Research On Process Monitoring Methods Based On Improved PCA

The key to ensuring the stability and safety of industrial processes is fast and accurate real-time monitoring.Principal component analysis(PCA)is one of the most representative methods in the multivariate statistical process monitoring.However,the standard principal component analysis assumes that the process is linear and stationary,and has many restrictions in pratical application.Based on the existing results,the following methods are proposed to remove the restrictions of standard PCA:Kernel PCA(KPCA)introduces kernel function into PCA,which enables PCA to deal with nonlinear data.However,KPCA does not consider the non-stationary characteristic of many real processes.To tackle this problem,this paper introduces the dynamic latent variable model(DLV)on the basis of KPCA and proposes a kernal dynamic latent variable model(KDLV).This method projects data into high-dimensional space through kernel functions,and then extracts the dynamic and static features of the data in high-dimensional space by using DLV.Simulations based on a numerical example and TE processe show that the fault detection rate of KDLV is higher than those of KPCA and DLV for dynamic nolinear process.PCA only considers global variance information and ignores local structure information in data.Global-local structure analysis(GLSA)reduces the dimension of data by considering both global variance information and local structure information,and has an improved fault detection effect compared with PCA.However,this method has the possibility to model noise information and does not consider the autocorrelation of a time sequence.This paper proposes a global-local structure analysis method based on data reconstruction space(GLSA-RS).GLSA-RS extracts principal components to reconstruct data;then performs GLSA on the reconstructed data.This method can not only preserve the advantages of GLSA but also avoid the issue of noise modeling.As an improvement to GLSA-RS,a dynamic GLSA-RS(DGLSA-RS)method is then proposed,which uses an augmented matrix as modeling data like that in DPCA.Simulations based on a numerical example demonstrate that GLSA has the possibility to model noise information,and GLSA-RS has a higher fault detection rate when there is noise or redundancy in the process data.Then,simulations based on TE process illustrate that,when the process is non-stationary,the fault detection rate of DGLSA-RS is higher than those of GLSA-RS,DPCA and three other methods.