WSCIConvNet: a dual-path time-frequency fusion network for bearing fault diagnosis under complex operating conditions

Bearing fault diagnosis (FD) is essential for the safe operation of rotating machinery (RM), but traditional FD often fails under heavy noise and varying operating conditions due to weak feature extraction and low robustness. This paper introduces an innovative FD model named wide-sample convolution and interaction convolution network (WSCIConvNet) to tackle these challenges. WSCIConvNet employs a dual-channel structure designed to extract features from both the time-domain (TD) and frequency-domain (FRD), and integrate a cross-attention mechanism (CAM) for effective feature fusion. Additionally, to improve the noise robustness and expressive capabilities of the sample convolution and interaction network (SCINet) architecture, we propose an enhanced method that combines the wide-sample convolution and interaction convolution (WSCIConv) module with convolutional and pooling techniques. This method allows the model to successfully extract long-term temporal connections while skilfully identifying features related to the bearing FD. The experiment results indicate that the proposed method achieves remarkable classification accuracy, even in varying noise levels and load conditions.