On-line monitoring several interrelated process (quality) variables in a manufacturing process is an essential task for maintaining and improving the quality and productivity of industrial products. This paper presents a hybrid artificial bee colony (ABC) algorithm/ classifier ensemble-based model for on-line accurate and quick recognition of mean shifts in multivariate processes. The objective of using classifier ensemble method is to improve the recognition capability of the computational intelligence-based classification model for the complicated multivariate process-monitoring task. Integrating the ABC algorithm with classifier ensembles aims for enhancing the classifier ensemble model by optimal selection of component classifiers. A fitness-scaling method is also introduced in this research to overcome the premature convergence problem of the typical ABC algorithm. A comparison between the proposed hybrid model and other computational intelligence-based approaches towards the trivariate process-monitoring scheme is presented. An application of the proposed model using a bivariate dataset from the published literature is also presented to demonstrate the model's on-line usage. The numerical results show that the classifier ensemble-based approaches significantly outperform the traditional single classifier-based approaches with respect to recognition accuracy. The proposed hybrid ABC/classifier ensemble-based model can also perform better than the approaches using merely the classifier ensemble method. With the same false alarm rate, the proposed computational intelligence-based model can outperform other published statistic-based algorithms with respect to the recognition speed.
線上即時監控並維持多個彼此相關的品質變數於一個穩定而可靠的水準上,是多變量統計製程管制 (multivariate statistical process control) 的一個重要功能。本研究整合蜂群演算法與集成式分類技術,提出一個線上即時辨識多變量製程平均值之變動的模式。整合多個單一分類器於一個分類任務的集成式分類技術,在本研究中被用來增進以計算型智慧 (computational intelligence) 為基之製程監控模式的辨識績效;本研究並應用蜂群演算法 (artificial bee colony algorithm) 於集成式分類 (classifier ensemble) 模式中,找出最適當的成份分類器組合,以最佳化多變量製程集成式分類模式的績效,本研究也進一步提出一個新型的適合度調整 (fitness scaling) 方法,以解決傳統蜂群演算法常遭遇的過早收斂 (premature convergence) 問題,以增進蜂群演算法在搜尋最佳值時的能力與速度。本研究所提出的模式在三變量 (trivariate) 製程中辨識平均值變動之模擬績效表現,除了遠優於單一分類器的績效外,也較典型的集成式分類器好。本研究也用一組文獻中的雙變量 (bivariate) 製程之真實數據,測試所提出的多變量製程監控模式,結果顯示,與傳統以統計為基之多變量管制技術相比較,在固定型一誤差 (Type I error) 的基礎下,本研究提出的整合蜂群演算法及集成式分類技術之模式,具有較短的平均串連長度,亦即具有較低的型二誤差 (Type II error)。