Research On Application Of Capsule Network In Transmission Line Fault Detection

The safe operation of transmission lines is one of the important links to ensure power supply,and regular inspection of transmission lines is an important measure to ensure power transmission.Most of the traditional inspection methods are manual inspection,which is easily affected by the inspection environment and has the characteristics of low efficiency and high risk.With the development of power grid intelligence and deep learning technology,UAV patrol and image processing technology have been combined to improve the detection efficiency in transmission line fault detection.Using advanced image classification algorithm,the transmission line fault detection is intelligent,which makes the detection efficient and safe.Based on the research of transmission line fault detection methods at home and abroad,the mechanism characteristics of different faults have been analyzed,and the types of transmission line faults have been summarized.In this paper,UAV was used to collect fault pictures of transmission lines and get pictures of transmission lines.The collected pictures were processed by adaptive histogram equalization with limited contrast,and the processed pictures were cut out to obtain standard pictures with resolution of 64×64.Use labelImg tool to mark pictures and label them with categories to get.xml format files,and use.xml files to generate.tfrecord files for network training.Finally,the data set has been expanded by using data expansion technology to generate transmission line fault data set(TLF data set).This paper first has described the principle and structure of capsule network,and listed the advantages and disadvantages of capsule network.According to the characteristics of complex background and few images of transmission lines,two improved capsule networks have been proposed,namely MFF-CapsNet network and Res-CapsNet network.The algorithm principles of the two networks have been introduced respectively,including their main structure,principles of each part and so on.Experiments have been carried out on four commonly used standard data sets:MNIST,Fashion-MNIST,SVHN and CIFAR-10.the experimental results showed that the two methods not only improved the classification accuracy,but also were competitive in terms of parameters.Compared with the existing improved capsule network,the classification accuracy of Res-CapsNet network on CIFAR-10 is the highest,reaching 93.33%,which indicates that Res-CapsNet network is effective for classification of complex background pictures.Finally,the two models have been tested on the TLF data set,and the classification accuracy reached 93.95%and 96.17%respectively,which was greatly improved compared with other classical models.Experiments show that the two capsule network models proposed in this paper are competitive and generalized compared with the existing models,which provide a certain experimental basis for intelligent inspection of power grid.Figure[54]table[12]reference[79]...