Research On Real-Time Detection Algorithms Of Power Components For Live Work In Distribution Network

With the rapid development of society,human life and social production require more and more stable power supply.With more than 4.4 million kilometers of distribution lines,China has formed the largest distribution network in the world.The application of robots for autonomous operation is an inevitable development trend because the inspection and maintenance tasks of distribution network lines are heavy.During the maintenance and repair of the power grid system,real-time recognition and positioning of insulators and fuses is of great significance for the motion planning of live working robots in the distribution network.The target detection algorithms based on deep learning have the advantages of high accuracy,fast speed and strong robustness.These algorithms can provide real-time feedback of the classification and position of the target to the robot.Combined with the binocular depth camera,the three-dimensional positioning of the power components can be realized,and then the autonomous motion planning can be carried out.This paper studies the one-stage object detection algorithm YOLOv4,and analyzes the complex working environment of the working robot and the shortcomings of the small target detection ability.An improved YOLOv4 is proposed for detecting insulators and drop fuses.The research work mainly done is shown as following:(1)Production of the dataset and data augmentation.The data set is used to train the object detection algorithm model.Image data is collected by drone.When the robots overhaul the power grid,the robots will face occlusion,cloudy,rainy and foggy weather,etc.Therefore,this paper introduces data augmentation methods to enrich the data set and improve the robustness of the model to cloudy,rainy and foggy weather and other conditions.Aiming at the problem of sample imbalance in the training set,the cost-sensitive learning is used to improve the loss function of object detection algorithms to alleviate the problem of sample imbalance.(2)Anchor boxes selection based on clustering algorithms.Anchor boxes provide prior information about size during object detection,which is used to generate detection boxes.In order to extract more suitable anchor boxes to provide more accurate prior knowledge for object detection,this paper will study two kinds of clustering methods for selecting anchor boxes:traditional clustering algorithms and unsupervised learning based on neural networks.YOLOv4 is trained by using the anchor boxes calculated by these clustering algorithms.After experimental comparison of different detection results,the Fuzzy C-Means clustering algorithm instead of K-Means clustering algorithm is used to select the anchors boxes to improve the detection accuracy.(3)Improvement of YOLOv4 network structure for small power components.In view of the lack of small target detection ability of YOLOv4,this paper attempts to improve the network structure of the object detection algorithm.The main improvements to the network structure are fusing more shallow features and adding the detection layer with larger size to improve the detection accuracy of small targets.Adding the detection layer will lead to a larger network structure,which reduces the detection speed.To solve this problem,some convolution blocks are replaced by residual blocks to make the network structure lightweight and keep the original detection speed.(4)Three-dimensional positioning of the power components for live work in distribution network.This paper introduces the principle of camera calibration and the principle of binocular ranging.The checkerboard paper and MATLAB calibration tool are used to calibrate the binocular camera.The three-dimensional positioning experiment is carried out on the simulation site of the indoor distribution network.Combined with the depth information,the positioning of the center point of the power component detection boxes and the three-dimensional coordinates of the target center point in the camera coordinate system have been obtained.