Fault Diagnosis of Chemical Processes Based on a novel Adaptive Kernel Principal Component Analysis

The kernel principal component analysis (KPCA) is widely applied in fault diagnosis of complex nonlinear chemical processes. However, the cumulative contribution rate which extracts the kernel principal is obtained based on the subjective judgment of expert opinions. Therefore, this paper presents a novel adaptive kernel principal component analysis (AKPCA) based on a moving window integrating the threshold method to adaptively extract the kernel principal. The covariance matrix is obtained based on the kernel function. Then the value of the covariance matrix is adaptively judged by using a moving window integrating the threshold to select the core principal component. Finally, the proposed method is applied in the fault diagnosis of the Tennessee Eastman (TE) process in complex chemical processes. Compared with the KPCA and the KPCA based on the threshold, the results verify that this proposed method can improve the cumulative contribution rate beyond 95%, which accurately find the main factor of the fault diagnosis in the complex chemical process.