針對時間議題的智慧型製程缺陷偵測

Abstract

近年來，製程缺陷偵測一直是製造業中一個很重要的議題。一般來說，產品良率越高意味著能有越高的獲利，為了提高產品的良率，一種最直接的方法就是找出造成良率下降的原因。然而，要找出缺陷的根本原因是很難的一件事，因為缺陷有許多不同的型態，而且有些缺陷還會互相影響。因此我們將焦點集中在與機器有相關處理時間的缺陷。 在這篇論文中, 我們提出一個針對時間因素來找出品質低落主因之評估函數，這函數由三個標準以及相對應的權重所組成，這三個標準分別為個別行為，內部相似度與整體趨勢。由於這三個標準在不同的應用領域有不同的重要性。因此我們利用基因演算法來替這三個標準產生合適的權重。我們亦提出矛盾分析的方法將基因演算法的學習資料集合中之矛盾的資料排除掉。最後，實驗結果證明我們提出的方法對製程錯誤偵測是有效而且正確性很高的。

In recent years, defect detection problem of the workshop has become an important issue for manufacturing domain. In order to raise the quality of the products, the root cause of the low-quality situations should be found out as soon as possible. In this thesis, the time issue problem for the manufacturing domain is formally modeled and defined. Accordingly, the manufacturing defect detection system using root cause evaluation function which can generate a ranked list of possible root causes for the given dataset is proposed. For the extensibility and reliability, some adaptive weights are embedded into the function. Besides, for the existing datasets with known root causes, a supervised learning approach using genetic algorithm and a contradiction analysis method using similarity measurement are proposed to learn the adaptive weights of our proposed evaluation functions and judge the quality of the given dataset. Finally, the experiments have been made and the results show the proposed method can ensure the efficiency and accuracy.