Bearing fault diagnosis of rolling machinery based on spatial pyramid channel attention network with depthwise separable convolution

Purpose Rolling bearings are critical components in rotating machinery, and their fault diagnosis is essential for ensuring safe equipment operation. However, traditional fault diagnosis methods face challenges including insufficient feature extraction, high computational complexity and difficulties in multi-scale information fusion. To address these issues, this paper proposes a spatial pyramid channel attention network with depthwise separable convolution for intelligent fault diagnosis of rolling bearings. Design/methodology/approach Continuous wavelet transform converts 1D vibration signals into 2D time-frequency images. An efficient feature extraction block integrates depthwise separable convolution, multi-scale dilated convolution and a cross-spatial multi-scale attention module to expand the receptive field while reducing complexity. A cross-spatial multi-scale attention mechanism uses parallel 1D global average pooling paths to capture long-range dependencies. Finally, a channel-spatial synergy module combines spatial pyramid structure and channel information interaction to establish inter-channel dependencies while preserving multi-scale spatial information. Findings Experimental results on three public datasets from Jiangnan University (JNU), Xi'an Jiaotong University (XJTU-SY) and Paderborn University (PU) demonstrate that the proposed method achieves diagnostic accuracies of 99.94%, 98.94% and 98.55%, respectively, showing superior diagnostic precision and better generalization capability compared to existing methods. Ablation experiments further validate the effectiveness of each core module. Originality/value This paper proposes a novel spatial pyramid channel attention network with depthwise separable convolution that synergistically integrates multi-scale dilated convolution, cross-spatial attention and channel-spatial synergy. The method effectively addresses challenges of insufficient feature extraction, high computational complexity and multi-scale fusion difficulties in rolling bearing fault diagnosis, demonstrating superior diagnostic accuracy and generalization capability across multiple public datasets.