GMDH神經網路之田口式實驗模式的建立與其在半導體品質設計上的應用

A novel application of a GMDH neural network for modeling Taguchi design process — a case of quality design for semiconductor manufacturing process

陳清風(Ching-Feng Chen)

指導教授 ： 鍾雲恭

元智大學/工程學院/工業工程與管理學系/碩士(2001年)

摘要

隨著日趨競爭的工商環境，如何提昇產品的品質與降低成本是提昇企業競爭力的兩項重要因素，特別是在半導體相關產業中，製程的改善，代表良率的提昇，而如何在不需大量增加成本與時間下，找出較佳的製程參數設定條件亦是本文主要研究目的。
本文利用GMDH的自組性網路架構結合田口實驗設計的直交表與SN比，以最少的實驗次數，找出製程中顯著的因子與因子間的交互作用，配適出符合製程的多項式模型。

關鍵字

田口方法 ； GMDH ； 神經網路 ； 穩健性 ； 直交表 ； S/N比

並列摘要

Taguchi method is popular means to design robust products and processes. Although many industries have successfully used it, its real benefits were not realized and fully understood in many cases. This lack of success could be attributed to some certain factors, but mainly because the experiments were treated in isolation and not integrated into a continuous improvement strategy. To overcome these disadvantages, this thesis presents the results of the novel application of the GMDH neural network to learn Taguchi methodology used to the optimal design process. As a design tool, the GMDH is trained to model an optimal representation of the design interrelationship between design factors and a responsive variable. The predictive performance of the proposed neural Taguchi system using the GMDH network has been proved to be superior and robuster to the traditional Taguchi method.

並列關鍵字

Taguchi Method ； Group Method of Data Handling ； Neural Network. ； Robust ； Orthogonal ； Signal - to - Noise ratio

參考文獻

linear Polynomials in Identifying Human Isometric Strength

Behaviour,” Computers and Industrial Engineering, Vol. 8,

to Predict Human Dynamic Strengths,” Computers and

Industrial Engineering, Vol. 9, No. 4, 371-377.

“Statistical Behaviour of the GMDH Algorithm,” Biometrics,