湖南电力 ›› 2024, Vol. 44 ›› Issue (6): 83-89.doi: 10.3969/j.issn.1008-0198.2024.06.012

• 新技术及应用 • 上一篇    下一篇

基于多模态图像信息的配电网部件定位方法

黄志鸿1,2, 颜星雨3, 陶岩4, 张辉3, 徐先勇1   

  1. 1.国网湖南省电力有限公司电力科学研究院,湖南 长沙 410208;
    2.湖南省湘电试验研究院有限公司,湖南 长沙 410208;
    3.湖南大学机器人学院,湖南 长沙 410082;
    4.长沙理工大学电气与信息工程学院,湖南 长沙 410114
  • 收稿日期:2024-07-29 修回日期:2024-09-04 出版日期:2024-12-25 发布日期:2024-12-25
  • 通信作者: 颜星雨(2002),女,硕士,研究方向为目标检测。
  • 作者简介:黄志鸿(1993),男,博士,高级工程师,研究方向为电力设备故障智能诊断。陶岩(1999),男,硕士,研究方向为缺陷检测。张辉(1983),男,博士,教授,博士生导师,研究方向为机器人视觉检测与控制、深度学习图像处理。徐先勇(1982),男,博士,高级工程师,研究方向为电力智能巡检、调频式谐振特高压电源等。
  • 基金资助:
    湖南省杰出青年科学基金(2021JJ10025); 国网湖南省电力有限公司科技项目(5216A522001Y); 湖南省科技人才托举工程-“小荷”科技人才项目(2023TJ-X48)

Distribution Network Component Positioning Methods Based on Multi-Modal Image Information

HUANG Zhihong1,2, YAN Xingyu3, TAO Yan4, ZHANG Hui3, XU Xianyong1   

  1. 1. State Grid Hunan Electric Power Company Limited Research Institute, Changsha 410208, China;
    2. Hunan Xiangdian Test and Research Institute Co., Ltd., Changsha 410208, China;
    3. School of Robotics,Hunan University, Changsha 410082, China;
    4. School of Electrical and Information Engineering, Changsha University of Science and Technology,Changsha 410114, China
  • Received:2024-07-29 Revised:2024-09-04 Online:2024-12-25 Published:2024-12-25

摘要: 针对配电网部件热故障判别前置任务,提出一种面向配电网巡检的多模态图像部件定位方法。该方法包括多模态图像信息协同与目标检测两个关键步骤:首先,针对高分辨率可见光图像与带有温度信息的红外图像信息不对齐问题,提出自适应图像配准方法,该方法能够高质量完成跨模态配准任务;其次,通过预测信息迁移方法将配准后可见光图像的预测信息迁移至红外图像,完成对红外图像的检测。结果表明,相比直接检测红外图像,提出的方法能够提高18.4%的检测精度,并在精确率与召回率上表现极佳。

关键词: 电力巡检, 图像配准, 目标检测, 多模态, 深度学习

Abstract: A multi-modal image component localization method for distribution network inspection is proposed for the task of thermal fault discrimination of distribution network components. The method includes two key steps that is multi-modal image information coordination and object detection. Firstly, a self-adaptive image registration method is proposed to address the issue of misalignment between high-resolution visible light images and infrared images with temperature information,which can complete high-quality cross modal registration tasks. Secondly,the prediction information of the registered visible image is transferred to the infrared image by the prediction information transfer method to complete the detection of the infrared image. The results show that compared with direct detection of infrared images,the proposed method can improve detection accuracy by 18.4% and performs extremely well in terms of precision and recall.

Key words: electrical inspection, image registration, object detection, multimodal, deep learning

中图分类号: