Research On Visual Detection Method Of Typical Components In Transmission Lines Based On Convolutional Neural Networks

Transmission line inspection is an important measure to ensure the safe and stable operation of the power grid system,and a large number of transmission line images are collected in the new inspection method.These images usually cover the typical components on the transmission line.Identifying and locating these typical components is very profound for the transmission line inspection and can replace the manual inspection to find faults on the transmission line in time.The traditional image processing technology has a large amount of prior knowledge on the target recognition,which results in poor detection performance on the transmission line images with wide field of view and complicated background.The convolutional neural network can automatically extract the feature information of the images and have strong feature expression ability for the images.This has been widely used in the object detection,which improves the accuracy and applicability of the detection algorithm.A visual detection method based on convolutional neural network to detect the typical components of the transmission line is proposed in this thesis,including damper bolt and broken wire,that provides a theoretical basis for intelligent inspection.Firstly,the feasibility and effectiveness of convolutional neural network are analyzed in this thesis as the image feature extractor from the basic principle and basic structure of convolutional neural network,and the two-stage object detection algorithm and one-stage object detection algorithm based on convolutional neural network are expounded.In order to ensure the accuracy of typical components detection on transmission line,a transmission line image dataset for training and testing the detection algorithm is established in this thesis.Aiming at the complex background of the transmission line image and the large scale change,YOLO V3 detection algorithm for detecting the typical components of the transmission line is proposed in this thesis.Firstly,the principle of YOLO V3 algorithm is studied,including its baseline,predictions across scales,anchor box and bounding box prediction.In order to further improve the classification accuracy and positioning accuracy of YOLO V3,the DropBlock layer and the GloU loss function are added to the YOLO V3 network.Through comparison of several experiments,the improved YOLO V3 algorithi improves the accuracy and recall rate of the transmission line by about 4.1%compared with the original YOLO V3 algorithm,and in the case of small targets and occlusion,the detection performance is significantly improved.And the improved YOLO V3 algorithm detects an image at a speed of about 31 FPS/s,that meets the requirements of real-time detection.Finally,using the detection results of damper and bolt by YOLO V3 to further detect the deformation of damper and the corrosion of bolt on transmission line that realizing the detection of typical components’ defects.The object detection algorithm is used to identify and locate the typical parts in the transmission line image,and detect whether there are defects in these parts,so as to realize the real-time monitoring of the transmission line condition and improve the intellectualization of the transmission line patrol inspection.