The traditional wavelet transform relies on a predefined wavelet kernel function, which is difficult to adapt to various signal types. The neural network can adaptively extract embedded features of signals, but the black-box nature limits its application in industry. This paper proposed an interpretable adaptive feature extraction distillation network (AFEDN) that integrates adaptive feature extraction with knowledge distillation. During the AFEDN training process, a teacher network integrated with the adaptive signal feature extraction layer is first trained to obtain complex interpretable time-frequency domain feature representations. In this layer, the original vibration signal is adaptively decomposed into different sub-bands, and each sub-band is weighted and fused to enhance the discrimination of significant features and the anti-noise performance. Subsequently, a knowledge distillation strategy combining soft and hard targets is employed to efficiently transfer the feature representation of the teacher network to a lightweight student network. The diagnostic effectiveness of the proposed AFEDN is verified in compound fault diagnosis tasks of the planetary gearbox dataset. Additionally, the Fourier spectra analysis demonstrates that the proposed AFEDN can accurately capture the fault characteristic frequency, which provides an intuitive physical explanation for the network’s decision-making.

Intelligent diagnosis, Compound fault, Adaptive feature extraction, Knowledge distillation, Interpretability.