在半導體的製造過程中,大量的製程資料會收集到工程資料庫中,以便進行製程監控、故障分析與品質管理。因爲半導體的製程複雜,影響良率的變因眾多且通常具有相互關係,工程師往往難以藉著本身的專業知識或是經驗法則,迅速且有效率的發覺導致製程異常的原因以及可能隱藏的資訊,進而迅速的處理事故問題。本研究建構半導體資料挖礦架構,並發展模糊規則決策樹方法,整合Kruskal-Wallis檢定、卡方交互影響檢測、目標變數模糊化與模糊規則、變異降低分支法則等方法,以減少當資料屬性存在交互作用時,建構模糊規則決策樹方法時所產生之變數選擇偏差的問題,以更有效地尋找可能造成製程變異的原因,提昇工程師在事故診斷和良率提昇的決策品質。本研究運用某半導體廠之實際資料,比較模糊規則決策樹方法與現行決策樹演算法的表現,並檢驗本研究的效度和可行性,做爲工程師及領域專家解決問題的參考依據,協助縮短事故診斷的時間,進而提昇半導體製程的良率,加強高科技產業的競爭力。
During semiconductor fabrication process, huge process data will be automatically or semi-automatically recorded and accumulated in database for monitoring the process, diagnosing faults, and managing manufacturing. However, the manufacturing factors that affect the wafer yield are mostly interrelated. Domain engineers cannot identify possible root causes of low yield efficiently and effectively on the basis of only their own domain knowledge. This study aims to construct a data mining framework for analyzing interrelated semiconductor manufacturing data and propose a fuzzy decision tree approach that integrates Kruskal-Wallis test, chi-square interaction detection, fuzzy set theory, and the variance reduction splitting criterion to analyze huge multi-dimensional semiconductor data to derive possible causes of faults for yield enhancement. We estimated the validity of the proposed approach with common database with existing algorithms and then conducted an empirical study in a semiconductor company for validation. The results demonstrated the practical viability of the proposed method to help the engineers to diagnose the faults efficiently and improve their decision quality effectively.