應用類神經網路與模糊分群法偵測化學氣相沈積之設備故障

Applying Artificial Neural Networks and Fuzzy Clustering Method for Fault Detection of CVD Equipment

柯佳賢(Chia-Hsien Ko)

指導教授 ： 張耀仁

中原大學/工學院/機械工程研究所/碩士(2010年)

https://doi.org/10.6840/cycu201000119

摘要

目前面板產業朝向大尺寸TFT-LCD(Thin Film Transistor-Liquid Crystal Display)面板發展，而半導體製造技術也遵循著摩爾定律(Moore's Laws)，縮小特徵尺寸(Feature Sizes)、加大晶圓尺寸、增加元件的密集度，而致使製程日趨複雜。當製程邁入ULSI紀元，製程控制也因而日益困難。製程故障偵測的主要目的是提升設備的使用率與產能、改善製程，提升產品品質，以維持具競爭性的生產能力及生產品質，製程故障偵測將成為一個相當重要的課題。
本論文針對TFT-LCD(Thin Film Transistor-Liquid Crystal Display)前段的Array薄膜製程，以PECVD設備之射頻供應系統的製程量測資料作為研究。故障偵測技術是利用類神經競爭網路中的Kohonen Network和模糊分群法(Fuzzy Clustering)中的Fuzzy C Means演算法則擷取資料的特徵關係，再利用Ellipsoidal Calculus訂定出安全區域與警戒區域，建構一個訓練模型來偵測半導體製程中的故障異常。
在實驗模擬中，利用22筆的正常資料做為Kohonen Network特徵神經元的訓練資料，再利用Fuzzy C Means對特徵神經元做選取的動作，將瀕臨異常的特徵神經元挑取出來，最後以Ellipsoidal Calculus訂定出訓練模型的安全區域與警戒區域。然後再以10筆正常資料測試訓練模型的歸納性，如果有誤判的正常資料，將重新加入誤判的資料到訓練資料加以修正模型的歸納性。最後以各為9筆的反應室、電磁閥、匹配器與射頻供應器等四種的故障型態資料當作Kohonen Network的測試資料，驗證此製程故障偵測與診斷工具在TFT-LCD薄膜製程上的可行性。

關鍵字

模糊C均值法 ； Kohonen 網路 ； 故障偵測 ； 橢圓面計算法

並列摘要

Currently larger panels are manufactured in the TFT-LCD (Thin Film Transistor-Liquid Crystal Display) fabs and, moreover, the semiconductor manufacturing technology, following Moore's Laws, is developed for reduced feature size, larger wafer size, and higher device integration. As the fabrication process comes into the ULSI era, process control becomes more important. The purpose of fault detection of fabrication process is to increase the utility rate of equipment and to improve the product quality. In order to maintain competitive production capacity and product quality, process fault detection is a very important issue.
This thesis studies the thin-film fabrication in the array process of TFT-LCD. Data obtained from the RF supply system of PECVD is studied. This proposed fault detection tool will capture the characteristics of RF signals using Kohonen competitive network and Fuzzy C Means. Then, the secure and warning areas of training model are defined by ellipsoidal calculus. This constructed model can be utilized to detect any malfunctions and faults occurred during the semiconductor manufacturing.
In our experiment, 22 sets of normal data were used as the training data for Kohonen network and, furthermore, these obtained feature neurons were classified into two groups by Fuzzy C Means in order to filter out the scattered neurons. Then, the secure and the warning areas of training model were calculated by the ellipsoidal calculus. Next, 10 sets of normal data were used to test the generalization of the training model. If any erroneous judgments occur, the model must be re-trained by adding these data for improving the generalization. Finally, four different kinds of faults, including chamber, solenoid valve, matching box and generator, were examined using 9 sets of data per faulty type. The feasibility of fault detection tool was verified.

並列關鍵字

Fault Detection ； Kohonen Network ； Fuzzy C Means ； Ellipsoidal Calculus

參考文獻

[12] 古典諭，“基於射頻訊號特徵萃取之電漿蝕刻設備故障偵測”，中原大學機械工程學系碩士學位論文。

[15] 陳源杰，“應用小波理論於薄膜製程之射頻功率故障偵測”，中原大學碩士論文，2004。

[1] C.J. Spanos, H.F. Guo, A. Miller, and J.L. Parrill, “Real-Time Statistical Process Control Using Tool Data,” IEEE Trans. on Semiconductor Manufacturing, Vol. 5, No. 4, Nov. 1992.

[4] X.A. Wang, and R.L. Mahajan, “Artificial Neural Network Model-Based Run-to-Run Process Controller,” IEEE Trans. on Semiconductor Manufacturing, Vol. 19, No. 1, Jan. 1996.

[5] B. Kim and G.S. May , “Real-time diagnosis of semiconductor manufacturing equipment using a hybrid neural network expert system” , IEEE Trans. Comp. Pack. Manufact. Technol. Vol20, No1, pp39-47, 1997.