SWKAN: An Ante-Hoc and Post-Hoc Explainable Spectral Enhanced Wavelet Kolmogorov–Arnold Networks for Reliable Fault Diagnosis

Deep learning has shown great promise in intelligent fault diagnosis of industrial equipment, yet its reliability in measurement-driven scenarios remains limited by three challenges: high sensitivity to measurement noise, poor representation of minority fault classes under class imbalance, and insufficient feature explainability. To address these challenges, an ante-hoc and post-hoc explainable spectral enhanced wavelet Kolmogorov-Arnold Network (SWKAN) is proposed for reliable fault diagnosis. In SWKAN, a spectral enhancement module (SEM) is designed to suppress noise and highlight informative spectral components of measured signals, and then the wavelet Kolmogorov-Arnold module (WKAM) with learnable scale parameter and translation parameter is proposed to achieve ante-hoc explainable feature extraction. After that, the Bayesian prior-informed LIME (BayLIME) is leveraged to achieve post-hoc feature explanation. Experiments on two measurement datasets are carried out to verify the effectiveness of the proposed methods. Experiments on two measurement datasets are carried out to verify the effectiveness of the proposed methods. The experimental results demonstrate that SWKAN can not only achieves robust diagnostic performance under noisy and imbalanced conditions but also offers both ante-hoc and post-hoc explanations for the diagnostic results. These results further confirm the effectiveness and reliability of SWKAN for measurement-based intelligent fault diagnosis. The library is available at: https://github.com/HazeDT/SWKAN.