Abstract As industrial systems become more complex, accurate and real‐time fault diagnosis is vital for maintaining production safety and efficiency. One significant challenge in fault diagnosis is the high number of false alarms caused by the smearing effect. While the reconstruction‐based approach effectively addresses this problem, it is often time‐consuming due to the need for multiple gradient descent calculations, especially in nonlinear systems. Therefore, this paper proposes a fast and stable fault diagnosis method for high‐dimensional nonlinear industrial processes. This method utilizes an autoencoder algorithm to build a process monitoring model based on normal operational data. It then applies the squared prediction error statistic to detect any abnormalities in the system. If a deviation from the normal state is identified, a reconstruction‐based indicator is used to isolate the fault. The innovation of this method lies in using non‐dominated sorting genetic algorithm II (NSGA‐II) to optimize both the faulty variables and their faulty magnitudes, effectively addressing the ‘smearing effect’. Additionally, it alleviates the high computational burden typically associated with gradient‐based optimization through its multi‐objective optimization strategy. Experimental results on a simulated case and an industrial system show that the proposed method achieves a fault detection rate (FDR) of approximately 90% and a false alarm rate (FAR) of around 2%, all while maintaining stable computational times at the millisecond level, regardless of the number of fault variables. This study offers a reliable solution for real‐time fault diagnosis in complex industrial systems, contributing both theoretical and practical advancements to the field.
Ying et al. (Mon,) studied this question.