Multivariate Statistical Process Monitoring Based On Feature Selection

The increasingly fierce market competition makes the industrial process have high requirements on production safety,production efficiency,product quality,energy consumption level and other indicators.How to make the normal work of industrial engineering continue to be stable is a problem that must be solved to ensure production safety and improve production efficiency.Therefore,process monitoring has important practical significance and theoretical research value.Distributed control system(DCS)and other technologies are widely used in the process industry,which makes the data information that can be measured and stored in the industrial process larger and larger.However,it is very difficult to get a precise model to describe the mechanism;how to extract useful information from the massive process data that can be used to monitor the running state of the process has become an urgent problem to be solved.Many scholars have invested in the research and solution of these problems,so the process monitoring based on data-driven is developing rapidly.In particular,multivariate statistical process monitoring(MSPM)has been widely concerned by academia and industry,and has become one of the research hotspots in the field of process monitoring.Although the research results of MSPM have been endless,MSPM method is still in development,there are many problems need to be further discussed and solved,such as how to select effective feature components,avoid the information of variation features being scattered or submerged;how to deal with process data variables being highly nonlinear,etc.based on the previous research work,this paper puts forward self-adaptive for these problems Adaptive principal component analysis(APCA)and adaptive kernel principal component analysis(AKPCA).In view of the problem that the principal component analysis(PCA)is applied to the process monitoring to choose the principal component subjectively,which will cause the fault related information to be scattered or submerged,this paper proposes an on-line adaptive method of selecting the principal component,namely APCA.Firstly,the mutation probability of on-line kernel principal component is calculated by kernel density estimation.The principal component with the largest mutation probability is selected as the dominant variation principal component(DV-PC).The principal components with similar load vectors have similar variation characteristics,so the similarity between kernel principal components is measured by calculating the reciprocal of Euclidean distance between load vectors.The non-dominant variation principal component(NDV-PC)is selected as the principal component with high similarity with DV-PC;DV-PC and NDV-PC are combined into adaptive principal component(APC)to construct T~2 statistics for monitoring.Finally,APCA is applied to numerical simulation and Tennessee Eastman(TE)process to verify its effectiveness.Aiming at the problem that the effective information in the nonlinear process monitoring based on KPCA(Kernel Principal Component Analysis)is dispersed or suppressed,an online adaptive method of selecting kernel principal components,akpca,is proposed.Akpca method can select the kernel principal components which may have variation characteristics online for process monitoring.Firstly,the mutation probability of online kernel principal component is calculated by kernel density estimation,and the kernel principal component with the highest mutation probability is selected as the dominant variant kernel principal component(DV-KPC).Then,the similarity is represented by the mutual information value of the other kernel principal components and DV-KPC.The greater the mutual information value is,the stronger the similarity is.The kernel principal component with larger mutual information value is selected as the non dominant variation kernel principal component(NDV-KPC);DV-KPC and NDV-KPC form the adaptive kernel principal component(AKPC);and use AKPC to construct T~2statistics for process monitoring;finally,verify its effectiveness in numerical simulation and TE process.