Scholarlink

<p>Wind turbine tower weld defects have such conditions as small size, unclear edges, heavy background interference, and multiscale distribution. It is difficult for existing detection methods to balance model parameter control, detection accuracy, and multiscale feature representation. Compressing parameters will weaken the presentation of fine-grained features, and strengthening attention and multiscale optimization will increase computational cost and enhance model complexity. This article focuses on the detection of welding defects in wind turbine towers, which involve small targets and weak boundaries. To address the limited memory on edge devices, YOLOv11-ECGA, a YOLOv11-based approach, is proposed. This method focuses on parameter efficiency, compressing parameters and the model size, allowing the trade-off of precision for acceptable FLOPs and inference speed. From the structural level, this model replaces the original YOLOv11 backbone network with EfficientNetV2, in this way, it not only reduces the volume but also strengthens the feature extraction ability; integrates CBAM into the backbone network, thereby improving the perception of defect target features in complex backgrounds; uses ghost convolution to compress redundant features; applies AFPN to the neck network to optimize the display of multiscale defects. Conduct ablation experiments and comparison experiments using the self-built welding defect dataset. The calculated value of ECGA-related parameters for YOLOv11 is 2.43 M; the model occupies 5.8&#xa0;MB, and there are also precision (P), recall (R), and mAP@0.5 metrics. It should be noted that these accuracy gains come at the expense of an approximate 70% increase in GFLOPs and a reduction in inference speed to 27.2 FPS. Nevertheless, with a typical sampling rate of 1–2 FPS during wind turbine tower inspections, the model still maintains ample speed redundancy, and its parameter compression offers a distinct advantage for deployment on edge devices with stringent memory constraints. Furthermore, the core contribution of this study lies in revealing the collaborative compensation principles among the four aforementioned modules, thereby providing a valuable reference for a collaborative mechanism integrating “parameter compression—attention enhancement—multiscale fusion” within the domain of industrial defect detection in resource-constrained scenarios.</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

A parameter-efficient attention-enhanced model for multiscale weld defect detection of wind turbine towers

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Laiqing Yan,
  • Rui Li,
  • Zia Ullah,
  • Renshuang Huang,
  • Ming Shi,
  • Tahir Khan,
  • Hasan Saeed Qazi

摘要

Wind turbine tower weld defects have such conditions as small size, unclear edges, heavy background interference, and multiscale distribution. It is difficult for existing detection methods to balance model parameter control, detection accuracy, and multiscale feature representation. Compressing parameters will weaken the presentation of fine-grained features, and strengthening attention and multiscale optimization will increase computational cost and enhance model complexity. This article focuses on the detection of welding defects in wind turbine towers, which involve small targets and weak boundaries. To address the limited memory on edge devices, YOLOv11-ECGA, a YOLOv11-based approach, is proposed. This method focuses on parameter efficiency, compressing parameters and the model size, allowing the trade-off of precision for acceptable FLOPs and inference speed. From the structural level, this model replaces the original YOLOv11 backbone network with EfficientNetV2, in this way, it not only reduces the volume but also strengthens the feature extraction ability; integrates CBAM into the backbone network, thereby improving the perception of defect target features in complex backgrounds; uses ghost convolution to compress redundant features; applies AFPN to the neck network to optimize the display of multiscale defects. Conduct ablation experiments and comparison experiments using the self-built welding defect dataset. The calculated value of ECGA-related parameters for YOLOv11 is 2.43 M; the model occupies 5.8 MB, and there are also precision (P), recall (R), and mAP@0.5 metrics. It should be noted that these accuracy gains come at the expense of an approximate 70% increase in GFLOPs and a reduction in inference speed to 27.2 FPS. Nevertheless, with a typical sampling rate of 1–2 FPS during wind turbine tower inspections, the model still maintains ample speed redundancy, and its parameter compression offers a distinct advantage for deployment on edge devices with stringent memory constraints. Furthermore, the core contribution of this study lies in revealing the collaborative compensation principles among the four aforementioned modules, thereby providing a valuable reference for a collaborative mechanism integrating “parameter compression—attention enhancement—multiscale fusion” within the domain of industrial defect detection in resource-constrained scenarios.