Research On Insulator Defect Detection Technology For High-voltage Transmission Lines Based On Improved YOLO Algorithm

China has a vast territory and uneven distribution of power generation resources.Combined with China’s "double carbon" goal,the demand for long-distance transmission of electric energy by transmission lines will be higher and higher in the future.The insulation equipment on the transmission line is exposed to various environments for a long time,which is prone to damage,defects and other faults,which seriously affects the power supply reliability of the power system.Therefore,the detection technology of insulators and defects on high-voltage transmission lines is studied.Aiming at the problem that the background area of insulator detection is too large under the horizontal rectangular frame detection caused by the shooting angle of the UAV,the rotating frame is used instead of the horizontal frame to detect the insulator.The YOLOV5 s algorithm with better detection speed and accuracy is used as the benchmark model.The CSL(Circular Smooth Label)is introduced on the original model,and an improved YOLO model(CSL-YOLO)is proposed.The experimental results of the rotating insulator data set constructed under the algorithm show that the mean Average Precision(m AP)of CSL-YOLO reaches 92.5%,which shows that the algorithm has a good effect on the target detection in the image and realizes the rotation detection of the target.It effectively solves the problem that the insulator under the horizontal bounding box will lead to strong overlap between the detection boxes and inaccurate representation of the target range.In order to improve the detection speed and detection accuracy,an improved YOLOV5 s algorithm is proposed by replacing the YOLOV5 s backbone feature network extraction network with Swin Transformer and adding attention mechanism.The experimental results trained on the data set show that the improved model further improves the detection accuracy of the model and has a better detection effect on defects on high-voltage transmission lines.The accuracy of the model is improved from94.4% to 96.2%.The size of the algorithm model is only 13.8MB,which lays a model foundation for the subsequent research on defect detection of high-voltage transmission lines that can be deployed on embedded hardware.Aiming at the problems of weak computing power and low model migration accuracy of embedded devices compared with large GPUs,the proposed improved YOLOV5 s model is compressed and accelerated by using Tensor RT,and transplanted into NVIDIA embedded device Jetson Nano.The detection speed on Jetson Nano is about 21 frames/s.Compared with 5 frames/s without acceleration,it can be seen that the model detection speed is greatly improved after acceleration,which is close to realtime detection(30 frames/s),which is suitable for the deployment of embedded edge computing platform.At the same time,the code interface faced by the power tester is converted into a graphical interface to facilitate the use and operation of the staff.