Monitoring And Abnormalities Diagnosis For Multivariate Process

This dissertation studies some problems of multivariate statistical process controlusing joint control technology, support vector machine, and cascade correlation neuralnetwork through computer simulation analysis. The problems include monitoringmean vector and covariance matrix of multivariate process, and quality diagnosis ofmultivariate process for finding the process abnormalities, and locating the factor ofthe out-of-control process. This can improve process products quality and reducequality defects. It is meaningful to up-grate market competitiveness of enterprise.Research will further improve and add some technologies and methods of multivariatestatistical process control and diagnosis. The main contents are summarized asfollows:1. Multivariate process mean vector and covariance matrix monitoring. It is veryimportant to design control charts simultaneously monitoring multivariate processmean vector and covariance matrix, and then identify process assignable factor. Basedon this point, in this dissertation, a single multivariate control chart is presented inorder to effectively monitor mean vector and covariance matrix of multivariateprocess in a big shift interval. This can provide a better application and referencevalue for enterprises with regard to quality control;2. For a special bivariate process in multivariate processes, this dissertationpresents a joint control chart based on T2and VMAX statistics to monitor the process,studies the ARL performance of the joint control chart, and compare it with the T2and|S|charts. The joint control chart can effectively monitor the mean vector andcovariate matrix; meanwhile a joint single variable control chart is designed. Based onnon-central chi-square distribution of the process and the joint chart, diagnosismethod of mean vector shift direction is proposed. This can provide more details ofprocess variation information for enterprise online operation personnel or qualityengineer, in order to recover the process in-control as soon as possible;3. Support vector machine optimization method for monitoring mean shift ofmultivariate control chart. Support vector machine can learn from small samples well,but selecting its kernel function parameter has largely effect on model of performance.On the one hand, using K folding cross validation method to select kernel function parameters of support vector machine, it constructs mean shift diagnosis model ofmultivariate control chart; on the other hand, using particle swarm optimizationmethod to select kernel function parameters of support vector machine, mean shiftdiagnosis model of multivariate control chart is proposed. Finally it analyzes theperformance of model. The results show the presented models can effectivelydiagnose out-of-control process;4. Cascade correlation neural network application in the manufacturing process.Considering advantages and disadvantages of back-propagation neural networkalgorithm, this dissertation proposes multivariate process abnormalities diagnosistechnology based on cascade neural network algorithm, and applies it in multivariatemanufacturing processes. Simulation and empirical analysis suggest that thetechnology can be effective for process abnormalities diagnosis, and diagnosticaccuracy is higher.