Insulator Damage Identification And Location Based On Res-CapsNet And Improved YOLOv4

In the wake of the development of China’s railway,greater demands have been placed on the safe operation of transmission lines.As an integral part of railway overhead lines,the integrity of insulators impacts the safety of transmission process,so it’s essential to enhance the regular inspection of insulators.At present,the inspection of insulators mainly relies on manpower,which is inefficient and cannot be timely found and dealt with.The traditional convolutional neural network(CNN)has some problems in insulator damage detection,such as poor recognition effect and slow speed.In contrast,the input and output of CapsNet are vectors,which can better retain the position characteristics of insulators,so as to detect insulators in complicated environments more accurately.Thus,this thesis proposes an algorithm based on ResCapsNet(residual capsule network)and improved YOLOv4,including insulator classification detection and damage location,to achieve accurate identification and location of insulator damage parts.As the damaged part of the insulator is small,the feature extraction is required to be higher.The traditional capsule network has only one convolutional layer,so it cannot obtain the features of the insulator more accurately,and increasing the convolutional layer will cause gradient vanish(explosion)and deterioration.The predicament of vanishing gradient(explosion)can be settled by batch standardization,degradation will make the model effect worse,and model degradation can be solved better by residual network,so the algorithm combining ResNet34 and CapsNet is proposed.Firstly,ResNet34 is used as the pre-training model to extract insulator characteristics.At the same time,the thick pooling layer and the full connected layer are removed,and then the convolution layer is added for dimension reduction,and the convolution features are converted into the capsule feature,and then transmitted through the dynamic routing mechanism.Finally,the algorithm proposed in this thesis is compared with SSD,Resnet,AlexNet and other networks.The experimental results show that the accuracy of insulator damage identification of the improved network reaches 97.98%,which is0.75% higher than that of the original network.It not only keeps the direction and Angle of output,but also can extract deeper features of insulators.Thus,it can identify the damaged insulators at a long distance in complex environment more accurately.In terms of insulator damage location,because the traditional YOLO network model has many parameters and a large amount of computation,the location accuracy of small insulator damage targets in complex environment is poor,so a damaged insulator location model based on improved YOLOv4 is established.Firstly,Res2Net residual unit is used to extract insulator fine features in CPSDarknet53,and then CBAM attention mechanism is introduced into the positioning network to pay attention to insulator profile and position features to improve the model accuracy.At once,due to the introduction of Res2Net residual unit,the model complexity increases.Therefore,the channel pruning compression network is adopted to reduce the amount of computation,which can realize the model training speed while maintaining a high precision.Finally,the ablation experiment is conducted on the improved network,and compared with SSD,Resnet,AlexNet and other networks.The experimental results showed that the localization accuracy of damaged insulators reached 96.57%,2.31%higher than the original network,and the number of model parameters is reduced by48.7M.It can quickly and accurately locate damaged insulators under different weather and distance,greatly improving the efficiency of intelligent inspection of insulators.