As the key component of rotating machinery, the bearing plays a crucial role in ensuring the safe and reliable operation of electric motor. Therefore, implementing fault diagnosis on rolling bearings is endowed with profound significance for improving mechanical system performance. In this paper, we propose an end-to-end dual-branch bearing fault diagnosis (EDBFD) framework that is based on cross-domain semantic pattern fusion (CSPF), referred to as the EDBFD-CSPF model. In the first branch, a one-dimensional dilated convolutional neural network (1D-DCNN) encodes sequential vibration data into time-domain feature maps. Subsequently, an attention-based bidirectional gated recurrent unit (BiGRU) functions as a decoder to extract fault-related semantic pattern from these feature maps. In the second branch, the vibration signal is initially converted into a frequency signal with the fast Fourier transform (FFT). Then, another 1D-DCNN generates frequency-domain feature maps from the amplitude spectrum. An attention-based Kolmogorov-Arnold network (KAN) decodes these feature maps to extract fault-related semantic pattern in the frequency domain. The semantic information from both branches is fused and then fed into a classifier to identify bearing fault categories. The design of dual-branch architecture effectively integrates cross-domain feature maps and fault-related semantic patterns, thereby enhancing the diversity of the feature representations. Using the CWRU and SEU datasets for evaluation, our model yielded impressive performance in bearing fault diagnosis with an accuracy of 100%. Meanwhile, the EDBFD-CSPF model exhibits strong robustness against noise.
Zhang et al. (Thu,) studied this question.