Small sample multi-objective integrated optimization design method and system for vehicle body structure

Abstract

The invention belongs to the technical field of automobile structure design, and particularly relates to a small sample multi-target integrated optimization design method and system for an automobile body structure, and the method comprises the steps: building an automobile body structure parameterized model, defining design variables, and obtaining a design variable combination; obtaining vehicle body structure performance data by utilizing finite element simulation, and constructing an integrated learning agent model; vehicle body structure design variables are optimized through a multi-objective optimization algorithm, and a Pareto leading edge and a solution set are obtained; and circularly adding a proxy model training data set, and iterating the whole optimization process to a termination condition. The technical problems that in the prior art, simulation calculation is long in time consumption, a single machine learning model is insufficient in prediction precision, and the multi-objective optimization effect is not ideal are solved, and the small sample integration optimization design method of the multi-objective optimization algorithm including the adaptive integration model and the fusion multi-strategy is constructed. And the multi-objective optimization efficiency of the vehicle body structure can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobile structure design, and in particular, relates to a small sample multi-objective integrated optimization design method and system for a vehicle body structure. Background Art

[0002] In the automotive industry, body structure optimization design is an important technical field that directly affects the overall performance, ride comfort and manufacturing cost of the vehicle. The body structure must not only meet the requirements of lightweight, but also excel in multiple performance indicators such as stiffness, strength, crash resistance, durability and NVH (noise, vibration and harshness). There are often conflicts between these performance indicators. For example, reducing the weight of the body may affect the crash resistance and NVH performance of the vehicle, making the optimization process complicated. Therefore, body structure optimization design is usually regarded as a multi-objective, multi-constrained optimization problem.

[0003] Traditional body structure optimization design methods mainly rely on finite element analysis (FEA). Through finite element simulation, engineers can simulate the force, deformation and vibration of the vehicle under different working conditions, so as to evaluate and optimize the performance of the body structure. However, finite element simulation calculation time is long, usually each simulation iteration takes tens of minutes to several hours, and the body structure optimization design often requires multiple iterations of simulation, resulting in low efficiency of the entire optimization process, which is difficult to meet the needs of rapid research and development.

[0004] To solve this problem, existing technologies have introduced agent models based on machine learning. These agent models can reduce the number of direct finite element simulation calculations by learning the relationship between input design parameters and output performance, thereby speeding up the optimization process. However, although a single machine learning model, such as a neural network or support vector machine, can improve the optimization efficiency to a certain extent, it requires a high amount of training data. In addition, most existing optimization algorithms have limited performance when faced with complex multi-objective and multi-constraint optimization problems, especially when coordinating conflicting performance requirements such as lightweight, crashworthiness, and NVH. It is difficult to obtain ideal optimization results.

[0005] Therefore, the existing technology has the following main problems in the optimization design of vehicle body structure:

[0006] 1. The simulation calculation is time-consuming and the optimization efficiency is low: The body structure optimization design is highly dependent on finite element simulation. The calculation time of each simulation iteration is long, and the optimization design often requires multiple iterative simulations, resulting in low overall optimization efficiency and difficulty in meeting the increasingly stringent R&D cycle requirements.

[0007] 2. Single machine learning models have a large demand for data and insufficient prediction accuracy: Although a single agent model can accelerate the optimization process, it has a large demand for training data and insufficient prediction accuracy when faced with strong nonlinear problems such as collision performance.

[0008] 3. The optimization effect of multi-objective optimization algorithm is poor: The optimization design of vehicle body structure involves multiple mutually restrictive performance indicators. Existing technologies make it difficult to ensure that the stiffness, strength, crash resistance and NVH performance are optimized while reducing the weight of the vehicle body, which hinders the improvement of overall optimization accuracy and efficiency.

[0009] In summary, how to reduce simulation time, improve the prediction accuracy of proxy models, and improve the accuracy and efficiency of multi-objective optimization have become key issues that need to be urgently addressed in the field of vehicle body structure optimization design. Summary of the invention

[0010] In view of the defects of the prior art, the present invention proposes a small sample multi-objective integrated optimization design method and system for vehicle body structure. The method constructs a multi-objective integrated optimization strategy based on a small sample to solve the problems existing in the optimization design of vehicle body structure, such as unsatisfactory optimization effect when the optimization algorithm faces multi-objective and multi-constraint problems, and a large amount of training data required to build an agent model. This is conducive to improving the efficiency of vehicle body structure optimization design and shortening the research and development cycle.

[0011] In one aspect, the present invention provides a small sample multi-objective integrated optimization design method for a vehicle body structure, the method comprising:

[0012] Step 1: Establish a parametric model of the vehicle body structure, define design variables, and obtain a small number of design variable combinations through experimental design based on the design space of the design variables;

[0013] Step 2: Use finite element simulation to obtain the body structure performance corresponding to different design variable combinations in step 1, and obtain a data set D containing body structure design variables and performance m (X,Y);

[0014] Step 3: Based on the dataset D in step 2 m (X,Y), taking the body structure design variables as input and the body structure performance as output, an integrated learning agent model combining multiple base learners is constructed;

[0015] Step 4: Based on the integrated learning agent model constructed in step 3, a multi-objective optimization algorithm is used to optimize the vehicle body structure design variables to achieve the goal of optimal vehicle body structure performance, and the Pareto frontier and its corresponding solution set are obtained;

[0016] Step 5: Determine whether the loop termination condition is met. If so, the loop terminates and the optimization result is output, that is, the optimal body structure performance set and its corresponding body structure design variables. If not, the Pareto solution set obtained is added to the data set D through finite element simulation calculation. m (X,Y), start the next loop until the loop termination condition is met.

[0017] Furthermore, the integrated learning agent model includes four base learners: Xgboost, RF, CatBoost and LightGBM.

[0018] Further, step 3 includes:

[0019] Step 3.1: Dataset D m (X,Y) is divided into training data and test data. Four base learners are trained separately in the training data to obtain four sets of prediction results.

[0020] Step 3.2: In the entire training data set, taking the actual value as the benchmark, calculate the similarity between the predicted value of each base learner and the actual value, and calculate the global weight through the similarity between the four groups of predicted results and the actual value;

[0021] Step 3.3: Take the four sets of prediction results of the base learner in step 3.1 as input and the actual performance values ​​in the training data set as output, train the linear regression model and obtain the local ensemble model;

[0022] Step 3.4: In the test set data or new data, the local weight of each test sample point is obtained by calculating the distance between the prediction value of the base learner and the prediction value of the local ensemble model;

[0023] Step 3.5: Generate adaptive weights by integrating local weights and global weights, and complete the construction of the integrated learning agent model based on the adaptive weights.

[0024] Furthermore, in step 4, the multi-objective optimization algorithm is a multi-objective marine predator optimization algorithm integrating multiple strategies. The specific steps of constructing the multi-objective marine predator optimization algorithm integrating multiple strategies include:

[0025] Step 4.1: Initialize algorithm parameters and population: Initialize the population based on the chaos strategy to optimize the spatial distribution of the initial population;

[0026] Step 4.2: Calculate the fitness value of each population;

[0027] Step 4.3: Construct the optimal solution set A1 by non-dominated sorting and calculating the congestion distance;

[0028] Step 4.4: Iterate and update the population position, calculate the fitness value of the updated population, and obtain a new optimal solution set A2;

[0029] Step 4.5: Merge sets A1 and A2, and obtain a new optimal solution set through non-dominated sorting and crowding distance calculation, and replace A1.

[0030] Furthermore, in step 4.4, the iterative update process is divided into three stages: initial exploration, mid-term development, and final convergence. Each stage corresponds to at least two update strategies. The specific update strategy selection is based on a random parameter when and When , the corresponding update strategies are adopted respectively;

[0031] Indicates the first stage: when When , the update strategy is when When , the update strategy is

[0032] Indicates the second stage: when When , the update strategy is in when When , the update strategy is

[0033] Indicates the third stage: when When , the update strategy is in when When , the update strategy is

[0034] In the update strategy formula: T max Indicates the maximum number of iterations; represents the decision factor, a random number between 0 and 1; t represents the current number of iterations; P i represents the i-th population set; E represents the elite population; M B represents the Brownian motion strategy; F represents the inferior population; M L represents the Levy motion strategy; q, r, c represent random parameters; k represents the adjustment factor; CF represents the adaptive correction coefficient; P k1 , P k2 Represents two randomly selected individuals from the population.

[0035] Furthermore, in step 4.4, the termination condition of the iterative update process is that the maximum number of iterations is reached or the change in the Pareto front solution set is less than a preset threshold.

[0036] Furthermore, in step 5, in a non-first optimization cycle, the Pareto solution set obtained in the previous optimization is used as the initial population for the next optimization iteration.

[0037] Furthermore, in step 3.2, the global weight is used to assign a stable weight to each base learner, and the global weight is determined by the following formula:

[0038]

[0039] In step 3.4, the local weight is used to dynamically adjust the weight of each base learner at different sample points to make the model more flexible at the local level. The local weight is determined by the following formula:

[0040]

[0041] In step 3.5, the adaptive weight is used to combine the global weight and the local weight so that the model has the characteristics of both global weight and local weight. The adaptive weight is determined by the following formula:

[0042]

[0043] Where: GW k represents the global weight of base learner k; S k Indicates the similarity between the prediction result of a base learner and the true value; Represents the local weight of the base learner k at the i-th sample point; represents the prediction error of the local ensemble model and the base learner at the i-th sample point; k represents four base learners, Xgboost (X), RF (R), CatBoost (C) and LightGBM (L); Represents adaptive weight.

[0044] Furthermore, the similarity is determined by the following formula:

[0045]

[0046] Where: m represents the number of samples; represents the predicted value of one of the four base learners for the i-th sample point, Represents the actual value of the i-th sample point.

[0047] On the other hand, the present invention also provides a small sample multi-objective integrated optimization design system for a vehicle body structure, comprising:

[0048] The parametric modeling module is used to establish a parametric model of the vehicle body structure, define design variables, and obtain a small number of design variable combinations through experimental design based on the design space of the design variables;

[0049] The simulation module is used to obtain the body structure performance corresponding to different design variable combinations using finite element simulation technology, and generate a data set D containing body structure design variables and performance. m (X,Y);

[0050] Integrated learning agent model module for m (X,Y), taking the body structure design variables as input and the body structure performance as output, an integrated learning agent model combining multiple base learners is constructed;

[0051] The optimization module is used to optimize the body structure design variables based on the integrated learning agent model and adopt a multi-objective optimization algorithm to obtain the Pareto frontier and its corresponding solution set to achieve the optimal body structure performance;

[0052] The control module is used to determine whether the loop termination condition is met. If the termination condition is met, the loop is terminated and the optimization result is output; if not, the solution in the Pareto solution set is passed to the simulation module and added to the data set D through finite element simulation calculation. m (X,Y), and repeat the optimization process until the loop termination condition is met.

[0053] The beneficial effects of the present invention are:

[0054] First, the present invention realizes efficient optimization of vehicle body structure design by constructing an integrated learning agent model and a multi-objective optimization algorithm. By integrating the advantages of multiple base learners, the accuracy of vehicle body performance prediction with a small amount of training data is improved; at the same time, an improved multi-objective optimization algorithm (such as a marine predator optimization algorithm that integrates multiple strategies) is constructed to quickly find a Pareto frontier with better performance. Compared with traditional methods, the present invention improves optimization efficiency and accuracy while reducing the number of finite element simulation calculations, making the vehicle body structure design more lightweight and able to meet multiple performance requirements such as stiffness, crashworthiness and NVH.

[0055] Second, in the preferred implementation, the present invention combines four base learners, Xgboost, RF, CatBoost and LightGBM, in an integrated learning agent model. The present invention fully utilizes the advantages of each algorithm and improves the prediction accuracy and robustness of the model. Xgboost and LightGBM are highly efficient in processing large-scale, high-dimensional data, RF is highly robust to noisy data, and CatBoost has outstanding processing capabilities for classification features. By combining these base learners, the integrated model can effectively avoid the limitations of a single model, and with a small amount of training data, can provide more accurate and stable performance predictions, thereby improving the efficiency and optimization effect in the body structure optimization process.

[0056] Third, in the preferred implementation, the present invention combines global and local weights to generate adaptive weights, which effectively enhances the robustness and generalization ability of the model and ensures better prediction results in complex data scenarios;

[0057] Fourth, in the preferred implementation, the present invention constructs a multi-objective marine predator optimization algorithm that integrates multiple strategies, which effectively improves the search efficiency and solution quality in the multi-objective optimization process. The population initialization of the chaos strategy enhances the diversity of the population. The non-dominated sorting and crowding distance ensure the uniformity and diversity of the solution set. At the same time, through iterative updating and solution merging, the Pareto optimal solution set is quickly found, and the optimization effect is significantly improved.

[0058] Fifth, in the preferred implementation, the present invention effectively improves the global exploration and local development capabilities of the optimization algorithm by dividing the optimization iterative update process into three stages: initial exploration, mid-term development, and final convergence. Two update strategies are adopted at different stages, and dynamic adjustments are made based on random parameters to enhance the diversity and adaptability of the population. The termination condition is set to the maximum number of iterations or the Pareto front change is less than a preset threshold, ensuring that the algorithm converges in time while ensuring the quality of the solution set, thereby improving the optimization efficiency and accuracy.

[0059] Sixth, in the preferred implementation, the present invention adds the Pareto solution set in the optimization cycle as new data to the training data of the proxy model after finite element simulation calculation, thereby enhancing the model's learning of data-specific domain features, which is beneficial to improving the prediction accuracy of the proxy model with a small amount of training data.

[0060] Seventh, in a preferred implementation, the present invention accelerates the convergence speed, avoids repeated searches for suboptimal solutions, and improves optimization efficiency by using the Pareto solution set in the optimization cycle as the initial population for the next iteration of the optimization algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1is a flow chart of a small sample multi-objective integrated optimization design method for a vehicle body structure of the present invention;

[0062] Figure 2 is a schematic diagram of a small sample multi-objective integrated optimization design method for a vehicle body structure according to an embodiment of the present invention;

[0063] Figure 3 is a flowchart of step 3 of an embodiment of the present invention for generating an integrated learning agent model;

[0064] Figure 4 is a flowchart of optimizing vehicle body structure design variables using a multi-objective optimization algorithm in step 4 of an embodiment of the present invention;

[0065] Figure 5 is a simulation diagram of the torsional stiffness of the optimized hood structure of an embodiment of the present invention;

[0066] Figure 6 is a simulation diagram of the bending stiffness of the optimized hood structure of an embodiment of the present invention;

[0067] Figure 7 is a first-order modal simulation diagram of the optimized hood structure of an embodiment of the present invention;

[0068] Figure 8 is a quality prediction graph of each model in the test data set of the embodiment of the present invention;

[0069] Fig. 9 is a torsional stiffness prediction diagram of each model of an embodiment of the present invention in a test data set;

[0070] Fig.10 It is a schematic diagram of a small sample multi-objective integrated optimization design system for a vehicle body structure according to the present invention. DETAILED DESCRIPTION

[0071] In order to enable those skilled in the art to better understand the technical solution of the present application, the present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0072] In the description of this application, the terms "first", "second", and "third" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance; the term "plurality" refers to two or more, unless otherwise expressly defined. Terms such as "installed", "connected", "connected", and "fixed" should be understood in a broad sense. For example, "connected" can be a fixed connection, a detachable connection, or an integral connection; "connected" can be a direct connection or an indirect connection through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to the specific circumstances.

[0073] In the description of the present application, it should be understood that the directions or positional relationships indicated by the terms "up", "down", "left", "right", "front", "back", etc. are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or unit referred to must have a specific direction, be constructed and operated in a specific orientation, and therefore, should not be understood as a limitation on the present application.

[0074] In the description of this specification, the description of the terms "one embodiment / method", "some embodiments / methods", "specific embodiments / methods", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment / method or example are included in at least one embodiment / method or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment / method or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments / methods or examples in a suitable manner.

[0075] As the instruction manual Figure 1-2 The object of the present invention is a small sample multi-objective integrated optimization design method for a vehicle body structure, comprising:

[0076] Step 1: Establish a parametric model of the vehicle body structure, define the design variables, and obtain a small number of design variable combinations through experimental design based on the design space of the design variables.

[0077] It should be noted that the body structure parametric model is a model that uses parametric methods to describe and optimize various design elements of the body during the body structure design process. It makes the design process more flexible by parametrically expressing the key components and geometric features of the body, and can automatically adjust the body structure to meet different design requirements and performance goals.

[0078] Specifically, the design variables in the body structure mainly include the thickness of the panels, crossbeams and longitudinal beams, welding points and connection methods, and shape parameters. The thickness of the panels is the thickness of each panel in the body structure (such as doors, roofs, floors, etc.), which is an important design variable that affects stiffness and collision performance. The crossbeams and longitudinal beams make up the body frame. The size, cross-sectional shape, material, etc. of the beams can be defined through parameterization, which directly affects the crash resistance and structural strength of the body. The connection method between the various parts of the body (such as welding and bolting) can also be described parameterized. The connection position and number will affect the overall stiffness and vibration characteristics of the structure. Shape parameters such as the curvature of the outer contour of the body, the bending radius and other geometric shapes can affect the aerodynamic performance and the aesthetics of the vehicle.

[0079] The design space is to limit the range of values ​​of each design variable to ensure that the generated design is physically feasible and meets the basic performance requirements. The scope of the design space should be reasonably constrained based on engineering experience and vehicle performance requirements. The design space of each design variable can be expressed as an interval, for example: panel thickness t∈[1.5,3.0]mm, beam section width w∈[50,150]mm, number of connection points n∈[5,15], body contour curvature radius r∈[500,1500]mm. Different combinations of these design variables form the entire design space. The dimension of the design space depends on the number of design variables. The larger the design space, the more combinations of design variables increase exponentially.

[0080] Since the design space is usually very large, it is impossible to simulate all the variable combinations one by one. Therefore, it is necessary to select a small number of representative design variable groups in the design space through experimental design methods to ensure that each design variable combination can represent the distribution of the entire design space, thereby reducing the amount of subsequent simulation calculations and ensuring that the sampling points can represent the design characteristics of different regions. Common experimental design methods include Latin hypercube design, full factorial design and orthogonal experimental design.

[0081] The process of generating design variable combinations includes: first, analyzing the physical characteristics and performance requirements of each design variable to determine a reasonable range of values, for example, defining the thickness of the plate as 1.5mm to 3.0mm and the number of welding points as 5 to 15. Then, according to the number of design variables and the size of the design space, select the appropriate experimental design method. For high-dimensional design spaces, Latin hypercube design can be selected; for situations where there are fewer variables and the design space needs to be fully covered, full factorial design can be selected. Based on the selected experimental design method, generate a small number of representative design variable combinations. For example, using Latin hypercube design can generate 10 to 20 design points, while using full factorial design may generate dozens or even hundreds of combinations.

[0082] Step 2: Use finite element simulation to obtain the body structure performance corresponding to different design variable combinations in step 1, and obtain a data set D containing body structure design variables and performance m (X,Y).

[0083] In the present invention, the goal of finite element simulation is to evaluate the structural performance of the design variable combination and establish the relationship between the design variable and the performance. The structural performance of the design variable combination is evaluated by using finite element simulation tools (such as Optistruct, ANSYS, ABAQUS, etc.) to simulate and calculate the body model corresponding to each set of design variable combinations to obtain its performance indicators. The relationship between the design variables and the performance is established by recording the corresponding relationship between different design variable combinations (inputs) and their performance (outputs) through simulation results, providing data support for subsequent agent model training and optimization.

[0084] Specifically, step 2 includes:

[0085] Step 2.1: According to the actual geometric dimensions of the vehicle body, establish an accurate three-dimensional model, divide the three-dimensional model of the vehicle body into a finite number of small units (grids), analyze each small unit through the finite element method, and calculate the mechanical response of the entire structure.

[0086] In this step, the 3D model includes the main structural components such as doors, roof, longitudinal beams, cross beams, floor, etc. The meshing should ensure high accuracy in key areas (such as locations where stress is concentrated).

[0087] Step 2.2: Determine the constraints of the vehicle body in actual use and the external load conditions of the vehicle under different working conditions.

[0088] Constraints such as the contact point between the wheel and the ground, the support point of the suspension system, etc. need to be reasonably constrained. External loads include static loads, dynamic loads and vibration loads. Static loads include static forces such as the vehicle's own weight and the passengers and luggage it carries. Dynamic loads include external forces that simulate dynamic conditions such as vehicle acceleration, deceleration or collision. In particular, in collision simulations, the definition of collision speed and direction needs to be considered. Vibration loads are used for NVH analysis to simulate vibrations caused by uneven road surfaces.

[0089] Step 2.3: Calculate the body stiffness, strength, crash resistance and NVH (noise, vibration and harshness) performance under different design variable combinations through simulation.

[0090] Stiffness is used to evaluate the body's ability to resist deformation when subjected to external loads, and is usually measured by analyzing the displacement, stress distribution and stiffness coefficient of various parts of the body. Strength is used to analyze the stress concentration of the body under external loads to ensure that it does not yield or break. Crash resistance is specifically aimed at vehicle collision conditions, and evaluates the body structure's ability to absorb energy and resist deformation in a collision, usually by simulating the body's stress conditions at different collision speeds and collision angles. NVH evaluates the noise and vibration of the vehicle during driving by analyzing the vibration mode and acoustic response of the body structure at different vibration frequencies.

[0091] Step 2.4: After each finite element simulation, extract the simulation results and record the corresponding performance data. Through the simulation of multiple design variable combinations, the data set D is finally formed. m (X,Y).

[0092] Dataset D m In the formula, X represents the combination of design variables, for example: X = [t1, t2, w1, w2, ...], where t represents the thickness of different panels and w represents the width of the beam. Y represents the performance index output, for example: Y = [Stiffness, Strength, Crashworthiness, NVH], where the values ​​represent the stiffness, strength, crashworthiness, and NVH performance of the vehicle body.

[0093] Step 3: Based on the dataset D in step 2 m (X,Y), taking the body structure design variables as input and the body structure performance as output, an integrated learning agent model combining multiple base learners is constructed.

[0094] In the present invention, the ensemble learning agent model is an adaptive ensemble model that uses a combination of Xgboost (Extreme Gradient Boosting), RF (Random Forest), CatBoost (Categorical Boosting) and LightGBM (Light Gradient Boosting Machine) as base learners.

[0095] Xgboost is an optimized gradient boosting algorithm that can handle high-dimensional and nonlinear problems, with good computational efficiency and strong nonlinear modeling capabilities. It is particularly suitable for dealing with data with multiple features and high dimensions, and performs well in complex multi-objective optimization. RF is an ensemble algorithm based on multiple decision trees, which improves accuracy and stability by building a large number of decision trees and combining their prediction results. The advantage of RF is that it is robust to noisy data, suitable for processing high-noise design data sets, and can significantly enhance the stability of the model. CatBoost is a gradient boosting algorithm optimized specifically for processing categorical features, which can effectively handle categorical variables and uneven data distribution. In vehicle body structure optimization, CatBoost can accurately handle discrete design variables, making it particularly suitable for complex scenarios with a large number of categorical features. LightGBM is an efficient gradient boosting framework designed for processing large-scale data. It has fast training speed and low memory usage, and is suitable for processing high-dimensional data and massive data sets. LightGBM can provide high-quality prediction results in an environment that requires efficient and fast training.

[0096] The present invention selects Xgboost, RF, CatBoost and LightGBM as the base learners of the integrated learning agent model because they have complementary advantages when processing complex data. By combining these base learners into an adaptive integrated model, their respective advantages can be fully utilized to solve the shortcomings that a single model may face. The high efficiency of Xgboost and LightGBM is suitable for processing large-scale data, CatBoost optimizes the processing of classification features, and RF can effectively deal with noisy data. After integrating these models, the predictive ability, robustness and stability of the model can be enhanced complementarily. Such an integrated strategy can effectively reduce the overfitting or underfitting problems caused by a single model, ensuring that the performance in diversified scenarios is more robust.

[0097] As the instruction manual Figure 2 , step 3 includes:

[0098] Step 3.1: Dataset D m (X, Y) is divided into training data and test data. Four base learners are trained separately in the training data to obtain four sets of prediction results.

[0099] Specifically, the training set and the test set can be divided into a common ratio of 80% training set and 20% test set, that is, the data set D m In (X, Y), 80% of the data is used for training the base learner, and 20% of the data is used for testing and evaluating the model performance. The training set and the test set can also be randomly divided to ensure that the training set and the test set have similar data distribution, and the repeatability of the division results is ensured by using random number seeds.

[0100] After the data set is divided, the four base learners (Xgboost, RF, CatBoost, LightGBM) are trained independently on the training data. The input of each base learner is the design variable X of the training set. train , the output is the body structure performance Y of the training set train After training, each base learner is trained on the test set X test Make predictions and generate four sets of prediction results. These prediction results are used to evaluate the performance of the base learners and serve as the input of the subsequent ensemble model. The prediction results of each base learner are compared with the true test set label Y test Compare and evaluate the prediction performance. This step can use the commonly used evaluation indicators: mean square error (MSE) and R square (R 2 ), the mean square error is used to measure the error between the predicted value and the true value, and the R square is used to measure the model's ability to explain the data.

[0101] Each of the four sets of predictions represents the predictions of a base learner, which are generated by predicting the test data set after each base learner is trained separately. Each base learner predicts the same test data set and generates a result that reflects the model's prediction of the input design variables (test set X test ) of the vehicle body structure performance (output Y test ) predictions.

[0102] Step 3.2: In the entire training data set, taking the actual value as the benchmark, calculate the similarity between the predicted value of each base learner and the actual value, and calculate the global weight through the similarity between the four groups of predicted results and the actual value.

[0103] The calculation of similarity is based on the inverse square of the prediction error, that is, if the prediction error of a base learner for a sample is smaller, the similarity will be greater, indicating that the performance of the model is better.

[0104] The similarity is determined by the following formula:

[0105]

[0106] Where: S k Indicates the similarity between the prediction result of a base learner and the true value, S k ∈{S X ,S R ,S C ,S L}; m represents the number of samples; represents the predicted value of sample i by one of the four base learners (Xgboost, Random Forest, CatBoost and LightGBM), Represents the actual value of the i-th sample point.

[0107] The global weight is obtained based on the training data set and is determined by the following formula:

[0108]

[0109] Where: GW k represents the global weight of base learner k, GW k ∈{GW X ,GW R ,GW C ,GW L}; S k Represents the similarity of base learner k; k represents four base learners, Xgboost (X), Random Forest RF (R), CatBoost (C) and LightGBM (L).

[0110] Formula (2) indicates that the global weight (GW) of a base learner k ) is equal to the similarity of the base learner (S k ) to the sum of similarities of all base learners. The greater the similarity of a base learner, the better its prediction performance, and thus the greater its global weight, and the more the ensemble model relies on the results of the base learner in the final prediction.

[0111] In this step, the weight of each base learner is calculated dynamically. Based on their prediction performance on the entire dataset, base learners with good performance will receive higher weights, thus having a greater impact on the final prediction results in ensemble learning. Similarity measures the performance of the model by calculating the difference between the predicted value and the true value. Using the inverse of the square of the error can amplify the effect of the model with a small prediction error and improve the overall accuracy of the ensemble model. By calculating the similarity of the four base learners, the ensemble learning model can adaptively select the base learner with better performance, avoid the dominance of a single model on the final result, and improve the robustness and generalization ability of the ensemble model.

[0112] Step 3.3: Take the four sets of prediction results of the base learner in step 3.1 as input and the actual performance value in the training data set as output, train the linear regression model (Linear Regression) and obtain the local ensemble model.

[0113] Specifically, use the four base learners in step 3.1 to train the data set X train The four sets of predicted values ​​generated As the input features of the linear regression model, these prediction results constitute the input feature matrix of the linear regression model, which can be expressed as:

[0114] m is the number of samples in the training set

[0115] Use actual performance values As the output label of the linear regression model, corresponding to each training sample, the actual performance value is used to guide the linear regression model on how to adjust the weights to optimize the overall prediction effect. The output can be represented as a vector:

[0116]

[0117] The local ensemble model uses Xgboost, RF, CatBoost and LightGBM as the first-layer base learners and the linear regression model as the second-layer model. In the training data set, the prediction results of the four base learners are weighted by the linear regression model to predict the performance value on new samples or test sets. The local ensemble model is expressed as follows:

[0118]

[0119] Where: represents the predicted value of the local integrated model; wX, wR, wC, wL represent the weight coefficients learned by the linear regression model; b represents the intercept term of the model; i represents the i-th sample.

[0120] Step 3.4: In the test set data or new data, obtain the local weight of each test sample point by calculating the distance between the prediction value of the base learner and the prediction value of the local ensemble model.

[0121] It should be noted that the local weight is obtained based on the test data set or new data and has adaptive characteristics. Different local weight values ​​will be generated for different data. The premise for calculating the local weight is to train a local ensemble model in the training data set.

[0122] Specifically, in the test data set or new data, the predicted value is obtained based on the trained local ensemble model by As a benchmark, define And the predicted values ​​of each base learner The absolute value of the difference is

[0123]

[0124] Where: Represents the prediction error of the local ensemble model and the base learner at the i-th sample point; Represents the predicted value of the local integrated model for the i-th sample; Represents the predicted value of the k-th base learner for the i-th sample.

[0125] Based on the error between each base learner and the local ensemble model, a normalized method is used to calculate the local weight of each base learner at the sample point. The local weight calculation is shown in the formula:

[0126]

[0127] Where: represents the local weight of the base learner k at the i-th sample point.

[0128] Step 3.5: Generate adaptive weights by integrating local weights and global weights, and build an integrated learning agent model based on the adaptive weights.

[0129] Specifically, the adaptive weight is calculated by multiplying the local weight by the global weight and normalizing it. The formula is as follows:

[0130]

[0131] Where: Represents adaptive weight; GW k represents the global weight of the kth base learner; represents the local weight of the k-th base learner at the i-th sample point.

[0132] The final prediction result of the integrated learning agent model is a weighted combination of the prediction results of each base learner based on the adaptive weight. The formula is as follows:

[0133]

[0134] Where: y i Represents the predicted value of the ensemble learning agent model for the i-th sample; Respectively represent the adaptive weights of Xgboost, Random Forest RF, CatBoost and LightGBM at the i-th sample point; They respectively represent the prediction results of the four base learners for the i-th sample.

[0135] It should be noted that the global weight of step 3.2 is calculated on the entire training data set, and the purpose is to assign a weight to each base learner, which represents the overall average performance of the base learner on all training data. The global weight provides the overall stability of the model to prevent the poorly performing base learners from affecting the overall prediction effect. The local weight of step 3.4 is calculated at each sample point or in a local range, and the purpose is to dynamically adjust the weight of each base learner at different sample points. The local weight is dynamically allocated through a linear regression model based on the specific prediction error of each sample point, so that the model can adapt to the characteristics of different samples more flexibly. This step constructs an adaptive weight by combining global weights and local weights, thereby generating a final integrated learning agent model. The model can not only maintain the stability of the prediction performance in a global range, but also can be dynamically adjusted at the local level to improve the prediction accuracy of different sample points. The adaptive weight ensures the reasonable distribution of the weights of each base learner through normalization processing, and has strong robustness and adaptability in different data distributions. This dual weight strategy improves the robustness, flexibility and prediction accuracy of the model, and adapts to different types of data distributions and sample characteristics, which is an important innovation and advantage of the present invention.

[0136] Step 4: Based on the integrated learning agent model constructed in step 3, a multi-objective optimization algorithm is used to optimize the body structure design variables to achieve the goal of optimal body structure performance and obtain the Pareto frontier and its corresponding solution set.

[0137] The optimization process of this step adopts a multi-objective marine predators optimization algorithm (MOMPA) that integrates multiple strategies.

[0138] As the instruction manual Figure 3 , the specific steps are as follows:

[0139] Step 4.1: Initialize algorithm parameters and population: Initialize the population based on the chaos strategy to optimize the spatial distribution of the initial population.

[0140] This step uses chaotic sequence, which is a common random number generation method used for population initialization to enhance the diversity of the algorithm. It can traverse the entire interval [0,1]. Its uncertainty and sensitivity make it suitable as a strategy for random number generation in optimization algorithms. It can better cover the design space and avoid the local concentration problem of population initialization.

[0141] The chaotic sequence is generated according to the following formula:

[0142]

[0143] Where: h t It represents the value of the t-th chaotic sequence, and the initial value h0=0.152 (based on the initial value, the sequence recursively generates subsequent chaotic values ​​through the above mapping rules).

[0144] Furthermore, the generated chaotic sequence is mapped to the initialized population as shown in the following formula:

[0145]

[0146] In the formula, H represents the matrix composed of chaotic sequences, with a dimension of n×d, where n represents the number of populations and d represents the dimension of the population (i.e., the number of design variables); each row h i,j Represents the chaotic value of the i-th population individual on the j-th design variable.

[0147] The population individuals are initialized by mapping the chaotic sequence to the domain of the design variables (i.e., the design space). The specific formula is as follows:

[0148] X0=X min +H·(X max -X min ) (10)

[0149] Where: X0 represents the initialized population individual matrix; X min and X max represent the minimum and maximum values ​​of the design variables respectively.

[0150] The design variable value of each population individual is calculated according to formula (10).

[0151] Furthermore, the n individuals of the initial population form a population group P, which is expressed as:

[0152]

[0153] In formula (11), P is a matrix composed of n individuals in the population, each of which has d design variable values ​​X i,j , X i,j represents the value of the jth design variable for the i-th population individual.

[0154] Step 4.2: Calculate the fitness value of each population.

[0155] It should be noted that the fitness value is an important criterion for measuring the quality of each individual (population solution) in the multi-objective optimization algorithm. For the multi-objective optimization problem in the present invention, the fitness value measures the performance of the population individuals on different objective functions.

[0156] The fitness value of each individual in the population is calculated based on the function value of multiple performance objectives. The optimization objectives of the body structure design are, for example, minimizing the body weight, maximizing the body stiffness or strength, optimizing the body collision safety performance, and minimizing the body noise and vibration.

[0157] Each population individual (solution) has multiple design variables. After performance evaluation, the corresponding multiple objective function values ​​are calculated:

[0158] F i ={f1(X i ),f2(X i ),……,f m (X i )} (12)

[0159] Where: F i Represents the fitness value vector of the i-th population individual, including m objective functions f1(X i ),f2(X i ),……,f m (X i ).

[0160] It should be noted that f1(X i ),f2(X i ),……,f m (X i )These objective functions are evaluated through the integrated learning agent model in step 3, which can quickly evaluate the objective function value of each individual in the population.

[0161] Step 4.3: Construct the optimal solution set A1 by non-dominated sorting and calculating the congestion distance.

[0162] It should be noted that non-dominated sorting is a commonly used sorting method in multi-objective optimization, which is used to determine which individuals in the population are non-dominated solutions, that is, Pareto front solutions. Non-dominated solutions refer to individuals that are not completely surpassed by other solutions on all objective functions.

[0163] Specifically, if solution A in the population is not dominated by any other solution, then A is a non-dominated solution. Find the non-dominated solutions in the population and form the first frontier solution set A1. These solutions form the optimal compromise solution set in multi-objective optimization. The first frontier solution set A1 is the goal of the algorithm. It contains solutions that are not dominated by other individuals in all objectives, that is, solutions on the Pareto front. Then remove the non-dominated solutions that have been found from the population, and then repeat the above process in the remaining population to find the second frontier, and so on, until all solutions in the entire population are assigned to different frontiers.

[0164] Furthermore, in order to maintain the diversity of the population, the algorithm not only needs to consider the fitness value of the solution, but also needs to calculate the crowding distance of each solution to prevent the solution from being too concentrated in a local area. The crowding distance is calculated by measuring the distance between a solution and its adjacent solutions in the target space, indicating the sparsity of the solution in the target space. The larger the distance, the sparser the solutions around the solution, which means that the population diversity is better.

[0165] Furthermore, after performing non-dominated sorting and crowding distance calculation, the crowding distance is finally used to screen the solutions to ensure the diversity and excellence of the first frontier solution A1. The screening criteria are: non-dominated solutions are retained first, and for those solutions that perform similarly on the target, solutions with larger crowding distances are retained to enhance population diversity.

[0166] Step 4.4: Iterate and update the population position, calculate the fitness value of the updated population, and obtain the new optimal solution set A2.

[0167] During the optimization process, the entire evolutionary optimization process is evenly divided into three stages: Phase 1, Phase 2, and Phase 3. Each stage corresponds to a different update strategy (in this step, there are two different update strategies for each stage). The specific update strategy selection is based on a random parameter when and When , different update strategies are adopted. The following is a detailed description of the scheme:

[0168] Table 1

[0169]

[0170] In Table 1, T max Indicates the maximum number of iterations; represents the decision factor, a random number between 0 and 1, which is used to control the selection of the update strategy in the optimization phase; t represents the current number of iterations; P i represents the i-th population set; E represents the elite population, that is, the set of solutions with the best performance in the current generation population; M B represents the Brownian motion strategy; F represents the inferior population, which is the worst individual in the population set; M L represents the Levy motion strategy; q, r, and c represent random parameters, which are used to control the range or step size of individual searches to ensure the randomness and diversity of individuals in the population. Usually, q takes a value of 0.5, r takes a value of a random number between 0 and 1, and c takes a value of a random number between 0 and 0.2; k represents the adjustment factor, which is used to scale the gap between the position of individuals in the population and the elite population to improve the adaptability of individuals in the design space. k takes a value of a random number between 0.3 and 0.7; CF represents the adaptive correction coefficient, which controls the position update between individuals, through time t and the maximum number of iterations T max Dynamic adjustment; P k1 , P k2 Represents a random integer between 0 and n for displacement. Two randomly selected individuals from the population are used to cross-update the positions between individuals to ensure the diversity of the population.

[0171] The first stage is the initial exploration stage. The main goal of the initial stage is global exploration, looking for a larger search space to find potential optimal solutions. The update strategy includes: When the i-th individual moves according to the position of the elite population E and the Brownian motion mode M B Update to guide the individual to approach the global optimal solution. The random factors q and r maintain the randomness of the individual to prevent it from falling into the local optimum. When , an elite reverse learning mechanism is adopted, and the position update of individuals is determined by the elite population E and the inferior population F, so that the individuals in the population can better adapt to the search space of the optimization problem, which is conducive to improving the global search ability of the algorithm.

[0172] The second stage is the mid-term development stage. The focus of optimization in the mid-term development stage shifts to local development. The goal is to narrow the search range and improve the accuracy of the solution. The update strategy includes: When , the adaptive correction coefficient CF dynamically adjusts the update step size of the individual position, and the individual adopts the Brownian motion mode M B Gradually approach the elite population and conduct local search. When the individual adopts the Brownian motion mode M L Update the position to avoid falling into the local optimum.

[0173] The third stage is the final convergence stage, where the population gradually converges to the optimal solution. The goal is to make local fine-tuning to ensure the accuracy of the optimization results. The update strategy includes: When , the adaptive correction coefficient CF is used again to adjust the individual position to ensure that the individual fine-tuning converges in the final stage. When a horizontal crossover mechanism is used, two random individuals P are introduced. k1 and P k2 To perform crossover updates, improve the diversity of the population, and avoid falling into local optimality. The adjustment factor c is used to adjust the crossover step size to prevent individuals from converging too quickly.

[0174] After each population position update, the performance of the individuals in the new position is evaluated and their fitness values ​​are calculated. The non-dominated sorting and crowding distance of the population are calculated based on the fitness values, and a new set of optimal solutions is selected to form a new elite population A2, which is retained for the next generation of optimization process.

[0175] This step achieves an effective combination of global exploration and local development through a three-stage optimization strategy. By dynamically adjusting the positions of individuals in the population, the optimization process can more efficiently explore the design space and gradually converge to the optimal solution. Combined with elite population guidance, the algorithm can not only widely explore the global solution, but also finely develop the local optimal solution area, improving the overall performance of the optimization algorithm.

[0176] Step 4.5: Merge solution sets A1 and A2, and obtain a new optimal solution set through non-dominated sorting and crowding distance calculation, and replace A1.

[0177] Solution set A1 is the optimal solution set obtained in the previous generation, that is, the Pareto front solution set obtained by non-dominated sorting and crowding distance calculation. Solution set A2 is the new solution set obtained after the population of the current generation is updated, which contains the updated population individuals. Specifically, solution sets A1 and A2 are merged to form a new solution set:

[0178] A combined =A1∪A2 (13)

[0179] Furthermore, for the combined solution set A combinedPerform non-dominated sorting to form the first Pareto front solution set F1, remove the non-dominated solutions that have been found from the solution set, and then find the second front F2 in the remaining solution set, and so on, until all solutions are sorted. Then, sort the solutions in the merged solution set according to the value of each objective function. For each boundary solution on the objective function, its crowding distance is set to infinity, indicating that these solutions should be retained first. For non-boundary solutions, the crowding distance is the sum of its distances from adjacent solutions in the target space. After non-dominated sorting and crowding distance calculation, select a new optimal solution set from the merged solution set. The selection process includes: giving priority to retaining those non-dominated solutions located on the first frontier (Pareto front), that is, selecting the maximum number of solutions from F1 until the preset population size is reached. When the number of non-dominated solutions exceeds the population size, use the crowding distance for sorting, and give priority to retaining solutions with large crowding distances to ensure the diversity of the solution set. The selected new solution set A new As the initial solution set of the next generation, it is used to replace the current optimal solution set A1.

[0180] Step 4.6: Determine whether the maximum number of iterations has been reached. If so, stop the loop iteration and output the optimal solution set. Otherwise, return to step 4.4 to update the population position.

[0181] Specifically, T max The maximum number of iterations preset for the algorithm, that is, the maximum number of generations that the optimization algorithm can execute. Before the optimization begins, this parameter is usually set based on the complexity of the problem and resource constraints (such as time, computing power, etc.). This parameter is used to control the termination condition of the optimization. In each iteration, the algorithm updates the position of the population and calculates the current number of iterations t. Once t reaches or exceeds the maximum number of iterations T max , the optimization process stops and enters the optimal solution output step. If the current number of iterations t has not reached T max , then return to step 4.4 and continue to update the population position and optimize iteratively.

[0182] The judgment logic of the number of iterations is:

[0183] Ift≥T max , then stop

[0184] Step 5: Determine whether the loop termination condition is met. If so, the loop terminates and the optimization result is output, that is, the optimal body structure performance set and its corresponding body structure design variables. If not, the Pareto solution set obtained is added to the data set D through finite element simulation calculation. m (X,Y), start the next loop until the loop termination condition is met.

[0185] In particular, in non-first cycles, the Pareto solution set obtained by the multi-objective optimization algorithm in the previous cycle will be used as the initialization population for the next cycle.

[0186] Specifically, after each round of multi-objective optimization algorithm is completed, it is necessary to determine whether the preset loop termination condition has been reached. The termination condition can be set based on the following: maximum number of iterations, convergence judgment, and performance index change. The maximum number of iterations is set by setting a maximum number of iterations T max , when the number of iterations of the algorithm reaches T max When , the loop terminates. Convergence judgment means that if the solution set of the Pareto front changes little after multiple iterations, it indicates that the solution set is close to the optimal one and the algorithm can be terminated early. Performance index change means that if the performance index of the current solution set (such as the weight, stiffness, and crashworthiness of the vehicle body structure) changes less than the preset threshold in multiple iterations, it can be considered that the solution is close to convergence and the iteration is stopped. If the termination condition is not met, the loop continues, and the current Pareto front solution set is added to the training data set of the proxy model through finite element simulation calculation for the next optimization iteration.

[0187] Example 1

[0188] The implementation process of the small sample multi-objective integrated optimization design method for the vehicle body structure of the present invention is explained through specific design parameters.

[0189] Step 1: Establish a parametric model of the hood structure, and use the design variables of the hood (such as the front beam size, the middle beam size, the side beam size, etc.) to establish a parametric hood model. Through experimental design (such as Latin hypercube design or full factorial design), generate a small number of representative design variable combinations to form an initial design space.

[0190] Using Latin hypercube sampling, the following design points were generated: Design point 1: front crossbar size 250mm, center crossbar size 60mm, side beam size 160mm. Design point 2: front crossbar size 255mm, center crossbar size 65mm, side beam size 170mm. Design point 3: front crossbar size 240mm, center crossbar size 70mm, side beam size 140mm.

[0191] (…other design points)

[0192] Step 2: Perform finite element simulation on these design points respectively, simulate the body structure performance under different loads and working conditions, and obtain the performance data corresponding to each design point (such as mass, torsional stiffness, bending stiffness, and first-order mode). The simulation tool can use Optistruct, ANSYS, or ABAQUS. Taking design point 1 as an example, the performance results of design point 1 are: mass 12.4kg, torsional stiffness 78N·m / deg, bending stiffness 130.12N / mm, and first-order mode 29.5Hz.

[0193] Step 3: Design variables and performance data set D generated based on step 2 m (X, Y), use machine learning models (such as Xgboost, random forest, CatBoost and LightGBM) to build a proxy model. The proxy model can quickly predict the performance of the vehicle body under different combinations of design variables. 80% of the data is used for training and 20% for testing. After training each base learner, a preliminary prediction model is obtained, and its performance is evaluated and combined into an integrated learning model.

[0194] Step 4: Use a multi-objective optimization algorithm (such as the marine predator optimization algorithm that integrates multiple strategies) to optimize the design variables of the vehicle body structure with the goal of optimal vehicle body performance. Through non-dominated sorting calculations, a Pareto front solution set is generated, in which the solutions achieve a compromise between different objectives (such as mass and torsional stiffness).

[0195] Step 5: If the termination condition (such as convergence or maximum number of iterations) is not met, the current Pareto solution set is added to the data set, and the next round of simulation and optimization is continued until the convergence criterion is met.

[0196] In the final output Pareto front solution set, the optimized hood structure has achieved significant improvements in multiple performance indicators. Original hood structure parameters: mass: 12.4kg, torsional stiffness: 80.89N·m / deg, bending stiffness 133.92N / mm, first-order mode: 29.81Hz. Optimized body design (a solution in the Pareto solution set): mass: 12.2kg (reduced by 0.2kg), torsional stiffness: 85.68N·m / deg, bending stiffness 139.51N / mm, first-order mode: 30.13Hz.

[0197] As the instruction manual Figure 5 , Figure 5 To optimize the torsional stiffness simulation diagram of the rear hood structure, Figure 5It can be seen that most areas are displayed in blue or green, which means that under the action of external force, the overall displacement of the hood is small, the deformation of the structure is relatively limited, and the displacement distribution in the figure is relatively uniform, and there is no large area of ​​red area, indicating that the overall stiffness distribution is reasonable and the torsional resistance is good. Compared with the torsional stiffness before optimization, the torsional stiffness after optimization is increased by 4.79N·m / deg, and the hood structure has a higher torsional stiffness.

[0198] As the instruction manual Figure 6 , Figure 6 To optimize the bending stiffness simulation diagram of the rear hood structure, Figure 6 It can be seen that the maximum displacement value is 1.111e-01mm, that is, 0.111mm, which means that the maximum displacement of the structure is very small when the load is applied. In the upper and lower bending areas, the color changes in the middle and top areas are more uniform, and the deformation amplitude is smaller, indicating that these areas have higher stiffness. Compared with the bending stiffness before optimization, the bending stiffness after optimization is increased by 6.41N / mm. The hood structure has higher bending stiffness.

[0199] As the instruction manual Figure 7 , Figure 7 To optimize the first-order modal simulation diagram of the rear hood structure, Figure 7 It can be seen that local vibration occurs near the hinge of the inner panel of the hood, and the area connected to the body is red, but the overall vibration amplitude distribution is relatively uniform and most areas are dominated by blue and green, which shows that the deformation of the structure is small, especially in the first-order mode, the vibration amplitude is well controlled, further indicating that the first-order modal frequency is high. Compared with the first-order mode before optimization, the frequency of the first-order modal after optimization is increased by 0.32Hz, indicating that the natural frequency of the hood is improved. The higher the frequency, the better the dynamic performance of the structure, and it can more effectively avoid resonance in the working state, indicating that the structural stiffness is higher and the modal performance is better.

[0200] Example 2

[0201] Compare the performance of the integrated learning agent model of the present invention with the single model in the prior art (such as XGBoost, RF, CatBoost or LightGBM) in the prediction of vehicle body structure performance. And compare the performance of the Marine Predators Algorithm (MPA) integrating multiple strategies of the present invention with the Genetic Algorithm (GA) in the prior art in the multi-objective vehicle body hood structure optimization problem.

[0202] Define the design variables of the body hood structure: front crossbar size t1∈[234,259]mm, middle crossbar size t1∈53,63]mm, side beam size t1∈[148,178]mm. Define the optimization goal to reduce the weight of the hood while ensuring the performance indicators: torsional stiffness greater than 85N·m / deg, bending stiffness greater than 138N / mm, and first-order mode greater than 30Hz.

[0203] Train single XGBoost, RF, CatBoost, and LightGBM models respectively to perform performance prediction, and compare the overall prediction results by combining the above four base learners into an integrated model. Figure 8-9 , Figure 8 The quality prediction graph of each model in the test data set is Figure 8 It can be seen that the ensemble model (red dashed line) performs the most stable, and the prediction results almost coincide with the true values ​​throughout the whole process. The ensemble model captures the trend of quality fluctuation very well, especially in the sample interval with large fluctuations (such as around sample 36), with small errors, which reflects excellent fitting ability. RF, XGBoost and LightGBM have large errors in some sample points with large fluctuations, and CatBoost performs second. Fig. 9 The torsional stiffness prediction diagram of each model in the test data set is shown in Figure 2. Fig. 9 It can be seen that the predicted value of the integrated model (red line) almost coincides with the true value at most sample points, with a small error. At the peak near sample number 36, although it fails to perfectly follow the large fluctuation of the true value, it still performs closer than other models. Overall, the stability and fitting ability of the integrated model are significantly better than other models.

[0204] Finally, the performance comparison between the integrated learning agent model and the single model is obtained:

[0205] Table 2

[0206]

[0207] It can be seen from Table 2 that the integrated learning agent model of the present invention is significantly better than any single model in terms of prediction accuracy. The improvement in prediction accuracy makes the integrated learning model have obvious advantages in complex vehicle body structure optimization problems.

[0208] Use traditional genetic algorithm to perform multi-objective optimization to find the Pareto optimal solution for car body design. Under the same optimization conditions, use the marine predator optimization algorithm that integrates multiple strategies to perform the same multi-objective optimization. The comparison of optimization algorithms is as follows:

[0209] Table 3

[0210]

[0211] It can be seen from Table 3 that the marine predator optimization algorithm of the present invention performs well in multi-objective optimization, and the optimized body design is superior to the genetic algorithm in terms of mass, torsional stiffness, bending stiffness and first-order mode. In terms of calculation time, the marine predator optimization algorithm takes 150 seconds, which is about 37.5% more efficient than the genetic algorithm's 240 seconds.

[0212] Through the comparison of experimental data in this embodiment, it can be seen that the method of the present invention is significantly better than the prior art in the performance of the integrated learning agent model and the optimization algorithm. The integrated learning model has higher prediction accuracy in complex small sample data scenarios through the combination of multi-base learners. The marine predator optimization algorithm that integrates multiple strategies performs well in multi-objective optimization, and can quickly find a vehicle body structure design solution with better performance, reduce calculation time, and improve optimization effect.

[0213] As the instruction manual Fig.10 Another object of the present invention is to provide a small sample multi-objective integrated optimization design system for a vehicle body structure, comprising:

[0214] The parametric modeling module is used to establish a parametric model of the vehicle body structure, define design variables, and obtain a small number of design variable combinations through experimental design based on the design space of the design variables.

[0215] The simulation module is used to obtain the body structure performance corresponding to different design variable combinations using finite element simulation technology, and generate a data set D containing body structure design variables and performance. m (X,Y).

[0216] Integrated learning agent model module for m (X,Y), taking the body structure design variables as input and the body structure performance as output, an integrated learning agent model combining multiple base learners is constructed.

[0217] The optimization module is used to optimize the body structure design variables based on the integrated learning agent model and adopt a multi-objective optimization algorithm to obtain the Pareto frontier and its corresponding solution set to achieve the optimal body structure performance.

[0218] The control module is used to determine whether the loop termination condition is met. If the termination condition is met, the loop is terminated and the optimization result is output; if not, the solution in the Pareto solution set is passed to the simulation module and added to the data set D through finite element simulation calculation. m (X,Y), and repeat the optimization process until the loop termination condition is met.

[0219] The above is only an embodiment of the present invention, and the common sense such as the known specific structure and characteristics in the scheme is not described in detail here. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present invention is defined by the attached claims rather than the above description, so it is intended to include all changes within the meaning and scope of the equivalent elements of the claims. Any figure mark in the claims should not be regarded as limiting the claims involved.

Claims

1. A small sample multi-objective integrated optimization design method for vehicle body structure, characterized in that: Methods include: Step 1: Establish a parametric model of the vehicle body structure, define design variables, and obtain a small number of design variable combinations through experimental design based on the design space of the design variables; Step 2: Use finite element simulation to obtain the body structure performance corresponding to different design variable combinations in step 1, and obtain a data set D containing body structure design variables and performance m (X,Y); Step 3: Based on the dataset D in step 2 m (X,Y), taking the body structure design variables as input and the body structure performance as output, an integrated learning agent model combining multiple base learners is constructed; Step 4: Based on the integrated learning agent model constructed in step 3, a multi-objective optimization algorithm is used to optimize the vehicle body structure design variables to achieve the goal of optimal vehicle body structure performance, and the Pareto frontier and its corresponding solution set are obtained; Step 5: Determine whether the loop termination condition is met. If so, the loop terminates and the optimization result is output, that is, the optimal body structure performance set and its corresponding body structure design variables. If not, the Pareto solution set obtained is added to the data set D through finite element simulation calculation. m (X,Y), start the next loop until the loop termination condition is met.

2. The small sample multi-objective integrated optimization design method for vehicle body structure according to claim 1 is characterized in that: The integrated learning agent model includes four base learners: Xgboost, RF, CatBoost and LightGBM.

3. The small sample multi-objective integrated optimization design method for vehicle body structure according to claim 1, characterized in that: Step 3 includes: Step 3.1: Dataset D m (X,Y) is divided into training data and test data. Four base learners are trained separately in the training data to obtain four sets of prediction results. Step 3.2: In the entire training data set, taking the actual value as the benchmark, calculate the similarity between the predicted value of each base learner and the actual value, and calculate the global weight through the similarity between the four groups of predicted results and the actual value; Step 3.3: Take the four sets of prediction results of the base learner in step 3.1 as input and the actual performance values ​​in the training data set as output, train the linear regression model and obtain the local ensemble model; Step 3.4: In the test set data or new data, the local weight of each test sample point is obtained by calculating the distance between the prediction value of the base learner and the prediction value of the local ensemble model; Step 3.5: Generate adaptive weights by integrating local weights and global weights, and complete the construction of the integrated learning agent model based on the adaptive weights.

4. The small sample multi-objective integrated optimization design method for vehicle body structure according to claim 1, characterized in that: In step 4, the multi-objective optimization algorithm is a multi-objective marine predator optimization algorithm integrating multiple strategies. The specific steps of constructing the multi-objective marine predator optimization algorithm integrating multiple strategies include: Step 4.1: Initialize algorithm parameters and population: Initialize the population based on the chaos strategy to optimize the spatial distribution of the initial population; Step 4.2: Calculate the fitness value of each population; Step 4.3: Construct the optimal solution set A1 by non-dominated sorting and calculating the congestion distance; Step 4.4: Iterate and update the population position, calculate the fitness value of the updated population, and obtain a new optimal solution set A2; Step 4.5: Merge sets A1 and A2, and obtain a new optimal solution set through non-dominated sorting and crowding distance calculation, and replace A1.

5. The small sample multi-objective integrated optimization design method for vehicle body structure according to claim 4 is characterized in that: In step 4.4, the iterative update process is divided into three stages: initial exploration, mid-term development, and final convergence. Each stage corresponds to at least two update strategies. The specific update strategy selection is based on a random parameter when and When , the corresponding update strategies are adopted respectively; Indicates the first stage: when When , the update strategy is when When , the update strategy is Indicates the second stage: when When , the update strategy is in when When , the update strategy is Indicates the third stage: when When , the update strategy is in when When , the update strategy is In the update strategy formula: T max Indicates the maximum number of iterations; represents the decision factor, a random number between 0 and 1; t represents the current number of iterations; P i represents the i-th population set; E represents the elite population; M B represents the Brownian motion strategy; F represents the inferior population; M L represents the Levy motion strategy; q, r, c represent random parameters; k represents the adjustment factor; CF represents the adaptive correction coefficient; P k1 , P k2 Represents two randomly selected individuals from the population.

6. The small sample multi-objective integrated optimization design method for vehicle body structure according to claim 4 is characterized in that: In step 4.4, the termination condition of the iterative update process is that the maximum number of iterations is reached or the change in the Pareto front solution set is less than the preset threshold.

7. The small sample multi-objective integrated optimization design method for vehicle body structure according to claim 1, characterized in that: In step 5, in non-first optimization cycles, the Pareto solution set obtained in the previous optimization is used as the initial population for the next optimization iteration.

8. The small sample multi-objective integrated optimization design method for vehicle body structure according to claim 3 is characterized in that: In step 3.2, the global weight is used to assign a stable weight to each base learner. The global weight is determined by the following formula: In step 3.4, the local weight is used to dynamically adjust the weight of each base learner at different sample points to make the model more flexible at the local level. The local weight is determined by the following formula: In step 3.5, the adaptive weight is used to combine the global weight and the local weight so that the model has the characteristics of both global weight and local weight. The adaptive weight is determined by the following formula: Where: GW k represents the global weight of base learner k; S k Indicates the similarity between the prediction result of a base learner and the true value; Represents the local weight of the base learner k at the i-th sample point; represents the prediction error of the local ensemble model and the base learner at the i-th sample point; k represents four base learners, Xgboost (X), RF (R), CatBoost (C) and LightGBM (L); Represents adaptive weight.

9. The small sample multi-objective integrated optimization design method for vehicle body structure according to claim 8, characterized in that: The similarity is determined by the following formula: Where: m represents the number of samples; represents the predicted value of one of the four base learners for the i-th sample point, Represents the actual value of the i-th sample point.

10. A small sample multi-objective integrated optimization design system for vehicle body structure, characterized in that: include: The parametric modeling module is used to establish a parametric model of the vehicle body structure, define design variables, and obtain a small number of design variable combinations through experimental design based on the design space of the design variables; The simulation module is used to obtain the body structure performance corresponding to different design variable combinations using finite element simulation technology, and generate a data set D containing body structure design variables and performance. m (X,Y); Integrated learning agent model module for m (X,Y), taking the body structure design variables as input and the body structure performance as output, an integrated learning agent model combining multiple base learners is constructed; The optimization module is used to optimize the body structure design variables based on the integrated learning agent model using a multi-objective optimization algorithm to obtain the Pareto frontier and its corresponding solution set to achieve the optimal body structure performance; The control module is used to determine whether the loop termination condition is met. If the termination condition is met, the loop is terminated and the optimization result is output; If it is not reached, the solution in the Pareto solution set is transferred to the simulation module and added to the data set D through finite element simulation calculation. m (X,Y), and repeat the optimization process until the loop termination condition is met.

Images