利用主成份分析與自適應提升演算法進行半導體之晶圓錯誤偵測與關鍵程序分析

Wafer Fault Detection and Key Step Analysis for Semiconductor Manufacturing Using Principal Component Analysis Adaboost and Decision Tree

林首志(Shou-Chih Lin)

指導教授 ： 范書愷

國立臺北科技大學/管理學院/工業工程與管理系碩士班/碩士(2014年)

摘要

在半導體製程中，訊號感測器會根據機台當下情況，傳送大量且種類繁多的數據資料。若是只透過傳統的監控方式，常常只能透過工程師個人的經驗法則去判斷關鍵參數(Key item)與關鍵程序(Key step)。因此發展出一套較客觀的製程錯誤偵測方法 (Fault Detection)，提供工程師發現製程異常的發生是極為重要的任務。本研究主要以晶圓製造之製程為例，本研究首先利用主成分分析法降低資料的維度，再利用分類演算法支持向量機 (Support Vector Machine) 或自適應提升演算法(AdaBoost)對及已知的 Good/Bab Wafer 進行各個參數(FD item)的辨識力比較，辨識力高的參數(FD item)即為關鍵參數；接著利用決策樹(Decision Tree)辨識與畫出樹狀圖，其關鍵結點即為關鍵程序。最後，依據本研究之方法進行錯誤偵測的測試與評估，建構快速且系統化的錯誤偵測監控能力！

關鍵字

半導體製程 ； 錯誤偵測 ； 主成份分析 ； 支持向量機 ； 自適應提升演算法 ； 決策樹

並列摘要

In the semiconductor manufacturing process, the signal sensors produce a huge amount of trace data according to the production states. In the traditional monitoring practice, frequently, the process engineer uses a variety of univariate statistics to seek the key parameters (items) and key steps with the help of subjective judgment. Therefore, developing a practicably effective way for fault detection in the semiconductor manufacturing process becomes an essentially important task. In this thesis, we take a real-world trace data set from a wafer manufacturing process for illustration. First, the principal component analysis (PCA) is utilized to reduce the dimension of the original data. Subsequently, the support vector machine (SVM), or the adaptive boosting algorithm (Adaboost) is employed to train the classification models by means of known data classes. These two models are used to classify the trace data with 38 parameters (items). The success rates of the classifiers with every parameter are evaluated. The parameters which give high accuracy are deemed as the key parameters. At last, the decision tree generation serving as a post hoc analysis is conducted to help the process engineer identify the corresponding key steps that might critically affect the yield of wafer manufacturing. The key parameters and key steps identified by using the proposed method are all confirmed by the experienced engineers. The proposed method is proved effective at finding key parameters and key items. Not surprisingly, the proposed method also allocates additional key parameters the process engineer cannot discover by using the univariate statistics method.

並列關鍵字

principal component analysis ； support vector machine ； adaptive boosting algorithm ； decision tree generation algorithm

參考文獻

1. Y. Freund & R. E. Schapire, "A desicion-theoretic generalization of on-line learning and an application to boosting. In Computational learning theory", Springer Berlin Heidelberg, 1995, January, pp. 23-37.

3. N. J. Belkin & W. B. Croft, " Information filtering and information retrieval: two sides of the same coin? ", Communications of the ACM, vol. 35, no. 12,1992, pp.29-38.

4. K. Pearson, "On Lines and Planes of Closest Fit to Systems of Points in Space". Philosophical Magazine, vol. 2, no. 11, pp.559–572.

5. H. Hotelling, "Analysis of a Complex of Statistical Variables into Principal Components", J. Educational Psychology, vol. 24, 1933, pp. 498-520. 52.

6. F. De la Torre, M.J. Black, "Robust principal component analysis for computer vision", International Conference on Computer Vision, 2001, pp. 362-369.