由於製程資料自動擷取系統己普遍運用於現代化之製程環境中，同時監控數個相關的製程（或品質）變數之多變量統計製程管制技術己普遍受到重視。近年來，機器學習技術（特別是類神經網路）已被用來偵測多變量製程中平均值變動之狀態，也獲致不錯的成果，但是類神經網路常有「過度學習」的困擾，而無法順利將訓練結果一般化。支援向量機是機器學習領域中，另一種較新的技術，在其學習過程中，採用結構風險最小化的原則，來避免過度學習的陷阱，所以常能有較佳的一般化能力。集成式分類器模仿人類在作重大決策前，會先諮詢多位專家意見的行為，其核心原理在於整合多個單一分類器的分類結果後，所作的決策，常比單一分類器的分類結果準確，在許多複雜的模式辨識問題中，集成式分類器的績效往往比單一分類器好。本研究應用以支援向量機為基之集成式分類技術構建一個在多變量製程中，線上即時監控平均值變動的模式。模擬數據顯示，本研究提出的模式可有效率地偵測到多變量製程中平均值的變動，而且能準確地指出那些變數的平均值已變動及其變動方向，與文獻中其他的類神經網路模式、支援向量機模式及傳統多變量管制圖相較，本研究提出的以支援向量機為基之集成式分類模式具有較佳的偵測速度（即較短的平均串連長度）。本研究提出的模式，可使品管人員更有效率且更準確地在多變量製程中，監控平均值的變動。

Using data acquisition systems and computers in on-line process control has led to increased interest in multivariate statistical process control (SPC) in which several interrelated quality variables are simultaneously monitored. Learning based techniques, especially neural networks, have been applied to detect mean shifts in multivariate processes with promising results. However, neural networks suffer from generalization problems due to overfitting. Support vector machines (SVMs) avoid the overfitting problem by adopting the structure risk minimization principle in the learning process. Classifier ensembles (i.e., combining of multiple classifiers) have been proven to be a method superior to single classifiers in many complex pattern recognition problems. With the SVM based classifier ensemble technique, this study proposes a straightforward and effective model to on-line recognize mean shifts in multivariate processes. Empirical results using simulation show that the proposed classifier ensemble model can not only efficiently recognize the mean shifts but also accurately identify the variables that have deviated from their original means. The shift direction of each of the deviated variables can also be simultaneously determined. Numerical comparisons in a bivariate scenario indicate that the proposed SVM based classifier ensemble model outperforms neural network models, SVM models, and conventional multivariate SPC approaches reported in the literature in terms of average run length. This study is useful for quality practitioners who seek efficient methods for on-line recognizing mean shifts in multivariate processes, where the investigation resulting from a false recognition is costly.