批次製程(batch process)在工業中扮演一個很重要的角色,許多多變量分析的方法亦被發展來監控批次製程。其中,多維主成份分析(Multiway Principal Component Analysis;MPCA)方法為最常使用之方法且顯現出很好的監控效果。MPCA須假設製程變數的主成份服從高斯分配,但此一假設往往與工業中的製程資料不服從高斯分配相違背,另一缺點則是傳統MPCA之監控方法有批次數據須等長之限制。為了改善上述之缺點,多維獨立成份分析(Multiway Independent Component Analysis;MICA)的方法因而被發展出來。MICA結合不同的展開法,將可避免批次數據須等長之限制,且監控的製程變數不須受服從高斯分配,因此,監控非高斯製程的成效較MPCA好。但因為傳統的MICA監控方法僅考慮目前觀測值的大小,對於先前觀測值之資訊卻未加以考慮。因此,傳統MICA的監控方法將會造成延遲偵測的問題,造成了成本上的損失及時間上的浪費。 基於上述,本研究擬以MICA監控方法為基礎,發展Enhanced MICA(EI)監控統計量,以提升傳統MICA監控方法於批次製程之偵測能力及改善延遲偵測的問題。本研究結合兩種不同的展開法,再以MICA方法選取獨立成份(Independent Components;ICs),搭配指數加權移動平均(Exponentially Weighted Moving Average;EWMA)法估計製程平均值之變動,最後將之整合成EI統計量。 本研究所提出之方法將被使用來偵測盤尼西林醱酵製程中的失效模式,並與傳統的MPCA及MICA做比較。實驗結果顯示本研究所提出之方法將能有效提升傳統MICA之偵測能力,並能提早偵測出製程產生變異。
Batch processes play an important role in many industries. Several methods of multivariate statistical analysis have been developed to monitor batch processes. Multiway Principal Component Analysis(MPCA)has shown a powerful monitoring performance in many batch processes. However, Principal components(PCs)of the process should be assumed to follow Gaussian distribution and traditional MPCA has a shortcoming of equal batch. In fact, the collected data from industrial processes rarely follow Gaussian distribution. In order to improve these drawbacks, Multiway Independent Component Analysis(MICA)has been developed. It combines different unfolding methods to overcome the mentioned disadvantage. Hence, MICA provides better monitoring performance than MPCA in cases with non-Gaussian variables. However, the traditional MICA based monitoring scheme only considers the magnitude of recent observations but ignores the information of previous observations. Hence, traditional MICA monitoring method will cause delayed fault detection problems. It also causes the cost of the loss and waste of time. As mentioned above, in order to enhance the detectability of traditional MICA based monitoring method, an Enhanced MICA(EI)statistic will be proposed. The study combines two unfolding methods, and then selects independent components(ICs). The Exponentially Weighted Moving Average(EWMA) method will be used to predict the changing direction of process mean and then EI statistic will be developed. The proposed method was used to detect faults in the fed-batch penicillin cultivation process. The simulation results clearly demonstrate the power and advantages of the proposed method in comparison to MPCA and MICA.