Research On Transmission Line Fault Classification Based On Multi-attention Network

The normal operation of transmission lines is a necessary condition for stable and reliable transmission of power,so regular inspection is an important task to maintain the line operation.The traditional method uses manual inspection,but it is not only inefficient but also has security risks.With the development of artificial intelligence,people use deep learning methods,by sending drones and inspection robots to collect images,combined with classification algorithms for image processing research.For fault classification of transmission lines,the work mainly includes the following aspects:1.Firstly,the transmission lines in a certain region are studied,and the fault images in the past seven years are sorted out.According to the different types of transmission lines,the fault causes are analyzed,and finally four types of images are screened out,and the transmission line fault dataset is made by image preprocessing.2.In view of the complex geographical background information of transmission line fault images and the large differences in type characteristics,ResNet50 was selected as the backbone network of transmission line fault classification,and the residual network was improved,and the DBONet50 model was proposed.Training was conducted on Cifar10,Cifar100 data sets and transmission line fault data sets respectively.The experimental results showed that the accuracy of DBONet50 model was improved on different data sets,which were 88.08%,66.15% and 98.49%,respectively.It solves the problem that the accuracy of network classification is not high and reduces the case of false detection.3.Through further analysis of the experimental results,it is found that the DBONet50 model can extract limited feature information when dealing with the situation that the feature difference of the category image is large.In order to obtain more useful feature information and improve the accuracy of fault classification,an attention mechanism is introduced.According to the features of fault images,two new attention modules,BAD and TAM,are introduced to improve the existing attention modules.On the one hand,the module has the advantages of small reference number,prominent weight of important feature information and improved feature extraction ability.On the other hand,as a plug-and-play module,it can be easily embedded into existing networks.The two attention modules are embedded in the DBONet50 network to form the fault classification model ADR50.The experimental results show that the accuracy rate of the network is improved to 99.3%.The effectiveness of the two attention modules was verified by further ablation experiment and comparison experiment.4.According to the characteristics of transmission line fault images,based on the transmission line fault classification model ADR50,a fault classification system software is designed by using Python language and Qt software to realize the fault classification of a single image and multiple images.Figure [57],Table [19],Parameter [76]...