BAO W, YANG Y N, et al． Multi-Label classification method for transmission line inspection via label-guided attention and pairwise focal loss[J]. Journal of Henan Polytechnic University( Natural Science) ,doi: 10.16186/j.cnki.1673-9787. (2025120094 ）..

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(网络首发)

鲍薇，杨亚男

国网河南省电力公司郑州供电公司,河南 郑州 450052

摘要: 目的 为针对输电线路巡检中设备类型多样、缺陷形态复杂，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 最高提升达7.15%，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、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);国网河南省电力公司郑州供电公司资助项目（B2171024K053）

收稿日期：2025-12-29

修回日期：2026-05-11

网络首发日期：2026-06-04

Multi-Label classification method for transmission line inspection via label-guided attention and pairwise focal loss (Online)

BAO Wei, YANG Yanan

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

Abstract: Objectives In view of the diverse equipment and complex defects in power line inspections, conventional detection approaches face issues such as high annotation costs and excessive computational demands. This study develops a robust and efficient intelligent recognition framework that achieves a superior trade-off between precision and speed, aiming to facilitate practical automated inspections in complex scenes, Methods In this work, transmission line equipment and defect detection is reformulated as a multi-label classification task, thereby eliminating the reliance on precise bounding box annotations. To address feature-level semantic ambiguity and label co-occurrence interference, we propose a unified framework that integrates label-guided attention with a pairwise focal loss. Specifically, the label-guided attention introduces learnable label prompts to steer the model toward discriminative regions relevant to each label, thus mitigating inter-label interference at the feature level. Meanwhile, the pairwise focal loss explicitly models co-occurrence and mutual exclusivity among label pairs, and incorporates a hard-example focusing mechanism to alleviate overfitting to statistical co-occurrence patterns, thereby enhancing the model’s ability to distinguish complex label combinations. Results The proposed approach consistently outperforms several state-of-the-art baselines on real-world inspection datasets. With a maximum mAP gain of 7.15%, our method demonstrates superior multi-label classification performance and maintains remarkable consistency across diverse architectural backbones. Conclusions Experimental results demonstrate that the proposed multi-label classification method effectively addresses the challenge of label co-occurrence in complex inspection scenarios, exhibiting strong generalization capability and practical applicability. It provides an effective solution for achieving low annotation cost, high accuracy, and robust performance in intelligent inspection systems.

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