Detection Of Damaged Insulators Based On Inception V4 And YOLOX Network

Insulators are one of the most widely used electrical equipment in the power system.If they are exposed outdoors for a long time,the insulators of the transmission line will appear damaged defects and pollution flashover,which will lead to the deterioration of insulation performance.Therefore,necessary inspection of insulators should be maintained regularly.However,manual inspection has the problems of low efficiency,high omission rate and false detection rate,which is difficult to meet the requirements of overhead line inspection.Nowadays,insulator detection technology based on deep learning has gradually replaced the traditional manual inspection method.Therefore,this thesis proposes an insulator fault identification and location algorithm based on improved Inception v4 and YOLOX networks.The main research contents are as follows:In view of the small number of damaged insulator samples in the existing data set,this thesis proposes a data enhancement algorithm based on improved SNGAN(Spectral Normalization GAN)to expand the damaged insulator samples.Firstly,the generated adversarial network satisfies the 1-Lipschitz constraint by means of spectral normalization,so that the generated model and the discriminant model adversarial training are more balanced,so as to improve the stability of the generated adversarial network and reduce the mode collapse.Secondly,the residual structure is introduced into the generating network and discriminant network respectively to deepen the depth of the network and acquire deeper features of the insulator image,so as to improve the quality of the image generation.Finally,the IS score of the improved SNGAN reaches 2.9,which is 0.5 higher than that before the improvement,indicating that the insulator image generated by the improved model has higher richness and better image quality.In order to solve the problem that traditional convolutional neural networks are difficult to extract effective features of insulators under complex background,an insulator damage recognition model based on improved Inception v4 network was established in this thesis.Firstly,by introducing CBAM(Convolutional Block Attention Module)attention mechanism,the model pays more attention to effective features under complex background,and improves the feature extraction ability of the model under complex background.Then,the original activation function Re LU is changed to Leaky Re LU to effectively prevent the situation that the weight cannot be updated because the gradient is zero when the input value is negative,thus solving the problem of neuron deactivation caused by Re LU activation function;Finally,the improved algorithm was compared with VGG,Res Net and other algorithms,and the recognition accuracy was 96.2%,11.7% and 4.8% higher than VGG and Res Net,respectively.Therefore,the improved Inception v4 network was more accurate in the classification of insulator faults.For the problems of excessive Anchor number,complicated calculation and poor positioning accuracy of damage detection in traditional positioning algorithms,this thesis proposed an improved YOLOX based insulator damage location algorithm.First of all,EIOU Loss was used as regression loss to combine three kinds of losses,namely the center point distance between the prediction box and the real box,the width and height of the prediction box and the degree of overlap,so as to make the regression of the prediction box more accurate.Secondly,coordinate attention mechanism is introduced between the trunk feature extraction network and the Neck layer to enable it to obtain feature information of different scales,so as to improve the ability of the network to detect minor insulator defects.Finally,the ablation experiment of the proposed algorithm was conducted and compared with Center Net,SSD and Faster R-CNN.The m AP of the improved YOLOX was 89.64%,4.41%higher than that of the original network,and the detection speed was 27.47 frames /s.The improved network model can meet the requirements of real-time insulator detection in recognition accuracy and detection speed.