Scholarlink

<p>In bearing surface defect inspection, traditional manual methods suffer from low efficiency and poor sensitivity to small-scale defects, particularly in scenarios with overlapping targets. To address these issues, this work develops a novel detection method based on modified YOLOv8 for bearing surface defects, termed RMB-YOLOv8.The backbone network is redesigned by replacing the C2f module with RCS-OSA, reducing redundant channels and enhancing spatial feature extraction capability. In addition, a Bidirectional Feature Pyramid Network (BiFPN) is introduced to replace PANet, strengthening multi-scale feature fusion through bidirectional information flow. Furthermore, a Multi-Scale Dilated Attention (MSDA) mechanism is incorporated to capture rich contextual information, thereby improving feature representation and inference efficiency. Experimental results demonstrate that RMB-YOLOv8 achieves accuracy improvements of 3.2%, 4.3%, and 3.1% on scuffing, scratch, and dent defects, respectively, reaching an overall mean Average Precision (mAP) of 92.3% (+ 3.6%) while maintaining a real-time inference speed of 125 FPS. The proposed method demonstrates strong robustness in complex scenarios involving overlapping targets and small-scale defects. These results indicate that RMB-YOLOv8 achieves superior detection performance while maintaining comparable model complexity and real-time inference capability, providing an efficient and practical solution for automated bearing surface defect detection in industrial applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

RMB-YOLOv8: a detection method for bearing surface defects

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Jun Guo,
  • GuanXi Wang,
  • JiaJun Jin,
  • ShiLong Zhang,
  • PingLin Gu

摘要

In bearing surface defect inspection, traditional manual methods suffer from low efficiency and poor sensitivity to small-scale defects, particularly in scenarios with overlapping targets. To address these issues, this work develops a novel detection method based on modified YOLOv8 for bearing surface defects, termed RMB-YOLOv8.The backbone network is redesigned by replacing the C2f module with RCS-OSA, reducing redundant channels and enhancing spatial feature extraction capability. In addition, a Bidirectional Feature Pyramid Network (BiFPN) is introduced to replace PANet, strengthening multi-scale feature fusion through bidirectional information flow. Furthermore, a Multi-Scale Dilated Attention (MSDA) mechanism is incorporated to capture rich contextual information, thereby improving feature representation and inference efficiency. Experimental results demonstrate that RMB-YOLOv8 achieves accuracy improvements of 3.2%, 4.3%, and 3.1% on scuffing, scratch, and dent defects, respectively, reaching an overall mean Average Precision (mAP) of 92.3% (+ 3.6%) while maintaining a real-time inference speed of 125 FPS. The proposed method demonstrates strong robustness in complex scenarios involving overlapping targets and small-scale defects. These results indicate that RMB-YOLOv8 achieves superior detection performance while maintaining comparable model complexity and real-time inference capability, providing an efficient and practical solution for automated bearing surface defect detection in industrial applications.