在本論文中,提出關於半導體設備的異常偵測方法,要偵測一個複雜系統中的異常,由於缺少適當的模型,所以是一個困難的任務;正因如此,這也是讓資料探勘技術具有吸引力的地方。對於離子植入機,提出了一個以分類為基礎的異常偵測與隔離方法。所提出的方法包含了兩部分:分類部分以及異常偵測與隔離部分。另外,對於模鑄型變壓器,提出一個以分類為基礎的監控分析。在分類部分,提出一個模糊群集決策樹(Fuzzy Clustering Decision Tree, FCDT),應用於較多類別和連續性質屬性的資料分類問題上。FCDT結合了分群演算法以及決策樹C4.5。所提出的分群演算法以可分離矩陣和模糊規則為基礎得到分割屬性,利用這些分割屬性將資料分群至終端群集。在終端群集裡,包含了超過一個以上的類別,將進一步以C4.5 進行分類。在錯誤偵測與隔離部分,則提出一個基於分類結果準確性,決定是否產生警報信號的標準,而異常隔離的機制則是依據模糊群集決策樹(FCDT)來隔離離子植入機的真正故障。並且將所提出的分類器應用於離子植入機裡正在運作的晶圓片及模鑄型變壓器所量測到的局部放電訊號進行分類測試,並和現有的商業軟體See5和CART來以實例進行比較分類結果與計算時間。
In this dissertation, we present related issues on the abnormality detection of semiconductor equipment, To detect the fault of a complex manufacturing system is a difficult task because of the lack of proper model; indeed, this is the key that makes the data mining technique attractive. We propose a classification based fault detection and isolation scheme for the ion implanter. The proposed scheme consists of two parts: the classification part and the fault detection and isolation part. Another, we propose a classification based monitoring analysis for cast-resin transformers. In the classification part, we propose a fuzzy clustering decision tree (FCDT) for the classification problem with large number of classes and continuous attributes. The FCDT combines a proposed clustering algorithm with the decision tree C4.5.The proposed clustering algorithm split the data set into terminal clusters using splitting attributes based on a separation matrix and fuzzy rules. The terminal clusters consisting of the data of more than one class will be further classified using the C4.5, and a k-fold cross validation error is treated as the accuracy of the classification result. In the fault detection and isolation part, we propose a warning signal generation criteria based on the classification accuracy to detect and fault isolation scheme based on the FCDT to isolate the actual fault of an ion implanter. We have successfully applied the FCDT to the classification problem of the working wafers in an ion implanter and partial discharge signals of cast-resin transformers, and compared the classification results and the computation time with the existing software See5 and CART for real cases.