Multi-Scale Temporal Coordinate Attention Network with Peak-Aware Mechanism for Rolling Bearing Fault Diagnosis Under Low Signal-to-Noise Ratio Conditions

Intelligent fault diagnosis of rolling bearings under high-noise industrial conditions remains a significant challenge. Traditional attention-based deep learning models often rely on global average pooling, which may inadvertently smooth out high-frequency transient impulses essential for fault identification, potentially leading to degraded performance in low signal-to-noise ratio (SNR) environments. To address this, we propose a Multi-Scale Temporal Coordinate Attention Network (MS-TCANet). The framework introduces a Peak-Aware Coordinate Attention (PACA) mechanism that combines max-pooling and average-pooling along directional coordinates. This dual-pooling design aims to better preserve transient impact features while maintaining a stable global representation, thereby mitigating the feature over-smoothing issue common in conventional attention modules. Additionally, an asymmetric multi-scale convolution block is incorporated to capture both short-term impacts and long-range periodic signatures. Experiments on three benchmark datasets (CWRU, Paderborn University, and XJTU-SY) indicate that the proposed MS-TCANet achieves favorable diagnostic accuracy compared to several representative and advanced methods, particularly under severe noise conditions (e.g., -10 dB SNR). t-SNE and Grad-CAM visualizations further suggest that the model can capture fault-related signatures more reliably than standard architectures in noisy environments.