鲍薇, 杨亚男, 张家毓,等.融合标签引导注意力与成对焦点损失的输电线路巡检多标签分类方法[J].河南理工大学学报（自然科学版）,2026,45(5):71-79.

BAO W, YANG Y N, ZHANG J Y,et al.Multi-label classification method for transmission line inspection integrating label-guided attention and pairwise focal loss[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(5):71-79.

融合标签引导注意力与成对焦点损失的输电线路巡检多标签分类方法

鲍薇, 杨亚男, 张家毓

国网河南省电力公司 郑州供电公司， 河南 郑州 450052

摘要: 目的针对输电线路巡检中设备类型多样、缺陷形态复杂，以及传统目标检测方法存在标注成本高、计算开销大及多标签场景性能受限等不足，进行融合标签引导注意力与成对焦点损失的输电线路巡检多标签分类方法研究。方法将输电线路设备与缺陷检测重构为多标签分类任务，避免对精确边界标注的依赖。针对特征语义混淆与标签共现干扰问题，提出融合标签引导注意力和成对焦点损失的方法：前者通过可学习的标签提示引导模型关注判别性区域，降低标签间干扰；后者在标签对层面建模共现与互斥关系，并结合难例聚焦，缓解对统计共现的过拟合，提升复杂标签组合的判别能力。  结果在真实巡检数据集上与多种主流基线网络的对比实验表明，所提方法在多标签分类性能上取得稳定且显著提升，mAP 最高提升7.65%，验证了其在不同网络结构下的有效性和一致性。  结论所提多标签分类方法能有效应对复杂巡检场景中的多标签共存问题，具备良好的泛化性和实用性，为实现低标注成本、高精度和高鲁棒性的智能巡检提供了有效方案。

关键词:输电线路巡检;多标签分类;标签引导注意力;成对焦点损失;轻量化

doi:10.16186/j.cnki.1673-9787.2025120094

基金项目:国家自然科学基金资助项目（61601172）

收稿日期:2025/12/29

修回日期:2026/05/11

出版日期:2026-07-24

Multi-label classification method for transmission line inspection integrating label-guided attention and pairwise focal loss

Bao Wei, Yang Ya’nan, Zhang Jiayu

Zhengzhou Power Supply Company， State Grid Henan Electric Power Company， Zhengzhou 450052， Henan，China

Abstract: Objectives In response to the diverse equipment types and complex defect patterns in transmission line inspection， as well as the limitations of conventional object detection methods， including high annotation costs， large computational overheads， and limited performance in multi-label scenarios， a multi-label classification method for transmission line inspection integrating label-guided attention and pairwise focal loss is proposed. Methods Transmission line equipment and defect detection are reformulated as a multi-label classification task to eliminate the dependence on precise boundary annotations. To address feature semantic confusion and label co-occurrence interference， a method integrating label-guided attention and pairwise focal loss is proposed. The label-guided attention module employs learnable label prompts to guide the model toward discriminative regions， thereby reducing interference among labels. The pairwise focal loss models co-occurrence and mutual exclusion relationships at the label-pair level and incorporates a hard-example focusing mechanism to alleviate overfitting to statistical co-occurrence patterns， improving the discriminative capability for complex label combinations.  Results Comparative experiments on real-world inspection datasets against several mainstream baseline networks demonstrate that the proposed method achieves stable and significant improvements in multi-label classification performance. The mAP is improved by up to 7.65%， verifying the effectiveness and consistency of the proposed method across different network architectures. Conclusions The proposed multi-label classification method effectively addresses the coexistence of multiple labels in complex inspection scenarios and exhibits strong generalization capability and practical applicability. It provides an effective solution for achieving low annotation cost， high accuracy， and robust intelligent inspection.

Key words:transmission line inspection;multi-label classification;label-guided attention;pairwise focal loss;lightweight