多變量自我相關製程之SPC與EPC整合流程

Abstract

統計製程管制(Statistical Process Control, SPC)是工業界最常應用的品質管制方法，此法主要是針對單品質特性而設計，然而現今產品的功能愈來愈複雜，管制產品╱製程所需考慮的品質特性也隨之增多，因此必須以多變量SPC來進行製程品質之管制。此外，品質特性本身常具有自我相關性，若以傳統的SPC手法來監控製程，往往會產生誤判的情況，故必須對資料進行先處理(pretreatment)以去除資料之自我相關性，再以SPC手法偵測出可歸屬原因(assignable cause)以進行改善。然而，製程除了受到可歸屬原因的影響外，也可能受到不易控制的機遇原因(chance cause)所影響。因此，當製程失控卻找不出可歸屬原因時，可透過工程製程管制(Engineering Process Control, EPC)來進行調整，以降低製程的變異。但使用EPC可能會有過度控制製程的問題，而對製程造成過度頻繁的調整。因此，本研究針對多變量自我相關之製程，提出一套完整有效之結合SPC與EPC之管制流程。此管制流程共分為兩階段進行，首先是建構多變量SPC流程，利用自組性演算法(Group Method of Data Handling, GMDH)構建預測模型，將資料轉換為無自我相關性之殘差，並以Hotelling T2管制圖和多變量累積和管制圖(multivariate cumulative sum control chart, MCUSUM)偵測製程平均值是否偏移。當製程失控卻找不出可歸屬原因時，則進行EPC流程，利用GMDH構建控制器對製程參數持續進行回饋調整，使製程之輸出變數值能夠回歸目標值。本研究最後以一化學機械研磨之實例驗證了本研究所構建之多變量自我相關製程之SPC與EPC流程的可行性及有效性。

Statistical process control (SPC) is a widely employed quality control method in industry. SPC is mainly designed for monitoring single quality characteristic. However, as the design of a product / process becomes complex, a process usually has multiple quality characteristics related to it. These characteristics must be monitored by multivariate SPC. When the autocorrelation is present in the process data, the traditional SPC may mislead the results. Hence, the autocorrelated data must be treated to eliminate the autocorrelation effect before employing SPC to detect the assignable causes. Besides, chance causes also have impact on the processes. When the process is out of control but no assignable cause is found, it can be adjusted by employing engineering process control (EPC). However, EPC may result in unnecessary adjustments. This study presents an integrated SPC and EPC procedure for multivariate autocorrelated process. The SPC procedure constructs a predicting model by group method of data handling (GMDH), which can transfer the autocorrelated data into uncorrelated data. Then, the Hotelling T2 and multivariate cumulative sum control charts are constructed to monitor the process. The EPC procedure constructs a controller utilizing GMDH to adjust the multiple quality characteristics to their target values. Finally, this study uses a set of simulated chemical mechanical polishing (CMP) data to verify the effectiveness of the proposed procedure.