Multi-block statistics local kernel principal component analysis algorithm and its application in nonlinear process fault detection

Abstract

It is vital for fault detection technology to extract features of industrial process data effectively. Local kernel principal component analysis (LKPCA) has proved its good performance in preserving global and local structural characteristics. However, it ignored useful high-order statistics of data, so multi-block statistics local kernel principal component analysis (MSLKPCA) algorithm integrating statistics pattern analysis (SPA) into LKPCA is proposed. The correlation coefficient matrix is first calculated and K-means clustering is adopted to divide the original variables into several blocks. Then the weighted SPA, which gives different weights to different samples in each window according to their distributions, is adopted to build statistic spaces containing both low-order and high-order statistics. After that, LKCPA is performed in each statistic space to realize feature extraction. To reduce the noise effect amplified by SPA, PCA is adopted in the residual space to remove noise. Bayesian strategy is used to fuse the results of each block and two monitoring statistics ET and ER are proposed to monitor the feature space and the residual space respectively. The Tennessee–Eastman (TE) process simulation shows the effectiveness and superiority of the proposed algorithm for process monitoring.