Scholarlink

<p>In the production process of aero-engine blades (AEBs), surface defect detection is essential. However, data scarcity and class imbalance in practical industrial scenarios make deep learning-based defect identification of AEBs challenging. In this article, we propose AWBN-YOLO, an enhanced YOLOv10n-based framework to address these challenges. Specifically, we design an Adaptive Sample Augmentation Method (ASAM) to synthesize photorealistic defect samples by adaptively aligning defect geometries with blade contours and optimizing background consistency. We also propose a Feature-driven Wavelet Downsampling (FWD) module to preserve critical spatial-frequency details through adaptive wavelet basis selection, enhancing sensitivity to fine-grained defects. Furthermore, we introduce BiFPN-Concat and Normalized Wasserstein Distance Loss (NWD-Loss) to optimize multi-scale feature fusion and small-defect localization. Experiments on the AeBAD-SL dataset, a sample-imbalanced benchmark for AEBs have proven that AWBN-YOLO can achieve state-of-the-art performance with 82.2% precision, 71.9% recall, and 71.7% mAP50, surpassing the baseline YOLOv10n by 2.8%, 1.2%, and 2.6%, respectively. AWBN-YOLO achieves superior detection accuracy while maintaining real-time performance (140 FPS), offering a robust solution for industrial quality inspection under practical constraints.</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

AWBN-YOLO: a surface defect detection method for aero-engine blades in sample-limited scenarios

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Weixuan Gao,
  • Nengbin Lv,
  • Fuzhou Du

摘要

In the production process of aero-engine blades (AEBs), surface defect detection is essential. However, data scarcity and class imbalance in practical industrial scenarios make deep learning-based defect identification of AEBs challenging. In this article, we propose AWBN-YOLO, an enhanced YOLOv10n-based framework to address these challenges. Specifically, we design an Adaptive Sample Augmentation Method (ASAM) to synthesize photorealistic defect samples by adaptively aligning defect geometries with blade contours and optimizing background consistency. We also propose a Feature-driven Wavelet Downsampling (FWD) module to preserve critical spatial-frequency details through adaptive wavelet basis selection, enhancing sensitivity to fine-grained defects. Furthermore, we introduce BiFPN-Concat and Normalized Wasserstein Distance Loss (NWD-Loss) to optimize multi-scale feature fusion and small-defect localization. Experiments on the AeBAD-SL dataset, a sample-imbalanced benchmark for AEBs have proven that AWBN-YOLO can achieve state-of-the-art performance with 82.2% precision, 71.9% recall, and 71.7% mAP50, surpassing the baseline YOLOv10n by 2.8%, 1.2%, and 2.6%, respectively. AWBN-YOLO achieves superior detection accuracy while maintaining real-time performance (140 FPS), offering a robust solution for industrial quality inspection under practical constraints.