104 / 2025-05-15 10:17:01

Differentiable time-frequency network for bearing fault diagnosis with small samples

Few samples,Short-time Fourier transform,Intelligent fault diagnosis.,Rolling bearing,Interpretable neural network

The existing intelligent fault diagnosis methods of rolling bearings based on deep learning focus on a lightweight network structure design of the diagnostic model to prevent over-fitting caused by insufficient samples, but lack an explanation of its decision-making mechanism, thereby failing to meet the high reliability requirements in actual fault diagnosis scenarios. To address the aforementioned issues, this study introduces a novel approach called interpretable Differential Time-Frequency Network (DTFN) dominated by Short-Time Fourier Transform (STFT). Firstly, a Differentiable Time-Frequency Representation (DTFR) method is developed based on time window adjustment factor. Secondly, a trainable frequency sensing factor is incorporated into the Differential Time-Frequency Kernel (DTFK) to enable simultaneous extraction of real and imaginary features. Then, the Differential Time-Frequency Layer (DTFLayer) constrained by DTFK is integrated into the front end of the selected 2-D prototype network to provide the interpretable feature representation embedded with the differentiable time-frequency knowledge. Finally, compared to other benchmark methods in publicly available bearing datasets, the results demonstrate that the proposed DTFLayer enables DTFN to exhibit exceptional accuracy in fault recognition and faster convergence speed even with limited samples.