Hierarchical Convolution-Transformer Framework for Gear Fault Diagnosis Under Severe Noise

To address the limitations of convolutional neural networks in capturing global fault features, the high computational cost and overfitting risk of Transformer models in gear fault diagnosis, and the feature degradation under strong noise, this study proposes a novel convolution–Transformer–channel attention network. The proposed method achieves a dynamic balance between local and global features through the joint design of multi-scale convolution and sparse self-attention. Multi-scale convolution extracts local features, while the enhanced sparse multi-head self-attention models long-range dependencies via adaptive sequence compression to reduce computational complexity. In addition, a dual-stage squeeze-and-excitation mechanism recalibrates feature channels to suppress noise interference and enhance robustness. This method is well-suited for gear fault diagnosis under strong noise conditions, offering improved generalization capability and promising potential for industrial applications.