CN115481488A - Multi-objective optimization method for crashworthiness of square cone energy-absorbing structures based on machine learning - Google Patents

Abstract

The invention provides a multi-objective optimization method for the crashworthiness of a square cone energy-absorbing structure based on machine learning, comprising the following steps: establishing a finite element simulation model of the square cone energy-absorbing structure of a subway train; based on the established finite element simulation The method of combining the model with the experimental design is used to extract the structural parameters and energy-absorbing characteristic curves of the energy-absorbing structure of the subway train; sampling is carried out according to the Latin super-cubic method, and the optimal design of the square-cone energy-absorbing structure of the subway train is determined through the Design of Experiments (DOE). Optimal energy absorption characteristic curve prediction model, as well as optimization variables and optimization objectives; according to the optimal energy absorption characteristic curve prediction model, optimization variables and optimization objectives, an optimization theoretical model is established; the optimization theoretical model is re-sampled by Hamersley sampling method , use the optimal energy absorption characteristic curve prediction model to calculate the corresponding energy absorption and peak force, and generate a new DOE; use the global response surface method (GRSM) for multi-objective optimization, and obtain the optimization result; based on the optimization result, get the Pa of the optimization target The reciprocal solution set uses the minimum distance method to make an optimal decision on the optimized Pareto solution set to obtain the optimal solution.

Description

Multi-objective optimization method for crashworthiness of square cone energy-absorbing structures based on machine learning

technical field

The invention relates to the technical field of train crashworthiness detection, in particular to a machine learning-based multi-objective optimization method for the crashworthiness of a square-cone energy-absorbing structure.

Background technique

During the collision process of the urban rail train, the energy is absorbed through the coupler device and the square cone energy-absorbing structure. Among them, the anti-climbing energy-absorbing structure is the last line of defense for train passive safety protection, and its crashworthiness is of great significance to the safety of the car body. Once the anti-climbing energy-absorbing structure fails and causes a collision accident, it will cause heavy casualties. Therefore, it is of great significance to study the crashworthiness characteristics of square cone energy-absorbing structures for subway trains.

The traditional way to study the energy-absorbing structure is mainly to use the finite element method, experimental method and multi-body dynamics method to study the crashworthiness of the train. The experimental method requires more material and financial resources, and the experiment has a large Uncertainty, the method of finite element method and multi-body dynamics is used for research, which requires high performance of the computer and takes a long time for calculation.

The square cone energy-absorbing structure is the main energy-absorbing element of the subway vehicle, and its energy-absorbing characteristic (force-displacement) curve has a great influence on the safety of the subway vehicle operation. Therefore, it is necessary to find a reasonable way to obtain the energy-absorbing The force-displacement curve of the structure is of great significance to the optimization of the crashworthiness of subway trains. At present, most of them use the method of combining finite element model and test to obtain the force-displacement curve of the square cone energy-absorbing structure, but this method requires a long time and expensive funds. Therefore, it is necessary to adopt a more effective method for multi-objective optimization of the crashworthiness of the square-cone energy-absorbing structure.

Contents of the invention

In order to reduce calculation time while ensuring calculation accuracy, the present invention proposes a multi-objective optimization method for crashworthiness of square cone energy-absorbing structures based on machine learning. By adopting the method of the invention, not only the prediction accuracy of the energy absorption characteristics is high, but also the calculation time can be greatly reduced, and the multi-objective optimization of the crashworthiness of the square-cone energy-absorbing structure of the subway vehicle is empowered with big data.

In order to achieve the above object, the present invention provides a multi-objective optimization method for the crashworthiness of square cone energy-absorbing structures based on machine learning, including the following steps:

Establish the finite element simulation model of the square cone energy-absorbing structure of the subway train;

Based on the method of combining the established finite element simulation model with experimental design, the structural parameters and energy-absorbing characteristic curves of the energy-absorbing structure of the subway train were extracted; according to the sampling method using the Latin superstructure, a virtual experimental design was carried out to determine the square-cone absorbing structure of the subway train. The prediction model of the optimal energy absorption characteristic curve of the energy structure, as well as the optimization variables and optimization objectives;

According to the optimal energy absorption characteristic curve prediction model, optimization variables and optimization objectives, an optimization theoretical model is established;

Using the Hamersley sampling method to resample the optimized theoretical model, use the optimal energy absorption characteristic curve prediction model to calculate the corresponding energy absorption and peak force, and generate a new DOE; use the global response surface method for multi-objective optimization, and get the optimization result;

Based on the optimization results, the Pareto solution set of the optimization target is obtained, and the optimal decision is made on the optimized Pareto solution set by using the minimum distance method to obtain the optimal solution.

Preferably, based on the same boundary conditions of the finite element simulation, the dynamic impact experiment of the full-scale square cone energy-absorbing structure is carried out, and the combination of experiment and simulation is adopted. By comparing the force-displacement curve and displacement- Energy curves and deformation sequence patterns verify the accuracy of the established finite element model.

Preferably, based on the method of combining the finite element simulation model established and the experimental design, the structural parameters and the energy-absorbing characteristic curve of the energy-absorbing structure of the subway train are extracted; The prediction model of the optimal energy-absorbing characteristic curve of the conical energy-absorbing structure, as well as the optimization variables and optimization objectives, specifically:

Use the five input variables of DOE as the input features of network training in machine learning, and use the output of DOE as the actual output of network training in machine learning;

Randomly divide the input and actual output of network training into training set and test set;

Normalize the input data and the actual output data;

Map the input geometry of the energy-absorbing structure to the required output response machine learning (ML) architecture, use the normalized training set and test set as samples, and use MLP, RNN, LSTM and GRU respectively The network model of machine learning predicts the energy-absorbing characteristic curve of the energy-absorbing structure, and compares and analyzes the prediction accuracy of different models. The network model with the highest accuracy of the predicted output curve is used as the optimal model for the square-cone energy-absorbing structure of the subway train. Energy absorption characteristic curve prediction model.

Preferably, the optimal energy-absorbing characteristic curve prediction model of the square-cone energy-absorbing structure of a subway train is a gated cyclic neural network model, and the outer wall thickness T _A , outer wall thickness T _B , partition thickness T _gb , and strength δ _A and the strength δ _B of the aluminum honeycomb B are used as optimization variables, and the energy absorption and peak force of the square cone energy-absorbing structure are the optimization objectives.

Preferably, the established optimization theoretical model is as follows:

PCF=max(F(s))

Among them, S is the compression displacement of the square cone energy-absorbing structure; F(s) is the axial force of the square cone energy-absorbing structure; T _A and T _B are the outer wall thicknesses of the square cone energy-absorbing structure; T _gb is the square The diaphragm thickness of the cone energy-absorbing structure; δ _A and δ _B are the strengths of aluminum honeycomb A and aluminum honeycomb B respectively; EA is the energy absorption of the square cone energy-absorbing structure; PCF is the peak force of the square cone energy-absorbing structure .

Preferably, based on the optimization result of GRSM, the Pareto solution set of the optimization target is obtained, and the minimum distance method is used to make an optimal decision on the optimized Pareto solution set to obtain the optimal solution, specifically:

The energy absorption EA, specific energy absorption SEA, average force F _mean , and energy absorption efficiency IFE are used as energy absorption evaluation indicators, and the optimal solution is compared with the experimental results using radar charts to verify the feasibility of the optimization method:

Among them, D is the distance between the knee joint point and the Pareto solution set point; m is the number of optimization objectives, and f _i ^k is the optimal point k of the i-th optimization objective;

Among them, m is the mass of the square cone energy-absorbing structure; s is the compression displacement of the square cone energy-absorbing structure; F _mean is the average force of the square cone energy-absorbing structure, and PCF is the peak force of the square cone energy-absorbing structure; IFE is the energy absorption efficiency of the square cone energy-absorbing structure.

The present invention has the following beneficial effects:

1. The present invention proposes a multi-objective optimization method for the crashworthiness of square cone energy-absorbing structures based on machine learning. Based on the method of combining experimental design and finite element simulation, the structural parameters and energy-absorbing structures of subway trains are extracted. characteristic curve. At the same time, by using four neural network models of multi-layer perceptron (MLP), recurrent neural network (RNN), long short-term memory method (LSTM) and gated recurrent neural network (GRU), the energy-absorbing characteristics of the energy-absorbing structure are respectively analyzed. The curve (force-displacement curve) is predicted, and the accuracy of the output curve predicted by the GRU model is high. Therefore, the present invention adopts the gated cyclic neural network model (GRU) as the optimal energy-absorbing characteristic curve prediction model of the square cone type energy-absorbing structure of the subway train, so that the accuracy of energy-absorbing characteristic prediction is high, and the reliability of the predicted characteristic curve is guaranteed . On the premise of ensuring the computational efficiency, it can also improve the prediction accuracy and ensure the credibility of the multi-objective optimization results.

2. The present invention proposes a multi-objective optimization method for the crashworthiness of square cone energy-absorbing structures based on machine learning, which adopts the method of combining global adaptive response surface (GRSM) and machine learning, and improves the energy-absorbing structure of subway trains. Multi-objective optimization for crashworthiness. When the data to be calculated is large, the calculation time can be greatly reduced, and the multi-objective optimization of the crashworthiness of the square cone energy-absorbing structure of the subway vehicle can be empowered by big data.

When the amount of data to be calculated is large, it takes a long time to generate data sets by using the traditional finite element and optimization algorithm combined method for multi-objective optimization; using machine learning methods can reduce the time to generate data sets, and GRSM Combined with machine learning methods, the parameter optimization range of the energy-absorbing structure can be expanded, the pattern recognition of the energy-absorbing structure can be carried out, the calculation efficiency can be greatly improved, the ability of product upgrading can be improved, and the cost of product development can be reduced.

The optimization algorithm cannot improve the calculation speed too much. The main reason is that machine learning can reduce the time to generate data sets. The optimization process is based on the generated data sets. Therefore, it is generally explained that the combination of machine learning and optimization algorithms can be used. Improve computational efficiency.

3. The present invention proposes a machine learning-based multi-objective optimization method for the crashworthiness of square cone energy-absorbing structures, providing a new idea for multi-objective optimization of energy-absorbing structures. At present, the square cone energy-absorbing structure mainly adopts the combination of finite element modeling and optimization algorithm to optimize the energy-absorbing structure. Provide a new way of thinking.

In addition to the objects, features and advantages described above, the present invention has other objects, features and advantages. Hereinafter, the present invention will be described in further detail with reference to the drawings.

Description of drawings

The accompanying drawings constituting a part of this application are used to provide further understanding of the present invention, and the schematic embodiments and descriptions of the present invention are used to explain the present invention, and do not constitute an improper limitation of the present invention. In the attached picture:

Fig. 1 is a technical flow chart of the present invention;

2 is a schematic diagram of the ML (Machine Learning: Machine Learning) framework of the prediction model of the present invention;

Figure 3 is a comparison between the predicted output curves of different deep learning models and the real output curves;

Fig. 4 is the comparison chart of experimental result of the present invention and optimized result radar;

Figure 5 is a time comparison diagram of DOE calculated by GRU and FE models with different sample sizes;

Fig. 6 is an experimental scene diagram of the present invention;

Fig. 7 is the contrast diagram of experiment and simulation of the present invention: (a) force-displacement curve; (b) energy-displacement curve;

Fig. 8 is a comparison diagram of the experiment and simulation deformation sequences of the present invention.

detailed description

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in various ways defined and covered by the claims.

The flow process of the present invention is as shown in Figure 1: at first adopt the mode that experiment and finite element simulation combine, verify the accuracy of finite element model, then utilize the experimental design (DOE) based on finite element model, change square cone type energy-absorbing structure Structural parameters to generate training sets and test sets for machine learning; introduce four types of multi-layer perceptron (MLP), recurrent neural network (RNN), long short-term memory method (LSTM) and gated recurrent neural network (GRU) The machine model predicts the energy absorption characteristics of the square cone energy-absorbing structure, and compares the prediction accuracy of the four models through different accuracy indicators. The comparison of the machine learning models is shown in Figure 2; finally, the Gated Recurrent Neural Network (GRU) As the most suitable model for predicting the energy-absorbing characteristics of square-cone energy-absorbing structures; finally, the gated recurrent neural network (GRU) is used as a proxy model, and the adaptive global response surface (GRSM) method is used for the crashworthiness of square-cone energy-absorbing structures Perform multi-objective optimization to obtain the optimal solution.

Specifically, the present invention proposes a multi-objective optimization method for the crashworthiness of square cone energy-absorbing structures based on machine learning, including the following steps:

Step 1, establishing a finite element simulation model of the square cone energy-absorbing structure of the subway train. And based on the same boundary conditions of the finite element simulation, the dynamic impact experiment of the full-scale square cone energy-absorbing structure was carried out, and the combination of experiment and simulation was adopted to compare the force-displacement curve and displacement-energy curve between the experiment and the simulation. and deformation sequence modes to verify the accuracy of the established finite element model.

Step 2. Taking the outer wall thickness T _A , outer wall thickness T _B , partition thickness T _gb , aluminum honeycomb A strength δ _A and aluminum honeycomb B strength δ _B of the square cone energy-absorbing structure as design variables, the square cone energy-absorbing structure The force-displacement output curve of the energy structure is used as the output, and the Latin hypercube method is used for sampling, and the virtual design of experiments (DOE) is carried out. The number of DOE is set to 1000 groups.

Step 3, use the five input variables of DOE as the input features of network training in machine learning, and use the output of DOE as the actual output of network training in machine learning; then, use the train_test_split() function to compare the input and actual network training The output is randomly divided into a training set and a test set, in which 75% of the data is used as a training set and 25% is used as a test set; finally, the input data and the actual output data are normalized using Pytorch's MinMaxscale() function, The normalized calculation equation is as follows:

是第d个样本的第i个输入特征的正则化值；X_d,i第d个样本的第i个输入特征的值；

是第d个样本的第i个输入特征对应输出的正则化值；Y_d,i是第d个样本的第i个输入特征对应的真实输出；X和Y表示输入和输出样本；min(X)和min(Y)分别为样本输入和真实输出的最小值；max(X)和max(Y)分别为样本输入和真实输出的最大值。Among them, i in the formula is the i-th input feature, i=1,...,5; d represents the number of samples, d=1,...,1000; t represents the dimension of the output result, t=1,...,121;

Is the regularization value of the i-th input feature of the d-th sample; X _d, the value of the i-th input feature of the d-th sample;

is the regularization value corresponding to the output of the i-th input feature of the d-th sample; Y _d,i is the real output corresponding to the i-th input feature of the d-th sample; X and Y represent the input and output samples; min(X ) and min(Y) are the minimum values of sample input and real output, respectively; max(X) and max(Y) are the maximum values of sample input and real output, respectively.

Step 4. First, map the input geometry of the energy-absorbing structure to the ML architecture of the required output response, use the normalized training set and test set as samples, and divide the training samples and test samples into batches , the number of samples in each batch is 10; then, using the mean square error (MSE) as the calculation method of the loss function MSE of the optimization process, as shown in formula 3, using the Adam algorithm as the deep learning optimizer, the mean absolute error (MAE), Root Mean Square Error (RMSE), Regression Coefficient (R ² ) and Average Relative Absolute Error (RAAE) are used as evaluation indicators, respectively using multi-layer perceptron (MLP), recurrent neural network (RNN), long and short time The memory method (LSTM) and the gated recurrent neural network (GRU) four neural network models predict the energy-absorbing characteristic curve (force-displacement curve) of the energy-absorbing structure respectively, and compare and analyze the prediction accuracy of different models, as shown in Fig. 2 is a schematic diagram of the ML framework of the prediction model.

It can be seen from Figure 3 and Figure 4 that the GRU predicted output curve has a high degree of fitting with the real output curve, and its loss function is also maintained at a small value, indicating that the GRU model predicts the output curve with high accuracy. The evaluation indicators of the four models during the training period, when the R ² is larger, and the MAE, RMSE and RAAE are smaller, it means that the trained model is more accurate, and the GRU is more stable than the other three models, and the MAE, RMSE and RAAE have been maintained At smaller values, R ² has been maintained at larger values, indicating that GRU missing is more accurate than the other three models.

为机器学习对应神经网络的预测输出，y_i为样本的实际输出。Among them, n is the number of samples,

is the predicted output of the neural network corresponding to machine learning, and _yi is the actual output of the sample.

Step 5. By comparing and analyzing the prediction effects of four machine learning network models: MLP, RNN, LSTM and GRU, GRU is used as the prediction model of the energy absorption characteristic curve of the square cone energy-absorbing structure of the subway train, and the outer wall thickness T _A , outer wall thickness T _B , separator thickness T _gb , strength δ _{A of aluminum honeycomb A} and strength δ _B of aluminum honeycomb B are used as optimization variables, and the energy absorption (EA) and peak force (PCF) of the square cone energy-absorbing structure are used as optimization targets. The established optimization theoretical model is shown in Equation 9; then, Hammersley sampling method (Hammersley) is used to resample the optimization, and the GRU network model is used to calculate the corresponding energy absorption (EA) and peak force (PCF), generating The new DOE; finally, the global response surface method (GRSM) is used for multi-objective optimization, and the parameters of the GRSM optimization algorithm are shown in Table 1.

PCF=max(F(s)) (8)

Among them, s is the compression displacement of the square cone energy-absorbing structure; F(s) is the axial force of the square cone energy-absorbing structure; T _A and T _B are the outer wall thicknesses of the square cone energy-absorbing structure; T _gb is the square The diaphragm thickness of the conical energy-absorbing structure; δ _A and δ _B are the strengths of aluminum honeycomb A and aluminum honeycomb B, respectively.

Among them, EA is the energy absorption of the square cone energy-absorbing structure; PCF is the peak force of the square cone energy-absorbing structure;

Table 1 Parameters of GRSM algorithm

Step 6. Based on the optimization results of GRSM, the Pareto solution set of the optimization target is obtained, and the minimum distance method (as shown in the following formula 9) is used to make an optimal decision on the optimized Pareto solution set to obtain the optimal solution . The energy absorption (EA), specific energy absorption (SEA), mean force (F _mean ) and energy absorption efficiency (IFE) were used as energy absorption evaluation indicators, and the optimal solution was compared with the experimental results using radar charts (as shown in Figure 4 ), to verify the feasibility of the optimization method

Among them, D is the distance between the knee joint point and the Pareto solution set point; m is the number of optimization objectives, and f _i ^k is the optimal point k of the i-th optimization objective.

Among them, m is the mass of the square cone energy-absorbing structure; s is the compression displacement of the square cone energy-absorbing structure; F _mean is the average force of the square cone energy-absorbing structure, and PCF is the peak force of the square cone energy-absorbing structure; IFE is the energy absorption efficiency of the square cone energy-absorbing structure.

In order to further illustrate the efficiency of the optimization process, it is necessary to compare the calculation time of the optimization process using the GRU model and the finite element model, respectively. In the process of model optimization, the training time of GRU is 100.5 hours. When the model training is completed, the single calculation time is 0.5 seconds, and the single calculation time of the finite element model is 1 hour. The GRU and FE models are required to calculate the DOE without the number of samples. The time comparison chart is shown in Figure 5. The calculation time of the optimization process of the GRU and FE models is shown in Equations 14-16.

T _ML =t _train +n×t _ml (14)

Among them, T _ML is the total calculation time required for calculating the DOE of the machine learning GRU model, t _train is the time required for GRU training, t _ml is the time required for each calculation of a sample after the model is successfully trained, and n is the DOE The number of samples

T _FE =n×t _FE (15)

Among them, T _FE is the total calculation time required for calculating the DOE of the finite element model, t _FE is the time required for each calculation of a sample of the finite element model, and n is the number of samples of DOE.

Among them, T _ML is the total calculation time required for calculating the DOE of the machine learning GRU model; T _FE is the total calculation time required for calculating the DOE of the finite element model; TIEF is the time required for calculating the DOE of the machine learning GRU model and the finite element model required time ratio.

It can be seen from Figure 5 that when n=500, T _ML and T _FE are 600.57 and 500 respectively, TIFE is 120.12%, and the time required by GRU is greater than the calculation time of the finite element model. When n=5000, T _ML and T _FE are 601.19 and 5000 respectively, TIFE is 5.58%, and the time required by GRU is less than the calculation time of finite element model. When n is small, the calculation time of GRU will be longer than that of FE. As the number of n increases, the calculation time of GRU will basically remain unchanged, while the calculation time of the finite element model will increase significantly, and TIFE will show an exponential decline. Therefore, it can be found that when the number of samples sampled in the optimization process is large, the calculation time can be greatly reduced by using the machine learning GRU model. The machine learning GRU model can enable the optimization process to achieve big data empowerment, and the square cone energy-absorbing structure model Optimization brings into a new realm.

The present invention is explained and illustrated below in conjunction with specific embodiments.

In order to observe the energy-absorbing characteristics and behavior mechanism of the end energy-absorbing structure during a high-speed collision, a full-scale impact test was carried out on the end energy-absorbing structure on a standard track. As shown in Fig. 6, the whole experimental system is mainly composed of a launching device providing a certain speed, a trolley, a force uniform plate for impact, a dynamic force sensor installed between the rigid wall and the force uniform plate, a recording A velocimeter for the incident velocity of the structure and a high-speed camera for capturing the impact process. The end energy-absorbing structure is fixed on the front end of the impact trolley. Drag the trolley to the far end of the impact point, and drive the trolley to hit the sample at an initial velocity of 17.9km/h through the motor drive device. The total weight of the trolley is 16.1t.

In order to verify the crashworthiness of the square cone energy-absorbing structure and verify the correctness of the finite element modeling method, the impact test was carried out using the impact test bench system. Meanwhile, the numerical and theoretical results can be further verified by experimental results. Fig. 7(a) shows the comparison of experimental and numerical results in terms of force-displacement curves. The results show that, in both experiments and simulations, the square cone energy-absorbing structure forms an initial peak force after contacting the rigid wall, and then drops rapidly. Considering that the experimental conditions are more complex than the simulated conditions, the force-displacement curves cannot be completely consistent. Nevertheless, it can be found that the number of initial peak forces of the two is the same, both form 12 force peaks, and the amplitudes of the initial peak forces are basically the same. As shown in Fig. 7(b), the comparison of the experimental and simulation results in terms of energy-displacement curves. The changing trend of the energy-displacement curves of the experiment and the simulation is consistent throughout the process. In addition, for better comparison, the values of crashworthiness index and error index are summarized in . For EA, IPCF, d and MCF, the experimental and numerical results are 220.38 kJ and 216.86 kJ, 549.69 kN and 526.67 kN, 702.09 mm and 711.92 mm, 313.89 kN and 304.61 kN, respectively. The data show that the simulated results are in good agreement with the experimental values.

Furthermore, the theoretical dynamic MCF of the end energy-absorbing structure predicted by the proposed theoretical model is 296.22 kN. The theoretically predicted dynamic steady-state force is in good agreement with the impact test, with a relative error of 5.63%, which meets the requirements of engineering calculation accuracy. Therefore, the proposed theoretical model and the constructed finite element model have sufficient precision to study the energy absorption characteristics of end energy-absorbing structures.

Table 2 Comparison of experimental and simulation results

Accurately predicting the shock response is not the only criterion, effective prediction of deformation modes is also required. The experimental and simulated deformation process of the square cone energy-absorbing structure is shown in Fig. 8. It can be seen that the deformation mode of the square cone energy-absorbing structure in the simulation is basically consistent with the experimental results during the collision. Whether it is an experiment or a simulation, from the collision end to the rear end, the deformation is stable and orderly, and finally forms a regular shape. Due to the presence of the septum, 12 folds occurred during both simulated and experimental deformations.

In summary, the force-displacement, energy-displacement and structural deformation modes in the numerical simulation are in good agreement with the impact experiments, indicating that the simulation model has good accuracy and can be used for subsequent research.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (6)

1. The multi-objective optimization method for the crashworthiness of the square cone type energy-absorbing structure based on machine learning is characterized in that, comprising the following steps: Establish the finite element simulation model of the square cone energy-absorbing structure of the subway train; Based on the method of combining the established finite element simulation model with experimental design, the structural parameters and energy-absorbing characteristic curves of the energy-absorbing structure of the subway train were extracted; according to the sampling method using the Latin superstructure, a virtual experimental design was carried out to determine the square-cone absorbing structure of the subway train. The prediction model of the optimal energy absorption characteristic curve of the energy structure, as well as the optimization variables and optimization objectives; According to the optimal energy absorption characteristic curve prediction model, optimization variables and optimization objectives, an optimization theoretical model is established; Using the Hamersley sampling method to resample the optimized theoretical model, use the optimal energy absorption characteristic curve prediction model to calculate the corresponding energy absorption and peak force, and generate a new DOE; use the global response surface method for multi-objective optimization, and get the optimization result; Based on the optimization results, the Pareto solution set of the optimization target is obtained, and the optimal decision is made on the optimized Pareto solution set by using the minimum distance method to obtain the optimal solution. 2. The multi-objective optimization method for the crashworthiness of square cone energy-absorbing structures based on machine learning according to claim 1, characterized in that, based on the same boundary conditions of finite element simulation, a full-scale square cone energy-absorbing structure is carried out The dynamic impact experiment adopts the combination of experiment and simulation, and verifies the accuracy of the established finite element model by comparing the force-displacement curve, displacement-energy curve and deformation sequence mode between the experiment and simulation. 3. the multi-objective optimization method for the crashworthiness of square cone type energy-absorbing structures based on machine learning according to claim 1, is characterized in that, based on the method that the finite element simulation model of establishment combines with experimental design, extract the subway train absorption Structural parameters and energy-absorbing characteristic curves of the energy-absorbing structure; according to the use of Latin super-cubic sampling, the virtual experiment design was carried out to determine the optimal energy-absorbing characteristic curve prediction model of the square-cone energy-absorbing structure of the subway train, as well as the optimization variables and optimization objectives. Specifically: Use the five input variables of DOE as the input features of network training in machine learning, and use the output of DOE as the actual output of network training in machine learning; Randomly divide the input and actual output of network training into training set and test set; Normalize the input data and the actual output data; Map the input geometry of the energy-absorbing structure to the required output response machine learning (ML) architecture, use the normalized training set and test set as samples, and use MLP, RNN, LSTM and GRU respectively The network model of machine learning predicts the energy-absorbing characteristic curve of the energy-absorbing structure, and compares and analyzes the prediction accuracy of different models. The network model with the highest accuracy of the predicted output curve is used as the optimal model for the square-cone energy-absorbing structure of the subway train. Energy absorption characteristic curve prediction model. 4. the method for multi-objective optimization of the crashworthiness of the square cone type energy-absorbing structure based on machine learning according to claim 1, wherein the optimal energy-absorbing characteristic curve prediction model of the square cone type energy-absorbing structure of the subway train is a door control loop neural network model, the outer wall thickness T _A , outer wall thickness T _B , partition thickness T _gb , strength δ _{A of aluminum honeycomb A} and strength δ _B of aluminum honeycomb B are used as optimization variables, and the square cone energy-absorbing structure Energy and peak force were the optimization goals. 5. The multi-objective optimization method for the crashworthiness of square cone type energy-absorbing structures based on machine learning according to claim 1, wherein the optimized theoretical model established is as follows:

PCF=max(F(s)) Among them, S is the compression displacement of the square cone energy-absorbing structure; F(s) is the axial force of the square cone energy-absorbing structure; T _A and T _B are the outer wall thicknesses of the square cone energy-absorbing structure; T _gb is the square The diaphragm thickness of the cone energy-absorbing structure; δ _A and δ _B are the strengths of aluminum honeycomb A and aluminum honeycomb B respectively; EA is the energy absorption of the square cone energy-absorbing structure; PCF is the peak force of the square cone energy-absorbing structure . 6. The multi-objective optimization method for the crashworthiness of square cone energy-absorbing structures based on machine learning according to claim 1, characterized in that, based on the optimization results of GRSM, the Pareto solution set of the optimization target is obtained, and the minimum The distance method is used to make an optimal decision on the optimized Pareto solution set to obtain the optimal solution, specifically: The energy absorption EA, specific energy absorption SEA, average force F _mean , and energy absorption efficiency IFE are used as energy absorption evaluation indicators, and the optimal solution is compared with the experimental results using radar charts to verify the feasibility of the optimization method:

Among them, D is the distance between the knee joint point and the Pareto solution set point; m is the number of optimization objectives, and f _i ^k is the optimal point k of the i-th optimization objective;

Among them, m is the mass of the square cone energy-absorbing structure; s is the compression displacement of the square cone energy-absorbing structure; F _mean is the average force of the square cone energy-absorbing structure, and PCF is the peak force of the square cone energy-absorbing structure; IFE is the energy absorption efficiency of the square cone energy-absorbing structure.

Images