The Research On Image Detection Technology Of Overhead Line Based On Deep Learning

The route routing method of UAV(Unmanned Aerial Vehicle)in overhead lines has been widely promoted.The line faults in aerial pictures observed by human eyes are easy to be affected by factors such as clarity,shooting angle and light,etc.The workload is large and the professional requirements are high.This paper presents an image detection model of overhead line based on deep learning,which can reduce the difficulty and improve the efficiency of line patrol.In order to solve the problem of unbalanced sample types in overhead line image samples,the existing fault detection algorithms are mostly single component detection.YOLO_v3 has the characteristics of considerable recognition rate and fast recognition speed.Focal loss can alleviate the impact of category imbalance by adjusting the learning weight of samples,and combining the two can alleviate the impact of category imbalance on YOLO_v3 when training samples.The model adopts the improved YOLO_v3 model based on Focal loss.Target detection is carried out on the poles,transformers and switches in aerial images of overhead lines,which is compared with the original YOLO_v3 model.The result shows that the proposed model can maintain a higher recognition rate at the same detection speed.Compared with the two-stage target detection algorithm Faster r-cnn,the recognition rate is close to that of Faster detection.Aiming at the problem that single component target detection is difficult to detect missing component in overhead line fault detection,a detection method based on fault feature calibration is proposed.In this method,the whole area where the fault occurs is taken as the sample calibration box,and the target detection algorithm of deep learning is used to identify the fault occurrence point directly.By this method,the model can identify the phenomena of insulator drop,arrester drop and pole tilt in the overhead line image.The model also identifies poles,transformers and switches in overhead lines.The purpose is to improve the recognition rate of pole tilt faults by comparing poles with tilted poles.The target detection of transformer and switch can be combined with image fusion or trained fault classifier for further fault judgment.Compared with the original model and Faster r-cnn,the proposed model is more practical.The results show that the improved YOLO_v3 model based on Focal loss proposed in this paper can better adapt to the aerial line fault online inspection of UAV.The research in this paper not only reduces the influence of unbalanced sample types of overhead line fault detection,but also increases the effect of automatic fault identification of multiple types,which can be applied to aerial line fault detection of UAV.