Lightweight rolling bearing fault diagnosis via dense connectivity and adaptive soft thresholding

Deep learning-based fault diagnosis of rolling bearings is frequently challenged by strong ambient noise in vibration signals and the high computational cost of deployable models. While deeper networks can enhance performance, they often lead to parameter redundancy and information loss in deep layers, hindering industrial application. To achieve a balance between noise robustness and model lightweightness, this paper proposes SE-SDCTNet, a novel architecture built upon Sparse Dense Compact Thresholding (SDCT) blocks and Squeeze-and-Excitation (SE) blocks. The SDCT blocks employ dense connections for efficient feature reuse, while incorporating sparsity constraints and an integrated soft-thresholding mechanism to actively suppress noise and reduce parameters. Subsequently, SE blocks adaptively recalibrate channel-wise features to compensate for potential information loss due to sparsity and to enhance discriminative power. Furthermore, dilated convolutions are embedded to preserve multi-scale contextual information throughout the network. Evaluated on the Case Western Reserve University (CWRU) bearing dataset, SE-SDCTNet demonstrates superior diagnostic accuracy (e.g., 93.1 % under severe 2 dB noise) and robustness across various signal-to-noise ratios, while containing only 0.32 million parameters, merely about 3 % of ResNet18. In summary, this work provides a lightweight, accurate, and robust solution that facilitates the transition of data-driven fault diagnosis from theoretical research to practical industrial deployment.