Scholarlink

<p>Bearing surface defect detection is imperative for ensuring the reliability and longevity of industrial machinery, as even minor flaws can precipitate catastrophic failures. Addressing the challenges of multi-scale and irregular texture characteristics inherent in bearing surface defect detection, coupled with the persistent high cost of GPU hardware, this paper proposes MBD-YOLOv8, a novel model engineered to achieve an optimal balance between accuracy and inference speed on CPU devices. The methodology commenced with comprehensive experimental comparisons, leading to the selection of YOLOv8n as the optimal baseline model due to its favorable speed–accuracy trade-off. Subsequent architectural enhancements included: (1) the efficient replacement of the backbone with MobileNetV4 and the implementation of channel pruning for redundant parameter reduction; (2) the proposal of a Dual Heterogeneous-Kernel Spatial Aggregation (DHK-SA) module, designed to effectively aggregate defect features through spatial-channel dual alignment while maintaining computational efficiency; and (3) the replacement of the concatenation layer with a Texture-Aware Additive Fusion (TAAF) module, specifically considering the texture characteristics of defects. Experimental results on bearing defect datasets demonstrated that MBD-YOLOv8 achieved a high mean Average Precision (mAP50) of 92.1% and a real-time detection speed of 99 frames per second. In comparison to the original YOLOv8n, mAP50 and mAP were improved by 2.0% and 2.8%, respectively, while the inference speed was enhanced by 24%. The finalized model was observed to outperform all mainstream counterparts in terms of inference speed, concurrently maintaining competitive detection accuracy. This demonstrates its particular effectiveness in defect detection. This work thus offers a practical solution for industrial inspection scenarios necessitating CPU-based deployment.</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

MBD-YOLOv8: a lightweight network for bearing surface defects detection via heterogeneous-kernel spatial aggregation

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Peiwang Liu,
  • Zhigang Ren

摘要

Bearing surface defect detection is imperative for ensuring the reliability and longevity of industrial machinery, as even minor flaws can precipitate catastrophic failures. Addressing the challenges of multi-scale and irregular texture characteristics inherent in bearing surface defect detection, coupled with the persistent high cost of GPU hardware, this paper proposes MBD-YOLOv8, a novel model engineered to achieve an optimal balance between accuracy and inference speed on CPU devices. The methodology commenced with comprehensive experimental comparisons, leading to the selection of YOLOv8n as the optimal baseline model due to its favorable speed–accuracy trade-off. Subsequent architectural enhancements included: (1) the efficient replacement of the backbone with MobileNetV4 and the implementation of channel pruning for redundant parameter reduction; (2) the proposal of a Dual Heterogeneous-Kernel Spatial Aggregation (DHK-SA) module, designed to effectively aggregate defect features through spatial-channel dual alignment while maintaining computational efficiency; and (3) the replacement of the concatenation layer with a Texture-Aware Additive Fusion (TAAF) module, specifically considering the texture characteristics of defects. Experimental results on bearing defect datasets demonstrated that MBD-YOLOv8 achieved a high mean Average Precision (mAP50) of 92.1% and a real-time detection speed of 99 frames per second. In comparison to the original YOLOv8n, mAP50 and mAP were improved by 2.0% and 2.8%, respectively, while the inference speed was enhanced by 24%. The finalized model was observed to outperform all mainstream counterparts in terms of inference speed, concurrently maintaining competitive detection accuracy. This demonstrates its particular effectiveness in defect detection. This work thus offers a practical solution for industrial inspection scenarios necessitating CPU-based deployment.