Research On Visual Detection Technology Of Pantograph-catenary Arc Based On Instance Segmentation

The pantograph–catenary system is an important part for electrified trains to obtain electrical energy from the traction power supply system.Its health status directly affects the quality of the electrified train’s current flow.The pantograph–catenary arc is an important factor in the health of the pantograph–catenary system.The realization of the visual detection and intelligent state analysis of the pantograph–catenary arc is of important significance to the intelligent operation and maintenance of the pantograph–catenary system.This paper constructs a pantograph–catenary arc dataset,and studies the main key technologies involved in the process of pantograph–catenary arc detection and analysis,including the instances segmentation of pantograph–catenary arc based on deep neural network and the analysis of pantograph–catenary arc based on foreground and background joint decision and centroid matching.In order to meet the data requirements for the instance segmentation algorithm training of the pantograph–catenary arc in this paper,the pantograph–catenary arc data set construction method is designed to solve the problems of huge amount of data collected on site,too much noise data,incomplete coverage of scenes,and sparse arcing instances in a single frame image.According to the noise characteristics of the pantograph–catenary data,automatic cleaning and manual cleaning are used to filter the noisy images in sequence,the image is segment annotated using Labelme,and the three data enhancement methods of digital image processing,Simple Copy-Paste and Self-Training are used to generate images data in multiple scenes improve the validity of the data,finally complete the construction of the pantograph–catenary arc data set.In order to realize the pixel-wise level fine-grained detection of pantograph–catenary arc,on the basis of Blend Mask instance segmentation network,a deep neural network model Arc Mask suitable for pantograph–catenary arc detection is established.Aiming at the problems of diversified arc shapes,high speed requirements,and complex and changeable backgrounds in the pantograph–catenary arc detection,various network improvements such as adding deformable convolution to the backbone network,replacing the feature pyramid network with a bi-directional feature pyramid network,and designing an attention-based multi-scale feature fusion module,so as to improve the effect and speed of the pantograph–catenary arc detection.In order to realize the analysis of the pantograph–catenary arc state based on the visible image,explore the method of detecting the pantograph–catenary arc energy level in the joint decision-making of the foreground and background,reduce the interference of the background on the judgment of the energy level,design the calculation method of arc centroid matching in the sequence frame,combining the pantograph–catenary arc energy level and the matching result can calculate the arc-related measurement values in the operation control section of the electrified trains.In order to verify the effect of the arc detection and status analysis methods of the pantograph–catenary arc designed in this paper,ablation experiments were carried out on Arc Mask,to evaluate and verify each improvement of the Arc Mask network,and to show the analysis results of the pantograph–catenary arc state.The experimental results show that the arc detection and status analysis method designed in this paper can not only realize the pixelwise level fine-grained detection of the pantograph–catenary arc,but also can better analyze the arc state,which provide a wealth of information for the automation and intelligent online monitoring of the pantograph–catenary system.