Arcing Detection Algorithm For Pantograph-catenary Based On Semantic Segmentation And Generative Adversarial Networks

Pantograph-Catenary system is an important part of high-speed railway traction power supply system.Its good current collection quality is an important guarantee for the safe operation of high-speed railway.During the high-speed operation of the train,the current collection quality of the Pantograph-Catenary system will be affected due to the vibration of the Pantograph-Catenary system,pantograph catenary defects and other factors,resulting in abnormal contact of the Pantograph-Catenary system and off-line sparks or even arcing,so as to reduce the service life of the catenary and pantograph,and even endanger the safety of train operation in serious cases.Therefore,studying the visual detection algorithm of Pantograph-Catenary arcing,dynamically monitoring Pantograph-Catenary arcing,and improving the automation and intelligence level of Pantograph-Catenary detection system is of great significance and practical value for analyzing the current collection quality during Pantograph-Catenary operation,and then monitoring the operation state of Pantograph-Catenary intelligently in real time.Aiming at the difficulty that the arcing detection algorithm based on deep learning is extremely dependent on arcing data,and the arcing data of pantograph and catenary is difficult to obtain,an unsupervised arcing detection algorithm for arcing target is proposed in this paper.Firstly,the Pantograph-Catenary image data is obtained by digital image processing operations such as video frame extraction and image clipping.In addition,the pantograph and catenary area in the image are labeled by the image annotation tool to obtain the Pantograph-Catenary semantic annotation image.These Pantograph-Catenary image data and Pantograph-Catenary semantic annotation image together constitute the Pantograph-Catenary dataset in this paper.Then,in order to obtain a more stable Pantograph-Catenary image,a Pantograph-Catenary segmentation model based on U~2-Net network is constructed.In this paper,by improving the loss function of U~2-Net network,the model can effectively identify the semantic information of Pantograph-Catenary image data,realize the semantic segmentation task of Pantograph-Catenary area in Pantograph-Catenary image data,and obtain the semantic image of pantograph catenary.Then the Pantograph-Catenary semantic image is input into the SPADE Pantograph-Catenary scene generation model integrating attention mechanism,so as to obtain the generated normal Pantograph-Catenary image.SPADE image generation model expands the traditional BN layer into three dimensions and retains more information of Pantograph-Catenary semantic image,so that the semantic information can be more effectively transmitted in the generative adversarial networks.In addition,in view of the difficulty of SPADE in generating small Pantographs,this paper adds an attention mechanism to the SPADE generation network,so that the network model can have a better generation effect for Pantographs-Catenary system.Finally,the improved difference method is used to make the difference between the Pantographs-Catenary image data and the generated normal Pantographs-Catenary image to obtain the arc difference image,and the arc detection area is obtained by digital image processing methods such as Gaussian filtering,ROI extraction,closed operation and maximum connected-domain determination.The experimental results show that the accuracy of the method proposed in this paper is as high as 99.93%,and the mean Intersection over union and F1-Score reach 61.55%and 69.63%,respectively.