應用資料探勘技術控制PI製程之膜厚變異

Abstract

對於生產製造業而言，製程良率的高低是導致ㄧ間公司成敗的關鍵因素。台灣某公司的polyimide film製程ㄧ直以來都無法有效控制其產品的膜厚變異，原因在於製造程序的複雜，不論在製程的前段或是後段，都會蒐集產品通過機台而自動化或人工記錄的參數資料，要從這些龐大的資料量中找出影響膜厚變異的關鍵因子，僅靠專業知識及傳統的統計方法是不夠用的。因此本研究利用資料探勘技術結合統計方法去建構一個可以改善製程良率的資料探勘新架構，包含先應用1R找出可疑參數來縮小製程範圍，然後利用相關分析與專業知識發現可控參數，最後建立迴歸方程式找出可控參數的設定值。結果有效的幫助公司找出影響膜厚變異的關鍵因子，使得製程良率從41.7 %提升到61 %，大幅改善約20 %。除此之外，本研究還比較四種1R版本及四種C4.5版本於該資料庫上的表現，結果指出1R不僅可使用在良率改善上，且表現不遜色於C4.5。

Process yield is a very important performance index for manufacturing. In order to enhance yield, process data will be automatically or semi-automatically recorded for diagnosing faults. Many production processes often involve hundreds of process and quality parameters. As a result, finding the root causes of process variation becomes a difficult problem. This study combined data mining techniques and statistical methods to develop a solution framework for identifying the critical process parameters. This framework applied 1R algorithm to identify the clue variables, and then used correlation analysis and domain knowledge to find the to-be-controlled variables. Finally, it utilized regression analysis to set the value of the to-be-controlled variables. We validated this framework with an empirical study about the control of thickness variation for polyimide film manufacturing. Results indicated that the process yield has been improved from 41.7 % to 61 %. Moreover, we compared four versions of 1R with four versions of C4.5, and our initial results show the promise of 1R algorithm to improve the process variation.