Fault diagnosis of mining rolling bearings based on Superlet Transform and OD-ConvNeXt-ELA

In response to the limitations of current fault diagnosis methods for mining rolling bearings, which suffer from limited feature extraction capabilities and poor generalization, a fault diagnosis method based on Superlet Transform (SLT) and OD-ConvNeXt-ELA was proposed. Built upon ConvNeXt-T, Batch Normalization (BN) technology was introduced to improve the network's generalization ability. Omni-dimensional Dynamic Convolution (ODConv) replaced the original depthwise separable convolution to enhance the adaptability of the network. Efficient Local Attention (ELA) was incorporated to focus the network on key feature locations. This formed the OD-ConvNeXt-ELA network model for fault diagnosis of mining rolling bearings. To fully leverage the image feature extraction ability of the OD-ConvNeXt-ELA model, SLT was used to convert the collected one-dimensional vibration signal of the rolling bearing into a two-dimensional time-frequency image, which was then input into the OD-ConvNeXt-ELA for model training. Fault diagnosis experiments were conducted using the bearing datasets from Case Western Reserve University (CWRU) and Paderborn University (PU). The results showed that for the CWRU bearing dataset under a single operating condition, the average fault diagnosis accuracy of OD-ConvNeXt-ELA was 99.65%, which was an improvement of 1.61% over ConvNeXt-T. For the CWRU bearing dataset under cross-operating conditions, the average fault diagnosis accuracy of OD-ConvNeXt-ELA was 87.50%, which was an improvement of 3.30% over ConvNeXt-T. For the PU bearing dataset under cross-operating conditions, the average fault diagnosis accuracy of OD-ConvNeXt-ELA was 89.33%, an improvement of 3.46% over ConvNeXt-T. The fault diagnosis method based on SLT and OD-ConvNeXt-ELA shows high accuracy and strong generalization ability under cross-bearing, cross-operating conditions, and noise interference.