Fault diagnosis of rolling bearings based on dual-channel Transformer and Swin Transformer V2

To address the challenge traditional one-dimensional fault diagnosis methods pose in accurately and efficiently handling nonlinear and non-stationary bearing fault signals, we propose a novel dual-channel deep learning architecture integrating a one-dimensional Transformer with two-dimensional wavelet time-frequency maps using Swin Transformer V2. This approach leverages the robust capabilities of the one-dimensional Transformer network to process raw vibration signals, effectively capturing temporal dependencies and latent fault patterns inherent in mechanical systems. Simultaneously, it employs continuous wavelet transformation to convert time-domain signals into two-dimensional time-frequency maps, which Swin Transformer V2 subsequently analyzes for image feature extraction. This dual-channel framework amalgamates information from time series and frequency-domain images and significantly enhances fault detection performance in mechanical systems. Experimental results demonstrate the superior diagnostic accuracy of the proposed model on bearing fault datasets, surpassing various current benchmark models and offering a novel approach to bearing fault diagnosis in mechanical engineering applications.