Multivariate Statistical Process Monitoring Incorporating Fault-Relevant Feature Ex-Traction And Selection

With increasing demand in plant safety and product quality,process monitoring is gaining increasing attention in both academic and industrial communities.Because of the rapid advancement of data gathering,transmitting,and storing techniques,process data become abundant.Data-based,especially multivariate statistical analysis-based process monitoring(MSPM)methods are of particular interests.Modern plant-wide processes are usually characterized by a large scale,multiple operation units,a large number of variables,and complex interactions,and monitoring of such processes becomes a new challenge.For monitoring large-scale processes,the following problems exist in using traditional MSPM methods: First,the existence of fault-irrelevant features may cause monitoring redundancy and degrade the monitoring performance.Second,peocess decomposition is the key step in monitoring large-scale processes,however,traditional methods does not take fault information into account.Third,the joint features throughout the entire process and individual features of a subprocess are not extracted in monitoring a large-scale process with multiple operation units.Based on the existing results,this dissertation makes some further researches and improvements on the data-driven multivariate statistical large-scale process monitoring and proposes several improved monitoring methods from the fault-relevant feature selection aspect.The main contributions are as follows.(1)Principal component analysis(PCA)is one of the most fundamental multivariate statistical process monitoring methods.However,for monitoring a large-scale process,the existence of fault-irrelevant principal components(PCs)may cause monitoring redundancy and degrade the monitoring performance.This dissertation analyzes the impact of PC selection on the PCA monitoring performance,and proposes a fault-relevant PC(FRPC)selection and subspace construction monitoring scheme,which improved the monitoring performance with the help of available fault information.Case studies on a numerical example and the Tennessee Eastman(TE)process demonstrate the efficiency of the proposed monitoring scheme.(2)Dynamic PCA(DPCA)has been widely used for dealing with dynamic pro-cesses.This dissertation analyzes the impact of PC selection on the DPCA monitoring performance,and extends the FRPC selection and subspace construction monitoring scheme to deal with dynamic processes.The DPCA monitoring performance for dynamic processes is improved with the help of available fault information.The proposed monitoring scheme is applied on a numerical example and the TE benchmark process,and monitoring results illustrate the efficiency.(3)Kernel PCA(KPCA)is one of the most popular techniques for dealing with nonlinear processes.This dissertation analyzes the impact of PC selection on the KPCA monitoring performance in the kernel feature space,and extends the FRPC selection and subspace construction method to deal with dynamic processes.The KPCA monitoring performance for nonlinear processes is improved with the help of available fault information.Application examples on a numerical example and the TE process demonstrate the efficiency of the proposed monitoring scheme.(4)Data-driven distributed monitoring is gaining increasing attention in dealing with large-scale processes.This dissertation analyzed the impact of variable selection and process decomposition on the distributed DPCA monitoring performance and the impact of feature selection on the Bayesian fault diagnosis system,and then proposes a performance-driven distributed DPCA monitoring scheme which achieves both efficient fault detection and diagnosis for large-scale dynamic processes.Application examples on a numerical example and the TE process demonstrate the efficiency of the proposed monitoring scheme.(5)A large-scale process is generally characterized by multiple operation units.Monitoring each unit individually may ignore the correlation with the other units whereas monitoring the whole process may involve monitoring redundancy for detecting a local fault.This dissertation proposes a joint-individual monitoring scheme for multiunit processes which incorporates multiset canonical correlation analysis to characterize both the joint feature throughout the process and the individual feature in each unit.The efficiency of the proposed monitoring scheme is theoretically analyzed,as well as experimentally verified.Finally,conclusions are drawn and some future studies on the large-scale process monitoring are discussed.