Dynamic kernel scatter-difference-based discriminant analysis for diagnosis of Tennessee Eastman process

Abstract

A dynamic kernel scatter-difference-based discriminant analysis (DKSDA) method, that addresses overlapping and auto-correlated data resulting from different types of abnormal situations, is proposed here for fault diagnosis of nonlinear chemical processes. The proposed method is based on scatter-difference-based discriminant analysis performed in a high dimensional nonlinear feature space that is obtained via nonlinear kernel transformation of a suitably lagged, dynamic representation of the variables. The DKSDA overcomes the singularity problem of within-class-scatter matrix that is encountered in kernel Fisher discriminant analysis (KFDA), by considering scatter difference form of the Fisher criterion. Fault diagnosis is performed by scores classification using the nearest neighbor classifier in DKSDA space. The performance of the proposed method is evaluated by applying it for the isolation of complex faults in the Tennessee Eastman process. The results demonstrate the superiority of the DKSDA over other recently reported nonlinear classification methods.