半導體產業為因應科技的日新月異,資訊化程度需求的提升,以及市場的競爭激烈,使得高良率成為所追求的目標之一。因此如再以人工目視辨識,不僅浪費時間且有時會因判定的主觀性,而增加誤判的可能性。本研究為提升偵測的準確性及製程良率,將運用機器視覺方法與資料採礦技術,發展一套晶圓錯誤圖樣分析架構。我們首先以模擬方式產生包含十六種錯誤類型以及不同等級隨機誤差的一維與二維晶圓圖資料,在第一階段將討論支持向量機分析方法的錯誤辨識能力,而在第二階段則使用兩種類神經網路方法與兩種分類樹方法,比較這幾種分類方法與資料型態在辨認晶圓圖錯誤樣式的能力優劣。本論文模擬結果顯示使用支持向量機分析法(SVM)的確能提高精準的辨識正確率,且類神經網路中MLP法的表現最佳。當晶圓資料型態轉換成二維時,類神經與分類樹分類方法均可以提升辨識正確率。
The human view-based methods are traditionally used in semi-conductor industry to trace production errors with the disadvantages such as time-wasting and subjectiveness. To enhance the accuracy of detection and the product rate, machine vision methods and Data Mining Techniques are applied in this study to develop a wafer-map analysis system. A two-phase method is adopted in our study. During the first phase, the ability of identification of the erroneous judgment by Support Vector Machine based method will be discussed. In the second phase, neural networks models and decision tree methos are adopted. Random samples of one-dimension and two-dimension wafer bin maps were generated from sixteen patterns with various levels of random noises to compare identification accuracy. Our study shows that the adoption of Support vector machine analysis increases the accuracy of identification. In the second phase, we find that mulit-layer perceptron neural network models functions best. Also, when the wafer data is converted to spatial data representation, both Neural Networks Model and Decision Tree Analysis Model increase the accuracy of identification.