Adaptive Process Monitoring of Online Reduced Kernel Principal Component Analysis

Abstract

In the case of dynamic systems, the traditional kernel principal component analysis (KPCA) method does not perform well. The moving window kernel principal component analysis method can adapt to the normal parameter drift of dynamic systems, but it needs a longer computation time when processing large number of samples. Therefore, an adaptive process monitoring method for online reduced kernel principal component analysis is proposed. In this method, a small training set is selected as the initial reduced set in a large number of samples for modeling, and the online real-time collected data are analyzed to judge whether the new sample is normal or not. If it is a normal sample, the method judges whether the sample is added to the reduced set, and updates the online KPCA model automatically when adding to the reduced set. The proposed method is applied to a numerical example and the Tennessee-Eastman (TE) process. The simulation results show that the proposed method is effective and feasible.

Keywords

• kernel principal component analysis (kpca)
• online reduced kernel principal component analysis (orkpca)
• moving window
• process monitoring