Identification And Location Of Suspension And Insulator Faults Based On Improved Capsule Network

With the development of China’s high-speed railways,the scale of the railway network has continued to expand.The higher requirements have been imposed on China’s railway operations and management.In order to ensure the normal and stable operation of the railway system,it is necessary to achieve high-efficiency inspection of high-speed railway lines.However,the corresponding inspection technology is relatively backward.Railway inspection mainly depends on the traditional remote sensing method of manual inspection,which has the problems of low efficiency and failure to deal with the problem in a timely manner.Due to its excellent feature expression ability,deep convolutional neural networks can achieve more robust performance than traditional image processing methods with the support of big data.Therefore,this paper mainly applies the capsule network model in deep learning to high-speed rail catenary fault detection,and realizes the identification and location of catenary suspension and insulator faults,while meeting high recognition rate and recognition rate.Firstly,a large number of suspension and insulator fault images are collected by the drone,and the more complete image edge information is extracted through the pre-processing methods of gray change,binarization,Gaussian filtering,and canny edge extraction.Secondly,in order to solve the problems of long training time and poor real-time performance of the traditional capsule network,an improved capsule network algorithm is proposed,which uses 3×3 convolution and layer 1×1 reduction layers to simplify the traditional 9×9 capsule neurons.At the same time,an intermediate capsule layer is added,and the parameter optimization algorithm is used to reduce the time to find weights.Experiments have proved that improving the capsule network has certain advantages,including reducing training time and improving recognition rate.In order to solve the most between-cluster variance insulator image segmentation method and clustering algorithm,the problem existing in the application CV model ofthe insulator image segmentation method for suspension and insulator image segmentation by using the Hough straight line detection to extract fault suspension and insulator line on fault location,its accuracy is 96.5% and frame rate is 45frames/s,so as to meet the requirement of the fault detection of real,it is more suitable for catenary suspension and insulator fault accurate positioning.