整 合 式 智 慧 型 最 佳 化 參 數 設 計 之 研 究

Abstract

參數設計乃是致力於選擇最佳控制因子的條件，以使得產品能夠對於一些不可控制因子(噪音因子:如刀具磨耗、老化與週遭溫溼度變化)的影響降至最低。由於製造業對於技術的提昇以及產品/服務市場的重視，因此參數設計一直是持續改善中的重要課題。長久以來，許多業界人士一直努力的利用實驗測試的方法，希望將真實的作業情況予以模式化，並且找出設計與績效的因果關係。在實務上，田口方法的已廣為產業界所使用，不過此種方法因為只能在已選定的控制因子水準中篩選最適的參數值，因此對於求取真正的最佳參數解仍有持續改善的空間。 在另一方面，雖然在許多文獻中已經對於處理多重反應值的問題提出解決方法，但由於多數的方法皆太過複雜而且需要許多的統計專業背景知識，因此無法廣為實務界所使用。 本論文應用三種整合式智慧型方法以達成參數最佳化的目的，此三種方法為：(1)整合類神經網路與田口方法，(2)整合類神經網路，基因演算法以及指數需求函數，與(3)整合類神經網路, 基因演算法，指數需求函數以及模糊理論。這些方法有下列幾項優點： (1) 第一種方法(整合類神經網路與田口方法)非常適合高度複雜且無法進行實驗設計的製程，該方法利用製程資料進行模擬，不但可以找出重要的製程參數更可以依據不同的客戶需求決定出個別的製程參數值，協助業界提昇產品的品質；(2)本研究所提的三種方法非常適合應用於具有高度複雜性的高科技產業 (例如半導體產業)參數最佳化；(3) 本研究所提的三項參數設計方法求得的最佳解皆屬於連續型數值，比田口方法只能獲取間斷型參數值為佳；(4) 三項參數設計方法都能夠有效的處理多重反應值問題，求出重要的製程參數以及獲得最佳化的參數組合；(5)對於只具備初級統計知識訓練的工程師而言，本研究所提的方法不僅容易應用，對於使用的數據也無需任何的假設條件；(6) 對於包含定性型與定量型的多重反應值問題上，本研究應用模糊理論以整合工程師在製程中的專業知識於參數設計的最佳化工作中。本研究並藉由台灣的三家半導體相關廠商的實際操作結果說明本研究所提的方法具有高度的可行性。

Parameter design is advocated to design a product by selecting the optimum conditions of control factors (factors which we can control) so that the product is least sensitive to uncontrollable factors (noise factors) such as wear, aging, and ambient conditions. The manufacturing interests which are created concern technological advances and the marketability of products and services. Thus, parameter design questions interact constantly with the realm of continuous improvement. Conventionally, Taguchi methods have been widely used in industry; however, this method can only get the optimal under discrete control factors that will lead the real optimum with uncertain. On the other hand, most published literature addressing the multiresponse problem is too difficult to be understood by engineers with limited statistical knowledge; therefore, it is difficult to apply on the shop floor. This dissertation proposes three types of integrated intelligent approach for the optimization of the parameter design: (1) neural networks and Taguchi methods (NN-Taguchi), (2) neural networks, genetic algorithms, and exponential desirability function (NN-GA-EDF), and (3) neural networks, genetic algorithms, exponential desirability function, and fuzzy theory (NN-GA-EDF-Fuzzy). The major advantages of the proposed methods are: (1) Neural networks and Taguchi methods provide a framework of application to help a manufacturing process that is infeasible to conduct the experimental design identifies major parameters and adjusts appropriate parameters settings according to different customer’s requirements; (2) The three proposed approaches of parameter design are very suitable for high-tech industry (ex. semiconductor industry) which involves very complicated process reaction; (3) The proposed approaches can obtain the optimal parameter settings in continuous values, (4) All of the proposed approaches have abilities to cope with multiple responses to identify the key process parameters, and locate multiple optimum, (5) The engineer with limited statistical knowledge is able to apply the proposed approach easily; neither do they need to make any assumption regarding the data set, and (6) The proposed approach using fuzzy theory is intuitively logical and allows the engineer to involve process information into analysis for evaluating the qualitative response problems. Three real examples that were carried out in the semiconductor related factories in Taiwan have demonstrated the practicability of the proposed procedures.