液晶顯示器產業出貨等級全賴人眼判斷，隨著新世代產線建立切割數增加， 及舊有電腦監視器所規畫產線，如今轉作中小尺寸面板，造成人力成本及需求倍 數增加。舊有產線更囿於空間，而有檢查人力不足的現象。為減少成本、降低不 良品、增加客戶滿意度，需要有效方法解決此一問題。 本研究將資料探勘實際運用於面板檢測，面板製程中會蒐集產品的電性參數 等資料。然而因變數眾多，無法從龐大資料中察覺可能導致品質不良的因素。使 用資料探勘技術，建立分類模型，以羅吉斯迴歸、類神經網路、C5.0決策樹，交 叉分析比較實際運用的可行性。研究發現利用C5.0決策樹，配合k-fold演算法， 可得出不良率檢出準確度最高、耗費人力最少之分類法則，成功達到人力成本減 少及品質提升之目的，發揮資料探勘幫助企業追求利潤、提高顧客滿意度之目標。 In the manufacturing process of TFT-LCD, the flat panel ranking mostly depends on human’s diagnosis. The output of flat panels increases because of the establishment of new generation factories and also because of the change of the market for more small-sized panels. As the result, more manpower is needed to handle the diagnostic jobs, and it is the manufacturer’s desire to figure out a way to provide an appropriate solution. The purpose of this research is to apply data mining technologies to propose an alternative that uses less manpower and satisfies the job requirement. The electrical characteristics of flat panels are measured and recorded in the manufacture process, but there are so many variables that it is not easy to find out the causes of poor quality. In this research, the goal is to replace manpower with automatically measured data and to create classification rules without incurring addition cost. Several models, including C5.0 decision trees, logistic regression, and Neural Network are considered in this study. The comparison results show that decision tree (C5.0) can achieve a higher accurate rate in ranking the panels. To apply the proposed model, the required manpower could be reduced, and still, satisfies the diagnosis job for ranking panels.
