應用計算智能技術於製程最佳化

Abstract

田口方法已廣泛地被應用在許多方面以改善產品及製程的品質，田口方法自發明至今，已受到全球工業界和學術界的肯定和尊崇，是一透過工程最佳化的方式來進行品質改善的方法，可用來改善目前的科技、產品與製程設計。其中參數設計（parameter design）更是一般公認最具貢獻的部分。 傳統的參數設計方法只能分析單一品質特性問題；在品質特性是望目特性時，欲使平均值與目標值一致與縮小變異此二目標有時是衝突的，而不一定可以使用田口的兩階段最佳化程序來獲得適當的解。當品質特性為動態特性時，系統敏感度與變異也不見得可以同時達成；且實驗是針對某一些特定水準進行，當因子是連續型的參數時，田口參數設計得到的最佳參數組合在製程上不一定是真正的最佳條件。 本研究針對多品質特性與動態特性問題提出整合步驟使參數設計的問題可以更有效被處理。為了建立更可靠的模式，利用階層式基因演算法（hierarchical genetic algorithm, HGA）演化神經網路以期能得到更可靠的預測結果；利用多目標演化式演算法（multiple objective evolutionary algorithm, MOEA）提供數組最佳解供決策者依據不同的情況選擇最適當的選擇。經過一些案例與實例之分析，結果顯示本研究所題的方法是有效的。

Taguchi’s parameter design is a powerful tool to improve product performance, and widely applied in industry. It is the largest contribution to quality methodology that Taguchi has made. However, Taguchi’s parameter design approach has limitations in practice, it can not obtain real optimal parameter combinations when the parameters are in continuous values, and only identify the most favorable solution among the pre-specified control factor levels. In a multi-responses system, the responses need to be optimized simultaneously. On the other hand, the sensitivity measure and variability measure need to be optimized simultaneously in a dynamic system, but in practice, these objectives are often conflict. In this study, we propose integrating a hierarchical genetic algorithm (HGA) based neural network and a multiple objective evolutionary algorithm (MOEA) to optimize the parameter design problem. Different sets of data either from other researchers' works or case study conducted in this research have been used to demonstrate the effectiveness of this proposed approach