Multivariate temporal process monitoring with graph‐based predictable feature analysis

Abstract

AbstractDynamic latent variable (DLV) methods have been widely studied for high dimensional time series monitoring by exploiting dynamic relations among process variables. However, explicit extraction of predictable information is rarely emphasized in these DLV methods. In this paper, the graph‐based predictable feature analysis (GPFA) algorithm is introduced for statistical process monitoring due to its explicit predictability, and a novel index, prediction information, is designed to determine the number of its principal components for dimensionality reduction and parameter optimization. A GPFA‐based dynamic process monitoring framework is proposed to differentiate among dynamic faults, normal operating condition changes, and break in relation in the normal data. Case studies on the Tennessee Eastman process and a high‐pressure feedwater heater are conducted to demonstrate the superiority of GPFA over other approaches in terms of fault detection performance.