Abstract
Bearing defect detection is essential for industrial equipment condition monitoring and predictive maintenance, yet identifying subtle, multi-scale, and morphologically diverse defects remains challenging. This study proposes MSF-YOLO (Multi-Scale Fusion YOLO) , an improved YOLOv8-based model enhanced by three core architectural innovations: a multi-scale convolutional module (MSPConv) for cross-scale feature extraction, a spatial-channel collaborative attention module (C2f_SCSA) for focusing on defect-critical regions, and a learnable-weight feature fusion module (L_FFM) for adaptive hierarchical aggregation of multi-scale features. Ablation experiments on a self-built bearing defect dataset show that successive integration of these modules consistently improves performance compared to the baseline YOLOv8n. The final MSF-YOLO achieves precision (P) 92.2% , recall (R) 83.1% , mAP50 88.1% , and mAP95 53.1% , representing 4.0%, 8.1%, 6.3%, and 6.4% improvements, respectively. These results confirm the effectiveness of the proposed multi-scale extraction, attention, and fusion strategies, providing a high-precision solution for industrial visual inspection of bearing defects. https://github.com/Hankerzao/MSF-YOLO-Bearings-Defect-Base-on-YOLOv8-.git