獨立成份分析(Independent Component Analysis;ICA)之主要目的為將具有高度相關性之原始資料經由線性轉換成各自具有獨立性之成份。其主要的轉換過程包括白化轉換、非高斯過程與成份之選取。近年來,ICA被廣泛應用於非高斯多變量流程之監控。然而選取成份的個數將影響其失效偵測之優劣。於ICA監控方法當中,常用之ICA成份選取方法為L2(Cardoso and Soulomica, 1993; Lee et al., 2004a) 與Durbin-Watson(Ammari et al., 2012; Bouveresse et al., 2012)。然而此二法仍須人員的主觀判斷選取成份個數。此將影響ICA異常偵測效率。有鑒於此,本研究擬發展一簡單且客觀的IC成份選取方法。本研究結合Durbin-Watson統計量與田口方法以選取顯著IC成份。首先利用田口直交表(Orthogonal Array;OA)進行ICs實驗組合,之後發展以Durbin-Watson為基之信號雜音比(Signal-to-Noise ; SN比)以決定顯著的ICs。此方法能以更客觀且明確的指出顯著之ICs。本研究所提之方法將應用於三個不同流程之監控以說明方法之有效性。第一個案例為模擬非線性流程之監控。第二個案例為台灣電力公司火力發電廠之流程監控。最後一個案例為田納西伊士曼流程之監控(Tennessee Eastman;TE)。實驗結果顯示所提方法能以較少的ICs獲得較好結果且能更客觀的選取IC成份。
Independent Component Analysis (ICA) is a multivariate technique aims at linearly transforming correlated variables into independent component. The transforming procedures include whitening, non-Gaussian and component selection. Recently, ICA has been widely used for monitoring non-Gaussian processes. However, the process fault detectability strongly dependents on the selected components. Traditionally, the Euclidean’s L2 norm and Durbin-Watson (DW) statistic were applied for opting the significant IC components. However, the main drawback of both methods still needs engineers’ effort to select the significant IC components. In this study, a simple and objective method will be developed to address the problem. This study proposed to integrate DW statistic and Taguchi method. First of all, the Orthogonal Array (OA) is adopted to experiment several possible combinations of selected components. After that, the DW embedded Signal-to-Noise (SN) ratio is used to measure the experiment results and ultimately in a bid to select significant IC components. The efficiency of the proposed method will be verified via three examples that included a non-Gaussian simulated process and two real case studies that came from Taiwan Power Company and Tennessee Eastman process, respectively. Experiments demonstrated that the proposed method can use lesser ICs to achieve superior performance.