Fine-tuning method for high-speed railway track irregularity intervals based on dynamics assessment

Abstract

The present invention relates to the technical field of railway track smoothness maintenance, and provides a method for fine-tuning the unevenness interval of a high-speed railway track integrated with dynamics evaluation, including: 1) establishing a potential track geometric feature extraction framework based on a deep Bayesian network; 2) establishing a proxy model mechanical module expression based on a vehicle-track coupling dynamics equation; 3) establishing a vehicle-track coupling dynamics proxy model based on a Fourier neural network; 4) establishing a vehicle body response loss active control function based on a track unevenness-vehicle body response parameter dense scale constraint algorithm; 5) designing a loss function that embeds adjustment cost and track smoothness quality optimization objectives; 6) embedding multiple track parameter constraint functions such as chord measurement, adjustment boundary, unit adjustment amount, etc.; 7) designing and selecting track fine-tuning schemes under specific constraints; 8) outputting track unevenness interval fine-tuning schemes based on Monte Carlo. The present invention can better achieve high-speed railway track unevenness interval fine-tuning.

Description

Technical Field

[0001] The present invention relates to the technical field of railway track smoothness maintenance, and in particular to a high-speed railway track unevenness interval fine-tuning method integrating dynamics evaluation. Background Art

[0002] Track irregularity refers to the deviation between the actual geometry and the designed geometry during line operation. Therefore, regular inspections and track maintenance are necessary to ensure train operation safety and passenger comfort. At present, maintenance scheme design mainly controls the smoothness of travel indirectly by controlling the static track geometry. In recent years, more and more scholars have focused their research on track geometry maintenance. For general railway lines, some scholars have proposed an iterative optimization algorithm based on static track irregularity data and constrained linear programming to achieve high-quality optimization of track geometry. Some scholars have implemented maintenance scheme design based on dynamic track irregularity data on the basis of the design target curve. On this basis, some scholars have considered the theoretical and actual adjustment effects and proposed a multi-objective optimization method based on Pareto optimal theory to achieve autonomous selection of multiple maintenance schemes. For railway lines on kilometer-level bridges, some scholars have studied the adjustment method by analyzing the changing characteristics of track geometry. Some scholars have considered the characteristics of bridge geometry and random track irregularity at the same time, proposed a method to optimize the existing geometry, and applied vehicle-track coupling dynamics simulation to scheme evaluation. In addition, some researchers have also proposed the use of wavelet transform for track irregularity optimization, however, they did not consider geometric parameters and the number of adjustable fasteners because they focused on achieving a smooth track alignment. Undoubtedly, the current maintenance scheme design is based on track irregularity data, while the scheme evaluation relies on vehicle-track coupled dynamics simulation. However, these two processes are strictly separated, and the interaction between the dynamic evaluation and scheme design process occurs after the scheme design. Due to the two pressing technical issues of design method solution and dynamic evaluation efficiency, no research has yet been conducted to achieve scheme design that integrates the evaluation and design processes. Summary of the invention

[0003] The content of the present invention is to provide a method for fine-tuning the unevenness interval of high-speed railway track integrating dynamic evaluation, which can solve the problems that the existing track smoothness maintenance scheme is only designed from the perspective of track parameter optimization, the influence of track unevenness uncertainty propagation is not considered in the scheme design process, and the dynamic response cannot be actively and efficiently controlled.

[0004] The method for fine-tuning the high-speed railway track irregularity interval integrating dynamics evaluation according to the present invention comprises the following steps:

[0005] 1) Establish a latent track geometry feature extraction architecture based on a deep Bayesian network;

[0006] 2) Establish the mechanics module expression of the proxy model based on the vehicle-track coupling dynamics equation;

[0007] 3) Establish a vehicle-track coupling dynamics proxy model based on Fourier neural network;

[0008] 4) Establishing the vehicle response loss active control function based on track irregularity-vehicle response parameter dense scale constraint algorithm;

[0009] 5) Design a loss function that embeds the adjustment cost and track smoothness quality optimization objectives;

[0010] 6) Embed multiple orbital parameter constraint functions, including chord measurement, adjustment boundary, and unit adjustment amount;

[0011] 7) Design and selection of track fine-tuning schemes under specific constraints;

[0012] 8) Output of Monte Carlo-based track irregularity interval fine-tuning solution.

[0013] Preferably, in the potential track geometry feature extraction architecture based on a deep Bayesian network in step 1), a Bayesian deep learning method is used to convert the point estimates of network weights and bias parameters into distribution estimates in each potential track geometry feature extraction layer, and multiple estimated parameters are obtained by repeated sampling; the Bayesian formula is as follows: ;

[0014] in It indicates the parameters that must be trained for each layer of the network after the input track is not smooth. express The posterior distribution of represents the likelihood function, represents the prior distribution;

[0015] The Monte Carlo Dropout method is used to obtain the posterior distribution. The potential feature extraction network using Dropout is shown below:

[0016] ;

[0017] in and Represent the number of layers and neurons respectively; Represents the potential track geometry feature extraction layer i The number of neurons in the layer; and Respectively represent the weights of each neuron before and after Dropout; represents the random activation or deactivation coefficient; Represents the potential track geometry feature extraction layer i Parameter distribution of layer neurons; Indicates application to l The random activation coefficient of the layer; and express Output values ​​before and after layer Dropout; and express layer The output value and bias of the neuron.

[0018] Preferably, in the proxy model mechanics module based on the vehicle-track coupling dynamics equation in step 2), the vehicle-track coupling system is described as a set of coupled ordinary differential equations , as shown in formula (3):

[0019] ;

[0020] in and Respectively indicate that and The physical parameters and system excitations of the vehicle-track coupling system of each element, including: vehicle body inertia, frame inertia, primary and secondary suspension stiffness and damping, vehicle characteristic length, fastener pad stiffness and damping, vehicle speed, and track irregularity parameters; express All solution functions in, including: vehicle body vertical vibration, vehicle body nodding, front and rear frame vertical vibration, front and rear frame nodding and vertical vibration of 4 wheel sets; Represents the time coordinate; Represents a bounded time domain.

[0021] Preferably, in the vehicle-track coupling dynamics proxy model based on Fourier neural network in step 3), a Fourier neural network is trained to approximate the ordinary differential equations of the vehicle-track coupling system Solver of the vehicle-track coupling system ordinary differential equations The solution operator By creating a mapping To approximate, as shown in formula (4):

[0022] ;

[0023] The dynamic system parameter combination and target dynamic response combination Both belong to finite-dimensional Banach function space ; represents the parameters of the Fourier neural network; and Representation mapping; represents a bounded set in the parameter space of the vehicle-track coupling system ; represents the corresponding dynamic response set, ;

[0024] Loss function of Fourier neural network Including data loss , equation loss and direct derivative loss ;

[0025] ;

[0026] in In formula (3), The average error between the solution and the network output; express The mean square norm in the time domain is obtained by inputting the network output, derivative values ​​and system parameters into the ordinary differential equations of the vehicle-track coupling system obtained in the system; Represent the 1st and 2nd order derivative losses respectively; and Respectively , and Weight coefficients of the three loss functions; N Represents the time domain T The length of the data within.

[0027] Preferably, in step 4), the vehicle body response loss active control function design based on the track irregularity-vehicle body response parameter dense scale constraint algorithm is mainly used to realize the conversion between the track fine-tuning scheme design and the dynamics evaluation task, specifically:

[0028] Step 4.1) Space-time transformation of track irregularity: treat the track irregularity on each track as a spatial discrete sequence , and convert it into a time series according to the average train speed v ;

[0029] ;

[0030] Step 4.2) Parameterization of discrete time series of track irregularity: Parameterize and assign parameter values ​​to it , as shown in formulas (7) to (9):

[0031] ;

[0032] in and Respectively represent the track unevenness sequence length and the position of each point; Represents the arc length between adjacent points; represents the cumulative arc length;

[0033] Step 4.3) Establish the continuous curve expression of track irregularity time series: Determine the node vector of track irregularity discrete sequence and basis functions And the continuity of the discrete time series of track irregularity is achieved by establishing B-spline expression;

[0034] Step 4.4) Dense scale sampling of track irregularity-vehicle response parameters: Densely sample the continuous curve of track irregularity time series at a high-frequency time sampling interval to determine the input of the vehicle-track coupling dynamic proxy model;

[0035] Step 4.5) Design and calculation of vehicle response loss active control function: The densely sampled track irregularity time series is used as an excitation input into the vehicle-track coupling dynamics proxy model to obtain the vehicle acceleration series; due to the distance difference between the front and rear wheel pairs, the processed track irregularity is applied to the four wheels respectively; on this basis, the acceleration active control loss of the design scheme is calculated , as shown in formula (10):

[0036] ;

[0037] in Represents the proxy model calculation process; and They represent the optimized track irregularity spatial sequence and the corresponding vehicle acceleration sequence respectively; Indicates the management threshold of the vehicle acceleration sequence.

[0038] As a preferred method, in step 5), a loss function is designed that embeds the adjustment cost and the track smoothness quality optimization objective to ensure that the design scheme meets the specification requirements from the track level; wherein the adjustment cost loss function is used To minimize the difference between the optimization result and the target curve, to ensure the minimum cumulative adjustment target, and use the track smoothness quality loss function to actively control the overall quality of the adjusted operation; these two losses are defined as follows:

[0039] ;

[0040] in, and Represent the adjusted and original track unevenness order columns respectively; represents the average value of the original track irregularity; N The length of the sequence representing the roughness of the original track.

[0041] As a preferred method, in step 6), by embedding multiple track parameter constraint functions, the model can be prompted to automatically generate a fine-tuning maintenance plan that meets management requirements; , Adjust Boundaries and unit adjustment For the trajectory parameter constraint loss, the constraint set is as follows:

[0042] ;

[0043] in, and Respectively, they represent the adjusted and original track irregularities; N Indicates the length of the track unevenness sequence; H represents the string measurement process matrix; and They represent the management thresholds of track irregularity chord measurement value and adjustment boundary respectively; Represents the linear rectification function calculation process; Represents the rounding function calculation process.

[0044] Preferably, in step 7), the orbit fine-tuning scheme is designed and selected under specific constraints. It is necessary to customize the weight ratio of the loss function under a single fixed constraint combination; execute the optimization model once, and reversely calculate each loss function based on the chain rule to obtain multiple estimation results as a set of fine-tuning design schemes for the fusion dynamics evaluation process, and make personalized selections based on needs.

[0045] As a preferred method, in step 8), the Monte Carlo-based track irregularity interval fine-tuning scheme output is performed, and it is necessary to determine the Monte Carlo Dropout ratio parameter in the deep Bayesian network. and the number of model repetitions ,Pick The mean value of the sub-results and the maximum-minimum envelope are output as the interval fine-tuning solution.

[0046] The present invention establishes a maintenance plan design framework and a dynamic proxy model based on a unified embedded physical neural network architecture. The vehicle-track coupling dynamic proxy model and the traditional physical constraint relationship are embedded in the neural network architecture through the "track irregularity-vehicle response parameter" intensive scale constraint method, thereby realizing a high-speed railway track section fine-tuning method integrating the dynamic evaluation process, ensuring the rapid formulation of long-line overhaul plans, and significantly reducing the workload of overhaul and maintenance personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of a method for fine-tuning a high-speed railway track irregularity section integrating dynamics evaluation in an embodiment;

[0048] Figure 2 It is a schematic diagram of the interval design result in the embodiment;

[0049] Figure 3 A schematic diagram showing a comparison of maintenance solutions in the embodiments;

[0050] Figure 4 This is a schematic diagram of the acceleration active control effect in the embodiment;

[0051] Figure 5 Schematic diagram of the convergence effect of different loss functions in the embodiments. DETAILED DESCRIPTION

[0052] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0053] Example:

[0054] like Figure 1 As shown, this embodiment provides a method for fine-tuning a high-speed railway track irregularity interval integrating dynamics evaluation, which includes the following steps:

[0055] 1) Establish a latent track geometry feature extraction architecture based on a deep Bayesian network;

[0056] In order to improve the robustness of the generation scheme, this embodiment proposes a feature network based on Bayesian deep learning. First, a basic three-layer fully connected neural network is established for the estimation task. Each layer of neurons calculates the input feature vector and connects the output to the next layer. In this process, slight weight changes will affect the output results. Therefore, this embodiment uses the Bayesian deep learning method to convert the point estimates of the network weights and bias parameters into distribution estimates in each potential track geometry feature extraction layer, and obtains multiple estimated parameters by repeated sampling. The Bayesian formula is as follows:

[0057] ;

[0058] in It indicates the parameters that must be trained for each layer of the network after the input track is not smooth. express The posterior distribution of represents the likelihood function, represents the prior distribution;

[0059] In order to obtain the posterior distribution, the denominator must integrate the entire set of real numbers, which is impossible. Therefore, this embodiment adopts the Monte Carlo (MC) Dropout method. The following formula explains the neural network with Dropout applied:

[0060] ;

[0061] in and Represent the number of layers and neurons respectively; Represents the potential track geometry feature extraction layer i The number of neurons in the layer; and Respectively represent the weights of each neuron before and after Dropout; represents the random activation or deactivation coefficient; Represents the potential track geometry feature extraction layer i Parameter distribution of layer neurons; Indicates application to l The random activation coefficient of the layer; and express Output values ​​before and after layer Dropout; and express layer The output value and bias of the neuron.

[0062] 2) Establish the mechanics module expression of the proxy model based on the vehicle-track coupling dynamics equation;

[0063] Different from the traditional numerical integration solution method, the vehicle-track coupled dynamics proxy model approximates the functional mapping relationship between the parameter space and the response space of the vehicle-track coupled dynamic system. After training, new working conditions considering parameter changes do not need to be re-solved using numerical integration, but only need to use the forward propagation of the neural network once. The entire workflow can be divided into a mechanics module and a deep learning module.

[0064] In the Agent Model Mechanics module, the vehicle-track coupled system is described as a set of coupled ordinary differential equations , as shown in formula (3):

[0065] ;

[0066] in and Respectively indicate that and Physical parameters and system excitations of the vehicle-track coupling system of each element, including: body inertia, frame inertia, primary and secondary suspension stiffness and damping, vehicle characteristic length, fastener pad stiffness and damping, vehicle speed, and track irregularity; express All solution functions in the system include: vertical vibration of the vehicle body, vehicle body nodding, vertical vibration of the front and rear frames, front and rear frame nodding and vertical vibration of the four wheel sets; Represents the time coordinate; represents the bounded time domain; it must be pointed out that the orbital system as a system of partial differential equations (PDE) needs to be converted into a system of ordinary differential equations (ODE) and participate in the calculation.

[0067] 3) Establish a vehicle-track coupling dynamics proxy model based on Fourier neural network;

[0068] In the vehicle-track coupled dynamics proxy model, a Fourier neural network is trained to approximate the ordinary differential equations of the vehicle-track coupled system. Solver of the vehicle-track coupling system ordinary differential equations The solution operator By creating a mapping To approximate, as shown in formula (4):

[0069] ;

[0070] The dynamic system parameter combination and target dynamic response combination Both belong to finite-dimensional Banach function space ; represents the parameters of the Fourier neural network; and Representation mapping; represents a bounded set in the parameter space of the vehicle-track coupling system ; represents the corresponding dynamic response set, ;

[0071] This embodiment uses the Fourier neural network operator (FNO) based on the graph convolution kernel network (GKN) as the backbone of the network, and then connects to a fully connected layer so that the module can process all ODEs of formula (3) at the same time. The loss function of the Fourier neural network is Including data loss , equation loss and direct derivative loss ;

[0072] ;

[0073] in In formula (3), The average error between the solution and the network output; express The mean square norm in the time domain is obtained by inputting the network output, derivative values ​​and system parameters into the ordinary differential equations of the vehicle-track coupling system obtained in the system; Represent the 1st and 2nd order derivative losses respectively; and Respectively , and Weight coefficients of the three loss functions; N Represents the time domain T The length of the data within.

[0074] 4) Establishing the vehicle response loss active control function based on the dense scale constraint algorithm of “track irregularity-vehicle response parameters”;

[0075] For the maintenance task, track irregularity is a discrete spatial curve with a sampling interval equal to the sleeper spacing (0.625m). For the vehicle response calculation task, track irregularity is densely sampled time series data (0.001s). Therefore, this embodiment proposes a "track irregularity-vehicle response parameter" dense scale constraint method to achieve the conversion between track fine-tuning scheme design and dynamic evaluation tasks. While improving the efficiency of maintenance scheme design, it also ensures the calculation accuracy of the proxy model.

[0076] Specifically:

[0077] Step 4.1) Space-time transformation of track irregularity: treat the track irregularity on each track as a spatial discrete sequence , and convert it into a time series according to the average train speed v ; ;

[0078] Step 4.2) Parameterization of discrete time series of track irregularity: Parameterize and assign parameter values ​​to it , as shown in formulas (7) to (9):

[0079] ;

[0080] in and Respectively represent the track unevenness sequence length and the position of each point; Represents the arc length between adjacent points; represents the cumulative arc length;

[0081] Step 4.3) Establish the continuous curve expression of track irregularity time series: Determine the node vector of track irregularity discrete sequence and basis functions And the continuity of the discrete time series of track irregularity is achieved by establishing B-spline expression;

[0082] Step 4.4) Dense scale sampling of “track irregularity-vehicle response parameters”: Densely sample the continuous curve of track irregularity time series at high-frequency time sampling intervals to determine the input of the vehicle-track coupling dynamics proxy model;

[0083] Step 4.5) Design and calculation of vehicle response loss active control function: The densely sampled track irregularity time series is used as an excitation input into the vehicle-track coupling dynamics proxy model to obtain the vehicle acceleration series; due to the distance difference between the front and rear wheel pairs, the processed track irregularity is applied to the four wheels respectively; on this basis, the acceleration active control loss of the design scheme is calculated , as shown in formula (10):

[0084] ;

[0085] in Represents the proxy model calculation process; and They represent the optimized track irregularity spatial sequence and the corresponding vehicle acceleration sequence respectively; Indicates the management threshold of the vehicle acceleration sequence.

[0086] 5) Design a loss function that embeds the adjustment cost and track smoothness quality optimization objectives;

[0087] Maintenance scheme design under the optimization framework essentially transforms the traditional constrained linear optimization problem into a network parameter optimization problem, and then uses backpropagation to achieve a fast approximate solution. Both the objective and the track parameter constraints are expressed in the form of loss functions. The difference is that the optimization objective is a non-threshold loss that reflects the direction of optimization, while the track parameter constraint loss is embedded in the specification management threshold as a hard constraint.

[0088] In the daily line maintenance process, the cumulative adjustment amount and the track quality index (TQI) are two important indicators of concern. This embodiment mainly uses the adjustment cost loss function To minimize the difference between the optimization result and the target curve, to ensure the minimum cumulative adjustment target, and use the track smoothness quality loss function To actively control the overall quality of the adjusted operation. The two losses are defined as follows:

[0089] ;

[0090] in, and Represent the adjusted and original track unevenness order columns respectively; represents the average value of the original track irregularity; N The length of the sequence representing the roughness of the original track.

[0091] 6) Embed multiple orbital parameter constraint functions such as string measurement, adjustment boundary, unit adjustment, etc.;

[0092] Multiple track parameter constraints are also expressed in the form of loss functions. The difference is that the track parameter constraint loss is embedded in the standard management threshold as a hard constraint. Track parameter constraints can promote the model to automatically generate a fine-tuning maintenance plan that meets management requirements. , Adjust Boundaries and unit adjustment For the trajectory parameter constraint loss, the constraint set is as follows:

[0093] ;

[0094] in, and Respectively, they represent the adjusted and original track irregularities; N Indicates the length of the track unevenness sequence; H represents the string measurement process matrix; and They represent the management thresholds of track irregularity chord measurement value and adjustment boundary respectively; Represents the linear rectification function calculation process; Represents the rounding function calculation process.

[0095] 7) Design and selection of track fine-tuning schemes under specific constraints;

[0096] Under a single fixed constraint combination, customize the weight ratio of the loss function. Execute the optimization model once, and reversely calculate each loss function based on the chain rule to obtain multiple estimation results as a set of fine-tuning design solutions for the fusion dynamics evaluation process, and make personalized choices based on needs.

[0097] 8) Output of track irregularity interval fine-tuning solution based on Monte Carlo;

[0098] Determine the Monte Carlo Dropout scale parameter and the number of model repetitions in deep Bayesian networks ,Pick The mean value of the sub-results and the maximum-minimum envelope are output as the interval fine-tuning solution.

[0099] This embodiment establishes a maintenance plan design framework and a dynamic proxy model based on a unified embedded physical neural network architecture. The vehicle-track coupling dynamic proxy model and the traditional physical constraint relationship are embedded in the neural network architecture through the "track irregularity-vehicle response parameter" intensive scale constraint method, realizing a high-speed railway track section fine-tuning method integrating the dynamic evaluation process, ensuring the rapid formulation of long-line overhaul plans, and significantly reducing the workload of overhaul and maintenance personnel.

[0100] Example Analysis

[0101] Taking a section of about 350m long measured track irregularity data as an example, the interval design effect, scheme comparison, acceleration active control and loss function convergence of the proposed method are demonstrated, as shown in the figure. Figures 2~5 shown.

[0102] like Figure 2 As shown in the figure, the proposed framework takes into account the uncertainty propagation in the program design process, and the design result is a maintenance interval. All measurement results within this interval meet the specification requirements, and the interval fluctuation is small, which proves that the framework is more robust. Considering that on-site fine-tuning maintenance can only be adjusted by adding a fixed thickness adjustment pad (0.5mm), this embodiment adds a unit adjustment amount constraint to ensure that the design solution can be directly used for maintenance. Taking the results under the average scheme as an example, the comparison with other scheme design methods is shown in the figure. Figure 3 shown.

[0103] like Figure 3 As shown, the existing method does not consider the unit adjustment amount constraint during the design process, and its adjustment amount is a continuous value, which is unacceptable in the actual application of the solution. In comparison, the method of this embodiment can directly output an integer multiple of the adjustment amount according to the unit adjustment amount, which is more in line with the actual situation on site.

[0104] High sampling frequency will not only cause high-frequency noise to interfere with the original track irregularity distribution, but also generate more decision variables in the scheme design process. This embodiment applies the proposed dense scale constraint method to appropriately discretize the track irregularity in the evaluation and decision-making process. The spatial domain track irregularity is considered in the decision-making process, and the time domain track irregularity is sampled in the evaluation process. In order to show some of the effects of dynamic constraints, only the minimum adjustment amount and the body acceleration threshold constraints are set here. The body acceleration constraint threshold is 0.05m / s 2 .

[0105] like Figure 4As shown in , applying this constraint allows the corresponding body response index to be evaluated at each iterative step of the solution design, so that the solution can meet the specified body response threshold requirements with the minimum cumulative adjustment. This means that there is no need for additional dynamic response evaluation after the solution design. The loss function converges as Figure 5 As shown, it is worth noting that the loss function is a normalized result.

[0106] like Figure 5 As shown in the figure, the vehicle acceleration and cumulative adjustment loss both achieve stable convergence within 100 steps. The acceleration loss curve goes through a process of “not exceeding the limit-exceeding the limit-not exceeding the limit”, mainly because the initial input is all 0, and its characteristics reflect the convergence after 50 steps.

[0107] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

Claims

1. A high-speed railway track irregularity interval fine-tuning method integrating dynamics evaluation, characterized by: The following steps are involved: 1) Establish a latent track geometry feature extraction architecture based on deep Bayesian network; 2) Establish the mechanical module expression of the proxy model based on the vehicle-track coupling dynamic equation; 3) Establish a vehicle-track coupling dynamics proxy model based on Fourier neural network; 4) Establishing the vehicle response loss active control function based on track irregularity-vehicle response parameter dense scale constraint algorithm; In step 4), the vehicle response loss active control function is designed based on the track irregularity-vehicle response parameter dense scale constraint algorithm to achieve the conversion between the track fine-tuning scheme design and the dynamics evaluation task, specifically: Step 4.1) Space-time transformation of track irregularity: The track irregularity on each track is regarded as a spatial discrete sequence m i =(x i ,y i ,z i ), and convert it into a time series T according to the average train speed v i =(x i ,y i ,t i ); Step 4.2) Parameterization of track irregularity discrete time series: For T i Parameterize and assign parameter value k to it i , as shown in formulas (7) to (9): Where n and i represent the length of the track unevenness sequence and the position of each point respectively; d i Indicates the arc length between adjacent points; L i represents the cumulative arc length; Step 4.3) Establish the continuous curve expression of track irregularity time series: Determine the track irregularity discrete sequence node vector U = {u0,u1,...,u i+j } and basis function N i,j (k) and realize the continuity of the discrete time series of track irregularity by establishing B-spline expression; Step 4.4) Dense scale sampling of track irregularity-vehicle response parameters: Dense sampling of continuous curves of track irregularity time series at high-frequency time sampling intervals is performed to determine the input of the vehicle-track coupling dynamic proxy model; Step 4.5) Design and calculation of vehicle response loss active control function: The densely sampled track irregularity time series is used as an excitation input into the vehicle-track coupling dynamics proxy model to obtain the vehicle acceleration series; due to the distance difference between the front and rear wheel pairs, the processed track irregularity is applied to the four wheels respectively; on this basis, the acceleration active control loss of the design scheme is calculated As shown in formula (10): in Represents the proxy model calculation process; y′ i and They represent the optimized track irregularity spatial sequence and the corresponding vehicle acceleration sequence respectively; ξ3 represents the management threshold of the vehicle acceleration sequence; 5) Design a loss function that embeds the adjustment cost and track smoothness quality optimization objectives; 6) Embed multiple orbital parameter constraint functions, including chord measurement, adjustment boundary, and unit adjustment amount; 7) Design and selection of track fine-tuning schemes under specific constraints; 8) Output of Monte Carlo-based track irregularity interval fine-tuning solution.

2. The high-speed railway track irregularity interval fine-tuning method integrating dynamics evaluation according to claim 1 is characterized by: In the potential track geometry feature extraction architecture based on deep Bayesian network in step 1), the point estimates of network weights and bias parameters are converted into distribution estimates in each potential track geometry feature extraction layer using Bayesian deep learning method, and multiple estimated parameters are obtained by repeated sampling; the Bayesian formula is as follows: Where ω represents the parameters that must be trained for each layer of the network after the input track is not smooth, p(ω|x,y) represents the posterior distribution of ω, p(x,y|ω) represents the likelihood function, and p(ω) represents the prior distribution; The Monte Carlo Dropout method is used to obtain the posterior distribution. The potential feature extraction network using Dropout is shown below: where i = 1, ..., L and j = 1, ..., K i-1 Respectively represent the number of layers and neurons; K i represents the number of neurons in the i-th layer of the potential track geometry feature extraction layer; W i and ω i Respectively represent the weights of each neuron before and after Dropout; r i,j represents the random activation or inactivation coefficient; p i represents the parameter distribution of the neurons in the ith layer of the potential track geometry feature extraction layer; r l represents the random activation coefficient applied to layer l; y l and Represents the output value before and after the l-layer Dropout; and Represents the output value and bias of neuron i in layer l+1.

3. The high-speed railway track irregularity interval fine-tuning method integrating dynamics evaluation according to claim 2 is characterized by: In the proxy model mechanics module based on the vehicle-track coupling dynamics equations in step 2), the vehicle-track coupling system is described as a set of coupled ordinary differential equations As shown in formula (3): Where p and f represent the number of n p and n f The physical parameters and system excitations of the vehicle-track coupling system of each element include: vehicle body inertia, frame inertia, primary and secondary suspension stiffness and damping, vehicle characteristic length, fastener pad stiffness and damping, vehicle speed, and track irregularity parameters; u represents All solution functions in , including: vehicle body vertical vibration, vehicle body nodding, front and rear frame vertical vibration, front and rear frame nodding and vertical vibration of four wheel sets; t represents the time coordinate; D represents the bounded time domain.

4. The high-speed railway track irregularity interval fine-tuning method integrating dynamics evaluation according to claim 3 is characterized by: In the Fourier neural network-based vehicle-track coupling dynamics proxy model of step 3), a Fourier neural network is trained to approximate the ordinary differential equations of the vehicle-track coupling system The solution operator of ; Ordinary differential equations for vehicle-track coupling system The solution operator By creating a mapping To approximate, as shown in formula (4): The dynamic system parameter combination and target dynamic response combination Both belong to finite-dimensional Banach function space θ represents the parameters of the Fourier neural network; → and Representation mapping; represents a bounded set in the parameter space of the vehicle-track coupling system u represents the corresponding dynamic response set, Loss function of Fourier neural network Including data loss Equation loss and direct derivative loss in represents the average error between the solution of u in formula (3) and the network output; express The mean square norm in the time domain is obtained by inputting the network output, derivative values ​​and system parameters into the ordinary differential equations of the vehicle-track coupling system obtained in the system; denote the first-order and second-order derivative losses respectively; ω1, ω2 and ω3 denote and Weight coefficients of the three loss functions; N represents the data length in the time domain T.

5. The high-speed railway track irregularity interval fine-tuning method integrating dynamics evaluation according to claim 4 is characterized by: In step 5), a loss function is designed that embeds the adjustment cost and track smoothness quality optimization objectives to ensure that the design scheme meets the specification requirements from the track level; among them, the adjustment cost loss function is used To minimize the difference between the optimization result and the target curve, to ensure the minimum cumulative adjustment target, and use the track smoothness quality loss function to actively control the overall quality of the adjusted operation; these two losses are defined as follows: in, and Represent the adjusted and original track unevenness order columns respectively; represents the average value of the original track irregularity; N represents the sequence length of the original track irregularity.

6. The high-speed railway track irregularity interval fine-tuning method integrating dynamics evaluation according to claim 5 is characterized by: In step 6), by embedding multiple track parameter constraint functions, the model can automatically generate a fine-tuning maintenance plan that meets management requirements; Adjust Boundaries and unit adjustment For the trajectory parameter constraint loss, the constraint set is as follows: in, and i Respectively represent the adjustment and original track irregularity; N represents the length of the track irregularity sequence; H represents the chord measurement process matrix; ξ1 and ξ2 represent the management thresholds of the track irregularity chord measurement value and the adjustment boundary; Relu(*) represents the linear rectification function calculation process; Represents the rounding function calculation process.

7. The high-speed railway track irregularity interval fine-tuning method integrating dynamics evaluation according to claim 6 is characterized by: In step 7), the track fine-tuning scheme is designed and selected under specific constraints, and the weight ratio of the loss function needs to be customized under a single fixed constraint combination; Execute the optimization model once, and perform reverse calculations on each loss function based on the chain rule to obtain multiple estimation results as a set of fine-tuned design solutions for the fusion dynamics evaluation process, and make personalized choices based on needs.

8. The high-speed railway track irregularity interval fine-tuning method integrating dynamics evaluation according to claim 7 is characterized by: In step 8), the Monte Carlo-based track irregularity interval fine-tuning solution output is performed. It is necessary to determine the Monte Carlo Dropout ratio parameter ρ and the model repetition number parameter in the deep Bayesian network. Pick The mean value of the sub-results and the maximum-minimum envelope are output as the interval fine-tuning solution.

Images