Research On Image Insulator Detection Technology Of UAV Patrol Inspection Based On YOLOv5

With the sustainable development of society and economy,the investment and construction of power system are accelerating day by day.In the power system,as the carrier of power transmission,transmission line is one of the most key links.As an important equipment in transmission link,insulator plays an important role in supporting fixed conductor and ensuring insulation distance.Most high-voltage transmission lines are mainly erected in non urban areas.Due to the large number,cross regional distribution and long-term exposure to the air,insulators are very prone to failure under the influence of harsh natural environment.With the expeditious construction of transmission projects,the traditional mode of relying on manual inspection has become more and more difficult to meet the requirements of high-quality operation and maintenance.With the increasing demand for intelligence of State Grid Corporation and the rapid application of UAV technology,the adoption of UAV intelligent patrol can greatly reduce the operation and maintenance personnel and time and improve the quality,so it has developed rapidly.The extensive application of deep learning technology and the continual improvement of computer computing performance have opened up a new solution for UAV to accurately identify and locate insulators and track and shoot in real time.In this paper,the insulators in the transmission line are identified and positioned,the deep learning technology is used,the target detection means based on YOLOv5 and its improved algorithm is adopted,and the pictures taken by the UAV are trained in combination with the characteristics of the insulator data set,so as to realize the accurate identification and positioning of insulators,and greatly improve the efficiency of accurate tracking and judgment of insulator equipment during UAV patrol inspection,It has very important application effect.The main work of this paper is as follows:(1)Firstly,this paper introduces the image acquisition and patrol status of UAV,introduces the patrol method,content,mode and requirements of UAV actual patrol in Xuzhou power supply company,and expounds the work flow of UAV patrol.Then it introduces the related knowledge,basic theory,classical convolutional neural network and common algorithm framework of deep learning.(2)The photographing,classification and naming of UAV images are described.The labeling software is used to calibrate the image,and the data enhancement and expansion are realized by means of image color adjustment,rotation,movement,scaling and filtering.(3)This paper introduces the current mainstream target detection algorithms CNN series algorithms and YOLO series algorithms.Considering from the two aspects of detection speed and detection accuracy,this paper focuses on YOLO algorithm,which is a new one-stage detection algorithm in view of deep learning model.It realizes the excellent detection speed and accuracy,and compares two typical algorithms experimentally.(4)This paper introduces the principle and changes of the newly released YOLOv5 algorithm,solves the existing problems,optimizes the YOLOv5 algorithm,and explores and upgrades the data enhancement,activation function,loss function,attention mechanism and so on.On the basis of self built data sets,experiments are carried out on YOLOv5 and improved YOLOv5 respectively,and the index values such as detection speed,average detection correctness and average recall are obtained.The results bring out that the improved algorithm based on YOLOv5 has better overall performance,high efficiency,good accuracy and small model,It can better adapt to the automatic patrol of UAV,and then it is applied in engineering.This paper provides a new solution for UAV image recognition.Through experiments,it can be found that compared with other algorithm models,the improved algorithm model proposed in this paper has obvious advantages.For insulators,especially small targets with high overlap,its recognition accuracy is higher,the recognition speed is also improved,and has a very excellent recognition effect.The research of this paper also provides a basis for the subsequent use of this technology to identify other normal or fault equipment in transmission lines.There are 57 figures,7 tables and 96 references in this paper.