Research On High-Speed Train Wheel Tread Flats Detection Algorithm Based On Rail Dynamic Response

With the continuous improvement of China’s high-speed railway mileage,it is very important to ensure the safety of train operation.Wheelset is a part that directly contact with rail.The health status of wheelset tread directly determines the quality and safety of train operation.The train with wheel tread flat will increase the impact between train and rail during operation,which has an important impact on the stability and safety of train operation.Realizing the real-time detection of wheel tread flat fault has important practical significance for ensuring the safety of high-speed trains.The scheme based on the vibration acceleration sensor installed on the rail to identify the wheel flat fault has the advantages of high detection efficiency and small equipment maintenance workload.However,due to the complex vehicle-rail coupling relationship and the complex components of rail vibration response signal,the detection algorithm has become the difficulty and key to the implementation of the scheme.Therefore,this paper studies the online detection algorithm of wheel tread flat fault based on rail dynamic response.In this paper,based on the vehicle-track vertical coupling model simulation,rail vibration response acceleration data under different running speeds and different wheel tread flat depth are obtained.Based on the network model of deep learning,the depth of wheel tread flat is realized by taking the acceleration of rail vibration response as input.Firstly,the vehicle-track coupling model is established,and the dynamics mathematical models of the vehicle system and the track system are established respectively.The wheel-rail force model is established using Hertz nonlinear contact theory to realize the coupling of the vehicle and the track system.The inverse Fourier transform method is used to simulate the track irregularity as the excitation of the vehicle-rail coupling system.The wheel tread flat model is established,and the vehicle-track coupling model is solved by numerical integration method,and the rail vibration response is obtained.Design of multi-sensor acquisition algorithm based on Entropy,and the rail vibration acceleration data containing the most abundant wheel flat fault information is extracted by calculating the data sample entropy.Hilbert-huang transform and wavelet transform are used to analyze the rail vibration response acceleration,and two time-frequency images are obtained to construct the dataset.Then a residual depth network(Res Net)algorithm based on Multi-Sensor Fusion(MSF)is proposed to detect wheel tread flat defect.The convolutional neural network is used to perform intelligent feature fusion of time-frequency images,and residual depth network is constructed to extract time-frequency image features of deep fusion,so as to realize the depth of wheel tread.Hilbert spectral image dataset and wavelet time-frequency image dataset are used as the network input respectively.The results show that the classification accuracy of scratch depth can reach 99.38 %.Then a wheel tread flat detection algorithm based on multi-stage feature fusion is proposed for different scale fault features in time-frequency images.Convolutional neural network realizes time-frequency image intelligent fusion,FPN module is used to realize feature extraction of different scales in different feature map scales,and channel attention mechanism is added and improved in each stage to improve the learning ability of the network.With the wavelet time-frequency image dataset as the network input,the classification accuracy of scratch depth can reach 99.82%.Finally,the wheel tread flat fault is simulated on the wheel/rail defect simulation test platform,and the track vibration acceleration data under different speed and wheel flat depth are collected according to the acceleration sensor installed on the track wheel.After the track vibration acceleration is transformed by wavelet,the algorithm in Chapter 5 is verified by using the wavelet time-frequency image data set.The results show that the classification accuracy of flat depth is 100%.