Research On Intelligent Fault Identification Method Of Catenary Parts Based On 3D Point Cloud Aided Analysis

Catenary is an important part of the electrified railway power supply system,which plays a key role in transmitting electric power to electric locomotive.In recent years,with the rapid development of non-contact detection technology,the fault detection accuracy and intelligent level of the catenary support device components has been greatly improved.However,2D image data has its own limitations,in complex catenary scenes,there are occlusions between arms and wrists,incomplete image features seriously affect the accuracy of intelligent positioning and recognition of parts.For small target of catenary,unacceptable local image pixels and unconspicuous image features,resulting in a decrease in recognition accuracy.In response to the above problems,this article comprehensively considers the 2D image data of the catenary wrist arm system combined with the synchronously collected 3D point cloud data,based on the deep learning method,the high-precision bad state detection of catenary arm parts is realized.First of all,the new generation of Time of Flight depth camera designed by Microsoft was used,a synchronous acquisition system of 3D point cloud and 2D image data of catenary was established at the same time.Field tests have verified the effectiveness of simultaneous3 D and 2D data collection of the cantilever structure of catenary system.The virtual point cloud simulation acquisition environment was constructed,provide data support for 3D point cloud deep learning,and overcome the insufficient amount of data collected on site.The measured/virtual point cloud data sample database is established,which lays a research foundation for the subsequent crack detection research based on 3D point cloud and 2D images.Secondly,in order to solve the problem of wrist and arm occlusion in complex catenary scene,the 3D point cloud depth information is used to realize the separation of the front and rear rods of the 2D image double-arm structure.Aiming at the problem of 3D point cloud data component segmentation,this paper attempts to apply a 3D Point Convolution Neural Network based on point convolution.The accurate segmentation of 16 kinds of components is realized,the average accuracy of segmentation can reach 96.5%,and the average error range is 2.0%～2.6%.Based on the synchronous acquisition correspondence between 3D depth image and 2D image,the mapping between 3D point cloud of catenary and 2D image was established,the segmentation results are mapped to 2D images to achieve accurate segmentation of various catenary parts.2D images of catenary components assisted by point cloud based on point convolution can achieve better segmentation effect.Then,in order to solve the problem of unacceptable local image pixels and unconspicuous image features of small parts in catenary,this article attempts to apply a natural realistic image super resolution neural network algorithm to achieve the cantilever structure of catenary system component image super-resolution,increase the image clarity of small parts and reinforce the local features of them.Finally,for the crack detection of the rotating ears,applying images processed by front end positioning.You Look Only Once Version5 is applied to realize crack detection of rotating ear,single frame image inference only takes 7ms.Compared with the existing methods of crack detection of rotating ear,the method used in this paper has higher detection accuracy and better detection timeliness.