Studying Of Quality Control Methods For Process Of Automatic Machining Based On SVM

Manufacturing process quality control is one of the important contents of TQC(Total Quality Control), which is the key mean to ensure final product quality. StatisticProcess Control (SPC) based on control chart has long been used as the main tools forprocess quality control. However, with the continuous improving of automation andcomplexity of manufacturing process, SPC no longer fairly meet the needs of processquality controling on modern production conditions. Support vector machine (SVM),which is newly developed in recent years, has been proved to have many advantagesover Artificial Neural Network (ANN). Introducing SVM into the quality control inmanufacturing process is now increasingly valued. This paper is to further exploreapplication of SVM to quality control in gear machineing process, with considering theneeds of the project ’Networked management and intelligent monitoring system for gearautomatic production line’ sponsored by Ministry of national science and technique. Themain contents of this paper are as follows:1）Control chart pattern recognition based on feature fusion and SVM. A newmethod of control chart pattern recognition for univariate process is proposed in thisstudy. In this method, the optimal fusion of raw features, shape features and statisticfeatures of control chart pattern is taken as classification feature, and SVM is used asrecognizer. The optimal feature fusion is got by reduction on the combination of variousfeatures to remove redunant and irrelevant features with reduction algorithms, i.e.KNNC and ReliefF. The control parameters of reduction algorithms are optimized alongwith the hyper-parameters of SVM by an intelligent optimization model includingParticle Swarm Optimization (PSO) and cross validation, which is to ensure the reducedfeature set more conveient for classification and the SVM have more generalizationability. The simulation experiment results show that, the proposed method can earnhigher precision.2) Estimation of process abnormality parameters based on optimal multi-kernelSupport Vector Regression (SVR). This study proposes to construct the kernel functionof SVR for estimating process abnormality parameters through convex combination ofthree types of basic kernel functions. The key parameters of these basic kernel functionsand the weight coefficient of each basic kernel function, along with the punishcoefficient of SVR, are optimized simultaneously through an intelligent optimization model including PSO and cross validation. The multi-kernel SVR is used with CUSUMchart to monitor mean shift and estimate its magnitude in process. Simulationexperiment has been conducted, and the result indicated that it perform mean shiftmagtitude estimation more precisely.3） Recognition of out-of-control sources in multivariate process based onoptimized Directed Acyclic Graph SVM (DAGSVM). Aiming at the difficult thatcommon multi-class SVMs can hardly provide high accuracy and efficiencysimutaneously for pattern recognition of multivariate process, which have lots ofpatterns when there are many kinds of abnormalities exist, the DAGSVM is studied forimproving its recognition accuracy. Two optimization approaches are proposed. One isheuristic optimization of the topology structure of DAGSVM based on the averagedifference measure between each pattern class with the others in kernel space, whichcan alleviate the error acumulating of DAGSVM to certain extent so as to improverecognition precision. The other approach is to construct selective ensembel withmultiple DAGSVMs through Binary Particle Swarm Optimization (BPSO) algorithma,which can get better generalization ability for improving recognition precision.Simulation experiment results verified the effectiveness of the proposed approaches.4) On-line monitoring and diagnosing of mean shift in multivariate process basedon hybrid SVM model. Aiming at the requirements of multivariate process control, i.e.,to detect mean shifts, to identify shift variables and to quantify variable shift magnitudesimultaneously, this study proposed a hybrid SVM model. This model contains SVMand severl SVRs. The SVM is used for recognition of the process state patterns, basedon which to detect abnormalities and identify abnormal variables. The SVRs are used toestimate shift magnitude of each variable. Simulation experiments results indicate thatthe proposed model perform better than conventional MSPC control chart andANN-based model for detecting mean shift with little size and identification of shiftvariables. In addition, this model can estimate mean shift magnitude precisely.Based on the above findings, and according to the needs of the project, a prototypesystem for intelligent process quality control is developed. It ensures the efficiency ofSVM for process monitoring and diagnosing through integration of.NET platform withMATLAB software. Its structure is designed as combination of C/S and B/S to meetvarious functional requirements of process quality control in the automatic gearproduction line. The system has been applied in certain gear factory and achieved initialresults.