Nonlinear Process Monitoring Based on Multi-block Dynamic Kernel Principal Component Analysis

Abstract

Kernel principal component analysis (KPCA) utilizes kernel trick to extract the nonlinear features and has demonstrated its effectiveness in many nonlinear process monitoring systems. However, the traditional KPCA ignores the dynamic property of process data and does not highlight local variable information. To provide better monitoring performance, a fault detection method based on multi-block dynamic KPCA (MDKPCA) is proposed to monitor nonlinear processes. Firstly, mutual information is calculated between different variables and kernel principal components (KPC), which is used to divide the monitored variables into several local variable blocks. Then, by using the idea of exponentially weighted moving average (EWMA), a multivariable dynamic KPCA model is established for each variable block. Finally, Bayesian theory is used to integrate the results of each blocks. The simulation results on the benchmark Tennessee Eastman (TE) process show that the proposed MDKPCA method is superior to the traditional KPCA method.