A multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network

Abstract

The present invention relates to the technical field of railway tracks. Specifically, it relates to a multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network, which includes the following steps: 1) determining the dimension and length of the initial input data according to the track irregularity index; 2) establishing a latent feature expression layer to perform multi-layer non-linear low-rank transformation and feature activation on the input vector; 3) embedding a multi-track parameter constraint layer to perform hard constraints according to the management values; 4) designing a weighted loss function that embeds track features and chord measurement formulas; 5) adaptively iteratively optimizing and selecting the output of the scheme. The present invention can preferably perform multi-objective fine-tuning of high-speed railway tracks.

Description

Technical Field

[0001] The present invention relates to the technical field of railway tracks, and more specifically, to a multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network. Background Art

[0002] The smoothness and safe operation of high-speed trains largely depend on high-quality track smoothness maintenance. Track fine-tuning is the most commonly used daily maintenance method by management departments. Operators control track smoothness by adjusting the fasteners and shims on each sleeper. The formulation of the fine-tuning plan is crucial because it determines the maintenance cost and effect. Some scholars have developed an optimization method aiming at minimizing the cumulative adjustment while ensuring that the track irregularity after adjustment meets the requirements of the chord measurement system specification. This method involves iterative optimization of constraint boundaries and can handle real conditions. Some scholars have developed a balanced track optimization algorithm, considering design alignment optimization and using a constrained linear programming algorithm in the track geometry optimization part. Some scholars have proposed an optimization method for track irregularity based on dynamic data, using long-wavelength features as the optimization target curve. This method ensures that the minimum cumulative adjustment meets the management value requirements. However, existing methods all use linear unit programming methods in the optimization solution process. Limited by the solution ability of this method, they are not applicable to large-scale track irregularity optimization and cannot fully consider the overall conditions of the line. Other scholars have also carried out some research on engineering and design issues in the track fine-tuning process, but there is still a blank in applying neural networks to the field of track fine-tuning maintenance plan design. There is an urgent need for a new design optimization method to achieve large-scale and rapid solution of the long-line fine-tuning plan design problem. Summary of the Invention

[0003] The content of the present invention is to provide a multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network, which can solve the problems of low efficiency in optimizing and solving large-scale data in the existing track smoothness maintenance plan and the lack of unified wavelength constraint management from the frequency domain range in the process of automatic plan formulation.

[0004] A multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network according to the present invention includes the following steps:

[0005] 1) Determine the dimension and length of the initial input data according to the track irregularity index;

[0006] 2) Establish a potential feature expression layer to perform multi-layer non-linear low-rank transformation and feature activation on the input vector;

[0007] 3) Embed a multi-track parameter constraint layer to perform hard constraints according to the management value;

[0008] 4) Design a weighted loss function that embeds track features and chord measurement formulas;

[0009] 5) Adaptive iterative optimization and scheme selection output.

[0010] Preferably, in step 1), single or combined indicators are freely selected from the left and right high-low and left and right alignment indicators of track irregularities to enter the optimization channel, so as to determine the initial input dimension and length.

[0011] Preferably, in step 2), any continuous function is approximated by full connection between neurons in different network layers, and potential structural data features are mined for feature activation; for each input X n×i Execute:

[0012] Y n×m = X n×i ·w i×m + b (1)

[0013] where n is the feature dimension; i and m are the input and output feature lengths respectively; w represents the parameters to be learned by the model; b represents the bias corresponding to the output feature; considering that the distribution characteristics of track irregularities conform to the characteristics of being close to the normal distribution, the hyperbolic tangent Tanh function is used for non-linear activation in the low-dimensional space.

[0014] Preferably, in step 3), each track parameter constraint layer in the multiple track parameter constraint layers includes a pair of threshold constraint layers and feature reconstruction layers, which are used to sort out the existing management indicators and constrain the track irregularity threshold, first-order difference threshold, sleeper check threshold, and adjustment amount boundary indicator.

[0015] Preferably, in step 3), when constraining the sleeper check threshold, the calculation process is as follows:

[0016]

[0017] where n is the feature dimension; i is the feature length; Y n×i and Y ' n×i represent the measured feature sequences before and after constraint respectively; X n×i and X' n×i represent the original feature sequences before and after restoration respectively; k represents the number of spaced sleepers, and X represents the management threshold.

[0018] Preferably, in step 4), considering the characteristics of track irregularity optimization, a cumulative adjustment amount Loss A , a track quality Loss S and a weighted loss function Loss based on the chord measurement formula overrun penalty F; The cumulative adjustment loss is used to minimize the difference between the optimization result and the target curve to ensure the minimum adjustment target; the track quality loss is used to actively control the overall riding quality after adjustment; the over-limit penalty loss is used to punish the over-limit points during the model convergence process to ensure that the final optimization plan is within the management value range; the three losses are defined as follows:

[0019]

[0020] where x i and respectively represent the estimated and target chord measurement values; represents the mean value of the target chord measurement; represents the estimated chord measurement value with chord length l; ξ l represents the management value of the chord measurement with chord length l; N and K respectively represent the track irregularity and the length of the chord measurement value sequence;

[0021] The final weighted loss function Loss F is in the following form:

[0022]

[0023] where p represents the number of chords, p = 1, 2,..., m; ω A , ω S and ω p respectively represent the weights of various loss functions.

[0024] Preferably, in step 5), the maximum number of iterations of the model and the loss stopping discrimination criterion are set to ensure the fast and accurate convergence stop of the model; among them, the maximum number of iterations does not exceed 500 times, and the iteration stops when the reduction of the weighted loss is less than 5%.

[0025] The present invention realizes the rapid formulation of the overhaul plan for long lines by embedding the traditional physical constraint relationship into the neural network architecture. Considering the cross-correlation of different constraints in the frequency domain, it realizes the unified control of different bands. It effectively guarantees the optimization efficiency of the track irregularity state under complex constraint conditions and greatly reduces the workload of overhaul and maintenance of railway workers. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flowchart of a multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network in Embodiment 1;

[0027] Figure 2 is a schematic structural diagram of the track parameter constraint layer in Embodiment 1;

[0028] Figure 3 is a schematic diagram of the loss function convergence curve in Embodiment 2;

[0029] Figure 4(a) is a schematic diagram of the effect of a single optimization scheme in Example 2;

[0030] Figure 4(b) is a schematic diagram of the effect of a set of multiple schemes in Example 2;

[0031] Figure 5(a) is a schematic diagram of the effect of the 10m chord index in Example 2;

[0032] Figure 5(b) is a schematic diagram of the effect of the 60m chord index in Example 2;

[0033] Figure 5(c) is a schematic diagram of the effect of the first-order difference index in Example 2. Detailed implementation manners

[0034] To further understand the content of the present invention, the present invention will be described in detail in combination with the accompanying drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention rather than limiting it.

[0035] Example 1

[0036] As Figure 1 shown, this embodiment provides a multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network, which includes the following steps:

[0037] 1) Determine the initial input data dimension and length according to the track irregularity index;

[0038] In step 1), select a single or combined index from the left high-low, right high-low, left alignment and right alignment indexes in the track irregularity to enter the optimization channel, so as to determine the initial input dimension and length. The optional channel combinations are shown in Table 1 below.

[0039] Table 1 Initial channel combination selection table

[0040]

[0041] Among them, 1 represents left high-low, 2 represents right high-low, 3 represents left alignment, and 4 represents right alignment.

[0042] 2) Establish a potential feature expression layer (a plurality of network layers together constitute the potential feature expression layer), and perform multi-layer non-linear low-rank transformation and feature activation on the input vector;

[0043] In step 2), approximate any continuous function through the full connection between neurons in different network layers, and mine the potential structure data features for feature activation; for each input X n×i Execute:

[0044] Y n×m = X n×i · w i×m + b (1)

[0045] Among them, n is the feature dimension; i and m are the input and output feature lengths respectively; w represents the parameters to be learned by the model; b represents the bias corresponding to the output features. Considering that the distribution characteristics of track irregularities conform to the characteristics of being close to a normal distribution, the hyperbolic tangent Tanh function is used for non-linear activation in the low-dimensional space. The specific network layer parameter settings are shown in Table 2 below.

[0046] Table 2 Structure Table of Latent Feature Expression Layer

[0047] layer input output activate input layer (1,k,m) - - latent feature representation layer 1 (1,k,m) (1,k,128) Tanh latent feature representation layer 2 (1,k,128) (1,k,64) Tanh latent feature representation layer 3 (1,k,64) (1,k,m) - output layer - (1,k,m) -

[0048] Among them, m represents the length of the input signal, and k represents the number of track irregularity channels.

[0049] 3) Embed the multi-track parameter constraint layer and perform hard constraints according to the management values;

[0050] In step 3), each track parameter constraint layer in the multi-track parameter constraint layer includes a pair of threshold constraint layers and feature reconstruction layers, as Figure 2 shown, which are used to sort out the existing management indicators and constrain the track irregularity threshold, first-order difference threshold, sleeper check threshold, and adjustment amount boundary indicator. The threshold constraint layer is responsible for converting the original data into specific indicators according to the index calculation formula and then controlling the threshold; the feature reconstruction layer is responsible for reversing the index calculation formula and restoring the constrained indicators to the original input.

[0051] In step 3), when constraining the sleeper check threshold, the calculation process is as follows:

[0052]

[0053] Among them, n is the feature dimension; i is the feature length; Y n×i and Y ' n×i respectively represent the measured feature sequences before and after the constraint; X n×i and X' n×i respectively represent the original feature sequences before and after the restoration; k represents the number of sleepers, and χ represents the management threshold.

[0054] 4) Design a weighted loss function that embeds the track feature and chord measurement formula;

[0055] In step 4), considering the characteristics of track irregularity optimization, establish the cumulative adjustment amount Loss A 、track quality Loss S and a weighted loss function Loss based on the chord measurement formula overrun penalty F; The cumulative adjustment amount loss is used to minimize the difference between the optimization result and the target curve to ensure the minimum adjustment target; the track quality loss is used to actively control the overall riding quality after adjustment; the over-limit penalty loss is used to penalize the over-limit points during the model convergence process to ensure that the final optimization scheme is within the management value range; the three losses are defined as follows:

[0056]

[0057] Among them, x i and respectively represent the estimated and target chord measurement values; represents the mean value of the target chord measurement; represents the estimated chord measurement value with a chord length of l; ξ l represents the management value of the chord measurement with a chord length of l; N and K respectively represent the track irregularity and the length of the chord measurement value sequence;

[0058] The final weighted loss function Loss F is in the following form:

[0059]

[0060] Among them, p represents the number of chords, p = 1, 2,..., m; ω A , ω S and ω p respectively represent the weights of various loss functions.

[0061] 5) Adaptive iterative optimization and solution selection output.

[0062] In step 5), set the maximum number of iterations of the model and the loss stop discrimination criterion to ensure the fast and accurate convergence stop of the model; among them, the maximum number of iterations does not exceed 500 times, and the iteration stops when the reduction of the weighted loss is less than 5%.

[0063] In this embodiment, by embedding the traditional physical constraint relationship into the neural network architecture, the rapid formulation of the overhaul plan for long lines is realized. Considering the cross-correlation of different constraints in the frequency domain, unified control in sub-bands is achieved. Effectively ensure the optimization efficiency of the track irregularity state under complex constraint conditions, and greatly reduce the workload of maintenance personnel for overhaul.

[0064] Embodiment 2

[0065] Taking the measured track irregularity data of a section about 1.5 km long as an example, the following is the application of the method. The specific implementation method is as follows:

[0066] 1) According to the track irregularity index, determine that the initial input data dimension is 4 and the length is 1600, and establish an input vector with a dimension of (1, 4, 1600);

[0067] 2) Establish the parameters corresponding to the potential feature expression layer as shown in Table 2 of Example 1;

[0068] 3) Embed a multi-track parameter constraint layer to impose hard constraints on four indicators, namely the track irregularity threshold, the first-order difference threshold, the sleeper check threshold, and the adjustment amount boundary, according to the management values. The thresholds are 2 mm, 0.5 mm, 0.5 mm, and 10 mm respectively;

[0069] 4) Specifically design the over-limit penalty loss of the chord measurement formula combined with 10 m and 60 m chords, and weight it with the adjustment amount and the track quality loss. The over-limit penalty thresholds are 1 mm and 3 mm respectively, and the weight coefficients are 0.4, 0.3, and 0.3 respectively;

[0070] 5) Set the maximum number of iterations to 500 times, and stop the iteration when the optimization is less than 5%;

[0071] 6) Output and view the convergence of the loss function, the adjustment plan, and the corresponding indicator effects as shown below.

[0072] As Figure 3 shown, although the maximum number of iterations is set to 500, according to the loss stop discrimination criterion, the model can converge quickly in only 230 times, which ensures the efficiency of the solution.

[0073] As shown in Figure 4(a), compared with two traditional single-objective optimizations, when the minimum track quality is the goal, the best track smoothness quality is guaranteed, but the adjustment cost is wasted. When the minimum cumulative adjustment amount is the goal, the situation is exactly the opposite. In contrast, the proposed method can freely choose between the two goals while ensuring control over both. Another advantage is that it can output multiple solutions simultaneously, as shown in Figure 4(b). Each point represents a maintenance plan, and the operator can choose a suitable optimization plan according to the scenario requirements. If the maintenance conditions are limited and the materials for maintenance operations are insufficient, the plan close to point A can be selected; if the materials are sufficient, the plan close to point B can be selected; when there is material redundancy and large-scale optimization of track quality is required, the plan close to point C can be selected.

[0074] Taking the elevation of the left rail as an example, the effects of the corresponding indicators for this plan are shown in Figures 5(a), 5(b), and 5(c). All the track parameter constraints added to the loss penalty can strictly comply with the management threshold constraints.

[0075] The above has schematically described the present invention and its embodiments. This description is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. Therefore, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, design similar structural modes and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. A multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network, characterized in that: It includes the following steps: 1) Determine the initial input data dimension and length according to the track irregularity index; 2) Establish a potential feature expression layer to perform multi-layer non-linear low-rank transformation and feature activation on the input vector; 3) Embed a multi-track parameter constraint layer to perform hard constraints according to the management value; In step 3), each track parameter constraint layer in the multi-track parameter constraint layer includes a pair of threshold constraint layers and feature reconstruction layers, which are used to sort out the existing management indicators and constrain the track irregularity threshold, first-order difference threshold, sleeper check threshold, and adjustment amount boundary index; In step 3), when the sleeper check threshold is constrained, the calculation process is as follows: where n is the feature dimension; i is the feature length; Y n×i and Y ' n×i represent the measured feature sequences before and after constraint respectively; X n×i and X' n×i respectively represent the original feature sequences before and after restoration; k represents the number of spacer pillows, and X represents the management threshold; 4) Design a weighted loss function that embeds the track feature and chord measurement formula; 5) Adaptive iterative optimization and solution selection output.

2. The multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network according to claim 1, characterized in that: In step 1), select single or combined indicators from the left and right high, left and right alignment indicators in the track irregularity to enter the optimization channel, so as to determine the initial input dimension and length.

3. A multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network according to claim 2, characterized in that: In step 2), approximate any continuous function through the full connection between neurons in different network layers, and mine the potential structural data features for feature activation; for each input X n×i Execute: Y n×m = X n×i · w i×m + b (1) Among them, n is the feature dimension; i and m are the input and output feature lengths respectively; w represents the parameter to be learned by the model; b represents the bias corresponding to the output feature; considering that the track irregularity distribution characteristic conforms to the characteristic of being close to the normal distribution, the hyperbolic tangent Tanh function is used for non-linear activation in the low-dimensional space.

4. A multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network according to claim 3, characterized in that: In step 4), considering the characteristics of track irregularity optimization, a cumulative adjustment amount Loss is established A , a track quality Loss S and a weighted loss function Loss based on the chord measurement formula overrun penalty F ; The cumulative adjustment amount loss is used to minimize the difference between the optimization result and the target curve to ensure the minimum adjustment target; The track quality loss is used to actively control the overall driving quality after adjustment; The overrun penalty loss is used to punish the overrun points in the model convergence process to ensure that the final optimization plan is within the management value range; The definitions of the three losses are as follows: where x i and represent the chord measurement values of the estimation and the target respectively; represents the mean value of the target chord measurement; represents the estimated chord measurement value with a chord length of l; ξ l represents the chord measurement management value with a chord length of l; N and K represent the track irregularity and the chord measurement sequence length respectively; The final weighted loss function Loss F is in the following form: where p represents the number of chords, p = 1, 2,..., m; ω A , ω S and ω p respectively represent the weights of various loss functions.

5. A multi-objective fine-tuning method for high-speed railway tracks driven by an embedded physical neural network according to claim 4, characterized in that: In step 5), set the maximum number of iterations of the model and the loss stop discrimination criterion to ensure the fast and accurate convergence stop of the model; among them, the maximum number of iterations does not exceed 500 times, and the iteration stops when the weighted loss reduction amount is less than 5%.

Images