Bearing defect detection algorithm and experiment based on YOLO-BDD

Abstract The detection of surface defects in bearings is crucial to mechanical stability, yet traditional methods suffer from inefficiency and insufficient accuracy. This study first analyzes the limitations of existing YOLOv5s models in bearing defect detection. To address these issues, we propose an improved YOLO-BDD network model with three core innovations that significantly enhance detection performance: (1) A lightweight network architecture using mobile-optimized MobileNet-v3 small as the backbone; (2) Incorporation of a Coordinate Attention (CA) mechanism to strengthen defect feature focusing capability; (3) Construction of a multi-scale detection module to improve micro-defect recognition. The experiments demonstrate that the improved model achieves significant lightweighting while maintaining high accuracy, with a mean average precision ([email protected]) of 83.1% (an increase of 5.5%), frames per second (FPS) reaching 33.5 (a growth of 50.2%), and a model size reduction of 37%, resulting in a final size of 8.67 MB. The solution maintains excellent performance even on low-end hardware, providing a cost-effective real-time detection solution for industrial settings. This advancement shows significant engineering application value and market potential.