Design And Implementation Of Catenary Foreign Body Detection System

Railway transportation is an important means of transportation for trade in China,and the safety of power supply of railway catenary is an issue that cannot be ignored.There are many kinds of foreign objects in railway catenary and they are not easy to be monitored in real time.Therefore,how to make a real-time and accurate detection of intrusions along complex railways has become the focus of scholars.In this paper,two methods,semantic segmentation and target recognition,are used to detect foreign objects in the catenary.Semantic segmented foreign object detection mainly aims at the real-time and accuracy of detecting unknown categories of railway intrusions.Catenary parts are taken as the research object,and the detection status of each part is analyzed to determine whether there are foreign objects in the catenary.The method of target recognition is to take the bird’s nest as an example,in order to detect whether the size of the bird’s nest is fixed and the edge is clear,as well as real-time and accuracy,this paper takes the bird’s nest of the catenary as the research content,and then improves the training of the model by analyzing the factors that affect the detection of bird nests of different sizes and their unclear edges,so as to achieve foreign object detection of the catenary.The main research contents are as follows.1)Design the SUBA-DeepLabV3s+ foreign object detection model.This model uses a semantically segmented approach to determine foreign objects of unknown class.The model first obtains the features of catenary parts by using the feature extraction network Mobile Netv3＿small model.Then,the SUBA structure is designed for the foreign object of catenary which has the characteristics of unfixed hanging position and random shape and size..The structure is composed of the improved spatial attention mechanism and the channel attention mechanism in parallel.The spatial attention mechanism is to replace ordinary convolution with void convolution.The channel attention mechanism is a convolution that replaces the MLP of a multilayer perceptor with a convolution core of c.Finally,the maximum continuous length algorithm is used to determine whether foreign objects exist in the catenary.In order to verify the performance of the SUBA-DeepLabV3s+model,experiments are compared with five models: FCN-8s,Seg Net,U-Net,etc.The results show that the average Intersection over Union and latency of this model are the best.2)Design the ESA4-YOLOV5 s foreign object detection model.The model uses target recognition to identify foreign objects in a bird’s nest.The model extracts the nest features using the Efficient Net-B4 model.The MSPP structure is designed to solve the problem of unclear edge characteristics of foreign object in the catenary.The structure incorporates the output of the upper ASPP layer into the input of the lower layer and introduces an unprocessed path,which enhances feature extraction while guaranteeing the original information.The structure of SUBA and MSPP is used to obtain abundant multiscale feature information.Finally,the prediction module is combined to determine whether foreign objects exist in the catenary.In order to verify the performance of ESA4-YOLOV5s+ model,the experiments are compared with four models,YOLOV4,SSD,Faster R-CNN,etc.The results show that this model performs best at average accuracy and frame rate.3)The design and implementation of the foreign object detection system for the catenary have been completed.The system designs and implements some functional modules including object detection and semantics segmentation.Visualizing the statistical analysis of foreign objects in the catenary with visualization tools makes it easier to serve the staff.