整合基因演算法與類神經網路於線上辨識管制圖異常形狀之研究

統計製程管制(statistical process control, SPC)是廣泛被工業界採用在製程監控的重要方法，而在SPC中用來做製程監控的主要工具為管制圖。在管制圖上常會出現異常的形狀(pattern)，此為造成製程失控的特定原因，而辨識與分析管制圖形狀(control chart pattern, CCP)就成了SPC的重要題。管制圖可以用來決定系統的狀態並偵測製程中隨時可能發生的異常情況。異常的管制圖形狀與製程變異中一些特殊的非機遇性原因有關聯，因此有效地辨識異常管制圖形狀能減少可能需要的檢查次數，並加速診斷搜尋。近年來類神經網路(Artificial Neural Networks, ANNs)已經成功地應用在管制圖形狀辨識上，但大多數的研究均以原始資料作為類神經網路的輸入向量(Raw data based, RB)；部分研究則是利用由原始資料所擷取出的特徵當作輸入向量(Feature data based, FB)以減少網路規模。但在類神經網路的架構上，不容易找出網路之最佳架構，所以本研究將利用基因演算法(Genetic Algorithm, GA)的最佳演化特性與ANNs做結合，使能將原本ANNs的辨識能力再提高。將使用的訓練範例與測試範例均利用蒙地卡羅模擬法產生出生產線製程數據並結合由製程數據所擷取出的統計特徵值一同當作倒傳遞類神經網路的輸入向量(HB)，再利用結合GA與ANNs 之NeuroGenetic Optimizer(NGO)這套系統做訓練與測試，並比較是否加入統計特徵值的原始製程數據之辨識準確率會比無加入統計特徵的辨識準確率來的較好。但實驗結果得知，依照我們對製程數據設定的干擾率( 0.1-1.0)的變化情況下有加入統計特徵值測試正確率為98.397%，而無加入統計特徵值測試準確率為98.575%。比較結果顯示出由於測試環境的關係，加入統計特徵值並沒有好的效用，反而辨識準確率比無加入統計特徵值的製程數據之辨識準確率要來的差。

關鍵字

類神經；基因演算法；管制圖；統計製程管制

並列摘要

Statistical process control (SPC) is an important method for control process in industry. Hence, control chart is an important tool at statistical process control. The regular session presents unusual pattern in the tube charting, creates the system regulation out of control specific reason, but recognizes and the analysis control chart patterns (CCPs) has become the SPC the important topic. CCPs can be used to determine the status of system. Unnatural CCPs can be associated with a particular set of assignable causes for process variation. In recent years, artificial neural networks (ANNs) have been successfully used in the CCP recognition task. In intelligent SPC, most of researches used raw data (RB) as input vector and the other researches have used statistical feature data extracted from raw data (FB) as input vector for reducing network size. On the kind of neural network’s construction, is not easy to discover the network the best construction. Therefore this study using Genetic Algorithm (GA) the best evolved characteristic makes the union with ANNs, will enable the kind of nerve network's identification ability to enhance again originally. We present an ANN-based approach, in which an improved hybrid training data (HB) integrates both the time series data (Raw data) and the statistical feature data (Feature data). This set of systems makes the training and the test using NeuroGenetic Optimizer(NGO), And compares whether to join the statistical characteristic value the primitive system number of better compared to not to join the statistical nature according to it identification rate of accuracy the identification rate of accuracy to come well. But the experimental result knew that the statistical characteristic value test accuracy according to us to the system number of passes according to the hypothesis disturbance rate ( 0.1-1.0 ) change situation is 98.397%, but not joins the statistical characteristic value test rate of accuracy is 98.575%. As a result of testing environment relations with the result regarding joins the statistical characteristic value not good effectiveness, instead recognizes the accuracy rate is better that the not joins the statistical characteristic value in this system.