Abstract

Gansu section of Qishao UHV DC line accounts for 52.4% of the total line length, and manual inspection is time consuming and laborious. This paper proposes an infrared image recognition method of hardware fittings based on YOLOv5 target detection. The image data set is constructed by using the infrared image of Qishao strain clamp of transmission operation inspection center, and the image annotation and data expansion are carried out. The YOLOv5 detection model was established, and trained by multistage transfer learning. Three methods were introduced to improve the training result, including Mosaic data enhancement, learning rate attenuation and label smoothing. The influences of different factors like anchor number, training method, and sample size, etc. on test results were analyzed, thus to obtain the optimal detection model. The image test set including insulator, equalizing ring and interval rods was detected, and the mean average precision (mAP) reaches 96.53%. Comparing the detection results of YOLOv5 with those of other commonly used models, it is found that the model in this paper has high accuracy and fast recognition speed, which provides a reference for fault inspection of Qishao UHV line.

中文翻译：

---

基于YOLOv5的特高压线拟合红外图像多目标检测算法

摘要

七少特高压直流线路甘肃段占线路总长度的52.4%，人工巡检费时费力。本文提出一种基于YOLOv5目标检测的硬件配件红外图像识别方法。利用输电运行检测中心七少应变钳红外图像构建图像数据集，并进行图像标注和数据扩展。建立YOLOv5检测模型，并通过多阶段迁移学习进行训练。引入了三种方法来改善训练结果，包括马赛克数据增强、学习率衰减和标签平滑。分析anchor数量、训练方法、样本量等不同因素对测试结果的影响，从而得到最优的检测模型。对包括绝缘子、均压环、间隔棒在内的图像测试集进行检测，平均精度（mAP）达到96.53%。将YOLOv5的检测结果与其他常用模型的检测结果进行比较，发现本文模型准确率高、识别速度快，为七少特高压线路故障检查提供参考。