Identifying the out-of-control variables of multivariate control chart using ensemble SVM classifiers

Abstract

Out-of-control signals in multivariate charts may be caused by one or few variables or a set of variables. Multivariate process control often encounters with the diagnosis or interpretation difficulty of an out-of-control signal to determine which variable is responsible for the signal. In this article, we formulate the diagnosis of out-of-control signal as a classification problem. The proposed system includes a shift detector and a classifier. The traditional multivariate chart works as a mean shift detector. Once an out-of-control signal is generated, an SVM-based ensemble classifier is used to recognize the variables that have shifted. We propose using subgroup data and extracted features (sample mean and Mahalanobis distance) as the input vectors of classifier. The performance of the proposed system was evaluated by computing its classification accuracy. We use the traditional decomposition method as a benchmark for comparison. The simulation studies indicate that the proposed ensemble classification model is a successful method in identifying the source of the mean change. The results also reveal that SVM using extracted features as input vector has slightly better classification performance than using raw data as input. The proposed method may facilitate the diagnosis of the out-of-control signal.

在多變量管制圖中 ， 管制外訊號的產生可能是由一個或多個變量所導致。 一般多變量管制圖的缺點在於當偵測到異常訊號時 ， 使用者很難判斷是由哪一個變量所造成。 在本研究中 ， 我們將辨識管制外訊號之來源視為一個分類問題。 此研究所提出之系統包含製程偏移的偵測器和分類器。 當傳統的多變量管制圖偵測出製程異常後 ， 我們將以 SVM 之整體式分類器辨識製程偏移之來源。 本研究提出一個創新之數據多樣性的方法來建構整體式分類器。 我們也提出並比較兩種輸入向量分別是原始數據和特徵值 (樣本平均數和馬氏距離)。 本研究以傳統的分解法作為比較基準 ， 利用分類之準確性來評估不同方法之績效。 根據實驗之結果得知 ， 整體式分類模型為一種成功的方法可應用於辨識異常之來源。 研究結果也顯示 ， 利用特徵值作為 SVM 之輸入向量的分類績效會優於使用原始數據作為輸入向量。 本研究所提出之方法有助於診斷多變量製程異常訊號之來源。

(*聯絡人 ： [email protected])

Keywords:

• multivariate control chart
• interpretation of signal
• ensemble classifiers
• support vector machine

Keywords:

• 關鍵詞 ： 多變量管制圖 ； 異常來源辨識 ； 整體式 ； 支援向量機

Notes

(*聯絡人 ： [email protected])