Industrial process monitoring based on optimal active relative entropy components

Abstract

The multivariate statistic method has been widely applied, but no clear mapping relationship exists between the latent variables and the fault information, which leads to various information in different latent variables. The traditional methods cause the loss of useful fault information and the decline of monitoring performance. In this paper, a process monitoring method based on the optimal active relative entropy component is proposed. Based on global and local preserving projection, relative entropy is selected as the monitoring statistic. The influence of latent variable selection on relative entropy calculation is analyzed, and a relative entropy activity index is proposed to select the relative entropy component beneficial to process monitoring. The parameters are optimized to generate the optimal active relative entropy component for process monitoring using the mayfly algorithm. Finally, a numerical case, Tennessee Eastman (TE) process, and three-phase flow facility are used to verify the superiority of the monitoring algorithm.