Catenary Foreign Object Detection Based MobileNetV3

With the rapid development of high-speed railways,electrified railways are also facing many problems in their development.In this regard,my country has gradually become the country with the fastest development and the most comprehensive technology.In the development of electrified railways,the role of the railway catenary is very important,and how to ensure the safe operation of the catenary is one of the important tasks.To conduct abnormal analysis and detection of common faults and abnormal operating conditions of the catenary,it is necessary to formulate preventive and problem-solving measures to ensure the safety of the operating status of the catenary.The catenary itself is an extremely complex system.In the process of its research,it faces problems such as excessive image data,complex image background,unbalanced data samples,and the influence of the natural environment.Using traditional manual identification and analysis of vehicle video,the detection accuracy is greatly affected by human factors,and the accuracy rate is not high.If the problem is to be detected accurately and timely,it is necessary to start research with automation and intelligence and adopt computer vision processing methods.In recent years,the use of deep learning for foreign object detection has become a research hotspot,but the uneven amount of collected data affects the detection effect of deep learning.At present,the detection of bird’s nests in the high-speed railway catenary mainly adopts the target detection method.Due to the problems of very few abnormal samples in bird’s nest data,unbalanced data,and special characteristics of bird’s nest samples,the automatic location,and recognition of bird nests is more difficult.In response to the above problems,this article analyzes and studies the characteristics of the entire image data of the bird’s nest,extracts the area where the bird’s nest is located,performs outlier detection,combines the classification network to extract the characteristics of the bird’s nest,and realizes the detection of foreign objects in the bird’s nest on the hard beam of the catenary.Through image processing technology and deep learning methods,relevant researches are carried out on the recognition of hard beam foreign objects.The task is transformed from a target recognition problem to a classification problem,and the bird’s nest detection is carried out with the idea of classification.Finally,the feasibility of the technical scheme was verified through experiments.This article discusses the following aspects:(1)In order to solve the problem of the unbalanced data of the bird’s nest on the hard beam,this paper analyzes the characteristics of the bird’s nest itself,realizes the detection algorithm of the branch point with the detection area,realizes the data acquisition through the point detection algorithm,and determines the correctness of the data set according to the principle of sample selection.Negative samples,make relevant data sets.Finally,data enhancement is performed on the acquired data to increase the amount of training data.(2)The detection of small samples of bird nests based on supervised learning is studied.Through the feature detection of the bird’s nest branch point,the problem of target recognition and positioning is transformed into a classification problem,combined with the classification network such as MobileNetV3 to realize the detection of the bird’s nest of the rigid beam,and the detection results are compared and analyzed.To verify the effectiveness and feasibility of the method of detecting foreign objects on the hard beam of the catenary with the classification idea.(3)Research on the classification and detection of bird’s nest data based on semi-supervised learning.The self-training method is used to process a large amount of unlabeled data to solve the problem of low efficiency of manual data classification,and the MobileNetV3 network is also used to realize the detection of the hard beam bird’s nest.Analyze the effect of using semi-supervised learning to classify data in the detection of foreign objects in the catenary hard beam.