Robust probabilistic predictable feature analysis and its application for dynamic process monitoring

Dynamic process monitoring with multivariate statistical methods has been widely researched and applied for anomaly detection in dynamic systems. However, most of them are designed under the assumption that the measurement noise follows a Gaussian distribution, which usually contradicts the actual situations. Alternatively, in this paper, a novel mixture of Gaussian and Student's t distributions is designed to account for the system noise, and a robust probabilistic predictable feature analysis (RPPFA) algorithm is proposed to capture both the process dynamics and the characteristic of the heavy tail from multivariate temporal data. Moreover, an improved Expectation-Maximization algorithm and a modified Kalman filter method are proposed for parameter optimization. Three monitoring statistics are designed to identify abnormal conditions, which are Hotelling's T2, squared prediction error and dynamic index. The superiority of RPPFA is demonstrated through two case studies, namely, a simulated three phase flow facility and an industrial medium speed coal mill. (c) 2022 Elsevier Ltd. All rights reserved.