A vehicle collision simulation optimization method based on machine learning

Abstract

The present invention discloses a vehicle collision simulation optimization method based on machine learning, which comprises the following steps: S1: establishing a whole vehicle collision simulation analysis model; S2: establishing a parameterized file of design variables and building a corresponding data acquisition workflow; S3: obtaining a calculation result file and sampling data; S4: obtaining collision sampling data containing collision deformation labels; S5: determining the deformation labels, sample points and value ranges of design variables involved in machine learning model training; S6: training the machine learning model; S7: judging whether the prediction accuracy of the trained machine learning model meets the set threshold and judging whether the number of sample points meets the requirements, and only proceeding to step S8 when both are satisfied; otherwise, returning to step S3; S8: building a corresponding numerical optimization workflow; S9: using an optimization algorithm to perform numerical optimization; S10: if the numerical optimization result does not meet the set numerical optimization accuracy, increasing the number of sample points and returning to step S3, and if it is satisfied, ending the step.

Description

Technical Field

[0001] The invention relates to a simulation method, in particular to a vehicle collision simulation optimization method. Background Art

[0002] In recent years, in order to improve the collision safety performance of the whole vehicle and save vehicle development costs, domestic and foreign OEMs have generally used Computer Aided Engineering (CAE) to perform collision simulation verification and optimization of the body structure in the early stages of vehicle development.

[0003] During the development of vehicle body structure collision safety, if the vehicle collision simulation calculation results show that the collision safety indicators are not met, the development engineer needs to analyze the collision simulation calculation results first, and then design optimization measures based on engineering experience (such as changing the sheet metal thickness of the vehicle body structure, modifying the shape of the vehicle body structure, and replacing the material model of the vehicle body structure, etc.) and modify the structure of the collision simulation model, recalculate and obtain new calculation results and evaluate the calculation results. Ultimately, it is necessary to obtain the optimization solution through continuous manual iteration and trial and error. This manual trial and error process often requires calculating dozens or even hundreds of solutions, which is time-consuming and labor-intensive, and extremely dependent on the engineering experience of engineers.

[0004] With the rapid development of computer technology, numerical optimization methods have emerged. At present, various mature commercial optimization software have appeared on the market, which can assist development engineers in completing the automatic numerical optimization of various simulation models. After the development engineer sets the design variables, the commercial optimization software can automatically call the simulation solver and use the various optimization algorithms built into the commercial optimization software for automatic iterative optimization, thereby obtaining the best optimization solution.

[0005] However, although the optimization algorithms provided by most commercial optimization software on the market can achieve good results in the numerical optimization of linear simulation conditions (such as body structure stiffness simulation conditions) and can quickly obtain global optimization solutions, these optimization algorithms still have great difficulties in the numerical optimization process of highly nonlinear simulation conditions (such as whole vehicle collision simulation conditions).

[0006] Traditional numerical optimization methods generally first sample through experimental design, establish an optimization proxy model using the response surface method, and finally perform numerical optimization on the proxy model to obtain an optimization solution. For users, this response surface optimization solution has poor adjustability for the optimization model and low predictive ability for unknown solutions. It can still obtain good results in linear simulation conditions, but it is difficult to obtain the correct optimization solution in highly nonlinear simulation conditions.

[0007] Taking the whole vehicle collision simulation condition as an example, in the iterative optimization process of collision safety performance, the change of the body structure state may cause the collision performance to reach a critical state. Among them, slight changes in design variables may lead to significant changes in collision results. For example: when the thickness of the collision key parts is reduced to a certain extent, it will cause the sheet metal to tear or even break completely, resulting in a rapid increase in the intrusion amount of the measurement point, thereby affecting the prediction accuracy of the response surface model.

[0008] In addition, in the current whole vehicle collision simulation, the currently commonly used quantitative measurement method is to evaluate the body structure deformation by measuring the maximum deformation at several locations on the collision key parts. This method is too simple and not comprehensive enough. On the one hand, this method cannot fully evaluate the body structure deformation of all collision key parts, and on the other hand, it cannot show the body structure tearing.

[0009] Based on this, in order to correctly evaluate the body deformation in the whole vehicle collision simulation analysis, the present invention designs a vehicle collision simulation optimization method based on machine learning. Summary of the invention

[0010] The purpose of the present invention is to provide a vehicle collision simulation optimization method based on machine learning. This method aims to solve the problems of insufficient accuracy of current collision simulation numerical optimization methods and incomplete vehicle body structure deformation evaluation methods, and builds a new vehicle collision simulation optimization process, which can quickly and accurately narrow the optimization space and finally quickly obtain a reasonable and effective optimization solution.

[0011] This vehicle collision simulation optimization method is rigorous and reasonable. It avoids many shortcomings and defects of existing traditional collision simulation optimization methods. It can achieve good application results in practical applications and is generally applicable to performance optimization of various collision simulation analysis conditions. It also has strong guiding significance and reference role for performance optimization of other nonlinear simulation conditions.

[0012] In order to achieve the above object, the present invention proposes a vehicle collision simulation optimization method based on machine learning, which comprises the steps of:

[0013] S1: Establish a vehicle collision simulation analysis model according to the vehicle collision test conditions;

[0014] S2: Based on the vehicle collision simulation analysis model, a parameterized file of design variables is established, and a corresponding data acquisition workflow is built;

[0015] S3: Based on the data acquisition workflow, determine the optimization goal, constraints, design variables and value ranges, experimental design sampling method and sampling quantity, and generate and run a sample simulation model to obtain calculation result files and sampling data;

[0016] S4: Based on the calculation result file and the sampling data, a collision deformation label is generated and collision sampling data containing the collision deformation label is obtained;

[0017] S5: Based on the collision sampling data, determine the deformation labels, sample points and value ranges of their design variables involved in the machine learning model training;

[0018] S6: performing machine learning model training, wherein the input parameters of the machine learning model are the design variables of the collected sample points, and the output parameters are the extracted collision feature data;

[0019] S7: Determine whether the prediction accuracy of the trained machine learning model meets the set threshold and whether the number of sample points meets the requirement. If both are met, proceed to the following step S8; if the prediction accuracy of the machine learning model does not meet the set threshold, return to step S3; if the number of sample points does not meet the requirement, increase the number of sample points and return to step S3;

[0020] S8: Building a corresponding numerical optimization workflow based on the parameterized file of the design variables established in step S2 and the machine learning model trained in step S6;

[0021] S9: defining optimization objectives, design variables and value ranges, and constraints, and performing numerical optimization using an optimization algorithm based on the numerical optimization workflow;

[0022] S10: If the numerical optimization result does not meet the set numerical optimization accuracy, increase the number of sample points and return to step S3. If it meets the requirement, end the step and obtain the final optimization solution.

[0023] In the above technical solution of the present invention, the inventors have adopted creative ideas to design a new vehicle collision simulation optimization method, which can solve the problems of insufficient accuracy of the current existing collision simulation numerical optimization methods and incomplete vehicle body structure deformation evaluation methods.

[0024] The process of this vehicle collision simulation optimization method is rigorous and reasonable. It can convert the deformation of the vehicle body structure into a set of deformation label lists to become structured data that can be recognized by the machine learning algorithm, and perform iterative calculations through machine learning automation and numerical optimization methods, thereby optimizing the performance of the whole vehicle collision simulation condition, so as to quickly and accurately narrow the optimization space and quickly obtain the optimal solution within the design space.

[0025] In practical applications, the vehicle collision simulation optimization method can achieve very good application results. It is generally applicable to the performance optimization of various collision simulation analysis conditions. It also has strong guiding significance and reference role for the performance optimization of other nonlinear simulation conditions, and has good promotion and application value.

[0026] Furthermore, in the vehicle collision simulation optimization method based on machine learning described in the present invention, in step S2, the data acquisition workflow is set to: be able to automatically change the model parameters of the vehicle body structure simulation model to generate a new vehicle collision simulation analysis model.

[0027] In step S2 described in the present invention, the data acquisition workflow can automatically change the model parameters of the vehicle body structure simulation model, such as part thickness, part shape, part material number, etc., to generate a new vehicle collision simulation analysis model, automatically submit it to the server and use the collision simulation solver to perform solution calculations, automatically download the calculation result file after the calculation is completed, and automatically call the CAE post-processing software to extract the collision simulation result data.

[0028] Furthermore, in the vehicle collision simulation optimization method based on machine learning described in the present invention, in step S2, the design variables include the thickness parameters and variation ranges of the B-pillar reinforcement plate, the B-pillar reinforcement plate patch, the front top beam, the front top beam patch, the front outer sill beam, and the front inner sill beam.

[0029] Furthermore, in the vehicle collision simulation optimization method based on machine learning described in the present invention, in step S4, a number of collision components are divided into multiple blocks, and a deformation label is assigned to the deformation result of each block to obtain a set of deformation label lists, and the deformation of the vehicle body structure is defined by the deformation label list.

[0030] Furthermore, in the vehicle collision simulation optimization method based on machine learning described in the present invention, in step S6, when the design variables include parameters of different types, the input parameters are normalized.

[0031] In the above-mentioned step S6 of the present invention, if the design variables contain parameters of different types, the input parameters need to be normalized to eliminate the influence of dimension, and machine learning training is performed by calling an open source machine learning toolkit. After the training is completed, the machine learning training model with the highest accuracy is saved as a machine learning result file for calling in subsequent numerical optimization steps.

[0032] Furthermore, in the vehicle collision simulation optimization method based on machine learning of the present invention, in step S7, the determination coefficient R is used. 2 To evaluate the prediction accuracy of the machine learning model, the calculation formula is:

[0033]

[0034] Where: y i is the actual output value of the i-th test sample, is the mean of the actual output values ​​of all test samples, is the predicted value output for the i-th test sample, is the mean of the predicted values ​​output by all test samples, N is the number of test samples, R 2 The value of is between 0 and 1. The closer its value is to 1, the higher the prediction accuracy of the machine learning model.

[0035] Furthermore, in the vehicle collision simulation optimization method based on machine learning described in the present invention, in step S7, when the number of sample points is 6-8 times the number of design variables, it is considered that the number of sample points meets the requirement.

[0036] Furthermore, in the vehicle collision simulation optimization method based on machine learning described in the present invention, in step S9, the preferred algorithm adopts adaptive evolution, simulated annealing or efficient global optimization.

[0037] Furthermore, in the vehicle collision simulation optimization method based on machine learning described in the present invention, in step S10, when the deviation between the numerical optimization result and the calculation result of the collision simulation software is less than 5%, it is considered that the numerical optimization accuracy meets the requirements.

[0038] Compared with the prior art, the vehicle collision simulation optimization method based on machine learning described in the present invention has the following advantages and beneficial effects:

[0039] The inventor has adopted a creative concept to design a new vehicle collision simulation optimization method, which can solve the problems of insufficient accuracy of the current collision simulation numerical optimization method and incompleteness of the vehicle body structure deformation evaluation method.

[0040] The process of the vehicle collision simulation optimization method described in the present invention is rigorous and reasonable. It can convert the deformation of the vehicle body structure into a set of deformation label lists to become structured data that can be recognized by the machine learning algorithm, and perform iterative calculations through machine learning automation and numerical optimization methods, thereby optimizing the performance of the whole vehicle collision simulation condition, so as to quickly and accurately narrow the optimization space, thereby quickly obtaining the optimal solution within the design space.

[0041] In practical application, the vehicle collision simulation optimization method designed by the present invention can obtain very good application effects, and has wide applicability and is generally applicable to various collision simulation analysis conditions performance optimization, including but not limited to frontal collision, frontal pole collision, side collision, side pole collision, rear collision, roof extrusion and door intrusion collision conditions,

[0042] Accordingly, the vehicle collision simulation optimization method described in the present invention also has strong guiding significance and reference role for the performance optimization of other nonlinear simulation conditions, and has good promotion and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The present invention is a flowchart of the steps of the vehicle collision simulation optimization method in one embodiment.

[0044] Figure 2 A schematic diagram of a vehicle side collision simulation working condition is shown schematically.

[0045] Figure 3 for Figure 2 The schematic diagram of design variables for the vehicle side collision simulation condition is shown.

[0046] Figure 4 for Figure 2 The schematic diagram of the structure of the key parts of the automobile side collision simulation is shown.

[0047] Figure 5 for Figure 4 Schematic diagram of the structure of the B-pillar reinforcement plate shown.

[0048] Figure 6 for Figure 4 The structural schematic diagram of the side roof beam assembly is shown.

[0049] Figure 7 for Figure 4 Schematic diagram of the structure of the outer door sill beam assembly shown.

[0050] Figure 8 The diagram schematically shows the deformation label classification of the sampled data obtained by the first round of data collection in one implementation manner of the vehicle collision simulation optimization method of the present invention.

[0051] Fig. 9 The figure schematically shows the value range after filtering the value range of the first round of data sampling.

[0052] Fig.10 The figure schematically shows the value range of the second round of data sampling.

[0053] Fig.11 The diagram schematically shows the deformation label classification of the sampled data obtained by the second round of data collection in one implementation of the vehicle collision simulation optimization method of the present invention. DETAILED DESCRIPTION

[0054] The vehicle collision simulation optimization method based on machine learning described in the present invention will be further explained and illustrated below in conjunction with the accompanying drawings and specific embodiments of the specification. However, such explanation and illustration do not constitute an improper limitation on the technical solution of the present invention.

[0055] In order to explain the vehicle collision simulation optimization method of the present invention in detail, in this embodiment, a side collision situation is taken as an example to explain the steps of the optimization method in detail.

[0056] Figure 1 The present invention is a flowchart of the steps of the vehicle collision simulation optimization method in one embodiment.

[0057] like Figure 1 As shown, in this embodiment, the vehicle collision simulation optimization method designed by the present invention specifically includes the following steps S1-S10:

[0058] S1: Establish a vehicle collision simulation analysis model based on the vehicle collision test conditions.

[0059] In step S1, the established vehicle collision simulation analysis model can correctly reflect the collision result performance of the corresponding working condition of the vehicle model. If the corresponding collision working condition has corresponding vehicle collision test data, it is necessary to ensure during the design that the error between the result data calculated by the established collision simulation model and the collision test data is within the allowable range.

[0060] It should be noted that, in this embodiment, the vehicle is described by taking the side collision as an example, and the side collision simulation diagram can be seen in FIG. Figure 2 . Figure 2 A schematic diagram of a vehicle side collision simulation working condition is shown schematically.

[0061] In this implementation, a vehicle side collision simulation model of a certain vehicle type can be established in the CAE pre-processing software according to the standard collision modeling process within the enterprise (such as Figure 2 The mobile deformable barrier is controlled to impact the test vehicle vertically from the side at a speed of 50 km / h.

[0062] The established collision simulation analysis model can be solved and calculated using a collision simulation solver, and a calculation result file can be obtained. The collision result data can be obtained through CAE post-processing software, including at least the body structure deformation and collision performance curve.

[0063] S2: Based on the constructed vehicle collision simulation analysis model, establish the parametric file of the design variables and build the corresponding data collection workflow.

[0064] In step S2, the constructed data acquisition workflow can automatically change the model parameters of the vehicle body structure simulation model, such as part thickness, part shape, part material number, etc., generate a new vehicle collision model, automatically submit it to the server and use the collision simulation solver to solve the calculation, automatically download the calculation result file after the calculation is completed, and automatically call the CAE post-processing software to extract the collision simulation result data.

[0065] Accordingly, in this embodiment, based on the constructed vehicle collision simulation analysis model, a design variable parameterization file can be first established, and the design variables are defined as the thickness parameters and variation ranges of the six key side collision components of the body in white (respectively, the B-pillar reinforcement plate DV01, the B-pillar reinforcement plate DV02, the front side top beam DV03, the front side top beam plate DV04, the front outer sill beam DV05, and the front inner sill beam DV06), such as Figure 3 shown. Figure 3 for Figure 2 The schematic diagram of design variables for the vehicle side collision simulation condition is shown.

[0066] In this embodiment, the initial values ​​and variation ranges of the designed six design variables are shown in Table 1. Table 1 lists the initial values ​​and variation ranges of the six design variables. Then, the design variable parameterization file is called in the CAE pre-processing software and the corresponding body-in-white structure simulation model is generated.

[0067] Table 1.

[0068]

[0069] After completing the parameterization file setting of the design variables, the commercial numerical optimization software known in the prior art can be used to integrate CAE pre-processing software, collision solver and CAE post-processing software, and establish a data collection workflow. After selecting the experimental design sampling method and the sample size, the data collection workflow can automatically generate all experimental design sample points and perform solution calculations to obtain collision simulation result files and collision performance data.

[0070] S3: Based on the data collection workflow, determine the optimization objectives, constraints, design variables and value ranges, experimental design sampling method and sampling quantity, and generate and run the sample simulation model to obtain the calculation result files and sampling data.

[0071] It should be noted that, in this embodiment, the sampling method for the first round of data collection is Latin Hypercube Sampling, which uses a uniform sampling method to sample the design variables, with a sampling number of 50 times. The collision performance data includes the total weight change (Parts_mass) of the left and right symmetrical parts corresponding to the six design variables and the survival space (BS_H_punkt) from the B-pillar to the H point of the dummy. Parts_mass and BS_H_punkt are both continuous parameters, which will be used as regression output parameters in the subsequent machine learning training.

[0072] After completing the data sampling, the obtained sampling data set is shown in Table 2, where the sample point with sequence number 0 is the sampling data of the initial state scheme.

[0073] Table 2 lists the sample data sets obtained in the first round of data collection.

[0074] Table 2.

[0075]

[0076] In this embodiment, the optimization goal is to make the weight of the parts involved in the optimization as light as possible while ensuring that the side collision performance remains unchanged, and the constraints include: Parts_mass≤38.5Kg, BS_H_punkt≥168mm.

[0077] S4: Based on the calculation result file and the sampling data, a side collision deformation label is generated and complete sampling data is obtained. The complete sampling data is collision sampling data containing the collision deformation label.

[0078] If the body structure is not designed properly, the body structure (such as the B-pillar reinforcement plate, the door sill beam and the side roof beam, etc.) is prone to cracking or tearing during a side collision. In the technical solution designed by the present invention, the deformation of the body structure in a side collision is evaluated by defining a deformation label. The specific method is to divide multiple collision key components into multiple blocks, assign a deformation label according to the deformation result of each block, and finally obtain a set of deformation label lists, and define the deformation of the body structure through the deformation label list.

[0079] Figure 4 for Figure 2 The schematic diagram of the structure of the key parts of the automobile side collision simulation is shown.

[0080] like Figure 4 As shown, in this embodiment, the body structure deformation can be evaluated by the deformation profiles of the B-pillar reinforcement plate, the side roof beam assembly (front side roof beam + rear side roof beam), and the outer rocker beam assembly (front outer rocker beam + rear outer rocker beam).

[0081] Figure 5 for Figure 4 Schematic diagram of the structure of the B-pillar reinforcement plate shown.

[0082] Figure 6 for Figure 4 The structural schematic diagram of the side roof beam assembly is shown.

[0083] Figure 7 for Figure 4 Schematic diagram of the structure of the outer door sill beam assembly shown.

[0084] like Figure 5 As shown, in this embodiment, the B-pillar reinforcement plate DV01 is divided into five blocks from top to bottom, and the distribution is represented by BS_01, BS_02, BS_03, BS_04 and BS05.

[0085] like Figure 6 As shown, in this embodiment, the side roof beam assembly is divided into five blocks from front to back, and the distribution is represented by DR_01, DR_02, DR_03, DR_04 and DR_05.

[0086] like Figure 7 As shown, in this embodiment, the outer sill beam assembly is divided into five blocks from front to back, and the distribution is represented by SW_01, SW_02, SW_03, SW_04 and SW_05.

[0087] In the present invention, according to the deformation condition of each block (intact, bent, partially torn and completely broken), a corresponding deformation label (0, 1, 2, 3) can be assigned. The deformation label is a discrete parameter, which will be used as a classification output parameter in the subsequent machine learning training.

[0088] It should be noted that, in this embodiment, by evaluating the collision calculation result file obtained by the first data collection, the deformation labels of the 15 blocks of each sample can be obtained. The deformation labels of the sampling data obtained by the first round of data collection are shown in Table 3, where the sample point with serial number 0 is the sampling data of the initial state scheme.

[0089] Table 3 schematically lists the deformation labels of the sampled data obtained in the first round of data collection.

[0090] Table 3.

[0091]

[0092]

[0093] Since the optimization goal of this implementation is to ensure that the side collision performance remains unchanged, according to actual engineering experience, the constraints of the deformation label can be added as follows: BS_01≤0, BS_02≤0, BS_03≤0, BS_04≤0, BS_05≤1, DR_01≤0, DR_02≤0, DR_03≤0, DR_04≤0, DR_05≤1, SW_01≤1, SW_02≤1, SW_03≤1, SW_04≤1, SW_05≤1.

[0094] S5: Perform data analysis on the collision sampling data obtained after processing in the above step S4 to determine the value ranges of deformation labels, sample points and design variables involved in machine learning model training.

[0095] In step S5, a plurality of collision components are divided into a plurality of blocks, and a deformation label is assigned to the deformation result of each block to obtain a set of deformation label lists, and the deformation of the vehicle body structure is defined by the deformation label lists.

[0096] By classifying the deformation label values ​​of the deformation blocks and judging the constraint conditions, the deformation labels that need to participate in machine learning training can be determined. By filtering the constraint conditions, sample points that meet some of the constraint conditions can be obtained, and the sample points and value ranges of the design variables that ultimately participate in machine learning training can be determined.

[0097] In this embodiment, the deformation label classification obtained by analyzing the sampling data obtained from the first round of data collection is as follows: Figure 8 As shown, Figure 8 The diagram schematically shows the deformation label classification of the sampled data obtained by the first round of data collection in one implementation manner of the vehicle collision simulation optimization method of the present invention.

[0098] from Figure 8 It can be seen that in this implementation, BS_02, BS_03, DR_05, and SW_05 each have only one deformation label value, and the deformation labels of these blocks do not need to participate in machine learning training, and all deformation label values ​​of SW_01, SW_02, and SW_04 are less than the constraint conditions, so the deformation labels of these blocks do not need to participate in machine learning training. It can be seen that the deformation labels that finally participate in machine learning training are: BS_01, BS_04, BS_05, DR_01, DR_02, DR_03, DR_04, and SW_03.

[0099] Accordingly, after determining the deformation labels that need to participate in machine learning training, according to the actual sampling situation (taking into account the increase of reasonable sampling space), the filtering constraints are determined as Parts_mass≤40Kg, BS_H_punkt≥160mm. Finally, the sampling data with 12 sample points satisfying the conditions are obtained, as shown in Table 4. These sampling data will participate in the next step of machine learning training.

[0100] Table 4 lists the sampled data after filtering the sampled data obtained in the first round of data collection.

[0101] Table 4.

[0102]

[0103] According to Table 4, the value range of the design variables that meet the filtering constraint conditions can be further narrowed to obtain the filtered value range of the design variables shown in the following Table 5.

[0104] Table 5 lists the ranges of design variables after the first round of data collection and filtering.

[0105] Table 5.

[0106]

[0107] Fig. 9 The figure schematically shows the value range after filtering the value range of the first round of data sampling. Fig. 9 As shown, after filtering the value range of the first round of data sampling points, the value range of the design variables can be further narrowed.

[0108] To summarize, in step S5, by classifying the deformation label values ​​of the deformation blocks and judging the constraint conditions, the deformation labels that need to participate in the machine learning training can be determined. By filtering the constraint conditions, sample points that meet some of the constraint conditions can be obtained, and the sample points and value ranges of the design variables that ultimately participate in the machine learning training can be determined.

[0109] S6: Perform machine learning model training, wherein the input parameters of the machine learning model are the design variables of the collected sample points, and the output parameters are the extracted collision feature data.

[0110] In step S6 of the present invention, machine learning training is performed based on the collected data of the sample points determined in step S5 and the value range of the design variables, and the accuracy of the machine learning model after training is determined; wherein the input parameters are the design variables of the collected sample points, and the output parameters are the extracted collision feature data. It should be noted that if the design variables contain parameters of different types, the input parameters need to be normalized to eliminate the influence of the dimension.

[0111] In the present invention, machine learning training can be performed by calling an open source machine learning toolkit (such as Python's machine learning toolkit scikit-learn). Commonly used supervised machine learning algorithms include: linear regression (LinR, Linear Regression), logistic regression (LogR, Logistics Regression), decision tree (DT, DecisionTree), random forest (RF, Random Forest), naive Bayes (NB, Naive Bayes), support vector machine (SVM, Support Vector Machine) and gradient boosting decision tree (GBDT, Gradient Boosting Decision Tree), etc.

[0112] Then, the collected data is divided into training sample data and test sample data according to a certain ratio, and one or more machine learning algorithm models are selected for machine learning training, and the best training effect is obtained through parameter tuning. After the training is completed, the machine learning training model with the highest accuracy is saved as a machine learning result file for subsequent numerical optimization steps. Among them, the regression output parameters use the machine learning regression algorithm model, and the classification output parameters should use the machine learning classification algorithm model.

[0113] In this embodiment, the output parameters Parts_mass and BS_H_punkt adopt a machine learning regression algorithm model, and the deformation labels (BS_01, ..., BS_05, DR_01, ..., DR_05, SW_01, ..., SW_05) adopt a machine learning classification algorithm model.

[0114] Step S7: Determine whether the prediction accuracy of the trained machine learning model meets the set threshold and whether the number of sample points meets the requirements. If both are met, proceed to the following step S8; if the prediction accuracy of the machine learning model does not meet the set threshold, return to step S3. If the number of sample points does not meet the requirements, increase the number of sample points and return to step S3, and continue to complete steps S4, S5 and S6.

[0115] In step S7 of the present invention, the determination coefficient R 2 To evaluate the prediction accuracy of the machine learning model, the calculation formula is:

[0116]

[0117] Where: y i is the actual output value of the i-th test sample, is the mean of the actual output values ​​of all test samples, is the predicted value output for the i-th test sample, is the mean of the predicted values ​​output by all test samples, N is the number of test samples, R 2 The value of is between 0 and 1. The closer its value is to 1, the higher the model accuracy. 2 All are above 0.8 and most are greater than 0.9, indicating that the machine learning models corresponding to each output parameter meet the accuracy requirements.

[0118] In order to improve the credibility of the machine learning model as quickly as possible, in addition to the machine learning model meeting the accuracy requirements, there must be sufficient data samples, that is, the size of the data sample exceeds several times the number of design variables, preferably more than 6-8 times.

[0119] In the new round of data sampling, the value range of the design variables is defined according to the range of design variables after filtering in the previous round of data analysis, so that more sample points can be collected in a smaller design space. This measure is conducive to improving the prediction accuracy of the machine learning model.

[0120] In this embodiment designed by the present invention, since too few sample points of the filtered data are obtained in the first round of data sampling, there are only 12 sample points of the filtered data, and there are 6 design variables. Relatively speaking, the sample points are too few, so it is necessary to increase the second round of data collection. The range of the design variable value of the new round of data collection is defined according to the range of the design variable after the first round of data analysis and filtering, as shown in Table 4.

[0121] In the second round of data sampling, 50 sample points were resampled according to the new design variable value range. After that, data processing was performed to extract the deformation labels of these sample points. The sample data (50) sampled in the second round and the sample data (12) sampled and filtered before constitute the new round of sample data (62) participating in machine learning training, as shown in Fig.10 shown. Fig.10 The diagram schematically shows the range of values ​​for the second round of data sampling. The number of sample data in this group exceeds 10 times the number of design variables, so the sampling quantity requirement is met and subsequent data analysis can be performed.

[0122] Accordingly, the deformation label classification of the second round of sampling data is as follows Fig.11 As shown, Fig.11 The diagram schematically shows the deformation label classification of the sampled data obtained by the second round of data collection in one implementation of the vehicle collision simulation optimization method of the present invention.

[0123] Depend on Fig.11 It can be seen that BS_02, BS_03, BS_04, BS_05, DR_05, and SW_05 all have only one deformation label value, and the deformation labels of these blocks do not need to participate in machine learning training. All deformation label values ​​of SW_01, SW_02, and SW_04 are less than the constraints, so the deformation labels of these blocks do not need to participate in machine learning training. Finally, the deformation labels participating in machine learning training are BS_01, DR_01, DR_02, DR_03, DR_04, and SW_03.

[0124] The sampled data after data analysis is shown in Table 6. These sampled data will participate in a new round of machine learning training. Table 6 lists the sampled data participating in the second round of machine learning training.

[0125] Table 6.

[0126]

[0127] Accordingly, the second round of sampling data is then subjected to machine learning training. In this embodiment, the second round of sampling data is automatically divided into training sample data and test sample data in a ratio of 7:3, and multiple machine learning algorithms are used for training and each machine learning algorithm is optimized to obtain the best training effect. After the training is completed, the machine learning training model with the highest accuracy is saved as a machine learning result file.

[0128] In this round of optimization, through algorithm benchmarking and parameter adjustment, the optimal machine learning training algorithm and prediction accuracy of each output parameter are obtained as shown in Table 7.

[0129] Table 7 lists the machine learning training algorithms and prediction accuracy of each output parameter.

[0130] Table 7.

[0131] Output Parameters Parts_mass BS_H_punkt BS_01 DR_01 DR_02 DR_03 DR_04 SW_03 Algorithm Type LinR RF GBDT GBDT GBDT GBDT GBDT GBDT Algorithm accuracy 0.999 0.914 0.846 0.985 0.906 0.922 0.999 0.968

[0132] It can be seen from Table 7 that the machine learning algorithm accuracy of all output parameters is greater than 0.8, but most of them are greater than 0.9. Therefore, the machine learning model trained by this round of data sampling meets the accuracy requirements, and the following step S8 can be performed to perform subsequent numerical optimization steps.

[0133] Step S8: When there are enough sample points and the prediction accuracy of the machine learning model meets the requirements, call the parameterization file based on the design variables established in step S2 and the trained machine learning model in step S6 (i.e., the machine learning result file obtained in step S6) to build the corresponding numerical optimization workflow.

[0134] In the present invention, the numerical optimization workflow can automatically call the machine learning result file and predict the values ​​of all output parameters according to the changes in the values ​​of each design variable in the parameterized file of the design variables.

[0135] It should be noted that, in this embodiment, a numerical optimization workflow is built using commercial optimization software, the design variables are the thickness changes of six parts (DV01, DV02, DV03, DV04, DV05 and DV06), and the value range of the design variables is the value range of the design variables after the first round of data acquisition and filtering, as shown in Table 5. The output parameters include Parts_mass, BS_H_punkt, BS_01, DR_01, DR_02, DR_03, DR_04 and SW_03. The corresponding machine learning result file is used to predict the value of the output parameter.

[0136] Step S9: define the optimization objective, design variables and value ranges, constraints, select the optimization algorithm, and perform numerical optimization based on the optimization workflow established in step S8.

[0137] In this embodiment, the optimization goal is to make the weight of the parts involved in the optimization as light as possible, and the constraint condition is that the side collision performance remains unchanged. The design variables and value ranges are shown in Table 6, and the corresponding optimization mathematical model is as follows:

[0138]

[0139] In this mathematical model: Parts_mass is the total mass of the optimized parts; BS_H_punkt is the living space from the B-pillar to the H-point of the dummy; BS_01, DR_01, DR_02, DR_03, DR_04 and SW_03 are the deformation labels of the body structure; is the design variable x i lower and upper limits.

[0140] After establishing the above content, by calling the global optimization algorithm provided by the commercial optimization software, iterative optimization calculations can be performed to finally obtain the global optimization solution.

[0141] Common global optimization algorithms in commercial optimization software include Self-Adaptive Evolution (SAE), Simulated Annealing (SA), and Efficient Global Optimization (EGO). In this embodiment, the SAE global optimization algorithm is specifically used, and the final optimization results are shown in Table 8.

[0142] Table 8 lists the optimization results predicted by the machine learning algorithm.

[0143] Table 8.

[0144]

[0145]

[0146] Accordingly, according to the values ​​of the design variables of the final optimization scheme, the collision results of the optimization scheme are solved in the collision simulation software and compared with the numerical optimization prediction results, as shown in Table 9.

[0147] Tables 9-1 and 9-2 list the comparison between the calculation results of the collision simulation software and the numerical optimization prediction results.

[0148] Table 9-1.

[0149]

[0150] Table 9-2

[0151] Evaluation indicators DR_03 DR_04 DR_05 SW_01 SW_02 SW_03 SW_04 SW_05 Simulation Results 0 0 1 1 1 1 0 1 Prediction results 0 0 1 1 1 1 0 1 Model Error 0％ 0％ 0％ 0％ 0％ 0％ 0％ 0％

[0152] Step S10: When the accuracy of the numerical optimization result does not meet the set numerical optimization accuracy requirement, it is necessary to increase the sample collection and return to step S3. Among them, if the difference between the numerical optimization result and the collision simulation software calculation result is less than 5%, it can be considered that the numerical optimization prediction accuracy meets the requirement.

[0153] Step S11: When the accuracy of the numerical optimization result meets the set numerical optimization accuracy requirement, the optimization solution obtained in step S8 is the final optimization solution.

[0154] In this embodiment, since the prediction accuracy of the machine learning algorithm for all output parameters meets the requirements, there is no need to perform step S10.

[0155] Accordingly, it can be seen from Table 9 that in this embodiment, the errors between the calculation results of the collision simulation software and the numerical optimization results are very small, both less than 2%, and the optimization results predicted by machine learning and numerical simulation optimization are the final optimization solutions. It can be seen from Table 8 that after numerical optimization based on machine learning, the obtained optimization simulation can achieve a weight reduction effect of 1.34Kg for the vehicle body structure while ensuring that the collision performance remains unchanged.

[0156] It should be noted that the prior art in the protection scope of the present invention is not limited to the embodiments given in the present application documents. All prior art that does not contradict the solutions of the present invention, including but not limited to prior patent documents, prior public publications, prior public uses, etc., can be included in the protection scope of the present invention.

[0157] In addition, the combination of the various technical features in this case is not limited to the combination described in the claims of this case or the combination described in the specific embodiments. All technical features recorded in this case can be freely combined or combined in any way unless there is a contradiction between them.

[0158] It should also be noted that the above-listed embodiments are only specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and similar changes or modifications made therewith can be directly derived or easily associated with by those skilled in the art from the contents disclosed in the present invention, and all should belong to the protection scope of the present invention.

Claims

1. A vehicle collision simulation optimization method based on machine learning, characterized in that: Includes steps: S1: Establish a vehicle collision simulation analysis model according to the vehicle collision test conditions; S2: Based on the vehicle collision simulation analysis model, a parameterized file of design variables is established, and a corresponding data acquisition workflow is built; S3: Based on the data acquisition workflow, determine the optimization goal, constraints, design variables and value ranges, experimental design sampling method and sampling quantity, and generate and run a sample simulation model to obtain calculation result files and sampling data; S4: Based on the calculation result file and the sampling data, a collision deformation label is generated and collision sampling data containing the collision deformation label is obtained; S5: Based on the collision sampling data, determine the deformation labels, sample points and value ranges of their design variables involved in the machine learning model training; S6: performing machine learning model training, wherein the input parameters of the machine learning model are the design variables of the collected sample points, and the output parameters are the extracted collision feature data; S7: Determine whether the prediction accuracy of the trained machine learning model meets the set threshold and whether the number of sample points meets the requirement. If both are met, proceed to the following step S8; if the prediction accuracy of the machine learning model does not meet the set threshold, return to step S3; if the number of sample points does not meet the requirement, increase the number of sample points and return to step S3; S8: Building a corresponding numerical optimization workflow based on the parameterized file of the design variables established in step S2 and the machine learning model trained in step S6; S9: defining optimization objectives, design variables and value ranges, and constraints, and performing numerical optimization using an optimization algorithm based on the numerical optimization workflow; S10: If the numerical optimization result does not meet the set numerical optimization accuracy, increase the number of sample points and return to step S3. If it meets the requirement, end the step and obtain the final optimization solution.

2. The vehicle collision simulation optimization method based on machine learning as claimed in claim 1, characterized in that: In step S2, the data acquisition workflow is configured to automatically change the model parameters of the vehicle body structure simulation model to generate a new vehicle collision simulation analysis model.

3. The vehicle collision simulation optimization method based on machine learning as claimed in claim 1, characterized in that: In step S2, the design variables include thickness parameters and variation ranges of the B-pillar reinforcement plate, the B-pillar reinforcement plate sticker, the front roof beam, the front roof beam sticker, the front outer sill beam, and the front inner sill beam.

4. The vehicle collision simulation optimization method based on machine learning as claimed in claim 1, characterized in that: In step S4, a plurality of collision components are divided into a plurality of blocks, and a deformation label is assigned to the deformation result of each block to obtain a set of deformation label lists, and the deformation of the vehicle body structure is defined by the deformation label lists.

5. The vehicle collision simulation optimization method based on machine learning as claimed in claim 1, characterized in that: In step S6, when the design variables include parameters of different types, the input parameters are normalized.

6. The vehicle collision simulation optimization method based on machine learning as claimed in claim 1, characterized in that: In step S7, the determination coefficient R 2 To evaluate the prediction accuracy of the machine learning model, the calculation formula is: Where: y i is the actual output value of the i-th test sample, is the mean of the actual output values ​​of all test samples, is the predicted value output for the i-th test sample, is the mean of the predicted values ​​output by all test samples, N is the number of test samples, R 2 The value of is between 0 and 1. The closer its value is to 1, the higher the prediction accuracy of the machine learning model.

7. The vehicle collision simulation optimization method based on machine learning as claimed in claim 1, characterized in that: In step S7, when the number of sample points is 6-8 times the number of design variables, it is considered that the number of sample points meets the requirement.

8. The vehicle collision simulation optimization method based on machine learning as claimed in claim 1, characterized in that: In step S9, the optimization algorithm adopts adaptive evolution, simulated annealing or efficient global optimization.

9. The vehicle collision simulation optimization method based on machine learning as claimed in claim 1, characterized in that: In step S10, when the deviation between the numerical optimization result and the result calculated by the collision simulation software is less than 5%, it is considered that the numerical optimization accuracy meets the requirement.

Images