在製程工程軟儀表開發中，可用於資料導向建模的數據量通常有限，使得傳統深度學習方法存在過擬合問題，且模型亦缺乏可解釋性。本研究擬結合第一原理模擬和遷移學習技術解決此等問題。我們先透過物理模型產生大量模擬數據，以提供類神經網絡學習正確的製程機制；再以工廠數據進行微調，使對實際系統的預測有定量準確性。本研究使用一選擇性氫化製程中的碳四分離塔說明此一方法。
我們以Aspen Plus及Aspen Plus Dynamics建立三個不同蒸餾塔的動態模擬，包含實際碳四塔、文獻的去丁烷塔和甲醇與水分離塔，並使用Aspen Dynamics和MATLAB-Simulink接口進生成大量數據。並以此等數據建立三個源域模型，再以實際數據及遷移學習產生三種微調前饋網絡；並與以實際數據直接訓練的簡單前饋網絡，及加入正則化的前饋網絡比較。比較的標準除了常用的預測數據均方根差外，還使用模型預測的兩個主要操作變量：回流量及塔底溫度設定對兩個主要質量變量：塔頂及塔底純度的四個增益之符號合理性作為模型物理可解釋性的指標。
研究結果發現，相對於簡單前饋網絡及正則化前饋網絡；FT-FFN能夠透過源域模型知識的嵌入，有效提升目標域模型的估計性能與可解釋性。對於一些次要製程機制解釋能力如交互作用，基於與真實製程較相似的源域模型的FT-FFN表現較好。

In the development of soft sensors for industrial processes, availability of data for data-driven modeling is usually limited. This leads to overfitting and lack of interpretability when deep learning models were used. In this study, first-principle simulations and transfer learning techniques are combined to address these problems. Source-domain models are obtained using large amount of simulation data provided by first-principle simulations. They were then fine-tuned by limited amount of real plant data to improve their prediction accuracies. An industrial C4 separation column operating at a selective hydrogenation unit was used as example to illustrate the effectiveness of this approach.
Three first-principle models, one that mimics the actual industrial C4 column, a debutanizer in literature, and a methanol/water splitter in literature, for three different distillation columns are built up by Aspen Plus and Aspen Dynamics. Large amount of data were generated by combining Aspen Plus Dynamics with MATLAB-Simulink interfaces. These data were used to train three source domain models using feedforward networks. These models were then fine tuned into the actual plant models using plant data. They are benchmarked against a simple feedforward network trained by plant data only and a feedforward network trained by plant data with regularization using L2 norm. The root mean square error of test data was used as a metric of accuracy. The sign consistency of the four gains, i.e. the effect of two main operating variables: reflux rate and temperature set-point of the reboiler on the two main quality variables: distillate and bottom impurities were used as a metric of physical interpretability.
Results showed that fine-tuned networks showed better accuracy and improved interpretability, compared to simple feedforward network with or without regularization, especially when the amount of actual data available is small. For some secondary effects such as interaction gain, fined-tuned networks based on physical model with greater similarity to the real process will have better interpretability.

摘要 I
Abstract II
致謝 III
目錄 IV
圖目錄 VI
表目錄 VIII
第一章 緒論 1
1-1 研究背景 1
1-2 小樣本 4
1-3遷移學習 5
1-4 研究動機 6
第二章 物理建模與數據生成 7
2-1 工業製程簡介 7
2-2 穩態物理建模 8
2-2-1 工業碳四塔模型 8
2-2-2 去丁烷塔模型 10
2-2-3 醇水分離塔模型 11
2-3 多變量擾動動態模擬 12
2-3-1 模擬方法簡介 12
2-3-2 動態模擬設定 13
第三章 數據處理與軟測量建模 15
3-1 數據預處理 15
3-1-1 工業數據庫簡介 15
3-1-2 輸入變量選擇 16
3-1-3 模型板數轉換 17
3-1-4 移動視窗法 18
3-2 建模與實驗 19
3-2-1 超參數設定 19
3-2-2 網絡結構 20
3-2-3 開發模式 21
3-2-4 網絡參數微調 22
第四章 模型可解釋性 23
4-1 物理模型穩態增益方向 23
4-2 動態增益定義與推導 24
4-3 事後解釋與可解釋性量化 25
第五章 結果與討論 26
5-1 源模型建模結果 26
5-1-1 工業碳四塔源模型建模結果 26
5-1-2 去丁烷塔源模型建模結果 29
5-1-3 醇水分離塔源模型建模結果 32
5-2 軟測量建模結果 35
5-2-1 塔頂 C5+濃度軟測量性能 35
5-2-2 塔底 C4濃度軟測量性能 36
5-2-3 增益方向一致性比較 37
第六章 總結與展望 41
參考文獻 42

[1] J.Mohd Ali, M.A.Hussain, M.O.Tade, J.Zhang, Artificial Intelligence techniques applied as estimator in chemical process systems - A literature survey, Expert Syst. Appl. 42 (2015) 5915–5931.
[2] C.Shang, F.Yang, D.Huang, W.Lyu, Data-driven soft sensor development based on deep learning technique, J. Process Control. 24 (2014) 223–233.
[3] P.Kadlec, B.Gabrys, S.Strandt, Data-driven Soft Sensors in the process industry, Comput. Chem. Eng. 33 (2009) 795–814.
[4] B.A.Jensen, J.Abonyi, Neural Networks for Process Modeling, in: Bela Liptak (Ed.), Instrum. Eng. Handbook, Vol. 2 Process Control Optim., 4th ed., Taylor & Francis, 2006: pp. 253–264.
[5] L.N.Smith, A disciplined approach to neural network hyper-parameters: Part 1 -- learning rate, batch size, momentum, and weight decay, US Nav. Res. Lab. Tech. Rep. (2018). http://arxiv.org/abs/1803.09820.
[6] S.Chakraborty, R.Tomsett, R.Raghavendra, D.Harborne, M.Alzantot, F.Cerutti, M.Srivastava, A.Preece, S.Julier, M.R.Raghuveer, T.D.Kelley, D.Braines, M.Sensoy, C.J.Willis, P.Gurram, Interpretability of deep learning models: A survey of results, 2017 IEEE SmartWorld, Ubiquitous Intell. Comput. Adv. Trust. Comput. Scalable Comput. Commun. Cloud Big Data Comput. Internet People Smart City Innov. (n.d.).
[7] D.Gunning, Explainable Artificial Intelligence (XAI), Def. Adv. Res. Proj. Agency. (2017).
[8] A.Adadi, M.Berrada, Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI), IEEE Access. 6 (2018) 52138–52160.
[9] W.J.Murdoch, C.Singh, K.Kumbier, R.Abbasi-Asl, B.Yu, Interpretable machine learning: definitions, methods, and applications, ArXiv Prepr. (2019).
[10] L.H.Gilpin, D.Bau, B.Z.Yuan, A.Bajwa, M.Specter, L.Kagal, Explaining explanations: An overview of interpretability of machine learning, 2018 IEEE 5th Int. Conf. Data Sci. Adv. Anal. (2018) 80–89.
[11] M.T.Ribeiro, S.Singh, C.Guestrin, “Why Should I Trust You?”: Explaining the predictions of any classifier, in: Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. - KDD ’16, ACM Press, New York, New York, USA, 2016: pp. 1135–1144.
[12] S.Ingrassia, I.Morlini, Neural network modeling for small datasets, Technometrics. 47 (2005) 297–311.
[13] L.Fortuna, S.Graziani, M.G.Xibilia, Comparison of soft-sensor design methods for industrial plants using small data sets, IEEE Trans. Instrum. Meas. 58 (2009) 2444–2451.
[14] X.Yuan, B.Huang, Y.Wang, C.Yang, W.Gui, Deep Learning-Based Feature Representation and Its Application for Soft Sensor Modeling with Variable-Wise Weighted SAE, IEEE Trans. Ind. Informatics. 14 (2018) 3235–3243.
[15] S.Park, C.Han, A nonlinear soft sensor based on multivariate smoothing procedure for quality estimation in distillation columns, Comput. Chem. Eng. 24 (2000) 871–877.
[16] L.Wang, C.Shao, H.Wang, H.Wu, Radial Basis Function Neural Networks-Based Modeling of the Membrane Separation Process: Hydrogen Recovery from Refinery Gases, J. Nat. Gas Chem. 15 (2006) 230–234.
[17] R.D.DeVeaux, J.Schumi, J.Schweinsberg, L.H.Ungar, Intervals for Neural Networks via Nonlinear Regression, Technometrics. 40 (1998) 273–282.
[18] L.Holmstrom, P.Koistinen, Using additive noise in back-propagation training, IEEE Trans. Neural Networks. 3 (1992) 24–38.
[19] Y.Grandvalet, S.Canu, S.Boucheron, Noise Injection: Theoretical Prospects, Neural Comput. 9 (1997) 1093–1108.
[20] R.M.Zur, Y.Jiang, L.L.Pesce, K.Drukker, Noise injection for training artificial neural networks: A comparison with weight decay and early stopping, Med. Phys. 36 (2009) 4810–4818.
[21] J.Zhang, E.B.Martin, A.J.Morris, C.Kiparissides, Inferential estimation of polymer quality using stacked neural networks, Comput. Chem. Eng. 21 (2003) S1025–S1030.
[22] G.-J.Qi, J.Luo, Small Data Challenges in Big Data Era: A Survey of Recent Progress on Unsupervised and Semi-Supervised Methods, ArXiv Prepr. (2019).
[23] S.J.Pan, Q.Yang, A survey on transfer learning, IEEE Trans. Knowl. Data Eng. 22 (2010) 1345–1359.
[24] Y.Guo, H.Shi, A.Kumar, K.Grauman, T.Rosing, R.Feris, SpotTune: Transfer Learning through Adaptive Fine-tuning, Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (2018) 4805–4814.
[25] J.Yosinski, J.Clune, Y.Bengio, H.Lipson, How transferable are features in deep neural networks?, Adv. Neural Inf. Process. Syst. (2014).
[26] Y.Li, N.Wang, J.Shi, J.Liu, X.Hou, Revisiting Batch Normalization For Practical Domain Adaptation, ArXiv Prepr. (2016).
[27] T.Streich, H.Kömpel, J.Geng, M.Renger, Process Engineering and Optimization Secure the best benefits from C 4 hydrocarbon processing-Part 1: Separation sequences, n.d.
[28] S.P.Bressa, J.A.Alves, N.J.Mariani, O.M.Martínez, G.F.Barreto, Analysis of operating variables on the performance of a reactor for total hydrogenation of olefins in a C3-C4-stream, Chem. Eng. J. (2003).
[29] R.Gani, C.A.Ruiz, I.T.Cameron, A generalized model for distillation columns—I. Model Description and Applications, Comput. Chem. Eng. 10 (1986) 181–198.
[30] R.K.Wood, M.W.Berry, Terminal composition control of a binary distillation column, Chem. Eng. Sci. 28 (1973) 1707–1724.
[31] Aspen DynamicsTM 12.1 Reference Guide, Aspen Technology, 2003. http://www.aspentech.com/.
[32] W.L.Luyben, Distillation Design and Control Using Aspen Simulation, John Wiley & Sons, 2013.
[33] H.Kaneko, K.Funatsu, Moving window and just-in-time soft sensor model based on time differences considering a small number of measurements, Ind. Eng. Chem. Res. 54 (2015) 700–704.
[34] J.Brownlee, Deep Learning with Python: Develop Deep Learning Models on Theano and Tensorflow Using Keras, 1.2, Machine Learning Mastery, 2016.
[35] Z.Lu, H.Pu, F.Wang, Z.Hu, L.Wang, The Expressive Power of Neural Networks: A View from the Width, Adv. Neural Inf. Process. Syst. (2017) 6231–6239.
[36] P.L.Bartlett, The Sample Complexity of Pattern Classification with Neural Networks: The Size of the Weights is More Important than the Size of the Network, IEEE Trans. Inf. Theory. 44 (1998) 525.
[37] D.E.Seborg, T.F.Edgar, A.M.Duncan, F.J.Doyle, Process Dynamics and Control, 4th ed., John Wiley & Sons, 2016.