工業4.0在2011年被提出之後，製造業推動產業結構調整，在現有的產線上安裝感測器累積了大量節點的數據，在很多傳統分析方法要實行，通常就會由專家或是工程師的經驗，從這麼多特徵資料挑選出來，所以我們主要想要探討如何在未經專家或是工程師介入的情況下，能找出資料中隱含的有用資訊。應用因果發現模型，使得因果發現方法能處理龐大的大量節點的數據，在處理過後會把從數據中挖掘出來的資訊，並且轉換成因果圖提供給不是專家的一般人也能理解。
本論文主要會探討的範圍主要會針對幾個因果發現方法，探討其能處理節點的上限與處理速度，因為能在出現異常之前提前預警在製造業非常重要，將Notears與Lingam方法應用於一個水循環系統。本論文釐清幾個因果發現方法在不同大小資料集下的準確度與效能，而在真實資料的在因果發現方法下產出的因果圖，提供了比起以往研究更容易理解的結果。

Industry 4.0 was introduced in 2011, the manufacturing industry has been promoted the adjustment of industrial structure. Edge data have been accumulated by the sensors on the existing production line. In many traditional analysis methods, characteristic data are selected by experienced experts or engineers. We mainly want to explore how to find useful information hidden in the data without the intervention of experts or engineers.The application of causal discovery model, it enables the causal discovery method to process a huge amount of data of many nodes, and after processing, the information mined from the data will be mined, and convert it into a directed acyclic graph for the person who is not an expert to understand.
The scope of this paper will mainly focus on several causal discovery methods and discuss the upper limit and processing speed of the nodes that can be processed, and because it is very important in the manufacturing industry to be able to early warn for abnormality, Notears and Lingam was then applied to a water circulation system. This paper clarifies the accuracy and performance of several causal discovery methods on datasets of different sizes, and causal diagram produced under the causal discovery method of real data provides results that are easier to understand than previous studies.

論文審定書 i
中文摘要 ii
Abstract iii
目錄 iv
圖次 vii
表次 ix
簡介 1
第一節 研究背景 1
第二節 研究動機 2
第三節 研究問題與目的 2
文獻探討 3
第一節 冷卻水循環系統 3
一、 消耗能源最佳化 4
二、 異常預警與故障排除 5
第二節 貝式網路模型（Bayesian Network Model） 6
一、 貝式網路（Bayesian Network） 6
二、 有向無環圖（DAG） 7
三、 完全部分有向無環圖（CPDAG） 7
第三節 基於統計方法識別貝式網路的演算法 9
一、 干預資料(Interventional Data)與觀察資料(Observational Data) 9
二、 僅使用觀察資料的演算法：PC、GES 10
三、 使用觀察和干預資料的演算法：GIES 13
第四節 深度學習模型 13
一、 NoTears 13
二、 DAG- GNN 14
研究方法 15
第一節 資料蒐集與前處理 15
第二節 深度學習模型 18
第三節 評估指標 19
一、 誤報率 false discovery rate (FDR) 19
三、 真陽性率 True Positive Rate (TPR) 20
四、 偽陽性率 False Positive Rate (FPR) 20
五、 結構漢明距離（Structural Hamming Distance SHD） 20
六、 預測次數（Predict Size） 21
七、 使用時間（Used Time） 21
研究成果 22
第一節 合成數據集的表現 22
第二節 真實數據集 25
第三節 研究結果總結 31
結論、討論與建議 32
第一節 結論 32
第二節 討論 32
第三節 未來建議 32
參考文獻 33
附錄：其他圖表 36

Chickering, D. (2002). Optimal Structure Identification With Greedy Search. Journal of Machine Learning Research, 3, 507–554. https://doi.org/10.1162/153244303321897717
Erdös, P., & Rényi, A. (1959). On Random Graphs I. Publicationes Mathematicae Debrecen, 6, 290–297.
Hauser, A., & Bühlmann, P. (2012). Characterization and Greedy Learning of Interventional Markov Equivalence Classes of Directed Acyclic Graphs. ArXiv:1104.2808 [Cs, Math, Stat]. Retrieved from http://arxiv.org/abs/1104.2808
Ho, W. T., & Yu, F. W. (2020). Determinants of low energy performance in a multi-chiller system serving an educational premise. International Journal of Refrigeration, 114, 47–53. https://doi.org/10.1016/j.ijrefrig.2020.02.019
Iurii D. Katser, & Vyacheslav O. Kozitsin. (n.d.). Skoltech Anomaly Benchmark (SKAB) [Data set]. Kaggle. https://doi.org/10.34740/KAGGLE/DSV/1693952
Jongh, M., & Druzdzel, M. (2009). A Comparison of Structural Distance Measures for Causal Bayesian Network Models.
Kingma, D. P., & Welling, M. (2014, May 1). Auto-Encoding Variational Bayes. arXiv. Retrieved from http://arxiv.org/abs/1312.6114
Pearl, J. (2000). Causality: Models, reasoning, and inference. Cambridge, U.K. ; New York: Cambridge University Press.
Pearl, J., & Mackenzie, D. (2018). The book of why: The new science of cause and effect (First edition). New York: Basic Books.
Shimizu, S., Hoyer, P. O., Hyvärinen, A., & Kerminen, A. (2006). A Linear Non-Gaussian Acyclic Model for Causal Discovery. J. Mach. Learn. Res., 7, 2003–2030.
Spirtes, P., Glymour, C. N., & Scheines, R. (2000). Causation, prediction, and search (2nd ed). Cambridge, Mass: MIT Press.
Verma, T., & Pearl, J. (1992). An Algorithm for Deciding If a Set of Observed Independencies Has a Causal Explanation. Proceedings of the Eighth International Conference on Uncertainty in Artificial Intelligence, 323–330. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc.
Xu, Y., Yan, C., Shi, J., Lu, Z., Niu, X., Jiang, Y., & Zhu, F. (2021). An anomaly detection and dynamic energy performance evaluation method for HVAC systems based on data mining. Sustainable Energy Technologies and Assessments, 44, 101092. https://doi.org/10.1016/j.seta.2021.101092
Yu, C.-W., Chen, J.-W., & Chen, Y.-L. (2020). Integration of IoT and Enhanced LSTM framework for water-cooled chiller COP forecasting. 2020 International Symposium on Computer, Consumer and Control (IS3C), 57–60. Taichung City, Taiwan: IEEE. https://doi.org/10.1109/IS3C50286.2020.00022
Yu, Y., Chen, J., Gao, T., & Yu, M. (2019, April 22). DAG-GNN: DAG Structure Learning with Graph Neural Networks. arXiv. Retrieved from http://arxiv.org/abs/1904.10098
Zheng, X., Aragam, B., Ravikumar, P., & Xing, E. P. (2018). DAGs with NO TEARS: Continuous Optimization for Structure Learning. ArXiv:1803.01422 [Cs, Stat]. Retrieved from http://arxiv.org/abs/1803.01422
Zhou, X., Wang, B., Liang, L., Yan, J., & Pan, D. (2019). An operational parameter optimization method based on association rules mining for chiller plant. Journal of Building Engineering, 26, 100870. https://doi.org/10.1016/j.jobe.2019.100870