Research On Defect Detection Technology Of Key Parts Of Pantograph-catenary Based On Deep Learning

In recent years,rail transit has developed rapidly in the process of urbanization,and the operating time and frequency of departures have continued to increase.While improving the efficiency of urban operation,it also requires higher-quality rail transit operation guarantees.As the power supply part of the vehicle,the pantograph-catenary system is essential to ensure the safe power supply and stable operation of the train.Therefore,the defect detection of the core components of the pantograph-catenary system is of great significance to ensure the safety of rail transit operations.At present,the defect detection of pantograph parts mainly has two methods based on traditional image processing and deep learning.Traditional image processing technology has high requirements for the stability of image quality,and it is difficult to adapt to the complex underground and above-ground lighting environment of urban rail transit,which causes the problem of low robustness for detection in massive video streams.However,the deep learning method is less effective in identifying parts in the pantograph-catenary environment with complex environmental backgrounds,and the parts in the pantograph are of different sizes and complex structural features,making it particularly difficult to detect multiple defects.Therefore,aiming at this problem,this paper proposes a defect detection method and idea based on improved network structure of the core components of the pantograph-catenary system.First of all,in view of the problem of low robustness and poor recognition effect of the YOLOv3 object detection algorithm in the detection of pantograph-catenary components,a method to improve the network structure is proposed.In this method,a residual module based on dense connections is designed to enhance the shallow feature extraction,and merge with the final convolutional layer feature of the module to improve the recognition ability of small components.At the same time,an uncertain prediction module based on Gaussian is added to the prediction box to reduce the error prediction results and improve recognition accuracy.Three groups of experiments show that the improved algorithm improves the recognition ability of small targets.and the detection accuracy is increased by 6.4% compared with the traditional YOLOv3 object detection algorithm.Secondly,aiming at the difficulty of defect detection caused by different sizes,complex structures,and numerous types of defects in key parts of the pantograph-catenary system,a method based on an improved NTS-Net detection network is proposed.Firstly,a fine-grained attention module CRAM is added to the NTS net network feature detector,which uses one-dimensional correlation convolution method to optimize the detection performance of various defects,thereby enhancing the detection ability of fine-grained defect features in the image.At the same time,a multi-scale feature reuse module is designed on the overall structure of the network,which enhances the detection ability of subtle defects through the multi-scale feature reuse of the bottleneck layer.Finally,three groups of experiments show that the improved algorithm improves the defect detection ability.Compared with the traditional fine-grained detection network NTS-Net,its defect detection accuracy is increased by 2.74%.