應用資料探勘技術於TFT-LCD面板邊框不均勻缺陷之偵測

Applying Data Mining Techniques for the Detection of TFT-LCD Frame Mura Defects

陳姿瑾(Chen, Tzu Chin)

指導教授 ： 蘇朝墩

國立清華大學/工學院/工業工程與工程管理學系/碩士(2016年)

摘要

薄膜電晶體液晶顯示器產業已進入新世代生產技術，由於產品生命週期時間短之特性，面對瞬息萬變且競爭強烈的市場，若能設法在新產品試產階段有效提升良率，將得以掌握關鍵的競爭優勢。然而，隨著製程技術日益先進，許多潛藏因子可能導致面板缺陷，使產品品質下降，單憑工程領域知識或經驗法則已無法有效釐清影響品質不良之原因。
本研究應用資料探勘技術，針對大尺寸面板邊框不均勻缺陷 (Frame Mura defects) 問題，提出一套屬性篩選流程，綜合倒傳遞類神經網路、決策樹、支持向量機與約略集合理論等四種方法來進行缺陷偵測。首先進行屬性篩選，找出影響面板良率之關鍵製程參數；再以分類器驗證篩選結果，並建構縮減模型，供未來診斷預測；最後彙總四種方法篩選之重要變數，依選取頻率來決定屬性相對重要度，作為改善優先順序之參考依據。
本研究以台灣某面板大廠為例，依照所提出之屬性篩選流程進行實證研究。研究結果從原始製程參數中篩選出5項重要屬性；另外，本研究比較四種方法之篩選前後模型分類績效，顯示經屬性篩選後的縮減模型仍保有不錯的分類能力，可在維持準確判斷下刪除不重要的屬性，證實本研究所提出偵測流程之可行性。

關鍵字

薄膜電晶體液晶顯示器 ； 邊框不均勻缺陷 ； 資料探勘 ； 屬性篩選 ； 倒傳遞類神經網路 ； 決策樹 ； 支持向量機 ； 約略集合理論

並列摘要

The manufacturing technology in the Thin Films Transistor-Liquid Crystal Display (TFT-LCD) industry has entered a new generation. Due to the short product cycle time, companies with the ability of maximizing the efficiency to improve yield during new product pilot run are most likely to have the key competitive advantage. However, many hidden factors in the advanced manufacturing processes could cause defects on the panel and result in low product quality. Simply relying on the domain knowledge or rules of thumb is unable to clarify the root causes of quality problems effectively.
This study applies data mining techniques for the detection of large-sized TFT-LCD Frame Mura defects issue, and proposes a general procedure for attribute selection. Four data mining techniques, including back-propagation neural network (BPNN), decision tree (DT), support vector machine (SVM), and rough set theory (RST), are used to select the important attributes from the data to detect Frame Mura defects. In the end, we aggregate the selected frequency of each attribute to determine the relative importance of attributes and provide improvement priorities for the decision makers.
The proposed process was employed to analyze the manufacturing data of a TFT-LCD company in Taiwan. The implementation results showed that five key attributes were identified from the original data and the reduced model still retained the ability to perform well in classification, which demonstrated the effectiveness of our proposed procedure.

並列關鍵字

TFT-LCD ； Frame Mura defects ； data mining ； attribute selection ； back-propagation neural network ； decision tree ； support vector machine ； rough set theory

參考文獻

[11] Hsu, C. Y., Chien, C. F., Lin, K. Y., & Chien, C. Y. (2010). Data mining for yield enhancement in TFT-LCD manufacturing: an empirical study. Journal of the Chinese Institute of Industrial Engineers, 27(2), pp. 140-156.

[1] Su, C. T. (2013). Quality Engineering: Off-Line Methods and Applications. CRC Press.

[2] Kim, M. M., & Lee, S. H. (2014). Automatic detection method for mura defects on display film surface using modified Weber’s law. Journal of Electronic Imaging, 23(4), 043019.

[4] Fayyad, U., Piatetsky-Shapiro, G., & Smyth, P. (1996). From data mining to knowledge discovery in databases. AI magazine, 17(3), pp. 37-54.

[5] Fayyad, U., Piatetsky-Shapiro, G., & Smyth, P. (1996). The KDD process for extracting useful knowledge from volumes of data. Communications of the ACM, 39(11), pp. 27-34.