Data-Driven Closed-Loop Process Fault Diagnosis With Enhanced Fault Dynamics

In a closed-loop system, feedback control may transfer process faults from source variables to other variables that obscure the diagnosis of the root cause of the fault. Taking the fused magnesium furnace (FMF) as an example, the effects of process faults, e.g., the semimolten condition, may propagate through the controlled system. To address this issue, this article proposes a new data-driven process fault diagnosis method for dynamic processes in the presence of feedback control. A new residual analysis (RA) method is first proposed to extract features of process faults in the feedback-invariant subspace. Moreover, a new fault diagnosis algorithm, which incorporates support vector machines (SVM) with leaky integrate-and-fire neurons (LIF), named LIF-SVM, is proposed. Unlike traditional diagnosis methods which classify each element independently, LIF-SVM efficiently takes into account the fault dynamics. The features of process faults extracted by RA are used by LIF-SVM for diagnosis to constitute the new RALIF-SVM method. Results from experimental studies on a simulated 4$\times$4 dynamic process and a real FMF show that the diagnosis accuracies using the proposed method increase by 10.9% and 9.15%, respectively, compared to the traditional subspace reconstruction method.