基于目标检测的电网涉鸟故障相关鸟种智能识别

Intelligent Recognition of Bird Species Related to Power Grid Faults Based on Object Detection

鸟类活动引起的电网故障呈现上升趋势,为了辅助输电线路巡检人员进行鸟类识别,该文提出一种基于YOLOv4目标检测的涉鸟故障相关鸟种图像识别方法。利用巡检图像与网络资源构建电网涉鸟故障典型危害鸟种图像数据集,并进行图像标注与数据增广。建立YOLOv4检测模型,采用多阶段迁移学习进行模型训练,并引入Mosaic数据增强、余弦退火衰减及标签平滑3种方法提升训练效果,分析先验框个数、训练方法、样本数量等因素对测试结果的影响,得到最优检测模型,对包含20类鸟种、1134个真实目标的图像测试集进行检测,平均精度均值可达92.2%。将YOLOv4检测结果与Faster RCNN、SSD、YOLOv3进行对比,其检测精度更高,误检数更低。研究结果表明,该文建立的YOLOv4模型能够有效检测输电线路巡检图像中的鸟类目标并实现鸟种识别,可为涉鸟故障差异化防治提供参考。

Power grid faults caused by bird activities reveal an upward trend. In order to assist transmission line inspector,recognize the bird species, in this paper, an object detection method based on YOLOv4 is presented for image recognition of birds related to power faults. An image dataset including the typical harmful bird species is constructed using the inspection images and network resources, and the image annotation and data augmentation are also realized. The YOLOv4 detection model is established and trained by the multi-stage transfer learning. Three methods are introduced to improve the training results, including the Mosaic data enhancement, the cosine annealing attenuation and the label smoothing. The optimal detection model is obtained by analyzing the influences of different factors like anchor number, training method, and sample size, etc. on the test results. The image test set including 20 bird species and 1134 true objects is detected, and the mean average precision(EmAP) reaches 92.2%. The detection results of the YOLOv4 are compared with those of the Faster RCNN,the SSD and the YOLOv3, which shows that the YOLOv4 model has a higher precision and less false detection number.This study indicates that the YOLOv4 model is able to detect the bird objects in transmission line inspection images and achieve bird recognition, which will provide reference for differential prevention of the bird-related outages.