Physics-inspired multiwavelet KAN for interpretable bearing fault diagnosis under harsh operating conditions

Deep learning models for rotating machinery diagnosis often lack interpretability and suffer severe performance degradation under strong noise. To address these limitations, this paper proposes a Physics-Inspired Multiwavelet Kolmogorov-Arnold Network (KAN) for robust and transparent bearing fault diagnosis. Building upon the Wav-KAN architecture, we develop a parallel multiwavelet expert framework utilizing Morlet, Laplace, and Mexhat activations to structurally capture complementary fault signatures. An adaptive evidence-weighting fusion mechanism is further introduced to dynamically balance expert contributions based on input signal characteristics. Extensive experiments across three bearing datasets, including a complex turbopump test rig, demonstrate the framework’s superior noise robustness. Notably, it achieves 92.05% accuracy on the PU dataset under severe −10 dB noise, significantly outperforming state-of-the-art deep learning, wavelet-based, and single-wavelet KAN baselines. Further analyses show that the learned wavelet scales remain close to theory-derived characteristic scales, that fusion weights vary across fault categories, and that decision attribution focuses on localized fault-related signal regions, providing supportive evidence for the interpretability of the proposed framework.