Canonical variate residual analysis for industrial processes fault detection

Abstract

With the rapid development of intelligent integration in industrial processes, a challenge emerges. Early failures cannot be detected in a timely manner, potentially leading to significant financial losses. While traditional canonical variate analysis (CVA) methods are effective for dynamic process monitoring, they may lack the flexibility required for early fault detection. To address this challenge, a fault detection method based on canonical variate residual analysis (CVRA) is proposed. CVRA introduces a distinctive residual statistic that preserves critical information about the data. It places heightened focus on the primary components of the data, capturing core features of system changes and enhancing sensitivity to early anomalies. Additionally, by incorporating the geometric properties of the Manhattan distance, it mitigates statistical data errors, thereby improving detection accuracy. Simulation results validate the method's effectiveness in the Tennessee Eastman (TE) process. Furthermore, the successful application of the three-phase flow facility provides a benchmark for evaluation using real process data.