Research On Defect Recognition Technology Of Urban Railway Track Based On Image And Sound

The safety status of urban railway tracks is an important guarantee for a safe and efficient operation of train services.Track defects will directly affect the safety of urban railway systems and would lead to accidents if those defects were serve.Therefore,it is crucial to adopt a timely and efficient track status detection system.Currently,the maintenance of tracks mainly relies on manual methods,which means a regular inspection is delivered via using hand-push test cars.But new challenges and higher inspection requirements follows with the increasing scale of urban railway network in a city.In recent years,new technologies,such as image processing technology,audio processing technology,and machine learning theory,has growingly dominated many applied aspects,which could provide new study directions and technical supports for realtime monitoring of tracks defects.Based on this soaring trend,Research on fastener defects and wheel-rail vibration noise in track defects,which take image and sounds as data sources,is conducted.This thesis also designs a track detection system which can support the rail intelligent and informatization detection.The main Research content covers the following aspects:An introduction of track defect detection content and related algorithms.In this chapter,the common types and manifestations of current track defects is introduced.This thesis mainly focuses on defections of Fastener Defects(FD)and Rail Surface Defects(RSD).The data source of FD and RSD are images and wheel-rail vibration noise(WRVN)Respectively.The cor Responding technique outlines are developed.Machine-learning algorithms and Deeping models are used,and Convolutional Neural Network is introduced.A comparison between VGG16 model and the deep Residual network model is given.The tree model algorithm(XGboost)in machine learning is introduced and cited in the wheel-rail vibration noise and classification.A recognition method of fastener defects based on images.In this chapter,the thesis has finished several tasks.Firstly,uneven sample distributions in anomaly detection is solved by adopting an image data feature enhancement method.Secondly,different fastener defects are identified via a trained VGG16 network.Thirdly,fasteners positioning,and fastener status is detected via a target network framework(Faster R-CNN)generated from VGG16,whose detecting speed and efficiency has ismproved significantly.Two rail defects identification methods based on WRVN.In this chapter,the thesis has finished several tasks.Firstly,the acoustic characteristics of WRVN and other common onboard sounds are analyzed.And the time domain and the frequency domain characteristics and differences among various types of sounds are stated.Secondly,two WRVN identification outlines are given via audio data enhancement method,which could add Ress insufficient audio data and manual operation difficulty.The first one is an audio feature extraction,which refers to short-term average energy in the audio,the short-term zero-crossing rate,the 1/3 octave spectrum sound p Ressure,and the log-mel spectrum coefficient and other featu Res,are achieved by training a XGboost Classifier.The second one is the log-mel spectrum of audio data extraction via a deep Residual network.Both methods are proved to be effective.A design of track defect detection system In this chapter,an integrated system,which covers the functions mentioned above,with support of Py Qt5 platform is developed.This system has embedded machine learning algorithms and deep learning algorithms to deliver the identification of fastener defects and WRVN defects.Also,this system can automatically generate test Result reports for references.