A Study On Manufacture Process Quality Control And Diagnosis Methodology For Customized Product

Customization has b een identified as a competitive strategy by an increasing number of companies. High degree of customization also requires higher quality. The number of custom ized products is few, even single. Com plex customized products also have high dim ensional critical quali ty characteristics and com plex relationship among influence factors, and it is difficu lt to identif y the source of variation . Consequently, trad itional qua lity co ntrol m ethos would no t be eas ily adapted to customization environments. It is natural to expect that quality control and diagnosis issues should be taken into account when deciding upon product customization.This dissertation introduces rough set (RS) theory into support vector m achine (SVM)--currently the b est m achine learning theory about sm all sa mple statistical learning. The purpose of the dissertation is to develop a feature extraction and pattern recognition approach based rough-S VM under few products with few transcendental information and high-dimensional feature. Specific research includes:1 .An algorithm based rough set and Pawlak attribute significance is proposed to extract critical control chart statistic feature. A model is described as a decision table with condition attributes--raw statistic feature of control chart and decision attributes--classification of control chart. The critical feature is extracted according to the importance of condition attributes which is described by'positive region'and Pawlak algorithm.2. A pattern classification m odel and rules are provided by introducing the rough set theory into SVC(support vector classification). A rough margin basedν-SVC(R-ν-SVC)is proposed to train separating hyper -plane and deal with the overfitting problem due to outliers. The R-ν-SVC searches for the se parating hyper-plane that maximizes the rough margin which is defined by upper margin and lower margin. The position of training sample will dec ide if they can be classif ied correctly or not. For binary classifier, the rules can be defined that those samples lie in the right side of the upper hyper -plane are positiv e class and thos e lie in th e left side o f the lower hyper-plane are negative class. For R-ν-SVC multi-classifiers, the 1-ν-1 approach is used for pairwise classifications between negative class and positive class. We also define some equivalence cl asses according to the rough concept of po sitive region, negative region and b oundary reg ion. In add ition, we s how that th e R-ν-SVC approach may reduced storage requirements and is useful to deal with the outliers and noisy data compared with conventionalν-SVC.The R-ν-SVC model also is used in contro l chart abnormal pattern re cognition under the conditions of the sm all sample and fuzzy information to estimate the trend of abnormal conditions and realize on-line quality control in time.3. Finally, a m odel about custom izied products m anufacture quality diagnosis based rough-set is provided which desc ribes the m anufacture processing and establishes the dynamic relationship among influence factor(scondition attributes)and product quality characteristics(decision attributes)in the use of decision tables. By calculating the importance of influence factors,the source of variation is identified.