Fusing Effective Channel Attention With Depthwise Separable Large Kernels for Bearing Fault Diagnosis Under Small Sample Conditions

Convolutional neural networks (CNNs) have gained widespread adoption in fault diagnosis applications owing to their exceptional ability to automatically learn discriminative features from complex data patterns. However, most existing methods based on CNNs use network frameworks that stack small-sized convolutional kernels. Such network models focus on extracting local features, thus limiting the ability to extract cross-domain invariant features, and perform poorly with small sample sizes. We propose an effective channel attention depth separable CNN (ECA-DSCNN). The method designs a feature extraction module based on a large convolutional kernel to extract high-quality cross-domain invariant features. By combining depth-separable convolution with an effective channel attention mechanism, the network parameters are reduced while compensating for the cross-channel interaction capability of depth-separable convolution. ECA-DSCNN achieves high-precision fault diagnosis under finite sample conditions without relying on domain adaptive methods and is robust to fluctuations in the amount of training data.