Research On Multivariate Statistical Process Monitoring Method Based On Data-Driven

With the rapid development of the times,the scale of industrial production is gradually becoming more and more complicated.In order to ensure the safe,efficient and stable operation of the modern process,prevent the safety accidents caused by minor faults in the process,and protect the safety of workers’lives and properties,the monitoring technology of chemical process becomes more and more important and gradually becomes an extremely important part of daily industrial production.The rapid development of computer technology combined with the continuous maturation of sensor technology has enabled the preservation of large batches of offline data produced by industrial processes,which has driven the development and application of multivariate statistical process monitoring technology.The process monitoring technology based on multivariate statistics can analyze the high latitude data generated in the process of process production,and establish the algorithm model through offline processing,and get the projection matrix.After the online monitoring phase,the online data can be monitored constantly for failure to ensure the normal operation of industry.In this paper,aiming at the limitations of the existing multivariate statistical monitoring algorithm,we improve it by combining some machine learning methods and data processing methods.The main contents are as follows:(1)There are many redundant feature variables for the high-dimensional data characteristics.A dynamic global local preservation projection method(GA-DGLPP)based on GA feature selection is presented,which combines genetic algorithm,feature selection and GLPP.This method effectively considers the dynamic characteristics of the data,the global and local characteristics of the data structure,and the integrity of the data.At the same time,a window sliding denoising algorithm is used to preprocess the original data.The monitoring effect and fault diagnosis rate are improved.(2)In view of the limitations of the PCA algorithm,the article combines the idea of hierarchical clustering analysis algorithm,DPCA method and intensive training,and puts forward an enhanced dynamic principal component analysis method(HCA-DPCA)based on hierarchical clustering analysis.This method effectively considers the dynamic characteristics of data,and improves the DPCA algorithm through the classified intensive training,which improves the monitoring effect and the fault diagnosis rate.(3)For the above two methods,before the establishment of the multivariate process monitoring model,the data are pre-processed,including extended dynamic matrix,GA feature selection and cluster analysis hierarchy,which makes the dimension of the input data change,the traditional contribution graph method fails.Therefore,this paper improves and designs a new contribution graph method for fault diagnosis.Finally,the improved method is validated by two simulation process objects.The comparison results show that GA-DGLPP method has better monitoring effect on T~2statistics,that is,it is more sensitive to the principal subspace feature variables.The HCA-DPCA method greatly improves the diagnostic performance of PCA and DPCA as a whole by improving the small faults that are not easily monitored by traditional methods.