應用類神經網路與乏晰理論於線外品質最佳化之研究

Abstract

製造廠商在競爭激烈的市場中若要佔有優勢的地位，就必須對製程或產品進行品質改善(quality improvement)。而參數最佳化法(parameter optimization)則是達成製程改善或品質改善的一項有效技巧。在進行參數最佳化時，實驗設計(experimental designs)和田口方法(Taguchi method)是目前工業界最常使用的兩種方法。有鑑於使用者之需求日漸增加，產品的功能也隨之增加，因此產品之設計也愈來愈複雜，在進行參數最佳化時，經常需考量到應用動態特性(dynamic characteristic)的分析方式。而田口動態方法是目前唯一可解決動態特性之參數最佳化法，應用田口動態方法來進行參數最佳化時只能找出最佳的參數水準設定條件，如果考量的參數是連續的型式時，應用田口動態方法來進行參數最佳化將無法找出最佳的參數設定值。此外，由於製造產品的複雜性增加，單一品質特性的好壞已不足以用來衡量產品的好壞，因此在進行參數最佳化時，經常會考量到同時最佳化多個品質特性(multiple responses)。一般在分析多個定量式(quantitative characteristic)的品質特性時，最常使用的方式是先針對個別品質特性進行研究，再利用工程經驗(engineering experience)與工程知識(engineering knowledge)來妥協(trade off)多個定量式品質特性的分析結果，利用這種方式有時將無法決定出最佳參數組合。此外，隨著產品複雜度的增加，所考量到的多個品質特性中可能會同時存有定量和定性式品質特性(qualitative characteristic)的情形，然而對於這種混雜著定量與定性式品質特性的問題時，目前的中外文獻中尚未見到有任何的解決程序提出。再則，若所考量的品質特性只是單一定性式之品質特性，目前只有少數的學者提出相關的研究，例如：、田口的累積分析法(accumulation analysis, 簡稱AA法)、Nair的分數設計法(scoring scheme, 簡稱SS法)、Jean和Guo的加權機率分數設計法(weighted probability scoring scheme, 簡稱WPSS法)，然而這些研究方法均未考慮到工程師的工程知識或經驗。但是從工程分析的觀點來看，利用工程知識與工程經驗來衡量定性式之品質特性確實可以提供許多有用的資訊。利用智慧型的技術(intelligent techniques)即可將工程知識或經驗納入此類型問題之最佳化的分析程序中。乏晰理論(fuzzy theory)和類神經網路(artificial neural networks)是最常使用的兩種智慧型技術，乏晰理論對於處理定性式的評價或是語意性措辭之評價是一種很有效的技巧，而且乏晰理論之技巧已經被廣泛地應用於解決有關工程、商業和醫學等科學之問題且具有相當的成效性。因此在本論文中，我們也引入乏晰理論來整合工程師對工程的主觀判斷，並將之納入參數最佳化之分析中。類神經網路目前已被廣泛地應用在許多領域上，應用範圍從傳統的分類(classification)和辨識(pattern recognition)以至於最佳化(optimization)和控制(control)，因為類神經網路特有的平行運算處理方式、特殊的記憶回想方式和容錯能力等，使得類神經網路在應用上較傳統的統計技巧來的容易。在本論文中，我們也引用了類神經網路的模式化(modelling)技巧來找出參數最佳的連續設定值(optimum continuous setting values)。本論文利用類神經網路或乏晰理論共發展五個線外品質最佳化之方法，同時輔以五個積體電路業界(integrated circuit, 簡稱IC)之實例來說明及驗證所提之方法的有效性和合理性。

Quality improvement is an essential work for manufacturing organizations competing in the global marketplace. Parameter optimization is an efficient technique to achieve quality improvement. Until now, experimental designs and Taguchi methods are two useful techniques to optimize the process parameters. Due to different requirements of customers and the complexity of a product, the dynamic analysis becomes popular for manufacturers to optimize their product/process. Taguchi proposed a dynamic method to resolve such dynamic problem. However, only optimum level settings can be determined since using Taguchi's dynamic method. Furthermore, customers normally consider more than one quality response in most manufactured products. During analyzing a multi-response problem involving several quantitative responses, the technique which separately optimize each response and then applying engineering judgments or engineering experience to make trade off the final decision is frequently used. However, it will lead to the uncertainties during making decision. As the product and process becoming increasingly complicated, multiple quality responses may involve simultaneously the qualitative and the quantitative quality responses. However, the fact that optimization of such a multi-response problem with qualitative and quantitative quality responses has not been studied. Besides, the interested quality response may only be a single qualitative characteristic in some cases. Optimization of a single qualitative quality response has seldom been mentioned due to the difficulty of analyzing the qualitative characteristic. From the viewpoint of engineering's analysis, the engineering's subjective information for evaluating the qualitative quality response should be involved into the analysis of parameter optimization. In lieu of above considerations, developing the techniques based on the artificial intelligence (AI) should be an important issue. Fuzzy theory is a well-known intelligent technique to handle the uncertainties of the qualitative type or linguistic description. Fuzzy theory is now applied to some problems in engineering, business, medical and related health sciences, and natural sciences. In this dissertation, fuzzy theory is employed to cooperate the engineering's subjective judgment into the analysis of parameter optimization. Equally, artificial neural networks (ANNS) is another well-known intelligent technique which has been used in a wide variety of applications, ranging from classification and pattern recognition to optimization and control. ANNS also provides a generic technique for the realization of continuous mapping, providing several advantages compared to conventional statistical techniques. In this dissertation, we will also employ ANNs technique to determine the optimum continuous setting values during parameter optimization both single response and multi-response problems. Five separate approaches based on ANNs or fuzzy theory are proposed in this dissertation and, especially, five illustrative examples are also employed to demonstrate the effectiveness of these proposed approaches. 英文摘要……………………………………………………………………..………. iii 誌謝…………………………………………………………………………..………. v 目錄…………………………………………………………………………..………. vi 表目錄………………………………………………………………………..………. ix 圖目錄………………………………………………………………………..………. xii CHAPTER 1 INTRODUCTION…………………………………………..……….. 1 1.1 Overview……………………………………………………………..…….. 1 1.2 Problem Statements……………………………………………..……….… 2 1.3 Research Objectives……………………………………………..……….... 4 1.4 Organization……………………………………………………..……….... 6 CHAPTER 2 LITERATURE REVIEW……………………………………..……... 7 2.1 Optimization Techniques for Taguchi's Dynamic Characteristics……..….. 7 2.2 Optimization Techniques for Multiple Quality Response…………..…….. 8 2.3 Optimization Techniques for a Single Qualitative Response………..…….. 12 CHAPTER 3 BACKGROUND INFORMATION…………………………..….….. 13 3.1 Taguchi's Dynamic Method………………………………………..……… 13 3.2 Artificial Neural Networks (ANNs)………………………………..……… 14 3.3 Fuzzy Theory…………………………………………………..….….. 17 CHAPTER 4 THE PROPOSED APPROACHES………………………..…….….. 21 4.1 ANNs Approach to Quality Improvement in Taguchi's Dynamic Method………………………………………………..…………. 21 4.2 ANNs Approach to Optimize the Multiple Response………………..……. 24 4.2.1 Involving Several Quantitative Quality Responses………..………… 24 4.2.2 Involving Quantitative and Qualitative Quality Responses…...…….. 25 4.3 Fuzzy Approach to Optimize a Single Qualitative Quality Response………………………………………………..…………. 29 4.4 Combining ANNs and Fuzzy to Optimize a Single Qualitative Quality Response…………………………………..………….. 33 CHAPTER 5 ILLUSTRATIVE EXAMPLES…………………………..………….. 40 5.1 Introduction……………………………………………………..………….. 40 5.2 Taguchi's Dynamic Experiment…………………………………..……….. 40 5.2.1 Introduction…………………………………………………...……… 40 5.2.2 The Analysis Result of the Proposed Approach…………..…………. 42 5.2.3 The Analysis Result of the Taguchi's Method ……………..………... 43 5.2.4 The Comparison……………………………………………..……….. 44 5.3 A Multiple Response Problem Involving Several Quantitative Quality Response………………………………………………..…………. 44 5.3.1 Introduction…………………………………………………..………. 44 5.3.2 The Analysis Result of the Proposed Approach……………………... 45 5.3.3 The Comparison…………………………………………….………... 46 5.4 A Multi-response Problem Involving the Quantitative and Qualitative Quality Response…………………………………….………... 47 5.4.1 Introduction…………………………………………………………... 47 5.4.2 The Analysis Result of the Proposed Approach……………………... 49 5.4.3 The Comparison……………………………………………………… 53 5.5 An Ion Implantation Process Involving a Single Qualitative Quality Response…………………………………………………………... 54 5.5.1 Introduction…………………………………………………………... 54 5.5.2 The Analysis Result of the Proposed Fuzzy Sets Approach…………. 55 5.5.3 The Analysis Result of Taguchi's AA……………………………….. 59 5.5.4 The Analysis Result of Jean and Guo's WPSS………………………. 60 5.5.5 The Comparison……………………………………………………… 61 5.6 A Lead Frame Manufacturing Process with a Single Qualitative 62 Quality Response…………………………………………………………... 62 5.6.1 Introduction…………………………………………………………... 62 5.6.1.1 Definition of Quality Characteristic……………………………. 64 5.6.1.2 Recognition of Current State…………………………………… 64 5.6.1.3 Determining Experimental Factors and Experimental Design…. 65 5.6.2 The Analysis Result of the Proposed Combined Approach………….. 65 5.6.3 The Comparison……………………………………………………… 71 CHAPTER 6 CONCLUDING REMARKS………………………………………… 73 REFERENCES……………………………………………………………………….. 76