Data-Driven Closed-Loop System Fault Diagnosis Using Feedback-Invariant Dynamic Residual Analysis

Residual analysis is widely used for fault diagnosis of closed-loop systems. However, traditional residual analysis-based closed-loop diagnosis methods mainly make use of static residual that may be obscured by the effects of feedback control. To this end, this article proposes a novel data-driven closed-loop system fault diagnosis method based on feedback-invariant dynamic residual analysis. First, a new stacked$\boldsymbol{d}$th-step prediction (SDSP) that effectively eliminates accumulated prediction errors of traditional multistep prediction methods is developed to extract dynamic residuals. Subsequently, the fault features are extracted from the dynamic residuals according to feedback-invariant principle. The corresponding fault isolation strategy is proposed. The experiments on a simulated 4$\boldsymbol{\times}$4 dynamic process and a real fused magnesium furnace demonstrate the effectiveness of the proposed method.