以資料探勘改善良率之研究

Abstract

在高科技的PI薄膜製造產業中，良率(yield)是決定企業成敗的關鍵指標。在複雜的製造程序中，靠著電腦或是人工記錄下來的資訊量非常龐大，要從大量的資料中找出影響製程的關鍵因子、提升良率，運用資料探勘技術(data mining)是最有效率且可靠的方式。本研究提出一個新的資料探勘演算架構，可以用來找出關鍵的製程參數(critical process parameters)並且計算製程控制區間(govern-interval)，此架構包含了四個主要的模組：資料整理、參數篩選、控制區間計算、資料驗證。本文引用了一個簡單且有效的參數篩選法— One-R演算法，並且發現不同的資料轉換法會顯著影響One-R演算法的參數篩選結果；進一步，本研究應用K-means分群演算法提出了一個可以推算參數控制區的技巧。最後，實地應用台灣一家專門生產PI薄膜工廠所提供的資料做驗證，結果顯示良率可從63%改善至93%。

In a PI film manufacturing process, process yield is the most important performance index. Hundreds of process parameters are monitored and analyzed for possible yield enhancement. However, it is difficult to identify the critical process parameters from the hundred ones. This thesis proposed a data mining technique for identifying the critical process parameters, and suggested an appropriate value-interval for each critical process parameter. The technique consists of four modules: data clean, feature selection, value-interval determination, and performance evaluation. We justified this technique by an empirical study in a PI film manufacturing factory. Test results indicated that the yield could improve from 63% to 93% by the proposed data mining technique.