Advanced Process Diagnosis in Complex Systems Using Nonlinear Variable Reconstruction

Abstract

This paper presents a new nonlinear multivariate statistical process control technique for identifying and isolating the root cause of abnormal process behavior. The new technique is a nonlinear extension to the variables reconstruction technique by (Dunia et al., 1996), based on nonlinear principal component analysis (NLPCA). This work demonstrates that the variable reconstruction (i) affects the geometry of the NLPCA model and (ii) alters the NLPCA based monitoring statistics. Incorporating such changes into the NLPCA model using reference data can address these issues. An industrial application study of a glass melter process shows that abnormal events can be identified and isolated earlier than conventional principal component analysis (PCA).