detection for time series using VAE-LSTM hybrid model,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2020), Barcelona, 2020, pp. 4322–4326.
[20]Y. S. Lee, S. K. Ooi, D. Tanny, and J. Chen, “Developing soft-sensor models using latent dynamic variational autoencoders,” 2021.
[21]Y. S. Lee and J. Chen, “Automatic Deep Extraction of Robust Dynamic Features for Industrial Big Data Modeling and Soft Sensor Application,” Neural Networks, vol. 154, pp. 455–468, 2022, doi: 10.1016/j.neunet.2022.06.010.
[22]K. Sohn, X. Yan, and H. Lee, “Learning structured output representation using deep conditional generative models,” Adv. Neural Inf. Process. Syst., vol. 2015-Janua, pp. 3483–3491, 2015.
[23]D. Bahdanau, K. H. Cho, and Y. Bengio, “Neural machine translation by jointly learning to align and translate,” in 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, SanDiego, 2015, pp. 1–15.
[24]X. Zhang and Z. Ge, “Automatic Deep Extraction of Robust Dynamic Features for Industrial Big Data Modeling and Soft Sensor Application,” IEEE Trans. Ind. Informatics, vol. 16, no. 7, pp. 4456–4467, 2020, doi: 10.1109/TII.2019.2945411.
[25]P. Xie, M. Gao, H. Zhang, Y. Niu, and X. Wang, “Dynamic modeling for NOx emission sequence prediction of SCR system outlet based on sequence to sequence long short-term memory network,” Energy, vol. 190, p. 116482, 2020, doi: 10.1016/j.energy.2019.116482.
[26]L. Feng, C. Zhao, and Y. Sun, “Dual attention-based encoder-decoder: A customized sequence-to-sequence learning for soft sensor development,” IEEE Trans. Neural Networks Learn. Syst., pp. 1–12, 2020, doi: 10.1109/tnnls.2020.3015929.
[27]X. Yuan, L. Li, Y. A. W. Shardt, Y. Wang, and C. Yang, “Deep learning with spatiotemporal attention-based LSTM for industrial soft sensor model development,” IEEE Trans. Ind. Electron., vol. 68, no. 5, pp. 4404–4414, 2021, doi: 10.1109/TIE.2020.2984443.
[28]Z. Ge, B. Huang, and Z. Song, “Nonlinear semisupervised principal component regression for soft sensor modeling and its mixture form,” J. Chemom., vol. 28, no. 11, pp. 793–804, 2014, doi: 10.1002/cem.2638.
[29]Z. Ge and Z. Song, “Semisupervised bayesian method for soft sensor modeling with unlabeled data samples,” AIChE J., vol. 57, no. 8, pp. 2109–2119, 2010, doi: https://doi.org/10.1002/aic.12422.
[30]Y. Tang, Y. Wang, C. Liu, X. Yuan, K. Wang, and C. Yang, “Semi-supervised LSTM with historical feature fusion attention for temporal sequence dynamic modeling in industrial processes,” Eng. Appl. Artif. Intell., vol. 117, no. September 2022, p. 105547, 2023, doi: 10.1016/j.engappai.2022.105547.
[31]G. D. Zhang, X. He, and X. Yang, “A fully decoupled linearized finite element method with second-order temporal accuracy and unconditional energy stability for incompressible MHD equations,” J. Comput. Phys., vol. 448, no. 110752, pp. 1–19, 2022, doi: 10.1016/j.jcp.2021.110752.
[32]A. Daw, A. Karpatne, W. D. Watkins, J. S. Read, and V. Kumar, “Physics-guided neural networks (PGNN): An application in lake temperature modeling,” 2021. doi: 10.48550/arXiv.1710.11431.
[33]S. Hwang and S. K. Choi, “Deep learning-based surrogate modeling via physics-informed artificial image (PiAI) for strongly coupled multidisciplinary engineering systems,” Knowledge-Based Syst., vol. 232, p. 107446, 2021, doi: 10.1016/j.knosys.2021.107446.
[34]Y.L. Zhuang, Y.X. Liu, A. Ahmed, Z.G. Zhong, E.A. del Rio Chanona, C. P. Hale, M. Mercang¨oz., “A hybrid data-driven and mechanistic model soft sensor for estimating CO2 concentrations for a carbon capture pilot plant,” Comput. Ind., vol. 143, no. March, 2022, doi: 10.1016/j.compind.2022.103747.
[35]T. N. K. Nguyen, T. Dairay, R. Meunier, and M. Mougeot, “Physics-informed neural networks for non-Newtonian fluid thermo-mechanical problems: An application to rubber calendering process,” Eng. Appl. Artif. Intell., vol. 114, no. February, p. 105176, 2022, doi: 10.1016/j.engappai.2022.105176.
[36]C. Zhang and A. Shafieezadeh, “Nested physics-informed neural network for analysis of transient flows in natural gas pipelines,” Eng. Appl. Artif. Intell., vol. 122, no. March, p. 106073, 2023, doi: 10.1016/j.engappai.2023.106073.
[37]T. A. Ihunde and O. Olorode, “Application of physics informed neural networks to compositional modeling,” J. Pet. Sci. Eng., vol. 211, no. 110175, pp. 1–10, 2022, doi: 10.1016/j.petrol.2022.110175.
[38]L. F. Nazari, E. Camponogara, L. S. Imsland, and L. O. Seman, “Neural networks informed by physics for modeling mass flow rate in a production wellbore,” Eng. Appl. Artif. Intell., vol. 128, no. November 2023, p. 107528, 2024, doi: 10.1016/j.engappai.2023.107528.
[39]K. S. Gyamfi, J. Brusey, and E. Gaura, “Differential radial basis function network for sequence modelling[Formula presented],” Expert Syst. Appl., vol. 189, no. September 2021, p. 115982, 2022, doi: 10.1016/j.eswa.2021.115982.
[40]J. Fernández, M. Corbetta, C. S. Kulkarni, J. Chiachío, and M. Chiachío, “Training of physics-informed Bayesian neural networks with ABC-SS for prognostic of Li-ion batteries,” Comput. Ind., vol. 155, no. November 2023, p. 104058, 2024, doi: 10.1016/j.compind.2023.104058.
[41]H. P. H. Luu, H. M. Le, and H. A. Le Thi, “Markov chain stochastic DCA and applications in deep learning with PDEs regularization,” Neural Networks, vol. 170, pp. 149–166, 2024, doi: 10.1016/j.neunet.2023.11.032.
[42]D. Chen, Y. Li, J. Guo, and Y. Li, “Estimation of hypersonic vehicle weight using Physics-Informed neural network supported by knowledge based engineering,” Expert Syst. Appl., vol. 195, no. December 2021, 2022, doi: 10.1016/j.eswa.2022.116609.
[43]M. S. F. Bangi, K. Kao, and J. S. Il Kwon, “Physics-informed neural networks for hybrid modeling of lab-scale batch fermentation for β-carotene production using Saccharomyces cerevisiae,” Chem. Eng. Res. Des., vol. 179, pp. 415–423, 2022, doi: 10.1016/j.cherd.2022.01.041.
[44]X. Sheng and W. Xiong, “Soft sensor design based on phase partition ensemble of LSSVR models for nonlinear batch processes,” Math. Biosci. Eng., vol. 17, no. 2, pp. 1901–1921, 2020, doi: 10.3934/mbe.2020100.
[45]X. Shi and W. Xiong, “Adaptive ensemble learning strategy for semi-supervised soft sensing,” J. Franklin Inst., vol. 357, no. 6, pp. 3753–3770, 2020, doi: 10.1016/j.jfranklin.2019.07.008.
[46]F. Shen, J. Zheng, L. Ye, and X. Ma, “LSTM Soft Sensor Development of Batch Processes with Multivariate Trajectory-Based Ensemble Just-in-Time Learning,” IEEE Access, vol. 8, pp. 73855–73864, 2020, doi: 10.1109/ACCESS.2020.2988668.
[47]A. S. Struchtrup, D. Kvaktun, and R. Schiffers, “Adaptive quality prediction in injection molding based on ensemble learning,” Procedia CIRP, vol. 99, no. March, pp. 301–306, 2021, doi: 10.1016/j.procir.2021.03.045.
[48]Q. Sun and Z. Ge, “Deep Learning for Industrial KPI Prediction: When Ensemble Learning Meets Semi-Supervised Data,” IEEE Trans. Ind. Informatics, vol. 17, no. 1, pp. 260–269, 2021, doi: 10.1109/TII.2020.2969709.
[49]J. M. Moreira de Lima and F. M. Ugulino de Araujo, “Ensemble deep relevant learning framework for semi-supervised soft sensor modeling of industrial processes,” Neurocomputing, vol. 462, pp. 154–168, 2021, doi: 10.1016/j.neucom.2021.07.086.
[50]X. Liu, L. Li, F. Zhang, and N. Li, “Research on soft sensor modeling method for complex chemical processes based on local semi-supervised selective ensemble learning,” Meas. Sci. Technol., vol. 35, no. 7, 2024, doi: 10.1088/1361-6501/ad3fd5.
[51]A. Gosavi, “A Tutorial for Reinforcement Learning,” Sci. Technol., no. 716, pp. 1–12, 2009, [Online]. Available: http://web.mst.edu/~gosavia/tutorial.pdf
[52]I. Bifulco, S. Cirillo, C. Esposito, R. Guadagni, and G. Polese, “An intelligent system for focused crawling from Big Data sources,” Expert Syst. Appl., vol. 184, no. June, p. 115560, 2021, doi: 10.1016/j.eswa.2021.115560.
[53]R. Xiong, J. Cao, and Q. Yu, “Reinforcement learning-based real-time power management for hybrid energy storage system in the plug-in hybrid electric vehicle,” Appl. Energy, vol. 211, no. 5, pp. 538–548, 2018, doi: 10.1016/j.apenergy.2017.11.072.
[54]X. C. Zhang, J. G. Gong, and F. Z. Xuan, “A physics-informed neural network for creep-fatigue life prediction of components at elevated temperatures,” Eng. Fract. Mech., vol. 258, no. 108130, pp. 1–13, 2021, doi: 10.1016/j.engfracmech.2021.108130.
[55]L. F. M. Zorzetto and J. A. Wilson, “Monitoring bioprocesses using hybrid models and an extended kalman filter,” Comput. Chem. Eng., vol. 20, no. SUPPL.1, pp. 689–694, 1996, doi: 10.1016/0098-1354(96)00124-x.
[56]D. P. Searson, S. C. Burnham, M. J. Willis, and A. R. Wright, “Identification of chemical reaction mechanism from batch process data,” in Proceedings of the 17th IASTED International conference on modelling and simulation, 2006, pp. 511–516. doi: 10.5555/1167113.1167204.
[57]J. Dai, N. Chen, X. Yuan, W. Gui, and L. Luo, “Temperature prediction for roller kiln based on hybrid first-principle model and data-driven MW-DLWKPCR model,” ISA Trans., vol. 98, pp. 403–417, 2020, doi: 10.1016/j.isatra.2019.08.023.
[58]C. D. Paternina-arboleda, J. R. Montoya-Torres, and A. Fábregas-Ariza, “Simulation-optimization using a reinforcement learning approach,” in Proceedings of the 2008 Winter Simulation Conference, Miami, 2008, pp. 1376–1383. doi: 10.1109/WSC.2008.4736213.
[59]F. Guo, X. Zhou, J. Liu, Y. Zhang, D. Li, and H. Zhou, “A reinforcement learning decision model for online process parameters optimization from offline data in injection molding,” Appl. Soft Comput. J., vol. 85, p. 105828, 2019, doi: 10.1016/j.asoc.2019.105828.
[60]S. Nikita, A. Tiwari, D. Sonawat, H. Kodamana, and A. S. Rathore, “Reinforcement learning based optimization of process chromatography for continuous processing of biopharmaceuticals,” Chem. Eng. Sci., vol. 230, p. 116171, 2021, doi: 10.1016/j.ces.2020.116171.
[61]M.I. Radaideh, I. Wolverton, J. Joseph, J.J. Tusar, U. Otgonbaatar, N. Roy, B. Forget, K. Shirvan, “Physics-informed reinforcement learning optimization of nuclear assembly design,” Nucl. Eng. Des., vol. 372, no. August 2020, p. 110966, 2021, doi: 10.1016/j.nucengdes.2020.110966.
[62]H. Liu, C. Yu, H. Wu, Z. Duan, and G. Yan, “A new hybrid ensemble deep reinforcement learning model for wind speed short term forecasting,” Energy, vol. 202, no. 117794, pp. 1–18, 2020, doi: 10.1016/j.energy.2020.117794.
[63]L. Bramlage and A. Cortese, “Generalized attention-weighted reinforcement learning,” Neural Networks, vol. 145, pp. 10–21, 2022, doi: 10.1016/j.neunet.2021.09.023.
[64]D. Dey, L. Ghosh, D. Bhattacharya, and A. Konar, “A 2-phase prediction of a non-stationary time-series by Taylor series and reinforcement learning,” Appl. Soft Comput., vol. 145, p. 110565, 2023, doi: 10.1016/j.asoc.2023.110565.