基于线性全局注意力与专家几何知识的防振锤缺陷级联检测
CSTR:
作者:
作者单位:

1.山东省新型配用电技术与装备重点实验室,山东理工大学;2.电气与电子工程学院,山东理工大学

作者简介:

通讯作者:

中图分类号:

TP391.4

基金项目:

国家自然科学基金(No. 52377110);山东省自然科学基金(ZR2022QE100)


A cascaded defect detection method for vibration dampers based on linear global attention and expert geometric knowledge
Author:
Affiliation:

Fund Project:

The National Natural Science Foundation of China (No. 52377110); The Natural Science Foundation of Shandong Province (No. ZR2022QE100)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    防振锤是输配电线路无人机航拍巡检中重点关注的金具之一.受制于无人机巡检安全距离与拍摄视角,巡线图像中防振锤目标具有像素占比少、缺陷特征差异不明显的特点,给防振锤的细粒度缺陷检测带来了严峻挑战.为此,本文提出基于线性全局注意力与专家几何知识的防振锤缺陷级联检测方法,将防振锤细粒度缺陷检测任务解耦为定位和分类两个阶段.在定位阶段,提出基于线性全局注意力的轻量化定位网络,在保证实时性的同时增强网络的全局上下文感知能力.在缺陷分类阶段,提出专家几何知识嵌入分类网络,将防振锤缺陷分类中的专家几何知识转化为显式特征表示,并嵌入深度学习网络,增强网络对缺陷类别的语义理解深度.最后,在自建数据集上实验表明:第一级定位网络实现95%召回率下84.1%的定位精确率,第二级缺陷分类准确率达93.42%,验证所提出方法在细粒度缺陷检测中的有效性.

    Abstract:

    Vibration dampers are critical hardware components and high-priority inspection targets in UAV-based power line surveys.Constrained by safety distances and camera perspectives, vibration dampers in inspection images often exhibit small pixel occupancy and subtle defect features, posing significant challenges for fine-grained defect detection.To address this, we propose a cascaded defect detection method for vibration dampers based on linear global attention and expert geometric knowledge, decoupling the fine-grained detection task into localization and classification stages.In the localization stage, a lightweight linear global attention-based network is developed to enhance global contextual awareness while maintaining real-time performance.In the defect classification stage, an expert geometric knowledge embedded network is presented to formalize geometric priors into explicit feature representations, augmenting the semantic comprehension of defect categories.Experimental results on a custom dataset show that the localization network achieves 84.1% precision at a 95% recall rate, and the classification accuracy reaches 93.42%, validating the efficacy of the proposed method for fine-grained defect detection.

    参考文献
    相似文献
    引证文献
引用本文
相关视频

分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2026-01-26
  • 最后修改日期:2026-04-28
  • 录用日期:2026-04-30
  • 在线发布日期: 2026-05-27
  • 出版日期:
文章二维码