Recursive ensemble canonical variate analysis for online incipient fault detection in dynamic processes

Abstract

Data-driven fault detection has made significant advancements. However, detecting incipient faults is still a challenging problem for traditional data-driven methods, because it has low amplitude and is easily masked by measurement noise. In this paper, we propose an approach for detecting incipient faults named recursive ensemble canonical variate analysis (RECVA) for dynamic processes. First, inspired by ensemble learning, multiple CVA-based incipient fault models are obtained based on sampling from the training data set using bootstrap. It generates two matrices based on the principal component and the residual subspace. Then, two sensitive detection metrics are created using the maximum eigenvalue of a one-step sliding window over each row of the two aforementioned matrices. Based on the first-order perturbation (FOP) theory, the eigenvalue can be recursively updated. Utilizing the two detection metrics, RECVA can effectively detect incipient faults, which cannot be detected by conventional statistics. The effectiveness of RECVA is fully validated with the incipient fault data collected from the continuous stirred tank heater (CSTH) and the Tennessee Eastman process (TEP) simulation benchmark. Compared with traditional PCA, CVA, and other incipient fault detection methods based on divergence, simulation demonstrated that incipient faults can be effectively detected at an early stage, which is very important for avoiding serious failures in the industrial processes, resulting in economic losses and casualties.