A rail transit wheel-rail wear prediction method and model fusing data

Abstract

The application discloses a kind of track traffic wheel rail wear prediction methods of fusion data, comprising the following steps: physical model data acquisition: obtain physical model data profile evolution matrix A;Measured data acquisition: obtain measured profile evolution matrix B;Data layer fusion: physical model data and measured data are fused, and profile evolution matrix C is obtained;Prediction model establishment: the convolution-multiple recurrent neural network architecture for wear prediction is established;Prediction model training: physical model data and field data are used as input, and prediction model is trained;Wear prediction: input a section of physical model data and field data fusion matrix C * , and profile wear matrix D is output by prediction model.The method proposed in the application can give full play to the advantages of physical model and field measured data, carry out data fusion in input layer, improve the generalization of prediction network model, widen the effective length of prediction time, and improve prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rail transit, and in particular to a rail transit wheel-rail wear prediction method and model fusing data. BACKGROUND

[0002] Wheel-rail wear prediction is one of the key technologies required in the field of rail transit, and is crucial for the safe service and orderly operation of trains. Traditional wear prediction methods are based on physical model calculations, which can easily overlook the impact of changes in environmental and passenger flow factors during service. Prediction methods based on field data have become popular in recent years, but their accuracy is low when data undergoes sudden changes or strong nonlinear transformations due to their excessive reliance on data integrity. SUMMARY

[0003] The present application aims to provide a rail transit wheel-rail wear prediction method and model that fuses physical model data and measured data to address the problems identified in the background.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] A rail transit wheel-rail wear prediction method fusing data, comprising the following steps:

[0006] (1) Physical model data acquisition: Establish a theoretical model to obtain the physical model data profile evolution matrix A of the wheel or rail;

[0007] (2) Measured data acquisition: Collect profile data of the wheel or rail using a profilometer to obtain the measured profile evolution matrix B;

[0008] (3) Data layer fusion: Fuse the physical model data and measured data to obtain the profile evolution matrix C;

[0009] (4) Prediction model establishment: Establish a convolution-multi-cycle neural network architecture for wear prediction;

[0010] (5) Prediction model training: Use the physical model data and field data as inputs to train the prediction model;

[0011] (6) Wear prediction: Input a matrix C that fuses a segment of physical model data and field data * , and the prediction model outputs the profile wear matrix D.

[0012] Further, the physical model can be established using multi-body dynamics software based on the actual mass and moment of inertia parameters, stiffness, and damping of the vehicle.

[0013] Further, the prediction model training includes the following steps:

[0014] First, set the neural network related parameter initial value; then, based on the prediction data and the field data to construct the comprehensive loss function; again, in a supervised learning way, set the training method parameters; finally, train the model parameters, until the model converges.

[0015] In the present application, a track traffic wheel rail wear prediction model fusing data adopts a convolutional neural network as a model front end for extracting spatial features of fused data; adopts n recurrent neural networks as a model back end for extracting time series information.

[0016] Further: the convolutional neural network and the recurrent neural network adopt a series structure, and multiple recurrent neural networks adopt a parallel structure, and the model has m network model parameters in total.

[0017] Compared with the prior art, the method proposed in the present application can fully exert the advantages of physical models and field measured data, carry out data fusion of the input layer, improve the generalization of the prediction network model, widen the effective length of the prediction time, and improve the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the embodiments will be briefly introduced below, the drawings herein are incorporated into the specification and form a part of the specification, the drawings show the embodiments consistent with the present disclosure, and are used to illustrate the technical solutions of the present disclosure together with the specification. It should be understood that the following drawings only show certain embodiments of the present disclosure, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0019] Figure 1 It is a flow chart of a track traffic wheel rail wear prediction method fusing data.

[0020] Figure 2 It is a structure schematic diagram of a track traffic wheel rail wear prediction model fusing data.

[0021] Figure 3 It is a schematic diagram of a physical model based on multi-body dynamics in the embodiments of the present application.

[0022] Figure 4 It is a comparison diagram of the wear profiles after training the measured data and the prediction data of the present method in the embodiments of the present application.

[0023] Figure 5 It is an error schematic diagram of the deepest wear amount and the measured wear amount in the embodiments of the present application. DETAILED DESCRIPTION

[0024] The technical solutions of the patent will be further described in detail in combination with specific implementation manners. The detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the claimed present disclosure, but only represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present disclosure.

[0025] Please refer to Figure 1 A rail transit wheel-rail wear prediction method and model fusing data, comprising the following steps:

[0026] 1) Physical model data acquisition: a theoretical model is established by rigid body dynamics, multi-body dynamics, etc., the same line conditions and load conditions as the measured conditions are set, the profile evolution data of the wheel or rail are calculated to form a measured profile evolution matrix A composed of n x t elements,

[0027]

[0028] Wherein, n represents the number of points measured along the transverse direction, and t represents the time point of sampling.

[0029] 2) Measured data acquisition: profile data of the wheel or rail are collected by a profile gauge to form a measured profile evolution matrix B composed of n x t elements,

[0030]

[0031] 3) Data layer fusion: the physical model data and the measured data are fused to construct a profile evolution matrix C composed of 2n x t elements,

[0032]

[0033] 4) Prediction model establishment: the present application proposes a convolution-multi-recurrent neural network architecture as shown in Figure 2 Firstly, a convolutional neural network is used as the front end of the model to extract the spatial features of the fused data. Then, n recurrent neural networks are used as the back end of the model to extract the time series information. The convolutional neural network and the recurrent neural network adopt a series structure, and the multiple recurrent neural networks adopt a parallel structure. The model has a total of m network model parameters.

[0034] 5) Prediction model training: the physical model data and the field data are used as inputs to train the prediction model. Firstly, the initial values of the neural network related parameters are set; then, a comprehensive loss function is constructed based on the prediction data and the field data; then, the parameters of the training methods such as the steepest descent method and the Adam method are set in a supervised learning manner; finally, the model parameters are trained until the model converges.

[0035] 6) Wear prediction: input a matrix C* of physical model data and field data fusion, and the profile wear matrix D is output by the prediction model.

[0036] In order to facilitate the public to understand, in the embodiment, the wheel tread profile wear prediction of rail transit vehicle in a city is taken as an example to illustrate the application effect of the application.

[0037] 1) According to the actual mass and moment of inertia parameters, stiffness and damping of the vehicle, as shown in the formula (1), a multi-body dynamics software is used to establish a physical model, a line state consistent with the actual measurement environment is designed, simulation is carried out, and wheel wear profile data after 10,000 km, 20,000 km, 30,000 km, 40,000 km, 50,000 km, 60,000 km, 70,000 km, 80,000 km and 90,000 km of operation mileage are obtained, which are recorded as A. Figure 3

[0038] 2) The profilometer is used to collect wheel profile data, and the wheel profile after 10,000 km, 20,000 km, 30,000 km, 40,000 km, 50,000 km, 60,000 km, 70,000 km, 80,000 km, 90,000 km, 120,000 km, 150,000 km, 180,000 km and 210,000 km of vehicle operation are collected respectively.

[0039] In the embodiment, the measuring instrument used for collection is CALIPRI non-contact wheel-rail profile measuring instrument, which is composed of a portable tablet computer and an optoelectronic measuring sensor.

[0040] 3) The physical model data and the measured data are fused, the data from 10,000 km to 90,000 km are selected as the training set, and the profile evolution matrix C is formed.

[0041] 4) A convolution-multiple cycle neural network is established, based on the profile evolution matrix C, the Adam method is used for supervised learning of the network parameters, until the model converges, and the trained prediction network model is obtained.

[0042] 5) As shown in the formula (5), the measured data and the physical model data from 30,000 km, 60,000 km and 90,000 km are fused into a matrix C* as input data, and the wheel wear profile at 120,000 km, 150,000 km, 180,000 km and 210,000 km is predicted, and the predicted profile wear matrix D is obtained. Figure 4 On this basis, the difference between the prediction result and the measured data obtained by collection is analyzed, as shown in the formula (6), it is found that the error of the prediction result gradually increases with the increase of the prediction time, but the overall prediction error is within 10%.

[0043] Figure 5

[0044] ​​​The preferred embodiments of the present patent have been described in detail, but the present patent is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present patent.

Claims

1. A rail transit wheel-rail wear prediction method fusing data, characterized in that, Comprise the following steps (1) Physical model data acquisition: establish a theoretical model to obtain the physical model data profile evolution matrix A of the wheel or rail; (2) Real-time data acquisition: track the profile data of the wheel or rail through the profiler to obtain the measured profile evolution matrix B; (3) Data layer fusion: fuse the physical model data and the measured data to obtain the profile evolution matrix C; (4) Prediction model establishment: establish a convolution-multiple recurrent neural network architecture for wear prediction; (5) Prediction model training: take the physical model data and the field data as input to train the prediction model; (6) Wear prediction: input a matrix C of physical model data and field data fusion, and the profile wear matrix D is output by the prediction model. * , The above steps specifically include: 1) Physical model data acquisition: establish a theoretical model through rigid body dynamics and multi-body dynamics method, set the same line conditions and load conditions as the measured conditions, calculate the profile evolution data of the wheel or rail to form a measured profile evolution matrix A composed of n x t elements, Wherein, n represents the number of points measured along the transverse direction, and t represents the sampling time point; 2) Real-time data acquisition: track the profile data of the wheel or rail through the profiler to form a measured profile evolution matrix B composed of n x t elements, 3) Data layer fusion: fuse the physical model data and the measured data to construct a profile evolution matrix C composed of 2n x t elements, 4) Prediction model establishment: first, adopt convolution neural network as the front end of the model to extract the spatial features of the fused data; then, adopt n recurrent neural networks as the back end of the model to extract the time sequence information; wherein, the convolution neural network and the recurrent neural network adopt a series structure, and the multiple recurrent neural networks adopt a parallel structure, and the model has m network model parameters in total; 5) Prediction model training: first, set the initial value of the neural network related parameters; then, construct a comprehensive loss function based on the prediction data and the field data; then, set the parameters of the steepest descent method and Adam method training method in a supervised learning manner; finally, train the model parameters until the model converges.

Images