應用類神經網路與共通式演算法於參數設計最佳化--- 以積體電路焊線製程為例

Abstract

在快速的市場環境變遷下，促使製造商必須迅速的開發出新產品並提供高品質的商品來滿足顧客善變的需求，以持續保有競爭力。對製造商而言，要同時縮短上市時間（time-to-market）及提昇產品與製程品質，參數設計是一個重要的課題。然而，參數設計最佳化的問題由於系統內存在著高度複雜的非線性關係以及參數間的交互作用，使得最佳參數設計不容易達成。工程師最近常運用田口方法（Taguchi methods）來求得最佳參數設計，藉由內外直交表配置因子以減少實驗次數，並計算訊噪比來估計輸出變異以降低產品損失，不過田口方法在實務應用上仍然有一些限制。本論文企圖運用類神經網加上三種共通式演算法（Meta-algorithm）的技術來進行改善。亦即，使用類神經網路學習田口實驗的結果，模擬系統輸入輸出之間的關係，以突破田口方法在非線性上的限制；再使用共通式演算法求取參數的最佳設計。 本研究提供一個二階段的求解程序並以積體電路焊線製程為例，利用類神經網路、模擬退火法、基因演算法與分散式搜尋法進行最佳化參數設計。最後，本研究亦比較這三個搜尋演算法的有效性及優缺點。

The fast change of environment makes manufacturers have to promptly develop new products and provide high quality products to meet customers’requirements so as to keep the competitive edges. Parameter design is critical for manufacturers to simultaneously achieve both the time-to-market reduction and the quality enhancement of the products and processes. However, the parameter design optimization problems are difficult owing to that nonlinear relationships exist in the system and interactions may occur among parameters. Engineers conventionally apply the Taguchi methods to optimize parameter design; however, the Taguchi methods has some limitations in practice. This thesis attempts to release those limitations by using neural network (NN), and three meta algorithms. First, neural network is used to simulate the relations between inputs and outputs; this procedure can resolve the problem of Taguchi’s nonlinear limitation. Next, three kinds of meta algorithms are utilized to obtain the optimization of parameter design. This thesis provides a two-step procedure and a case of IC wire bonding process using NN, simulating annealing (SA), genetic algorithm (GA), and scatter algorithm (SS) to optimize the parameter design. Finally, these three meta algorithms are also compared .