SREFNet: A Spectral Residual–Based Evidence Fusion Network for Industrial Fault Diagnosis

Abstract

Fault detection and diagnosis technology is crucial for ensuring operational safety and improving product quality. In recent years, various deep learning methods have achieved high-precision performance in fault diagnosis. However, real-world process data often contains inevitable noise, which significantly degrades the generalization ability of conventional models. Therefore, this paper proposes a spectral residual–based evidence fusion network for industrial fault diagnosis, reducing the influence of randomness, noise, and variable coupling. Specifically, a learnable spectral residual module is first used to mitigate the influence of irrelevant patterns in the frequency domain, facilitating the subsequent capture of temporal and cross-channel dependencies. Furthermore, an evidence fusion module is employed to model the uncertainty of the process data and integrate the latent distribution features adaptively. Finally, extensive experiments on the Tennessee Eastman benchmark process demonstrate the effectiveness of the proposed method, with its robustness and interpretability further validated through visualization and comparative analysis.