Research On Automatic Detection Of Power Line Insulator Defects In Aerial Images Based On Convolutional Neural Networks

Transmission lines are an important part of China’s infrastructure construction and energy supply guarantee.Due to the complexity of the transmission lines,the timely and safe detection of the transmission lines has become a difficult challenge.With the application of unmanned aerial vehicles,high-speed communication networks and parallel computing technology,the inspection methods of transmission lines have also undergone tremendous changes.In order to meet the needs of smart grid construction,this paper aims to propose a solution for the automatic identification of insulator defects in transmission lines,using deep learning technology to automatically identify and detect the images of overhead line insulators collected by drones.Focusing on the autonomous detection of aerial images of transmission line insulators,this paper studies a computer vision method based on deep learning techniques.Deep convolutional neural network is a stacked structure and contains a large number of parameters with hierarchical features.By simulating the activation working mode of biological neurons,convolutional neural networks can extract nonlinear characteristics of data,learn the distribution of data in high-dimensional space and establish a mapping between input samples and expected output.In this paper,the cascaded convolutional neural network structure is used for stepwise detection.First,the target detection network is used to extract the insulator region in the image to initially filter out the interference of the complex background environment,and the network that fuses the scale structure and the residual module is used as the feature extraction of the network.The experiment proves that the detection accuracy has been improved.At the same time,the structure of the network is improved through the receptive field theory and feature visualization methods to speed up the calculation speed of the model.The fast and lightweight YOLO2 model is used as a secondary detection network to improve the overall detection speed of the model.The size of the pre-selection box for regression is selected through clustering experiments to further improve the detection speed and regression accuracy.In order to avoid overfitting of the network and improve the robustness of the network,the insulator image data set is enhanced to increase the number of samples while simulating insulator images taken under different sampling conditions.Compared this method with other detection methods under the same evaluation index,the experimental results show that the cascading target detection structure and the model improvement strategy proposed in this paper ultimately achieve better results than other types of detection methods,and the accuracy exceeds the single-level detection method,and the speed and accuracy are better than other cascaded algorithms,achieved the optimization on the accuracy and speed.