Scholarlink

<p>The inspection of polyethylene (PE) sheaths on cable-stayed bridge cables is critical for maintaining structural integrity. However, significant challenges in efficiency, safety, and precision arise when traditional methods are employed. To address the limitations, a lightweight deep learning model named You Only Look Once-Lightweight, Yield and Edge-optimized (YOLO-LYE) is proposed. The model has been optimized for deployment on edge-based climbing robots to enable real-time, high-accuracy defect detection. Three core innovations are integrated: (1) A novel hybrid backbone is proposed, introducing the newly designed Cross-Scale Feature Fusion Module (CCFM) into a streamlined ShuffleNetV2 framework to achieve an optimal balance between computational efficiency and feature representation capability. (2) The feature alignment pathway is re-engineered through the strategic replacement of standard up/down-sampling operations with the dynamic upsampler (DySample) and a deformable convolution-based downsampling module (C3k-LD), forming a novel combination that is specifically designed to enhance sensitivity to small-scale defects. (3) The complete model is optimized for deployment on edge devices via the high-performance ncnn framework, facilitating real-time robotic inspection. Computational complexity is reduced by 52.4%, and the number of parameters is decreased by 64.1% compared to YOLOv11, while robust feature representation is maintained. Field validation was conducted on 236 stay cables. Detection accuracy exceeding 90% was achieved across defect categories under operational conditions. The work establishes a scalable framework for automated, high-precision infrastructure inspection in edge computing environments.</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

Lightweight YOLO-LYE enables edge intelligence for climbing robotic inspection of the bridge cable sheath defects

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Xinfeng Yin,
  • Chenhao Wang,
  • Yang Quan,
  • Yong Liu,
  • Jun He,
  • Zhou Huang

摘要

The inspection of polyethylene (PE) sheaths on cable-stayed bridge cables is critical for maintaining structural integrity. However, significant challenges in efficiency, safety, and precision arise when traditional methods are employed. To address the limitations, a lightweight deep learning model named You Only Look Once-Lightweight, Yield and Edge-optimized (YOLO-LYE) is proposed. The model has been optimized for deployment on edge-based climbing robots to enable real-time, high-accuracy defect detection. Three core innovations are integrated: (1) A novel hybrid backbone is proposed, introducing the newly designed Cross-Scale Feature Fusion Module (CCFM) into a streamlined ShuffleNetV2 framework to achieve an optimal balance between computational efficiency and feature representation capability. (2) The feature alignment pathway is re-engineered through the strategic replacement of standard up/down-sampling operations with the dynamic upsampler (DySample) and a deformable convolution-based downsampling module (C3k-LD), forming a novel combination that is specifically designed to enhance sensitivity to small-scale defects. (3) The complete model is optimized for deployment on edge devices via the high-performance ncnn framework, facilitating real-time robotic inspection. Computational complexity is reduced by 52.4%, and the number of parameters is decreased by 64.1% compared to YOLOv11, while robust feature representation is maintained. Field validation was conducted on 236 stay cables. Detection accuracy exceeding 90% was achieved across defect categories under operational conditions. The work establishes a scalable framework for automated, high-precision infrastructure inspection in edge computing environments.