Gearbox fault diagnosis based on spatial channel reconstruction residual network with attention mechanism

As a critical component of rotating machinery, gearbox fault diagnosis is essential to ensure equipment safety. However, vibration signals are susceptible to multi-source noise interference, and traditional diagnosis methods relying on manual feature extraction suffer from inefficiency and poor generalization. To address these challenges, this paper proposes a Spatial Channel Reconstruction Residual Network with Efficient Channel Attention (ESCRNet). The model takes raw vibration signals as input and first employs a wide-kernel convolution layer to expand the receptive field for initial feature extraction. Subsequently, Spatial Channel Reconstruction Convolution (SCConv) is integrated into residual blocks, leveraging a multi-scale dynamic feature separation and fusion strategy to suppress redundant features and enhances the feature expression ability. Finally, an Efficient Channel Attention (ECA) module is introduced to dynamically recalibrate channel-wise features through lightweight 1D convolution, thereby improving the saliency of critical characteristics. Experimental results on the BJTU-RAO bogie dataset and the SEU dataset show that the proposed method still maintains a high average accuracy in a strong noise background, with strong noise robustness and feature expression capability.