Edge-Oriented Bearing Fault Diagnosis via Triple-Lightweight Network With Adaptive Pruning

Real-time diagnosis of motor bearing faults is essential for fault detection and timely maintenance, minimizing operational risks. However, most existing approaches often depend on cloud platforms or high-performance local servers, which introduce latency and potential security vulnerabilities due to raw data transmission. To overcome these limitations, this paper presents an adaptive pruning triple-lightweight network (APTL-net), which achieves model lightweighting for edge deployment across three stages: model construction, feature extraction, and structural pruning. First, a lightweight fault diagnosis framework optimized for edge environments is proposed, enabling parallel training and recursive inference to eliminate the traditional trade-off between training efficiency and inference performance. Second, a frequency-domain decomposition-based, weight-sharing multiscale convolution module is designed to efficiently extract multiscale features from periodic signals while harmonizing scale variations. Moreover, a fully automated adaptive pruning strategy is proposed to dynamically group and remove redundant structures during training without manual intervention. Finally, APTL-net is deployed on the edge device and validated on the physical platform equipped with real hardware. Experimental results indicate that APTL-net achieves a diagnostic accuracy of 99.54% while reducing parameters by 47.82%, FLOPs by 49.98%, and edge inference latency by 20.59%, thereby delivering superior performance over existing lightweight fault diagnosis networks.