Improved dynamic kernel principal component analysis for fault detection

Abstract The dynamic kernel principal component analysis (DKPCA) has attracted significant attention with regards to the monitoring of nonlinear and dynamic industrial processes. However, DKPCA generally observes decreased performance when the number of samples is large. To address this problem, we propose to use feature vector selection (FVS) to improve the original DKPCA (FVS–DKPCA) in order to enhance its performance and reduce computational complexity in fault detection. FVS is used to solve the redundancy problem of the observation variable, while preserving the geometric structure of data, which is directly related to monitoring performance. Through extensive simulations on the Tennessee Eastman process simulator, it is demonstrated that the proposed fault detector can significantly improve the detection accuracy with less computational cost.