Research Of Multivariate Process Monitoring Based On Support Vector Machine

In order to meet the market demand for diversified products,the manufacturing process is developing towards complexity,which leads to a sharp increase in the amount of information in the manufacturing process.In such a multivariate process,the problems of traditional control charts are becoming more prominent.In recent years,with the development of information technologies and data processing methods,machine learning has developed rapidly and has become an important method to process large amounts of data.The control charts based on machine learning methods have also become a research hotspot.In researches of control charts based on machine learning methods,the SVM-based control chart has advantages of small sample size,continuous statistics and distribution free.However,it has the problems of selecting the reference data randomly and being affected by the model parameters significantly.The parameters of individual support vector machine are fixed and cannot guarantee a good performance in all tests.In this thesis,the multivariate process monitoring problem is taken as the research object.Based on the support vector machine,combined with the reference data screening and ensemble learning method,a systematic multivariate process monitoring scheme is proposed.The control chart model proposed in this thesis is divided into three stages.In the first stage,support vector data description method is used to select the representative reference data from the phase I in-control population to replace the original randomly-selected method.Then,the SVM is used as the classifier in RTC method to classify the reference data and window data,and get the statistics by some conversion methods.In the third stage,an ensemble SVM control chart is constructed by combining the ensemble learning method with SVMs of different parameters to achieve a better performance for different degrees of process shifts.The experiments of simulations and a real case show that the control chart model proposed in this thesis has a better monitoring performance than the random forest and the original SVM control charts.