整合J-B Test、隨機森林及貝氏推論於多變量製程監控

Integrating J-B Test, Random Forest and Bayesian inference for the multivariate process monitoring

陳建隆(CHIEN-LUNG CHEN)

指導教授 ： 許俊欽

朝陽科技大學/理工學院/工業工程與管理系/碩士(2020年)

摘要

近年，以管制圖來進行整廠製程監控(Plant-wide Process Monitoring)之技術越來越受到關注。然而，傳統之管制圖建模方式是使用製程之全變量來建構管制圖，此意味著可能使用了多餘變量(Redundant Variables)來進行建模，此將降低異常偵測之能力。此外，使用全變量直接建模之另一缺點在於無法分析掌握流程局部變化之行為，此將降低異常之偵測率。最後，整廠監控所收集到之資料，通常同時存在常態(Normal)及非常態(Abnormal)性之資料，因此，如何發展一有效之整廠製程監控方法為一重要之議題。基於上述，本研究提出了一種整合主成份分析法(Principal Component Analysis, PCA)、獨立成份分析法(Independent Component Analysis, ICA）和貝氏推論(Bayesian Inference)之方法。
本研究主要目的在於提出一分散式(Distributed)方法於多變量(Multivariate)之流程監控。首先，利用Jarque-Bera(J-B test)統計量，將原始資料變數劃分為常態子集與非常態子集。之後，分別利用隨機森林(Random Forest, RF)來進行變數篩選，本研究採用之篩選適應性函數(Fitness Function)為最小化無法偵測異常之比例(Non-Detection Rate)。隨後，分別利用主成份分析法及獨立成份分析法於保留變數的常態與非常態子集以建立監控模式。最後，各子集的監控結果，再利用貝氏推論整合最後之監控結果。
本研究所提的方法將應用於三個實際的製程分別為田納西伊士曼(Tennessee Eastman, TE)、台灣電力公司(Taiwan Power Company, TPC)和SECOM之多變量製程監控，第一個案例數據類型包括常態分配和非常態分配，用於將數據分為兩個子集，並使用貝氏推論將兩個子集融合進行評估、台電火力發電廠為一個非常態分配數據，用於研究本方法進行非常態數據分析以及SECOM製程為一個常態分配數據，使用本方法進行常態數據分析，比較傳統之監控方法，如：PCA、ICA等，將用於驗證所提方法之有效性，結果顯示所提之方法有效提升監控效果。

關鍵字

多變量統計製程監控 ； 常態性檢定 ； 主成份分析 ； 獨立成份分析 ； 隨機森林 ； 貝氏推論

並列摘要

In recent years, the technology of plant-wide process monitoring based on control charts has attracted more and more attention. However, the traditional control chart modeling method uses the full variables of the process to construct the control chart, which means that redundant variables may be used for modeling, which will reduce the ability of anomaly detection. In addition, another disadvantage of direct modeling using full variables is that it is impossible to analyze and grasp the behavior of local changes in the process, which will reduce the abnormal detection rate. Finally, the data collected by the whole plant monitoring usually contains both normal and abnormal data. Therefore, how to develop an effective whole plant process monitoring method is an important issue. Based on the above, this research proposes a method that integrates Principal Component Analysis (PCA), Independent Component Analysis (ICA), and Bayesian inference Method.
The purpose of this research is to propose a distributed method for multivariate process monitoring. First, the Jarque-Bera (J-B Test) statistic is used to divide the raw data variables into normal and abnormal subsets, and then random forest is used to filter the variables, and the fitness function of the screening is to minimize the non-detection rate. Thereafter, principal component analysis and independent component analysis were used to maintain the normal and abnormal subsets of the variables to establish a monitoring mode. The monitoring results of the last subsets are then combined with the final monitoring results using Bayesian inference.
The method proposed in this research will be applied to three actual manufacturing processes: Tennessee Eastman (TE), Taiwan Power Company (TPC) and SECOM multivariable process monitoring. The first data Types include normal distribution and abnormal distribution, which are used to divide the data into two subsets, and use Bayesian inference to fuse the two subsets for evaluation. TPC assigns data to an abnormal distribution, which is used to study the method for abnormal distribution. Data analysis and SECOM process are a normal distribution of data. Use this method for normal data analysis. Compared with traditional monitoring methods, such as PCA, ICA, etc., it will be used to verify the effectiveness of the proposed method. The results show that the proposed method is effective to improve monitoring results.

並列關鍵字

Multivariate Statistical Process Monitoring ； Normality Test ； Principal Component Analysis ； Independent Component Analysis ； Random Forest ； Bayesian Inference

參考文獻

[1] Breiman, L., "Random Forests", Machine Learning, Vol.45, pp.5-32, (2001).

[2] Chen, F.H., Chi, D.J., and Wang, Y.C., "Detecting biotechnology industry's earnings management using Bayesian network, principal component analysis, back propagation neural network, and decision tree", Economic Modelling, Vol.46, pp.1-10, (2015).

[3] Ge, Z.Q., Zhang, M.G., and Song, Z.H., "Nonlinear process monitoring based on linear subspace and Bayesian inference", Journal of Process Control, Vol.20, pp.676-688, (2010).

[4] Ge, Z.Q., Chen, J.H., "Plant-Wide Industrial Process Monitoring: A Distributed Modeling Framework", IEEE Transactions on Industrial Electronics, Vol.12, No.1, pp.310-321, (2015).

[5] Ghosh, K., Ramteke, M., and Srinivasan, R., "Optimal variable selection for effective statistical process monitoring", Computers & Chemical Engineering, Vol.60, pp.260-276, (2014).