摘要
为解决图像识别技术在电力工作中应用不佳的问题,本文基于各类电力部件的图像识别方法,选用高质量图片识别分析,并给出发展建议。研究结果如下:(1)YOLO算法检测快、精度高,广泛应用于无人机智能巡检中;(2)理论研究的故障识别率超过90%,实际应用却不足50%,因为理论样本集质量高,且算法参数仅对理论样本集有效。实际工作中每天上传不同的样本,样本可能包含多种电力部件,且待测部件不明显,算法无法适应差异性大的样本集;(3)改善图片质量可有效提升图像识别效果,本文识别率达75%,但考虑电力安全,算法会矫枉过正,识别出很多不存在的缺陷。可基于当前激光点云建模的工作模式,辅以图像识别技术,综合利用激光、红外等数据,实现无人机自主巡线,通过精细化巡检提升图片质量。
In order to solve the problem of poor application of image recognition technology in electric power work, this paper selects high-quality image recognition and analysis based on image recognition methods of various power components, and gives development suggestions. The research results are as follows:(1)The YOLO algorithm has fast detection and high precision, and is widely used in UAV intelligent inspection.(2)The fault recognition rate of theoretical research exceeds 90%, but the actual application is less than 50%, because the theoretical sample set is of high quality and the algorithm parameters are only valid for the theoretical sample set. In actual work, different samples are uploaded every day. The samples may contain a variety of electrical components, and the components to be tested are not obvious, so the algorithm cannot adapt to the sample sets with large differences.(3)Improving the image quality can effectively improve the image recognition effect. The recognition rate of this paper reaches 75%, but considering electrical safety, the algorithm will overkill and identify many flaws that do not exist. It can be based on the current working mode of laser point cloud modeling, supplemented by image recognition technology, and comprehensively use laser, infrared and other data to realize the autonomous line inspection of the UAV, and improve the picture quality through refined inspection.
作者
司小庆
束庆霏
SI Xiaoqing;SHU Qingfei(State Grid Zhangjiagang Power Supply Company of Jiangsu Electric Power Co.,Ltd.,Zhangjiagang 215600,Jiangsu,China)
出处
《电力大数据》
2022年第10期37-44,共8页
Power Systems and Big Data
关键词
无人机
电力巡检
图像识别
深度学习
激光点云
unmanned aerial vehicle
power inspection
image recognition
deep learning
laser point cloud
