A deep learning-based optimization control method for instability of heavy haul locomotive coupler

Abstract

The application relates to a deep learning-based optimization control method for instability of a heavy haul locomotive coupler, comprising the following steps: firstly, sensitivity analysis and dynamic simulation are respectively performed, high-sensitivity parameters and key response dynamic data are obtained, and preprocessing is performed to obtain a sample set; then, a long heavy haul train coupler stability proxy model is constructed; then, the sample set is used to train the long heavy haul train coupler stability proxy model; finally, a multi-parameter optimization problem with coupler lateral stability as the optimization objective is constructed, an improved NSGA-III algorithm based on SOBOL and SHAP analysis is used for global optimization based on the proxy model, and an optimal parameter combination is obtained to generate a coupler lateral instability optimization control strategy. A high-precision long heavy haul train coupler stability proxy model is constructed, and joint optimization design between complex systems is realized. Meanwhile, the improved NSGA-III algorithm is combined to efficiently realize optimization control of coupler instability.

Description

Technical Field

[0001] This application relates to the field of optimization and control technology for coupler instability in heavy-duty locomotives, and in particular to an optimization and control method for coupler instability in heavy-duty locomotives based on deep learning. Background Technology

[0002] Heavy-haul railway transportation, as a core mode of bulk cargo transportation, boasts significant advantages such as large capacity, low energy consumption, and environmental friendliness. With the profound adjustment of the "road-to-rail" transportation structure, the demand for railway freight continues to grow, making it inevitable to expand transport capacity by increasing train formation length and improving traction quality. However, many of my country's heavy-haul railways are located in areas with complex terrain, facing harsh conditions such as a high proportion of long gradients, small curve radii, and large elevation differences. This results in the coupler system enduring severe longitudinal impacts and lateral coupling loads during braking and speed regulation. When the lateral force on the coupler exceeds the critical threshold, it can easily lead to instability phenomena such as coupler deflection and abnormal opening of the coupler tongue, which in severe cases can cause major safety accidents such as train derailment and coupler breakage.

[0003] Existing methods for coupler stability control mainly rely on physical models based on multibody dynamics, conducting stability analysis by establishing coupled dynamic equations between the train and the track. However, this approach has significant limitations: firstly, heavy-haul train systems contain hundreds of degrees of freedom, and the dynamic models are highly nonlinear, resulting in low simulation efficiency (a single full-train dynamic simulation takes more than 2 hours), making it difficult to meet real-time optimization requirements; secondly, under complex operating conditions, model parameters exhibit uncertainty, making it impossible to achieve high-precision instability early warning and control.

[0004] Therefore, there is an urgent need in related technologies for a coupler stability control method that can improve the safety and transportation efficiency of heavy-haul trains. Summary of the Invention

[0005] Therefore, it is necessary to provide an optimization and control method for coupler instability of heavy-haul locomotives based on deep learning, which can improve the operational safety and transportation efficiency of heavy-haul trains and address the aforementioned technical problems.

[0006] Firstly, this application provides an optimized control method for coupler instability in heavy-haul locomotives based on deep learning. The method includes:

[0007] Sensitivity analysis and dynamic simulation were performed separately to obtain high-sensitivity parameters and key response dynamic data, and preprocessing was performed to obtain a sample set;

[0008] A stability proxy model for couplers of long and heavy-haul trains is constructed. The model includes a dual-channel feature extraction module, a multi-dimensional feature fusion module, a dual-channel convolutional feature regression module, and an adaptive data loss and physical constraint module.

[0009] The sample set was used to train the stability proxy model of the coupler of the long and heavy-haul train;

[0010] A multi-parameter optimization problem with the lateral stability of the coupler as the optimization objective is constructed. An improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis is used to perform global optimization based on the surrogate model to obtain the optimal parameter combination and generate the coupler lateral instability optimization and control strategy.

[0011] Optionally, in one embodiment of this application, the step of performing sensitivity analysis and dynamic simulation to obtain high-sensitivity parameters and key response dynamic data includes:

[0012] Sensitivity analysis was used to screen for highly sensitive parameters affecting the lateral instability of the coupler and to determine the range of parameter values.

[0013] Simulation calculations were conducted based on the multibody dynamics model of heavy-haul trains to obtain dynamic data of key responses.

[0014] Optionally, in one embodiment of this application, the preprocessing includes:

[0015] The high-sensitivity parameters and key response dynamic data are normalized and divided into training set, test set and validation set according to a preset ratio.

[0016] Optionally, in one embodiment of this application, the dual-channel feature extraction module and the dual-channel convolutional feature regression module are composed of convolutional neural networks and fully connected layers, and the multi-dimensional feature fusion module is composed of a cascaded encoder-decoder network.

[0017] Optionally, in one embodiment of this application, the adaptive data loss and physical constraint module includes:

[0018] Calculate the forward propagation loss based on the physical constraints of the output channel weights and output parameters;

[0019] The backpropagation gradient is determined based on the forward propagation loss and model parameters.

[0020] Optionally, in one embodiment of this application, the multi-parameter optimization problem with the lateral stability of the coupler as the optimization objective is as follows:

[0021]

[0022] Among the hook and buffer parameters, the radius of the hook tail arc surface is... r , front plate arc radius R Coefficient of friction between hook tail arc surface μ And the stiffness of the secondary spring in the suspension parameters K sy Second-stage stop lateral free clearance hsf and the lateral stiffness of the second-stage stop K z It can be represented as a vector. x = ( x 1, x 2, x 3, x 4, x 5, x 6) = ( r , R , μ , K sy , h sf , K z ), designated as design variables, Let any one of these be an objective function. To represent the total number of functions that need to be minimized in the objective function, For inequality constraints, Let be the total number of inequality constraints. Represents any one of these variables. The total number of design variables, To design the lower bound of the variable, To design the upper limit of variables, To minimize the objective function, The constraints in the optimization problem are defined. Find the variable values ​​that satisfy the objective function and constraints;

[0023] The optimization objective is:

[0024]

[0025]

[0026] in, For the coupler swing angle, For the lateral force of the wheel and axle, and Each represents its maximum limit;

[0027] The vehicle dynamics performance standard is used as a constraint, expressed as follows:

[0028]

[0029]

[0030] in, For derailment coefficient constraints, For wheel load reduction rate constraints, For the lateral force of the wheel and rail, For the vertical force of the wheel and rail, To reduce the load on the wheels, This represents the average static wheel weight.

[0031] Optionally, in one embodiment of this application, the step of using the improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis to perform global optimization based on the surrogate model includes:

[0032] Find the index of the expression with the largest difference value based on the function expression in each iteration;

[0033] The first-order SOBOL effect and interaction effect of the expression with the largest difference value are used to dynamically regulate the size of the genetic population as well as the mutation rate and magnitude.

[0034] Based on the dynamic changes of the parameters and the different features provided by SHAP analysis, the specific impact of each feature on the model output under different values ​​is determined.

[0035] The aforementioned deep learning-based optimization and control method for coupler instability in heavy-haul locomotives firstly involves sensitivity analysis and dynamic simulation to obtain high-sensitivity parameters and key response dynamic data, which are then preprocessed to obtain a sample set. Next, a surrogate model for the stability of couplers in long, heavy-haul trains is constructed. This model includes a dual-channel feature extraction module, a multi-dimensional feature fusion module, a dual-channel convolutional feature regression module, and an adaptive data loss and physical constraint module. The surrogate model is then trained using the sample set. Finally, a multi-parameter optimization problem is constructed with the lateral stability of the coupler as the optimization objective. An improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis is used to perform global optimization based on the surrogate model, obtaining the optimal parameter combination and generating an optimized control strategy for lateral coupler instability. In other words, a high-precision surrogate model for the stability of couplers in long and heavy-haul trains was constructed using six high-sensitivity parameters from the coupler and suspension systems as inputs and four dynamic response peaks as outputs. The surrogate model includes a dual-channel convolutional feature extraction module, which can fully extract features from the coupler and suspension systems. CEDNet achieves multi-dimensional feature fusion in a cascaded manner, effectively overcoming the problem of poor fitting results. By introducing error differences, the predictive performance of each channel is adaptively learned, and combined with the physical range limitations of the output parameters, the loss of each channel is ensured to be in a balanced state, thus achieving high-precision prediction of the surrogate model. It has good adaptability and realizes joint optimization design between complex systems. Simultaneously, combined with the improved NSGA-Ⅲ algorithm, the optimized control of coupler instability is efficiently achieved. Attached Figure Description

[0036] Figure 1This is a flowchart illustrating an optimized control method for coupler instability in heavy-duty locomotives based on deep learning, as shown in one embodiment.

[0037] Figure 2 This is a schematic diagram of a heavy-haul train simulation and the location of key parameters in one embodiment;

[0038] Figure 3 This is a schematic diagram illustrating the dynamic response of the output performance index under unstable and non-instable conditions in one embodiment.

[0039] Figure 4 This is a schematic diagram of the structure of a stability proxy model for the coupler of a long, heavy-haul train in one embodiment.

[0040] Figure 5 This is a schematic diagram of the dual-channel feature extraction module in one embodiment;

[0041] Figure 6 This is a schematic diagram of the structure of a multi-dimensional feature fusion module in one embodiment;

[0042] Figure 7 This is a schematic diagram of the Pareto front for multi-objective optimization based on a surrogate model for the stability of couplers in a long and heavy-haul train, as shown in one embodiment.

[0043] Figure 8 This is a schematic diagram illustrating the range of multi-objective optimization results based on a stability proxy model for couplers of long and heavy-haul trains in one embodiment.

[0044] Figure 9 This is a schematic diagram illustrating the validity verification of the specific simulation response of the multi-objective optimization results in one embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] In one embodiment, such as Figure 1 As shown, a deep learning-based optimization control method for coupler instability in heavy-haul locomotives is provided, including the following steps:

[0047] S101: Perform sensitivity analysis and dynamic simulation respectively to obtain high sensitivity parameters and key response dynamic data, and perform preprocessing to obtain a sample set.

[0048] In this embodiment of the application, firstly, sensitivity analysis and dynamic simulation are performed to obtain high-sensitivity parameters and key response dynamic data, and preprocessing is performed to obtain a sample set. The sensitivity analysis is based on detailed historical research and experience to screen high-sensitivity parameters that affect the lateral instability of the coupler, and the dynamic simulation is carried out by establishing a dynamic model of the heavy-haul train to perform simulation calculations.

[0049] Specifically, in one embodiment of this application, the step of performing sensitivity analysis and dynamic simulation to obtain high-sensitivity parameters and key response dynamic data includes:

[0050] S201: Sensitivity analysis is used to screen high-sensitivity parameters that affect the lateral instability of the coupler and to determine the range of parameter values.

[0051] S203: Conduct simulation calculations based on the multibody dynamics model of heavy-haul trains to obtain key response dynamic data.

[0052] In one embodiment of this application, such as Figure 2 The diagram shows the locations of key systems affecting the lateral instability of the coupler. Sensitivity analysis was used to screen high-sensitivity parameters affecting the lateral instability of the coupler and to determine the range of these parameters. As shown in Table 1, the six screened high-sensitivity parameters are: the radius of the coupler tail arc surface of the coupler-flush system, the radius of the front trailing plate arc surface, the coupler tail friction coefficient, and the lateral stiffness of the secondary suspension spring, the free clearance of the secondary suspension lateral stop, and the stiffness of the secondary suspension lateral stop.

[0053] Table 1

[0054]

[0055] Simulation calculations were conducted based on an established multibody dynamics model of heavy-haul trains. Different combinations of six high-sensitivity parameters were obtained through Latin hypercube sampling. These parameters were then adjusted and calculated within the multibody dynamics model of the heavy-haul train, ultimately yielding key dynamic response data including coupler sway angle, wheel-rail lateral force, derailment coefficient, and wheel load reduction rate. Figure 3 As shown, a comparative diagram illustrating the response results of the coupler under lateral instability and non-instability is presented based on simulation results. The peak value of the response signal at the moment of instability is collected as the final key dynamic response data.

[0056] In one embodiment of this application, the preprocessing includes:

[0057] The high-sensitivity parameters and key response dynamic data are normalized and divided into training set, test set and validation set according to a preset ratio.

[0058] In one embodiment of this application, min-max normalization is used to normalize the high-sensitivity parameters and key response dynamic data, mapping the data values ​​to the range [-1, 1] to eliminate the influence of dimensions. The six normalized high-sensitivity parameters (key parameters) serve as the model input, and the four dynamic response peaks (important performance) serve as the model output, forming a sample set. The sample set is then divided into a training set, a test set, and a validation set according to a preset ratio.

[0059] S103: Construct a stability proxy model for couplers of long and heavy-haul trains. The model includes a dual-channel feature extraction module, a multi-dimensional feature fusion module, a dual-channel convolutional feature regression module, and an adaptive data loss and physical constraint module.

[0060] In this embodiment of the application, a construction is performed as follows: Figure 4 The stability proxy model for couplers of long and heavy-haul trains shown includes a dual-channel feature extraction module, a multi-dimensional feature fusion module, a dual-channel convolutional feature regression module, and an adaptive data loss and physical constraint module. The dual-channel feature extraction module is used to extract features of key parameters, the multi-dimensional feature fusion module is used to mine multi-dimensional features of key parameters, the dual-channel convolutional feature regression module is used to process the feature information obtained by the multi-dimensional feature fusion module, and a high-precision regression module is used to predict the specific output value. The adaptive data loss and physical constraint module is used to generate corresponding losses to adjust the parameters of other modules.

[0061] Specifically, in one embodiment of this application, the dual-channel feature extraction module and the dual-channel convolutional feature regression module are composed of convolutional neural networks and fully connected layers, and the multi-dimensional feature fusion module is composed of cascaded encoder-decoder networks.

[0062] In one embodiment of this application, the dual-channel feature extraction module consists of a convolutional neural network and a fully connected layer. Its basic units include Conv (convolutional layer), BN (batch normalization), ReLU (activation function), etc., and the internal module connections are as follows: Figure 5 As shown. The multi-dimensional feature fusion module uses a CEDNet cascade encoder-decoder to mine multi-dimensional features of key parameters. Its basic units include Conv (convolutional layer), ReLU (activation function), Fc (fully connected layer), etc., and the specific internal module connections are as follows. Figure 6 As shown.

[0063] In one embodiment of this application, the adaptive data loss and physical constraint module includes:

[0064] S301: Calculate the forward propagation loss based on the physical constraints of the output channel weights and output parameters;

[0065] S303: Determine the backpropagation gradient based on the forward propagation loss and model parameters.

[0066] In one embodiment of this application, a custom regression network layer is first used to obtain the weights of each output channel, which are then assigned to each channel. Combined with the physical constraints of the output parameters, the forward propagation loss is obtained.

[0067]

[0068]

[0069]

[0070]

[0071] in, For the first n Total forward propagation loss, covering For data loss and For physical loss; For each channel of data loss, a loss weight is assigned, and Similarly, The loss weight for each channel is the physical loss, and ; This is the first training session c Prediction results for each channel, This is the first training session c The actual results of each channel and For the first time in this training c The upper and lower limits of the output characteristics of each channel.

[0072] Then, obtain the backpropagation gradient:

[0073]

[0074]

[0075] in, The gradient is used for backpropagation. These are model parameters; For the first n The penalty of +1 training iterations is used for backpropagation gradient return.

[0076] S105: The sample set is used to train the stability proxy model of the coupler of the long and heavy-haul train.

[0077] In this embodiment, the stability surrogate model of the coupler of a long and heavy-haul train is trained using training set data from the sample set. A nonlinear mapping relationship between key parameters of lateral instability of the coupler and important performance is established. The model hyperparameters are adjusted and the model structure is optimized using validation set data. At the same time, the prediction accuracy of the surrogate model is verified using test set data. The reliability of the model is evaluated by indicators such as root mean square error and coefficient of determination. Based on the verification results, the model performance is evaluated to obtain the final surrogate model.

[0078] S107: Construct a multi-parameter optimization problem with the lateral stability of the coupler as the optimization objective. Use the improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis to perform global optimization based on the surrogate model, obtain the optimal parameter combination, and generate the coupler lateral instability optimization and control strategy.

[0079] In this embodiment, to balance the overall dynamic performance of the vehicle, considering coupler stability, vehicle curve handling performance, and operational safety, a constrained multi-objective optimization control problem for coupler instability is constructed, using coupler sway angle and wheel axle lateral force as multi-objective functions and safety indicators such as derailment coefficient and wheel load reduction rate as constraints. An improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis is employed to perform global optimization using a surrogate model, obtaining the optimal parameter combination and generating a coupler lateral instability optimization control strategy. Specifically, the surrogate model can be trained with a small number of six high-sensitivity samples to predict a wider range of dynamic combinations, including those not included in the sample. This surrogate model can then perform multi-objective optimization for all combinations, finding the combination with a better dynamic response than the original combination—the combination that prevents coupler instability—which constitutes the coupler lateral instability optimization control strategy.

[0080] In one embodiment of this application, the multi-parameter optimization problem with the lateral stability of the coupler as the optimization objective is as follows:

[0081]

[0082] Among the hook and buffer parameters, the radius of the hook tail arc surface is... r , front plate arc radius R Coefficient of friction between hook tail arc surface μ And the stiffness of the secondary spring in the suspension parameters K sy Second-stage stop lateral free clearance h sf and the lateral stiffness of the second-stage stop K z It can be represented as a vector. x = ( x 1, x 2,x 3, x 4, x 5, x 6) = ( r , R , μ , K sy , h sf , K z ), designated as design variables, Let any one of these be an objective function. To represent the total number of functions that need to be minimized in the objective function, For inequality constraints, Let be the total number of inequality constraints. Represents any one of these variables. The total number of design variables, To design the lower bound of the variable, To design the upper limit of variables, To minimize the objective function, The constraints in the optimization problem are defined. Find the variable values ​​that satisfy the objective function and constraints;

[0083] The optimization objective is:

[0084]

[0085]

[0086] in, For the coupler swing angle, For the lateral force of the wheel and axle, and Each represents its maximum limit;

[0087] The vehicle dynamics performance standard is used as a constraint, expressed as follows:

[0088]

[0089]

[0090] in, For derailment coefficient constraints, For wheel load reduction rate constraints, For the lateral force of the wheel and rail, For the vertical force of the wheel and rail, To reduce the load on the wheels, This represents the average static wheel weight.

[0091] In one embodiment of this application, the global optimization based on the surrogate model using the improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis includes:

[0092] S401: Find the index of the expression with the largest difference value based on the function expression in each iteration.

[0093] S403: Dynamically regulates the size, variation rate, and magnitude of the genetic population based on the SOBOL first-order effect and interaction effect of the expression with the largest difference value.

[0094] S405: Based on the dynamic changes of the parameters and the different features provided by SHAP analysis, determine the specific impact of each feature on the model output under different values.

[0095] In one embodiment of this application, the NSGA-III multi-objective optimization algorithm is improved based on SOBOL and SHAP analysis. SOBOL analysis provides global sensitivity analysis, which can identify the parameters that have the greatest impact on the model output and their interactions. SHAP analysis provides local sensitivity analysis, which can show in detail the specific impact of each feature on the model output under different values. The implementation method is as follows:

[0096] First, find the index of the expression with the largest difference value based on the function expression in each iteration, as shown in the following formula:

[0097]

[0098] Where index is the index (1-4), argmax represents the index of the expression that makes the expression reach its maximum value, and Cost... i (x) are respectively { f 1(x), f 2(x), g 1(x), g 2(x)}.

[0099] Then, based on the SOBOL first-order effect and interaction effect of the expression with the largest difference value, the genetic population size (crossover and mutation), as well as the mutation rate and amplitude, are dynamically adjusted to achieve a larger crossover and mutation effect for inputs with larger effects.

[0100]

[0101]

[0102]

[0103]

[0104] in, , , , The first j The 1st gene, i.e., the 1st gene j The output includes the number of crossover populations, the number of mutant populations, the probability of individual gene mutations in the mutant populations, and the length of the mutant population. and These represent population crossover and mutation probabilities, respectively. The initial mutation probability, The initial variation amplitude, For population size, and These are the upper and lower limits of the output, respectively; and S This is the set of SOBOL effect value matrices. This is the floor function.

[0105] Then, based on the dynamic changes of the parameters and the different features provided by SHAP analysis, the specific impact of each feature on the model output under different values ​​is determined. The specific improvement methods are crossover and mutation.

[0106]

[0107]

[0108] In the formula, , Paternal line, , For offspring lines formed through genetic crossover, As the crossover factor, the influence of eigenvalues ​​is assigned based on SHAP analysis, with the aim of allowing the corresponding genes in the daughter lines to crossover within a certain range. For the individual before the mutation, Similarly, for the mutated individuals, It is a variable factor, the purpose of which is to cause the genes of the offspring line to mutate in a positive or negative direction. Generate random numbers that follow a standard normal distribution.

[0109] In one specific embodiment of this application, a set of Pareto optimal solutions were obtained through a multi-objective optimization method. These solutions are... Figure 7 The diagram is presented in the form of a Pareto front, which represents the set of optimal solutions that achieve a balance among multiple objectives. Figure 8This further details the impact of the six input parameters on the optimization range of the four output parameters. Each sub-graph corresponds to one output parameter; the horizontal axis represents different values ​​of the input parameter, and the vertical axis represents the optimization range of the corresponding output parameter. This allows for a direct observation of how each input parameter affects the optimization effect of the output parameter. In the dynamic model, a 100-second measured coupler force is provided for coupler instability verification. The coupler tension-compression transition occurs at 69 seconds, at which point the coupler force reaches its maximum value. According to the multi-objective optimization results, when the coupler reaches its maximum value, no lateral instability occurs; the maximum coupler swing angle is only 0.14°, far below the 6° limit. Other response results are also below the instability limit. The verification results of the multi-objective optimization scheme are as follows: Figure 9 As shown.

[0110] In the aforementioned deep learning-based optimization and control method for coupler instability in heavy-haul locomotives, firstly, sensitivity analysis and dynamic simulation are performed to obtain high-sensitivity parameters and key response dynamic data, which are then preprocessed to obtain a sample set. Next, a surrogate model for the stability of couplers in long, heavy-haul trains is constructed. This model includes a dual-channel feature extraction module, a multi-dimensional feature fusion module, a dual-channel convolutional feature regression module, and an adaptive data loss and physical constraint module. Then, the surrogate model is trained using the sample set. Finally, a multi-parameter optimization problem is constructed with the lateral stability of the coupler as the optimization objective. An improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis is used to perform global optimization based on the surrogate model, obtaining the optimal parameter combination and generating an optimized control strategy for lateral coupler instability. In other words, a high-precision surrogate model for the stability of couplers in long and heavy-haul trains was constructed using six high-sensitivity parameters from the coupler and suspension systems as inputs and four dynamic response peaks as outputs. The surrogate model includes a dual-channel convolutional feature extraction module, which can fully extract features from the coupler and suspension systems. CEDNet achieves multi-dimensional feature fusion in a cascaded manner, effectively overcoming the problem of poor fitting results. By introducing error differences, the predictive performance of each channel is adaptively learned, and combined with the physical range limitations of the output parameters, the loss of each channel is ensured to be in a balanced state, thus achieving high-precision prediction of the surrogate model. It has good adaptability and realizes joint optimization design between complex systems. Simultaneously, combined with the improved NSGA-Ⅲ algorithm, the optimized control of coupler instability is efficiently achieved.

[0111] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0112] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0113] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An optimized control method for coupler instability in heavy-haul locomotives based on deep learning, characterized in that, The method includes: Sensitivity analysis and dynamic simulation were performed separately to obtain high-sensitivity parameters and key response dynamic data, and preprocessing was performed to obtain a sample set; A stability proxy model for couplers of long and heavy-haul trains is constructed. The model includes a dual-channel feature extraction module, a multi-dimensional feature fusion module, a dual-channel convolutional feature regression module, and an adaptive data loss and physical constraint module. The sample set was used to train the stability proxy model of the coupler of the long and heavy-haul train; A multi-parameter optimization problem with the lateral stability of the coupler as the optimization objective is constructed. An improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis is used to perform global optimization based on the surrogate model to obtain the optimal parameter combination and generate a coupler lateral instability optimization and control strategy. The multi-parameter optimization problem with the lateral stability of the coupler as the optimization objective is as follows: Among the hook and buffer parameters, the radius of the hook tail arc surface is... r , front plate arc radius R Coefficient of friction between hook tail arc surface μ And the stiffness of the secondary spring in the suspension parameters K sy Second-stage stop lateral free clearance h sf and the lateral stiffness of the second-stage stop K z It can be represented as a vector. x = ( x 1, x 2, x 3, x 4, x 5, x 6) = ( r , R , μ , K sy , h sf , K z ), designated as design variables, Let any one of these be an objective function. To represent the total number of functions that need to be minimized in the objective function, For inequality constraints, Let be the total number of inequality constraints. Represents any one of these variables. The total number of design variables, To design the lower bound of the variable, To design the upper limit of variables, To minimize the objective function, The constraints in the optimization problem are defined. Find the variable values ​​that satisfy the objective function and constraints; The optimization objective is: in, For the coupler swing angle, For the lateral force of the wheel and axle, and Each represents its maximum limit; The vehicle dynamics performance standard is used as a constraint, expressed as follows: in, For derailment coefficient constraints, For wheel load reduction rate constraints, For the lateral force of the wheel and rail, For the vertical force of the wheel and rail, To reduce the load on the wheels, This represents the average static wheel weight.

2. The method for optimizing and controlling coupler instability in heavy-haul locomotives based on deep learning according to claim 1, characterized in that, The process of performing sensitivity analysis and dynamic simulation to obtain high-sensitivity parameters and key response dynamic data includes: Sensitivity analysis was used to screen for highly sensitive parameters affecting the lateral instability of the coupler and to determine the range of parameter values. Simulation calculations were conducted based on the multibody dynamics model of heavy-haul trains to obtain dynamic data of key responses.

3. The method for optimizing and controlling coupler instability in heavy-haul locomotives based on deep learning according to claim 1, characterized in that, The preprocessing includes: The high-sensitivity parameters and key response dynamic data are normalized and divided into training set, test set and validation set according to a preset ratio.

4. The method for optimizing and controlling coupler instability in heavy-haul locomotives based on deep learning according to claim 1, characterized in that, The dual-channel feature extraction module and the dual-channel convolutional feature regression module are composed of convolutional neural networks and fully connected layers, and the multi-dimensional feature fusion module is composed of cascaded encoder-decoder networks.

5. The method for optimizing and controlling coupler instability in heavy-haul locomotives based on deep learning according to claim 1, characterized in that, The adaptive data loss and physical constraint module includes: Calculate the forward propagation loss based on the physical constraints of the output channel weights and output parameters; The backpropagation gradient is determined based on the forward propagation loss and model parameters.

6. The method for optimizing and controlling coupler instability in heavy-haul locomotives based on deep learning according to claim 1, characterized in that, The improved NSGA-Ⅲ algorithm based on SOBOL and SHAP analysis performs global optimization based on the surrogate model, including: Find the index of the expression with the largest difference value based on the function expression in each iteration; The first-order SOBOL effect and interaction effect of the expression with the largest difference value are used to dynamically regulate the size of the genetic population as well as the mutation rate and magnitude. Based on the dynamic changes of the parameters and the different features provided by SHAP analysis, the specific impact of each feature on the model output under different values ​​is determined.

Images