Combined optimization method of large-scale machine tamping and track fine-tuning based on neuron

Abstract

The present invention relates to the field of railway and rail transit technology, specifically, to a neuron-based large-scale machine tamping and track fine-tuning combined optimization method, comprising the following steps: step 1, for track ledger data, select the smoothness index constraint amplitude; step 2, set the adjustment amount constraint amplitude of large-scale machine tamping and track fine-tuning; step 3, establish a track geometry adjustment constraint model; step 4, establish the shared layer part of the neuron solver NNS; step 5, construct a loss function based on the constraint model; step 6, calculate the gradient of the neural network, realize the model solution, and obtain the optimization result of the adjustment amount of tamping and fine-tuning at each sleeper; step 7, use the dynamic programming method to distribute the adjustment amount to the adjustment amount corresponding to the large-scale machine tamping and fine-tuning. The present invention can better optimize the track unevenness.

Description

Technical Field

[0001] The present invention relates to the technical field of railways and rail transit, and in particular to a neuron-based combined optimization method for large-scale machine tamping and track fine-tuning. Background Art

[0002] Track irregularity is a disturbing factor in high-speed railway operations. It is the main cause of vibration and wheel-rail interaction, and significantly affects the safety, stability, comfort and service life of the vehicle. Under conditions of high speed, heavy load and foundation settlement, the impact of track unevenness is more significant, becoming a major limiting factor in train speed. High-speed railways require high track smoothness, which is also a key technical requirement for controlling the operation and maintenance design of railway infrastructure. It plays a decisive role in the success or failure of high-speed railway projects. Maintaining track irregularity is an important manifestation of the overall performance and bearing capacity of the track structure, and is also the goal of railway large-scale machine tamping and fine-tuning.

[0003] Current optimization methods are slow and cannot keep up with the maintenance demands arising from the increase in total railway mileage; moreover, they are unable to effectively optimize track irregularities. Summary of the invention

[0004] The content of the present invention is to provide a neuron-based combined optimization method for large-scale machine tamping and track fine-tuning, which can overcome certain or some defects of the prior art.

[0005] The neuron-based combined optimization method for large-scale machine tamping and track fine-tuning according to the present invention comprises the following steps:

[0006] Step 1: Select the constraint amplitude of the smoothness index for the track inventory data;

[0007] Step 2: Setting the adjustment amount constraint amplitude of large machine tamping and track fine adjustment;

[0008] Step 3: Establishing a track geometry adjustment constraint model;

[0009] Step 4: Establish the shared layer part of the neural solver NNS;

[0010] Step 5: Construct a loss function based on the constraint model;

[0011] Step 6: Calculate the gradient of the neural network to solve the model and obtain the optimization result of the adjustment amount of tamping and fine tuning at each sleeper;

[0012] Step 7. Use dynamic programming to distribute the adjustment amount to the adjustment amount corresponding to the large machine tamping and fine adjustment.

[0013] Preferably, in step one, the inventory data is the operating speed of the line, and the smoothness indicators include the amplitudes of the chord measurement method and the vector distance difference method. The 10m chord measurement method, 60m chord measurement method, 5 / 30m vector distance difference method, and 150 / 300m vector distance difference method are used to perform smoothness evaluation and line smoothness optimization.

[0014] Preferably, in step 2, the geometric adjustment amount of the large-scale tamping is its corresponding track starting value, and the starting value is selected as an integer multiple of 10, which is {0, 10, 20, 30, 40, 50, 60}; for track fine-tuning, fastener adjustment is affected by the combination of pads, and the types of pads include under-rail pads and adjustment pads. The specification stipulates that the number of under-rail pads is 1, and the thickness is 5mm or 10mm; the number of adjustment pads is between 0 and 2, and the thickness is 1, 2, 5 or 8mm. It can be seen that the total thickness range of the fastener pads is 5-26mm; based on this, the adjustment range of the combination of large-scale tamping and track fine-tuning is 5-86mm.

[0015] Preferably, in step 3, the process of establishing the constraint model is:

[0016] String measurement formula:

[0017]

[0018] Among them, e MCO is the chord measurement value, c is the ratio of chord length to sleeper spacing, m1 is a point inside the chord, n is the starting point of the chord, p n is the absolute deviation data at sleeper n, p n+c is the absolute deviation data at the sleeper n+c;

[0019] Vector difference formula:

[0020]

[0021] Among them, e ACO is the vector distance difference, c is the ratio of the reference chord length to the sleeper spacing, w is the ratio of the detection chord length to the sleeper spacing, m2 is the starting point of the detection chord, n is the starting point of the reference chord, is the absolute deviation data at sleeper m2, is the absolute deviation data at the sleeper m2+w, k is the ratio of the reference chord to the test chord;

[0022] The calculation of vector distance difference and chord measurement is regarded as multi-channel convolution. The convolution process is as follows:

[0023]

[0024] I represents the vector distance difference / chord measurement matrix, B is the multi-channel convolution kernel, P is the absolute deviation data, AL is the horizontal deviation, Δ is the constraint amplitude converted from the original deviation to the chord measurement method and vector distance difference system;

[0025] in:

[0026]

[0027] P=[p1p2…p n ] T

[0028] AL=[AL1AL2…AL n ] T

[0029] Subtract the original absolute deviation data from the thickness of the fastener before adjustment to obtain the absolute deviation data excluding the thickness of the fastener before adjustment, recorded as Establish track fine-tuning constraint model:

[0030]

[0031] Among them, δ i is the adjustment amount of the ith sleeper, B is the multi-channel convolution kernel, To adjust the vector, is the upper bound of the adjustment amount, The lower bound of the adjustment amount.

[0032] Preferably, in step 4, a multilayer perceptron is used to randomly initialize the geometric adjustment amount at each sleeper, the input part of the multilayer perceptron is a random number, the output dimension matches the number of sleepers in the track adjustment part, and the output vector serves as the geometric adjustment amount for each sleeper.

[0033] As a preferred method, in step 5, based on the multi-layer perceptron established in step 4, combined with the constraint model obtained in step 3, the objective function and the constraint are embedded into the neural network, that is, the network output is connected to the objective function neuron and the constraint neuron, and a RELU activation layer is embedded after the constraint neuron, and then added to the objective function; gradient normalization is embedded in the back propagation process, and the optimization adjustment amount is obtained after multiple rounds of iterations; for the track fine-tuning optimization constraint model, its loss function is written as:

[0034]

[0035] Preferably, in step six, the gradient of the neural network is calculated and the adjustment amount is updated; as the number of neural network iterations increases, the network output will gradually approach the optimal solution under the current adjustment amplitude constraint and smoothness evaluation index constraint, and the optimized adjustment amount will be obtained when convergence.

[0036] As a preferred method, in step seven, the obtained adjustment amount is distributed to the large machine tamping track starting amount, the track fine-tuning track pad thickness and the adjustment pad thickness by using the dynamic programming method; the combination is solved by the knapsack problem, and the calculated adjustment amount solved by NNS is used as the capacity of the knapsack, and the four adjustments are regarded as four commodities. There are multiple choices for the space occupied by each commodity, but each commodity must be selected and can only be selected once.

[0037] The present invention provides a neural network based solver (NNS) to solve the track fine tuning linear programming problem and obtain the adjustment plan. NNS takes into account the regularity criteria and the constraints of the corrective adjustment to ensure that the plan meets the technical specifications. At the same time, the present invention applies a gradient normalization process to ensure that the gradient is normalized in a way that balances the influence of various constraints, promoting more effective and efficient optimization. The solution speed of the model can reach 20 seconds per kilometer.

[0038] The present invention uses dynamic programming to allocate cumulative adjustment amounts to compaction and fine adjustment at each sleeper, thereby obtaining a compaction adjustment and fine adjustment pad combination that can effectively optimize track irregularities. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flowchart of a neuron-based large-scale machine tamping and track fine-tuning combined optimization method in an embodiment;

[0040] Figure 2 It is a schematic diagram of the combination of the adjustment pad and the under-rail pad in the embodiment;

[0041] Figure 3 A schematic diagram of a chord measurement in an embodiment;

[0042] Figure 4 A schematic diagram of the vector distance difference in the embodiment;

[0043] Figure 5 is a schematic diagram of a neuron solver in an embodiment;

[0044] Figure 6 It is a schematic diagram of the training results in the embodiment;

[0045] FIG. 7( a ) is a schematic diagram of the fine-tuning tamping distribution result in the embodiment;

[0046] FIG7( b ) is a schematic diagram of high and low deviation results in the embodiment;

[0047] FIG. 7( c ) is a schematic diagram of a dynamic simulation in an embodiment. DETAILED DESCRIPTION

[0048] 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.

[0049] Example

[0050] like Figure 1 As shown, this embodiment provides a neuron-based combined optimization method for large-scale machine tamping and track fine-tuning, which includes the following steps:

[0051] Step 1: Select the constraint amplitude of the smoothness index for the track inventory data;

[0052] Step 2: Setting the adjustment amount constraint amplitude of large machine tamping and track fine adjustment;

[0053] Step 3: Establishing a track geometry adjustment constraint model;

[0054] Step 4: Establish the shared layer part of the neural solver NNS;

[0055] Step 5: Construct a loss function based on the constraint model;

[0056] Step 6: Calculate the gradient of the neural network to solve the model and obtain the optimization result of the adjustment amount of tamping and fine tuning at each sleeper;

[0057] Step 7. Use dynamic programming to distribute the adjustment amount to the adjustment amount corresponding to the large machine tamping and fine adjustment.

[0058] In step 1, the inventory data is the operating speed of the line. The smoothness indicators include the amplitudes of the chord measurement method and the vector distance difference method. The 10m chord measurement method, 60m chord measurement method, 5 / 30m vector distance difference method, and 150 / 300m vector distance difference method are used for smoothness evaluation and line smoothness optimization. The amplitude selection is shown in Table 1, recorded as Δ.

[0059] According to the "High-speed Railway Track Engineering Construction Quality Acceptance Standard", the vector distance difference and chord measurement constraint amplitude are shown in Table 1 below:

[0060] Table 1 Constraint amplitude table

[0061]

[0062]

[0063] In step 2, the geometric adjustment of the tamping machine is the corresponding track start value, and the start value is selected as an integer multiple of 10, which is {0, 10, 20, 30, 40, 50, 60}; for track fine adjustment, the fastener adjustment is affected by the pad combination, and the pad types include rail pads and adjustment pads. Take WJ-8 fasteners as an example, Figure 2As shown, the number of pads under the rail is 1, and the thickness is 5mm or 10mm; the number of pads is adjusted between 0 and 2, and the thickness is 1, 2, 5 or 8mm. It can be seen that the total thickness of the fastener pads ranges from 5 to 26mm; based on this, the adjustment range of the combination of large-scale tamping and track fine-tuning is 5 to 86mm.

[0064] In step 3, the process of establishing the constraint model is:

[0065] like Figure 3 As shown, the chord measurement formula is:

[0066]

[0067] Among them, e MCO is the chord measurement value, c is the ratio of chord length to sleeper spacing, m1 is a point inside the chord, n is the starting point of the chord, p n is the absolute deviation data at sleeper n, p n+c is the absolute deviation data at the sleeper n+c;

[0068] like Figure 4 As shown, the vector difference formula is:

[0069]

[0070] Among them, e ACO is the vector distance difference, c is the ratio of the reference chord length to the sleeper spacing, w is the ratio of the detection chord length to the sleeper spacing, m2 is the starting point of the detection chord, n is the starting point of the reference chord, is the absolute deviation data at sleeper m2, is the absolute deviation data at the sleeper m2+w, k is the ratio of the reference chord to the test chord;

[0071] The calculation of vector distance difference and chord measurement is regarded as multi-channel convolution. The convolution process is as follows:

[0072]

[0073] I represents the vector distance difference / chord measurement matrix, B is the multi-channel convolution kernel, P is the absolute deviation data, AL is the horizontal deviation, Δ is the constraint amplitude converted from the original deviation to the chord measurement method and vector distance difference system;

[0074] in:

[0075]

[0076] P=[p1p2…p n ] T

[0077] AL=[AL1AL2…AL n ] T

[0078] Subtract the original absolute deviation data from the thickness of the fastener before adjustment to obtain the absolute deviation data excluding the thickness of the fastener before adjustment, recorded as Establish track fine-tuning constraint model:

[0079]

[0080] Among them, δ i is the adjustment amount of the ith sleeper, B is the multi-channel convolution kernel, To adjust the vector, is the original absolute deviation data, Δ is the constraint amplitude, is the upper bound of the adjustment amount, The lower bound of the adjustment amount.

[0081] In step 4, a multilayer perceptron is used to randomly initialize the geometric adjustment amount at each sleeper. The input part of the multilayer perceptron is a random number, and the output dimension matches the number of sleepers in the track adjustment part. The output vector is used as the geometric adjustment amount for each sleeper.

[0082] In step 5, based on the multi-layer perceptron established in step 4, combined with the constraint model obtained in step 3, the objective function and the constraint are embedded into the neural network, that is, the network output is connected to the objective function neuron and the constraint neuron, and the RELU activation layer is embedded after the constraint neuron, and then added to the objective function; the gradient normalization is embedded in the back propagation process, and the optimized constraint model is obtained after multiple rounds of iterations, such as Figure 5 shown.

[0083] For track fine-tuning optimization, the loss function is written as:

[0084]

[0085] In step six, the gradient of the neural network is calculated and the adjustment amount is updated; as the number of neural network iterations increases, the network output will gradually approach the optimal solution under the current adjustment amplitude constraint and smoothness evaluation index constraint, and the optimized adjustment amount will be obtained when convergence occurs.

[0086] The first few steps only obtained the adjustment amount for each sleeper. Step seven uses the dynamic programming method to distribute the obtained adjustment amount to the large machine tamping track start amount, track fine-tuning rail pad thickness and adjustment pad thickness. The knapsack problem is used here to solve the combination. The calculated adjustment amount solved by NNS is used as the capacity of the knapsack. The four adjustments are regarded as four commodities. Each commodity has multiple choices for space occupation, but each commodity must be selected and can only be selected once. The input of the algorithm is the adjustment amount selection corresponding to the four combinations and the calculated adjustment amount generated by NNS. The pseudo code corresponding to the specific algorithm is as follows:

[0087]

[0088]

[0089]

[0090] Unlike traditional neural networks, which process predefined inputs by learning weights to produce outputs, neural networks operate without explicit input data or a traditional learning phase. Instead, neural networks start with random variables at the input of a multilayer perceptron (MLP). This approach is fundamentally different because the network is designed to directly optimize the decision variables rather than learning from a dataset.

[0091] experiment

[0092] The test was conducted using a section of absolute deviation data from K148+201 to K151+340 in a certain section. Due to the settlement effect of the roadbed, the absolute deviation at some locations has decreased. The long and short wave irregularities of the track have shown varying degrees of deterioration. The number of sleepers in this section is 5016, and the section length is 3.1 kilometers. The 10m chord measurement, 60m chord measurement, 5 / 30m vector distance difference, and 150 / 300m vector distance difference of this section are 2.94, 15.55, 6.20, and 61.90, respectively, which exceed the constraint amplitude limit. The design speed of this section is 250 km / h.

[0093] Due to the back-propagation effect of the objective function neuron on the variables, the constraint neurons cannot be lowered below the limit, which may lead to an infeasible solution. In order to obtain an optimal and feasible solution, a two-step training method is used. In the first stage, the loss function consists of five components, including the loss of the objective function neuron and the loss of the four regularity constraint neurons. In the second stage, the coefficients of the objective function are set to zero so that the standard parameters can be automatically adjusted to a level below the constraints.

[0094] The hyperparameter settings of NNS are shown in Table 2:

[0095] Table 2 Hyperparameter settings

[0096] Input size 5016 Hidden Layer 120 Output size 5016 Phase 1 Iterations 600 Phase 2 Iterations 200 Learning Rate 20e-3 Gradient Normalization Parameter 2.0 Constraint Quantity 13040956

[0097] The model is trained and the training results are as follows: Figure 6 As shown in the figure, as the number of training times increases, the model gradually converges and obtains a solution that satisfies the constraints after the second stage.

[0098] The model results are visualized and dynamic simulation is performed at the same time, as shown in Figure 7(a), Figure 7(b) and Figure 7(c). It can be found that compared with the manual method, the method of this embodiment can better reduce the body response and improve the train comfort.

[0099] 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 neuron-based combined optimization method for large-scale machine tamping and track fine-tuning, characterized by: The following steps are involved: Step 1: Select the constraint amplitude of the smoothness index for the track inventory data; Step 2: Setting the adjustment amount constraint amplitude of large machine tamping and track fine adjustment; Step 3: Establishing a track geometry adjustment constraint model; In step 3, the process of establishing the constraint model is: String measurement formula: Among them, e MCO is the chord measurement value, c1 is the ratio of chord length to sleeper spacing, m1 is a point inside the chord, n1 is the starting point of the chord, is the absolute deviation data at sleeper n1, is the absolute deviation data at the sleeper n1+c1; Vector difference formula: Among them, e ACO is the vector distance difference, c2 is the ratio of the reference chord length to the sleeper spacing, w is the ratio of the detection chord length to the sleeper spacing, m2 is the starting point of the detection chord, n2 is the starting point of the reference chord, is the absolute deviation data at sleeper m2, is the absolute deviation data at the sleeper m2+w, k is the ratio of the reference chord to the test chord; The calculation of vector distance difference and chord measurement is regarded as multi-channel convolution. The convolution process is as follows: I represents the vector distance difference or chord measurement matrix, B is the multi-channel convolution kernel, P is the absolute deviation data, AL is the horizontal deviation, Δ is the constraint amplitude converted from the original deviation to the chord measurement method and vector distance difference system; in: P=[p1p2…p n ] T <h2 style=";text-align:left;direction:ltr">AL=[AL1AL2…AL<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr"> ]<h2 style=";text-align:left;direction:ltr"> T Subtract the original absolute deviation data from the thickness of the fastener before adjustment to obtain the absolute deviation data excluding the thickness of the fastener before adjustment, recorded as Establish the track fine-tuning constraint model: Among them, δ i is the adjustment amount of the ith sleeper, B is the multi-channel convolution kernel, To adjust the vector, is the upper bound of the adjustment amount, is the lower bound of the adjustment amount; Step 4: Establish the shared layer part of the neural solver NNS; Step 5: Construct a loss function based on the constraint model; Step 6: Calculate the gradient of the neural network to solve the model and obtain the optimization result of the adjustment amount of tamping and fine tuning at each sleeper; Step 7. Use dynamic programming to distribute the adjustment amount to the adjustment amount corresponding to the large machine tamping and fine adjustment.

2. The neuron-based combined optimization method for large-scale machine tamping and track fine-tuning according to claim 1 is characterized in that: In step one, the inventory data is the operating speed of the line, and the smoothness indicators include the amplitudes of the chord measurement method and the vector distance difference method. The 10m chord measurement method, 60m chord measurement method, 5 / 30m vector distance difference method, and 150 / 300m vector distance difference method are used to perform smoothness evaluation and line smoothness optimization.

3. The neuron-based combined optimization method for large-scale machine tamping and track fine-tuning according to claim 2 is characterized in that: In step 2, the geometric adjustment amount of the large-scale tamping is its corresponding track start value, and the start value is selected as an integer multiple of 10, which is {0, 10, 20, 30, 40, 50, 60}; for track fine-tuning, the fastener adjustment is affected by the combination of pads. The types of pads include under-rail pads and adjustment pads. The specification stipulates that the number of under-rail pads is 1, and the thickness is 5mm or 10mm; the number of adjustment pads is between 0 and 2, and the thickness is 1, 2, 5 or 8mm. It can be seen that the total thickness of the fastener pads ranges from 5 to 26mm; based on this, the adjustment range of the combination of large-scale tamping and track fine-tuning is 5 to 86mm.

4. The neuron-based combined optimization method for large-scale machine tamping and track fine-tuning according to claim 3 is characterized in that: In step 4, a multilayer perceptron is used to randomly initialize the geometric adjustment amount at each sleeper. The input part of the multilayer perceptron is a random number, and the output dimension matches the number of sleepers in the track adjustment part. The output vector is used as the geometric adjustment amount for each sleeper.

5. The neuron-based combined optimization method for large-scale machine tamping and track fine-tuning according to claim 4 is characterized in that: In step 5, based on the multi-layer perceptron established in step 4, combined with the constraint model obtained in step 3, the objective function and constraints are embedded into the neural network, that is, the network output is connected to the objective function neuron and the constraint neuron, and the RELU activation layer is embedded after the constraint neuron, and then added to the objective function; the gradient normalization is embedded in the back propagation process, and the optimization adjustment amount is obtained after multiple rounds of iterations; for the track fine-tuning optimization constraint model, its loss function is written as:

6. The neuron-based combined optimization method for large-scale machine tamping and track fine-tuning according to claim 5 is characterized in that: In step six, the gradient of the neural network is calculated and the adjustment amount is updated; as the number of neural network iterations increases, the network output will gradually approach the optimal solution under the current adjustment amplitude constraint and smoothness evaluation index constraint, and the optimized adjustment amount will be obtained when convergence occurs.

7. The neuron-based combined optimization method for large-scale machine tamping and track fine-tuning according to claim 6 is characterized in that: In step seven, the dynamic programming method is used to distribute the obtained adjustment amount to the large machine tamping start amount, the track fine adjustment of the track pad thickness and the adjustment of the pad thickness; the knapsack problem is used to solve the combination, and the calculated adjustment amount solved by NNS is used as the capacity of the knapsack. The four adjustments are regarded as four commodities. Each commodity has multiple choices for the space it occupies, but each commodity must be selected and can only be selected once.

Images