.png){kind=icon}
.png){kind=icon}
Return
A multi-scale defect detection network for wind turbines utilizing margin aware features
.png){kind=icon}
PRE
AI
DOI：10.1088/1361-6501/ae08db.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The long-term operation of wind turbines (WTs) leads to multi-scale surface defects that critically compromise operational reliability. Drone-based defect detection offers a viable approach for real-time assessment of WT operational status. However, the current deployment of UAV-based detection systems struggles to simultaneously achieve both sensitivity and positioning accuracy for such multi-scale defects. To address this limitation, we propose a novel defect marginal-aware and multi-scale collaborative attention network (DMCA-Net). First, we propose a defect marginal detail transfer backbone to enhance edge information in shallow features, which can be fused with multi-scale features. Second, a triple-layer anchor attention feature selection and fusion pyramid network is introduced to optimize channel-space interactions, which can dynamically balance local details and global features, thereby improving defect localization accuracy. In addition, a histogram-based synergistic attention head encoder is designed to detect small object defects by co-optimizing frequency-domain split-box attention and cross-box attention to enhance the feature intensity of small object defects. Finally, the Normalized Wasserstein Distance–Inner Distance–IoU (NWD-InnerDIoU) loss is introduced to enhance model generalization and mitigate severe data imbalance, effectively reducing performance fluctuations resulting from interactions among multi-scale targets. Experimental results demonstrate that DMCA-Net achieves state-of-the-art performance with 83.1% mAP50, representing a 3.1% improvement over baseline, while maintaining real-time detection capability at 81.3 frames-per-second on the WT defect dataset. Especially, it outperforms commonly used detection models in terms of detection performance.
Journal
IF:
3.4
Papers:
2.6K
Citations:
2.3W