建構類神經網路辨識器於ICA管制圖異常信號之診斷

Constructing Neural Network Recognizer for Diagnosing the ICA Control Chart Out-of-Control Signals

鄭仲宏(Chung-Hung Cheng)

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(2011年)

摘要

獨立成份分析法（Independent Component Analysis或簡稱ICA）被廣泛用於非高斯（non-Gaussion）多變量流程之異常監控（monitoring），如：化學流程、生物科技流程等。基本上，ICA為一資料轉換之過程，即將原始資料透過線性轉換之方式，轉成各自互相獨立之成份，即ICA成份（Independent Component）。其主要轉換過程包含：白化過程、非高斯化迭代演算及成份個數之選取。由此可知，當ICA管制圖發生異常信號（out-of-control signal）時，針對此ICA異常信號欲回溯辨識哪一原始變數發生異常，此非容易之事，因為轉換過後之原始資料已被扭曲。
傳統上，針對ICA之異常信號之辨識，常使用之工具為貢獻圖（Contribution Plots）。然而，貢獻圖深受成份個數之選取與人員主觀判斷之影響，因此並無法準確地辨識哪一原始變數發生異常。此種情況將導致品管工程人員無法正確矯正流程以排除異常。基於上述，本研究將以倒傳遞類神經網路（Back-Propagation Neural Network或簡稱BPNN）建構ICA異常信號辨識器。本研究將用兩個案例研究說明方法之有效性，包含一模擬流程與一實際案例：田納西伊士曼流程。透過此兩案例比較貢獻圖與所提方法之診斷異常能力。實驗結果顯示，本研究所提方法能有效辨識異常之品質特性。

關鍵字

獨立成份分析 ； 類神經網路 ； 異常信號辨識器 ； 管制圖 ； 貢獻圖

並列摘要

Recently, Independent Component Analysis (ICA) has been widely used for non-Gaussian multivariate process monitoring, such as chemical and biomedical processes. In fact, the ICA is a data transformation technique and its aim is to linearly combine original variables into several Independent Component (ICs) and these ICs are mutually independent. The data transformation procedures mainly involved whitening, non-Gaussianlize and component selection. Thus, when the ICA based control charts triggered out-of-control signals, it is difficult to exactly trace back which original variables are in abnormal circumstance. Because the original dataset has been distorted once the ICA applied.
Traditionally, the contribution plot is developed to diagnose the ICA out-of-control signal. However, the diagnosed accuracy strongly dependents on the number of ICs selected. Furthermore, the use of contribution plot also needs engineers’ subjective judge which means different engineers will conclude different results. As mentioned above, contribution plot may fail to precisely single out the abnormal variables which will mislead engineers to rectify the process. Thus, the objective of this study aims to design Back-Propagation Neural Network (BPNN) based intelligent identifier to interpret the ICA out-of-control signals.
The efficiency of proposed BPNN identifier will be evaluated via implementing two examples, including a simulated process and a real case of Tennessee Eastman (TE) chemical process. Results demonstrated that the proposed method processes satisfied identifying ability when compared to contribution plot.

並列關鍵字

Neural Network ； out-of-control identifier ； Independent Component Analysis ； contribution plot ； control chart

參考文獻

[1] Aparisi, F., and Sanz, J., “Interpreting the Out-of-Control Signals of Multivariate Control Charts Employing Neural Networks,” World Academy of Science, Engineering and Technology, Vol. 61, pp. 226-230 (2010).

[3] Chiang, L. H., Russell, E. L., and Braatz, R. D., Fault Detection and Diagnosis in Industrial Systems, Springer, London (2001).

[5] Dong, D., and McAvoy, T. J., “Nonliner principal component analysis based on principal curves and neural networks, ” Computers and Chemical Engineering, Vol. 20, No. 1, pp. 65-78 (1996).

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

[7] Hyvarinen, A., and Oja, E., “A fast fixed point algorithm for independent component analysis,” Neural Computation, Vol. 9, pp. 1483-1492 (1997).