應用資料探勘技術建構面板異常問題之處理模式

Application of Data Mining Techniques for Building TFT Abnormal Failure Processing Mode

劉志洪(Chih-Hung Liu)

指導教授 ： 鄭春生

元智大學/工程學院/工業工程與管理學系/碩士(2012年)

摘要

由於目前機器尚無法有效靈活地搜尋與判別面板缺陷種類，因此 TFT-LCD 面板製造廠對於客戶退貨異常品的缺陷檢測，仍然仰賴大量人工以繁瑣的檢查、拆解與分析來判斷面板異常之成因。面板製造廠為了維持企業競爭力而減少檢測人員，使得檢測工作緩慢，然而若是未能掌握品質異常之原因並且及時改進，將會造成大量不良品的產生，不利企業之競爭力。
本研究針對台灣某家 TFT-LCD 面板製造廠所提供之異常現象與異常真因數據，提出一種以分群演算法K-means 為基的分類方法。首先利用K-means 分群法將異常現象數據分群，再利用不同的分類方法針對各群資料進行分類，以辨識準確率作為績效指標來比較不同分類方法的績效。研究結果顯示，利用本研究所提出之以分群演算法K-means 為基的分類方法來處理面板異常現象數據，可以有效地提昇分類之績效。本研究之結果能協助業界達成追求利潤、提高顧客滿意度之目標。

關鍵字

K -means ； 分類迴歸樹 ； 倒傳遞類神經網路 ； 徑向基函數網路

並列摘要

Currently, TFT-LCD industry relies heavily on human to manually repair the defected panels returned by the end users. The lack of more advanced machine to assist on quick and accurate trouble-shooting can lead to slow and false responses and inefficient rework. In addition, such inefficient, costly and time consuming manual repairs may often be eliminated by the company for cost-down purpose to stay competitive in the business. The overall result will blind the company in production line correction to improve overall product quality.

The main purpose of this study is to develop a cluster algorithm-based (Used K-means) classification model with the failure data from one of the TFT-LCD makers. First, the study used K-means method to group the failure data. Then, we used different classification methods to classify data of each group. The performance of different classification methods is compared based on accuracy.

According to the test result, the cluster algorithm-based (Used K-means) classification model could improve the accuracy of classification model. Thus, this research offers the industry a mean to assist in improving the customer satisfaction and company profit.

並列關鍵字

K-means ； CART ； RBFN ； BPN

參考文獻

4.Cybenko, G., 1989, Approximation by superposition of a sigmoidal function, Mathematics of Control, Signals and Systems, 2(4), 303-314.

6.Davies, P. C., 1994, Design issues in neural network development, Neurovest, 5(1), 21-25.

8.Eksioglu, M., Fernandez, J. E. and Twomey, J. M., 1996, An artificial neural network prediction model for determining peak pinch strength, International Journal of Industrial Ergonomics, 18(5), 431-441.

9.Fayyad, U. M. and Piatetsky, G. S., 1996, The KDD process for extracting useful knowledge from volumes of data, Commun ACM, 39(11), 27-34.

10.Fish, K. E., Barnes, J. H. and Aiken, M. W., 1995, Artificial neural networks: a new methodology for industrial market segmentation, Industrial Marketing Management, 24(5), 431-438.