A physically interpretable sparse representation-based auto-encoder network for impact fault extraction
Zheng Yuan, Guolin He, Weihua Li, Zheng Chen, Zhuyun Chen · 2025
Extraordinary feature extraction ability renders neural network a powerful weapon for fault diagnosis. However, the output of the intelligent model is unreliable because the extracted feature does not own explicit physical meaning. To counter, a physically interpretable sparse representation-based auto-encoder (SRAE) network for credible impact fault feature (IFF) extraction is proposed. Firstly, an SRAE module is designed by combining auto-encoder with sparse representation principle. The adaptive optimization enables SRAE to extract the IFF parameters from the time-domain signal. The IFF parameters extraction and reconstruction process own explicit physical meaning and hence can be interpreted from the fault mechanism perspective, which provides the basis for the network performance to ensure its reliability. Secondly, combined with the module, the screening process of natural frequency and damping ratio is added to achieve the high-precision IFF reconstruction and parameters extraction. Finally, simulation and experiment are conducted for verification. The results have demonstrated that the proposed method yields less reconstruction error and higher accuracy of parameters recognition than the comparable sparse representation method, reflecting the effectiveness and superiority of the proposed network.