Research On Insulator Fault Detection Of Railway Catenary Based On Convolutional Neural Network

With the advancement of electrified railways,the requirements for safe operation of electrified railways have also increased,and the operation and maintenance of the locomotive traction power supply system is particularly important.As an important component of the traction power supply system,insulators work in outdoor environments for a long time,and are also affected by strong electric fields and strong mechanical tensions,which will cause varying degrees of damage,which will affect the safe and stable operation of railways.In the task of monitoring and detecting railway catenary insulators,the complex background,small size and unobvious failure of the insulators will cause the detection and classification of faulty insulators to be a challenging task.At present,the main method of insulator detection is manual detection,supplemented by traditional image processing algorithms.Among them,manual detection has low efficiency,easy to miss detection and large errors.Traditional image processing algorithms are also difficult to completely replace manual detection due to problems such as poor generalization ability and insufficient robustness.Therefore,the current large number of catenary image detection methods need to be further studied.Deep learning technology has developed rapidly with the support of massive data and advanced computer hardware.In this paper,in order to meet the needs of insulator detection speed and accuracy,two railway insulator fault detection network models based on convolutional neural networks are proposed.One is the cascaded network fault detection model,and the other is the lightweight YOLOv3 model.In the research project,the following work was mainly done:(1)Constructed an insulator failure training database to process a large amount of catenary image data.First,select a large number of samples that meet the needs according to the sample selection rules;then,in order to ensure the performance of the trained model,enrich the image data through data enhancement algorithms;finally,use the professional data labeling software Label Image to mark the target and mark the result Save in PASCAL VOC format for model training.(2)Propose a cascaded network detection model based on convolutional neural algorithm.In the cascaded network detection model,the location detection network and the fault classification network are cascaded to solve the problem of large-resolution images with a large amount of calculation and low fault classification accuracy for small-resolution images.In the location detection network,a multi-layer fusion module and a multi-region adaptive module are designed to achieve different levels of feature map fusion and adaptive weighting of target objects in the feature map,which improves the detection accuracy of the location detection network.Finally,through experiments,the accuracy of target detection and fault classification of the cascaded network model are verified.(3)A lightweight YOLOv3 target detection model is designed based on the YOLOv3 target detection model.In the lightweight YOLOv3 model,asymmetric convolution and grouped convolution structures are used to achieve the lightweight of the YOLOv3 network model and improve the efficiency of network model detection.In the lightweight YOLOv3 network model,the dilated convolution is added to expand the receptive field without increasing the amount of network model parameters to ensure the accuracy of network model detection.And perform experimental verification to detect the detection efficiency and detection accuracy of the lightweight YOLOv3 model on the catenary data set.Through experimental verification,the following results and conclusions are obtained:(1)The fusion of feature maps of different levels can enrich the expression ability of the cascaded network and improve the accuracy of detecting the network in the cascaded network.The AP value is 94.23%.(2)The multi-region adaptive module designed based on the attention mechanism proposed in this paper can improve the accuracy of the detection network in the cascade network by adaptively enhancing the weight value of the target region.(3)The cascaded insulator fault detection network m AP designed in this paper reaches 93.46%,which is better than other classic convolutional neural network models.(4)The lightweight YOLOv3 model proposed in this paper uses asymmetric convolution and grouped convolution structures in the classic YOLOv3 model to achieve a lighter network structure and improve detection efficiency.(5)The dilated convolution used in the lightweight YOLOv3 model,without changing the amount of network parameters,not only expands the receptive field of the convolutional layer,but also improves the accuracy of model detection,and finally reduces the weight of the YOLOv3 model.The detection speed is 40 fps.After experimental verification,the following conclusions are drawn: The two catenary insulator detection models proposed in this paper can accurately detect the position of the insulator and correctly classify the state of the insulator.The cascaded network fault detection model performs better in detection accuracy,and the lightweight YOLOv3 model performs better in detection speed.