Research On Complex Process Monitoring Based On Partial Least Squsres And Slow Feature Analysis

With the advent of the era of big data,research on data-driven process monitoring algorithms has attracted much attention.Considering that the data usually comprise information in the form of multiple variables,multivariate methods are widely used to capture the relations of variables.Particularly,multivariate statistical process monitoring(MSPM)technique is effective for large-scale complex process monitoring using multivariate characteristics and statistical principles.Among MSPM approaches,partial least squares(PLS)is widely used for monitoring key performance indicators(KPIs).The core idea of PLS is to reduce the data dimensions by performing linear transformations on the process measurement and KPIs for extracting the correlation between process measurement and KPI as much as possible.Slow feature analysis(SFA)is a popular MSPM approach for distinguishing real faults incurring dynamic anomalies from normal changes in operating conditions.The core idea of SFA is to reduce the data dimensions by performing linear transformations on the process measurement for extracting the slowly changing information.However,the traditional PLS and SFA rely on the assumption that processes are linear and static,so they cannot address nonlinear and dynamic characteristics in industrial processes,besides,the traditional PLS has difficulties in decompose measurements in to KPI-related and KPI-unrelated parts accurately.To address the above-mentioned issues,an improved nonlinear PLS and a modified dynamic SFA are proposed in this thesis,which are summarized as follows:1.This thesis proposes KPI-based kernel PLS(KPI-KPLS)for KPI-related nonlinear process monitoring.The core idea of KPI-KPLS is a reasonable decomposition of measurements considering the useful information of KPIs,and then measurement space is divided into KPI-related and KPI-unrelated parts based on the general singular value decomposition(GSVD)of calculable loadings.Besides,the fault detectability analysis is proposed for kernel methods,and a criterion for classifying KPI-related faults is proposed.2.To address dynamic issue,the two-step dynamic SFA(TS-DSFA)is proposed for dynamic process monitoring.TS-DSFA uses variable auto-regression model to describe the dynamic processes,the operations are divided in two steps:first,estimate the dynamic structure and extract dynamic components form process data;second,dynamic SFA is applied to monitoring the dynamic components.TS-DSFA inherits the advantage of slow feature analysis that can distinguish real faults incurring dynamic anomalies from normal changes in operating conditions,and maintaining good monitoring performance for dynamic process with noise fluctuations.