電漿化學氣相沉積製程動態學習與線上診斷系統

Dynamic Learning and On-line Prognostics System in PECVD Process

黃舜君(Shun-Jyun Huang)

指導教授 ： 章明

中原大學/工學院/機械工程學系/碩士(2007年)

https://doi.org/10.6840/cycu200700811

摘要

本研究主要是結合類神經網路與專家系統組成一個具動態學習功能的智慧型線上診斷系統，用來改善製程中機台參數之漂移（Drifting）現象所造成的誤差。我們主要是利用類神經網路主動學習且可重複訓練的特性並透過模糊規則萃取將類神經網路學習到的知識以規則的型式（IF…THEN…）表現出來，而儲存於專家系統知識庫中。由於系統包含有類神經網路的緣故，所以類神經網路神經元的權重會隨著新資料不斷加入而導致神經元權重向量改變；而經萃取出新規則後，再加上與規則知識庫合構成一智慧型專家系統。我們可以依據萃取出之新規則及其信任函數，進而提供配方調整建議。
在動態學習部分，我們提出一個類神經網路重新訓練的策略，以此來解釋類神經網路中權重向量改變的現象，並以電漿化學氣相沉積機台之製程參數為例，詳細描述此系統的各個步驟。我們總共蒐集了1425筆資料作為類神經網路動態學習訓練樣本，並以此智慧型線上診斷系統來對膜厚作出量化預測，進而對工程師提出配方調整的建議。由實驗結果顯示，誤差率由原先的10.14％經由加入新訓練資料後下降為6.64％，證實類神經網路動態學習能有效改善預測誤差及達到提升良率之效。

關鍵字

專家系統 ； 類神經網路 ； 模糊規則萃取 ； 動態學習

並列摘要

This research mainly combines artificial neural networks (ANNs) and an expert system to constitute an intelligent on-line diagnosis system that has dynamic learning function. The system is used to improve the error margin of the machine parameter which arises due to the drifting during the manufacturing process. We use active learning and re-usable training characteristics of ANNs to extract the knowledge. ANNs has learnt this knowledge by Fuzzy rule extraction method in the (IF...THEN...) form and store them in the knowledge base of the expert system. Because of the system contains ANNs, the weight of the neuron of ANNs will increase continuously with new data. After extracting the new rules, we constitute an intelligent expert system with knowledge base rule. We can provide the adjustment suggestion of the recipe by the new extracted rules and its trust value.
In the part of the dynamic learning, we refer a strategy about the retraining of ANNs. We explain the phenomenon of the weight vector changed in ANNs by this strategy. We take the machine parameter in PECVD process as example to describe each step of this system in detail. We collected 1425 data of the training sample of dynamic learning in ANNs and predicted of the membrane thickness with the intelligent on-line diagnosis system, then putting forward the suggestion for recipe adjustment to the engineers. The experimental results show that the error rate in average membrane thickness measurement decreases from 10.14% to 6.64% when new training data are combined. These confirm that ANNs dynamic learning can improve the error rate of the prediction and the result of yield rate.

並列關鍵字

Artificial Neural Network ； Dynamic Learning ； Expert System ； Fuzzy Rule Extraction

參考文獻

【1】 W. S. McCulloch and W. Pitts,“A Logical Calculus of Ideas Immanent in Nervous Activity,”Bull, 1943.

【2】F. Rosenblatt,“The Perceptron︰A Probabilistic Model for Information Storage and Organization in The Brain,” Cornell Aeronautical Laboratory, Psychological Review, v65, No. 6, pp. 386-408,1958.

【5】 陳宥佐,“應用類神經網路預測電漿化學氣相沉積膜厚與參數調整建議系統,＂中原大學機械所碩士論文, 2005.

【11】 張瑞文,“電漿化學氣相沉積膜厚預測類神經網路模型與規則庫建立,＂中原大學機械所碩士論文, 2005.

【13】R. Setiono, “Extracting Rules from Neural Networks by Pruning and Hidden-unit Splitting,” Neural Computation, Vol.9, Issue 1, pp.205-225, 1997.