摘要
针对电力设备异常发热故障诊断过程中识别目标设备单一、红外数据集样本数目庞大、平均识别准确率较低和识别速率较低的问题,提出一种基于深度卷积神经网络的改进YOLOv3目标检测方法,识别和定位绝缘子、隔离开关触头、套管、线夹4类电力设备及其异常发热区域。在改进YOLOv3算法的训练过程中,网络将数据集图片裁剪为416×416像素大小,使用Yolomark工具对图像进行标注,得到的标签和样本集一起送入深度学习卷积神经网络进行训练,经历多轮迭代后得到最终模型,最后采用运检部门用红外热像仪现场采集的电气设备红外图谱数据进行效果测试。实验结果表明,训练得到的改进YOLOv3模型相比于YOLOv3和快速区域卷积神经网络(faster region convolution neural network,Faster R-CNN)算法,识别定位的准确率较高,检测速度更快,可基本实现实时检测,可有效应用于变电站电力设备的红外巡检工作。
In order to solve the problems of identifying a single target device,large numbers of infrared data set samples,low average recognition accuracy and low recognition rate in abnormal heating fault diagnosis process of the power equipment,this paper proposes an improved YOLOv3 target detection method based on the deep convolutional neural network(CNN)to identify and locate four types of electrical equipment including the insulators,the disconnect switch contacts,the bushings and the wire clips,as well as their abnormal heating areas.In the training process of the improved YOLOv3 algorithm,the network cuts the data set picture to 416×416 pixels in size,uses yolomark to label the images and inputs the labels and sample set into the deep learning CNN for training.After several turns of iterations,the finally model is obtained and the infrared spectrogram data of the power equipment collected by the thermal infrared imager used by the State Grid operation and maintenance department is used for effect tests.The test results show that the improved YOLOv3 model,compared with the YOLOv3 and the Faster R-CNN algorithm,has a higher accuracy of identification and localization,faster detection speed,and can basically realize real-time detection,which is effective in infrared inspection of substation power equipment.
作者
陈同凡
刘云鹏
裴少通
CHEN Tongfan;LIU Yunpeng;PEI Shaotong(Department of Electrical Engineering,North China Electric Power University,Baoding,Hebei 071003,China)
出处
《广东电力》
2021年第6期21-29,共9页
Guangdong Electric Power
基金
中央高校基本科研业务费项目(2020MS093)。
