應用以田口方法最適化之倒傳遞類神經網路於TFT-LCD Cell製程缺陷分類之研究

TFT-LCD是目前應用最為廣泛之平面顯示器，然而其生產製造過程卻不如銷售般那樣順遂。通常TFT-LCD的誕生是需要經過多道繁瑣的製造程序，因此每當產品有不良缺陷產生時，便很難在短時間內斷定其不良缺陷的成因。此外，這些不良缺陷的分析常仰賴有經驗的工程師，對於新進的職員而言，短時間學會如何分辨不良缺陷著實是一項艱難的任務。從管理層面的觀點來看，訓練及解析問題所耗之時間與成本顯然不符合成本經濟效益，因此，發展或建立一套缺陷分類系統或模型是勢在必行的。本研究經由文獻之探討，發現近年來倒傳遞類神經網路已成功地應用在分類及預測領域上，因此本研究經仔細評估與判斷後，決定採用倒傳遞類神經網路來研究在TFT-LCD Cell製程缺陷分類之問題。為了使網路的表現最適化，本研究採用田口實驗來決定關鍵性之參數，進而選出最適之參數組合。根據數值分析之結果顯示網路參數-轉換函數、學習速率及世代大小之主效應及彼此之交互作用對於網路的表現皆有顯著性之影響，因此在選擇最適參數組合時，需透過二維或一維之參數平均表現反應表或Duncan’s Multiple Range Test來做決定。最後，將最適參數組合應用在倒傳遞類神經網路中，提供給廠商做為未來缺陷分類之依據。

關鍵字

直交表；田口方法；倒傳遞類神經網路；缺陷分類

並列摘要

TFT-LCDs have become one of the most popular flat panel display devices in these days and applied to various fields in the world. However, manufacturing the TFT-LCD panel indeed requires passing through many complex processes. It would be a tough task to analyze the causes of defects when the defects occur. As a result, manufactures need an efficient method to help employees quickly clear the causes of defects and then do the appropriate treatments to avoid resulting in a mass profit loss. In this study, we use BPNN to approach the relationship between defective characteristics and causes of defects so as to solve the defects classification problem in the TFT-LCD cell process. In order to make BPNN perform well, we adopt Taguchi method to study not only the significant parameters but the optimum parameter settings in BPNN. Numerical analysis results show the main and interaction effects of parameters- transfer function, learning rate, and epoch size have significant effects on the performance of BPNN in our case. Besides, the corresponding optimal levels of each significant parameter can be determined by two-way, one-way table, or Duncan’s Multiple Range Test. At last, we extract the final weights and biases from trained BPNN at the optimal condition and then provide this network model with the optimum designs for manufacturers to apply to defects classification problem in TFT-LCD cell process.

並列關鍵字

Taguchi method ； BPNN ； Orthogonal array ； Defect classification

參考文獻

[1] Bishop, C.M. (1995), Neural networks for pattern recognition. Clarendon Press, Oxford, UK.

[2] Bounds, D.G., Lloyd, P,J. (1988), A multilayer perceptron network for the diagnosis of low back pain. Proc. Second IEEE Int''l. Conf. Neural Networks, San Diego, July 24-27, II-481-II-489.

[4] Castillo, P.A., Merelo, J.J., Prieto, A., Rivas, V., Romero, G. (2000), G-Prop: Global optimization of multilayer perceptrons using GAs. Neurocomputing 35, 149-163.

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

[8] Feng, C.X.J., Gowrisankar, A.C., Smith, A.E., Yu, Z.G.S. (2006), Practical guidelines for developing BP neural network models of measurement uncertainty data. Journal of Manufacturing System 25(4), 239-250.

國際替代計量

應用以田口方法最適化之倒傳遞類神經網路於TFT-LCD Cell製程缺陷分類之研究

未授權

主題瀏覽