Automobile part optimization design method based on machine learning

Abstract

The invention discloses an automobile part optimization design method based on machine learning, and relates to the field of automobile part design optimization, and the method comprises the steps: constructing and simplifying a geometric model of an automobile part; collision simulation is carried out through finite element analysis, and a stress and deformation cloud picture is obtained; selecting sample points in a design parameter space by adopting Latin hypercube sampling; performing simulation on the sample points and extracting preset node displacement data to construct a data set; carrying out dimension reduction processing on the data set by utilizing an intrinsic orthogonal decomposition method; dividing the data after dimension reduction into a training set and a test set, constructing a Kriging agent model and performing precision verification; taking the minimization of the deformation displacement of the part in the specified direction as a target, taking the maximum stress as a constraint, and adopting a sequence least square programming algorithm to carry out optimization solution to obtain an optimal design parameter combination; and finally updating the part design. According to the method, finite element simulation, data dimension reduction and an agent model are combined, and the optimization design efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of automotive component design optimization, specifically a machine learning-based automotive component optimization design method. Background Technology

[0002] In the field of mechanical design and manufacturing, especially in the automotive industry, structural optimization design is a crucial step in improving product performance, safety, and economy. Traditional optimization design methods mainly rely on a combination of mathematical optimization algorithms and computer-aided engineering (CAE) technology. Engineers establish geometric models of the structure and use finite element analysis (FEA) software for simulation calculations to iteratively optimize the structure's topology, shape, and dimensions. This method can, to some extent, replace costly and time-consuming physical experiments, solving complex structural analysis problems that are difficult to perform manually with high computational accuracy. This improves design reliability and significantly reduces testing costs in the early stages of product development.

[0003] However, with the rapid development of modern engineering technology, engineering structures such as automotive parts are evolving towards functional integration and complex forms, with their scale and analytical dimensions increasing daily. Against this backdrop, the limitations of traditional CAE optimization methods are becoming increasingly apparent. First, the optimization process heavily relies on the engineer's experience. Designers must pre-define the design area, set optimization objectives, and accurately define loads and boundary conditions. This process involves a large amount of subjective judgment and repetitive human work, resulting in low levels of intelligence and automation. Second, for complex engineering problems, the number of design parameters involved in optimization design is often numerous, and these parameters may have strong nonlinear coupling relationships. To find the globally optimal or satisfactory solution within the vast design space, massive finite element simulation iterations are required. Each high-fidelity finite element simulation consumes significant computational resources and time. When the number of iterations reaches hundreds or thousands, the entire optimization cycle becomes extremely long, severely restricting design efficiency and making it difficult to meet the rapid iteration development needs of modern products.

[0004] Specifically, the design of automotive safety components, such as energy-absorbing boxes and crash beams, directly impacts occupant safety through collision performance. Optimizing these components typically aims to maximize energy absorption or minimize intrusion while meeting stringent strength constraints. Traditional simulation-based optimization methods, such as parametric scanning or conventional optimization algorithms, become computationally prohibitively expensive when dealing with multiple design variables and highly nonlinear collision dynamics problems, making thorough optimization exploration within the project lifecycle impractical.

[0005] In recent years, machine learning and artificial intelligence technologies have demonstrated enormous potential across various industries. In the field of engineering optimization, researchers have begun exploring the introduction of machine learning techniques into structural optimization design processes to overcome the bottlenecks of traditional methods. Machine learning models, especially surrogate models (also known as meta-models or approximate models) such as Kriging models, can construct high-dimensional nonlinear mappings between design variables and system responses by learning and fitting a finite number of sample data. Once a high-precision surrogate model is established, the performance of new design schemes can be predicted within a fraction of a second, thus replacing computationally expensive finite element simulations and making it possible to complete the exploration and optimization of large-scale design spaces in a short time.

[0006] While surrogate model-based optimization methods demonstrate advantages, their effectiveness is highly dependent on the quality and representativeness of the sample data, as well as the accuracy of the model itself. The key to reducing upfront computational costs lies in constructing a high-accuracy surrogate model using as few simulation samples as possible. Furthermore, for problems containing a large amount of output response data (such as displacement and stress data for all nodes in the entire structure), directly using all the data for machine learning model training faces the "curse of dimensionality," leading to difficult training, low efficiency, and poor stability. Therefore, while developing efficient machine learning optimization methods, it is essential to combine them with effective data dimensionality reduction techniques, such as intrinsic orthogonal decomposition (POD), to reduce the dimensionality of high-dimensional simulation data. This significantly reduces the data dimensionality while preserving most of the original data's feature information, laying a solid foundation for subsequently constructing accurate and efficient machine learning surrogate models.

[0007] In summary, developing an automotive component optimization design method that integrates intelligent sampling, data dimensionality reduction, machine learning model construction, and rapid optimization algorithms to address the problems of high computational cost, long cycle, and low efficiency in traditional CAE optimization has become a pressing technical challenge in this field. Summary of the Invention

[0008] The purpose of this invention is to provide a machine learning-based optimization design method for automotive parts. By combining finite element simulation, intrinsic orthogonal decomposition dimensionality reduction and Kriging proxy model, a high-precision approximate model is efficiently constructed, and the sequential least squares programming algorithm is used for optimization solution. Thus, while ensuring the structural strength of the parts (such as satisfying the maximum stress constraint), the deformation displacement in the specified direction is significantly reduced, and finally, the simultaneous optimization design of lightweight automotive parts and crashworthiness is achieved.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based automotive component optimization design method, comprising the following steps:

[0010] S1. Construct geometric models of automotive parts and simplify the models;

[0011] S2. Based on the simplified geometric model, a finite element model of the automotive component is established using finite element analysis software. Material parameters, load conditions, and constraint conditions are set, and a collision simulation analysis is performed to obtain stress cloud map and deformation cloud map data of the automotive component during the collision process.

[0012] S3. Using the Latin hypercube sampling method, multiple sample points are selected within the design parameter space of the automotive parts;

[0013] S4. For each sample point selected in step S3, perform the collision simulation analysis described in step S2, extract the displacement data of the preset nodes, and construct a dataset for machine learning.

[0014] S5. The dataset obtained in step S4 is subjected to dimensionality reduction using the intrinsic orthogonal decomposition (POD) method. The POD method is implemented through singular value decomposition (SVD), specifically including:

[0015] Construct the response matrix Q;

[0016] Perform singular value decomposition on Q: Q = U × ∑ × V T , where U is the left singular matrix, V is the right singular matrix, and is the singular value matrix;

[0017] The order of reduction, k, is determined according to the formula, where

[0018] For eigenvalues, For singular values, p ≥ 0.95 is required;

[0019] S6. Divide the dimensionality-reduced dataset from step S5 into a training set and a test set. Use the training set to construct a Kriging surrogate model. The expression of the Kriging surrogate model is as follows:

[0020] Y(x) = β0 + Z(x), where β0 is a deterministic value of the global trend model.

[0021] Z(x) is a function with mean 0 and variance . static stochastic process

[0022]

[0023] S7. Use the test set to verify the accuracy of the Kriging agent model established in step S6, and calculate the decision coefficients R. 2 and root mean square error (RMSE), the R 2 The calculation formula is: The formula for calculating RMSE is:

[0024]

[0025] S8. Based on the Kriging surrogate model verified in step S6, with the objective function being the minimization of the deformation displacement of the automotive component in a specified direction, and the constraint being the maximum stress of the automotive component, the Sequential Least Squares Programming (SLSQP) algorithm is used to optimize and solve the problem to obtain the optimal combination of design parameters; the optimization model is expressed as:

[0026] MinF(x)

[0027]

[0028] Where F(x) is the objective function, representing the displacement in the specified direction; Smax is the maximum stress borne by the automotive component; Smin is the minimum stress borne by the automotive component; x1 is the thickness design parameter; x2 is the elastic modulus design parameter;

[0029] S9. Update the design of the automotive parts based on the optimal design parameter combination obtained in step S8.

[0030] Furthermore, in step S1, the simplification process includes removing bolts and welded nuts from the automotive parts.

[0031] Further, in step S2, the automotive component is an automotive energy-absorbing box, and the material parameters include an elastic modulus of 200 GPa and a density of 7.22 g / cm³. 3 The load condition is set such that the energy-absorbing box impacts the planar rigid wall at a speed of 8.94 m / s.

[0032] Furthermore, the number of sample points is 9.

[0033] Furthermore, in step S4, the preset node is node number 259, and the extracted data is the displacement data of this node in the Z-axis direction.

[0034] Further, in step S6, the covariance of the random variable Z(x) at different spatial locations in the Kriging proxy model is expressed as:

[0035] cov(Z(x),Z(x′))=σ 2 H(x,x′), where the correlation function

[0036] The parameter θ is determined by the maximum likelihood estimation method.

[0037] Further, in step S7, the decision coefficient R 2 The requirement is greater than 96%.

[0038] Furthermore, including:

[0039] The geometric modeling and simplification module is used to perform step S1;

[0040] The finite element simulation analysis module is used to perform collision simulation analysis in steps S2 and S4.

[0041] The sample sampling module is used to execute step S3;

[0042] The data dimensionality reduction module is used to execute step S5;

[0043] The proxy model construction and verification module is used to execute steps S6 and S7;

[0044] The optimization solution module is used to execute step S8;

[0045] The design output module is used to execute step S9.

[0046] Furthermore, when the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 7.

[0047] Furthermore, when the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

[0048] This invention provides a machine learning-based method for optimizing the design of automotive components, which has the following advantages:

[0049] 1. Improved optimization efficiency and reduced computational cost: This method constructs a dataset by selecting representative sample points through Latin hypercube sampling and uses intrinsic orthogonal decomposition (POD) to reduce the dimensionality of the data, effectively reducing the amount of data required for subsequent modeling. Based on this, a Kriging surrogate model is established to replace complex finite element simulation, transforming the optimization solution process from time-consuming physical simulation to efficient mathematical model calculation, significantly shortening the design cycle and saving computational resources.

[0050] 2. Ensuring model accuracy and optimization reliability: The method explicitly includes a validation step for the surrogate model, which is achieved by calculating the decision coefficients (R²). 2 Quantitative metrics such as R0 and root mean square error (RMSE) are used to evaluate the model's predictive accuracy, and R0 is required to be... 2 Greater than 96%. This stringent standard ensures that the surrogate model can accurately reflect the behavior of the real physical system, providing a reliable foundation for subsequent optimization solutions, thereby guaranteeing the accuracy and effectiveness of the optimization results.

[0051] 3. Achieving Multi-Objective Collaborative Optimization Design: The optimization model of this method explicitly aims to minimize deformation displacement in a specific direction, while using the maximum stress of components as a constraint. This setting allows the optimization process to strictly ensure structural strength (controlling stress levels) while pursuing lightweighting or performance improvement (reducing deformation), achieving collaborative optimization of performance and safety, and contributing to obtaining a design scheme with better overall performance.

[0052] 4. Standardized and Systematized Process: The claims define in detail the complete steps from geometric modeling, simulation analysis, data sampling and processing, surrogate model construction and verification to final optimization solution. This structured process makes the optimization design process of automotive parts clear and repeatable, reduces over-reliance on the experience of individual experts, and facilitates the standardized application of this method and its promotion in different part design projects.

[0053] 5. Enhanced Engineering Practicality and Guiding Value: The methodology specifies key steps in detail, such as removing bolts and welded nuts during simplification, clarifying the material parameters and impact conditions of the energy-absorbing box, and specifying the nodes from which data is to be extracted. These specific provisions make the methodology more aligned with engineering practice, provide clear operational guidance, and enhance its practicality and direct guiding value in the design of specific automotive components (such as energy-absorbing boxes). Attached Figure Description

[0054] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0055] Figure 1 This is a flowchart of the overall method of the present invention;

[0056] Figure 2 Flowchart of data preparation and dimensionality reduction for this invention;

[0057] Figure 3 This is a flowchart illustrating the proxy model construction and verification process of the present invention.

[0058] Figure 4 The flowchart for optimizing the solution process in this invention is shown below;

[0059] Figure 5 This is a flowchart illustrating the system module structure of the present invention. Detailed Implementation

[0060] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0062] How to use:

[0063] Geometric Model Construction and Simplification (S1): Users first need to accurately draw the three-dimensional geometric model of the component to be optimized in professional CAD (Computer-Aided Design) software. Subsequently, the model must be simplified as necessary, such as removing local features that have little impact on the overall mechanical properties (such as bolt holes, weld nuts, etc. as described in claim 2 below), in order to improve the efficiency and stability of subsequent simulation analysis.

[0064] Finite Element Model Establishment and Simulation (S2): Import the simplified geometric model into finite element analysis software (such as Abaqus, ANSYS, etc.). In this step, the user needs to strictly define the material properties of the components (such as density, elastic modulus) and accurately set their boundary conditions (such as fixed constraints) and load conditions (such as impact velocity, force) under actual working conditions. After completing the settings, perform a collision simulation analysis and obtain key physical field data, namely stress contour maps and deformation contour maps.

[0065] Experimental Design Sampling (S3): Determine the design parameters to be optimized (such as thickness, material properties) and their range of variation. Users need to use the Latin hypercube sampling method to scientifically select a certain number of sample points within this design space. This method ensures that the sample points uniformly cover the entire parameter space, obtaining the most representative data with fewer simulations.

[0066] Sample point data extraction (S4): For the geometric model corresponding to each sample point selected in step S3, repeat the collision simulation analysis in step S2. After the simulation is completed, the user needs to extract the displacement data of the pre-set representative observation points (the displacement of node 259 in the Z direction as described in claim 5) from the result file, and organize the data of all sample points into a structured dataset.

[0067] Data Dimensionality Reduction (S5): Since the extracted displacement data may have high dimensionality, directly using it for modeling is inefficient. Therefore, users need to use intrinsic orthogonal decomposition (IOD) to reduce the dimensionality of the dataset. Specifically, this involves constructing a response matrix and performing singular value decomposition on it. By calculating the cumulative energy contribution rate (typically setting a threshold p ≥ 0.95), the number of principal components to be retained, k, is automatically determined, thus using a few principal components to efficiently represent most of the features of the original data.

[0068] Proxy Model Construction (S6): The dimensionality-reduced dataset is divided into training and test sets. A Kriging surrogate model is constructed using the training set data. This model establishes a mathematical mapping relationship between the design parameters and the response after dimensionality reduction. Users need to adjust the relevant function parameters in the model according to the data characteristics (as described in claim 6) to obtain the optimal model fit.

[0069] Proxy Model Validation (S7): The accuracy of the constructed Kriging proxy model is validated using reserved test set data. The decision coefficients R are calculated. 2 The two metrics, R and RMSE (as described in claim 7), are: 2 Typically, a score greater than 96% is required. Only validated models can be used for subsequent optimization; otherwise, it is necessary to return to check the previous steps or adjust the model parameters.

[0070] Optimization Solution (S8): After the surrogate model has been validated, a clear optimization mathematical model is established. This model uses the performance indicators of the components (such as minimizing deformation displacement in a specified direction) as the objective function and their mechanical performance requirements (such as the maximum stress not exceeding the allowable value) as constraints. The user then calls optimization algorithms such as sequential least squares programming to quickly find the optimal combination of design parameters that satisfies all constraints.

[0071] Design Update (S9): The optimal parameter combination (such as optimal thickness and elastic modulus) obtained in step S8 is fed back into the original CAD geometric model to complete the final update of the design.

[0072] When simplifying the model in the first step (S1) of the method, small connecting structures such as bolts and welded nuts should be explicitly removed from the geometric model.

[0073] When the optimized component is an automotive energy-absorbing box, in the second step (S2), the material parameters should be set to an elastic modulus of 200 GPa and a density of 7.22 g / cm³. 3 The load conditions should be set so that the energy-absorbing box impacts the rigid wall at a speed of 8.94 m / s.

[0074] When performing Latin hypercube sampling in the third step (S3), the number of sample points should be set to 9.

[0075] In the fourth step (S4) of extracting simulation data, it is necessary to focus on and extract the displacement data of node 259 in the Z-axis direction.

[0076] In step 6 (S6), when constructing the Kriging agent model, the spatial correlation of the stochastic process Z(x) should be described using a specified exponential correlation function, the parameter θ of which needs to be determined by the maximum likelihood estimation method.

[0077] In step seven (S7), when verifying the accuracy of the surrogate model, the decision coefficient R... 2 The qualification criterion should be set to greater than 96%.

[0078] Users sequentially invoke the functions of each module according to the system's modular structure. First, the "Geometric Modeling and Simplification Module" prepares the model; next, the "Sample Sampling Module" generates the experimental design scheme; then, the "Finite Element Simulation Analysis Module" automatically or semi-automatically performs simulation calculations for all sample points; the "Data Dimensionality Reduction Module" then processes the massive amounts of data generated; the processed data is sent to the "Proxy Model Construction and Verification Module" for model training and verification; the verified model is then passed to the "Optimization Solution Module" for parameter optimization; finally, the optimization results are used to generate a new design scheme through the "Design Output Module." Users must follow this modular operating system process.

[0079] A computer program (such as a software installation package) that implements the method steps described in any one of claims 1 to 7 is installed on a device with computing capabilities (such as a workstation or server). When the processor in the device executes the program, it automatically implements the aforementioned optimization design method. The user can complete the entire optimization process by operating the software interface, inputting initial parameters, and starting the calculation.

[0080] Example:

[0081] Example 1: Lightweight Design of Automotive Energy Absorbing Box

[0082] This embodiment focuses on the lightweight design of the energy-absorbing box in the front longitudinal beam of a certain model of automobile. First, engineers used CAD software to construct a detailed geometric model of the energy-absorbing box and, following the guidance of claim 2, removed the mounting bolt holes on the model to simplify mesh generation. Subsequently, in the finite element analysis software, based on the settings of claim 3, the model was assigned an elastic modulus of 200 GPa and a density of 7.22 g / cm³. 3The material properties were determined, and the boundary condition of impacting a rigid wall at a speed of 8.94 m / s was set. Initial collision simulation was performed, and large deformation displacement was observed. Next, the engineers used the wall thickness (x1) and elastic modulus (x2) of the energy-absorbing box as design variables, and selected 9 sample points in the design space using the Latin hypercube sampling method according to claim 4. After simulation for each sample point, the displacement data of node 259 specified in claim 5 in the Z-axis direction was extracted to construct a dataset. After dimensionality reduction using the POD method, a Kriging surrogate model was established, and its accuracy was verified to meet R... 2 The requirement was >96%. Finally, with the goal of minimizing the Z-direction displacement and the constraint that the maximum stress does not exceed the material's yield strength, the SLSQP algorithm was used for optimization. The optimization results yielded a new combination of thickness and elastic modulus parameters, achieving a weight reduction of approximately 8% while maintaining energy absorption performance.

[0083] Example 2: Vibration performance optimization of engine mount

[0084] This embodiment aims to optimize the engine mount of a commercial vehicle to improve its vibration characteristics. Engineers first simplified the complex geometry of the mount, removing unnecessary rounded corners and grooves. Then, a finite element model was established, and harmonic response analysis was performed using periodic loads from engine operation to obtain dynamic strain data of the mount at different frequencies. The design parameters were selected as the thickness and width of the main support arms of the mount. Twenty sample points were selected using the Latin hypercube sampling method for simulation, extracting vibration displacement data at the connection point between the mount and the engine. Due to the large amount of dynamic response data, an intrinsic orthogonal decomposition method was used for effective dimensionality reduction, retaining 99% of the energy information. A high-precision Kriging proxy model was constructed using the dimensionality-reduced data, with decision coefficients R0. 2 The accuracy rate reached 97.5%. During the optimization phase, the objective function was minimizing the amplitude at a specific key frequency, with the maximum equivalent stress of the support as the constraint. The final optimized design reduced engine vibration transmission by 15% during idling, improving cabin comfort.

[0085] Example 3: Improved side-impact performance of door anti-collision beams

[0086] To improve the performance of a certain vehicle model in side-impact safety tests, this embodiment optimizes its door anti-collision beam. After simplification, the initial model was simulated in finite element software to analyze its bending performance. Design parameters included the cross-sectional shape parameters of the anti-collision beam (such as height and width) and the yield strength of the material. Fifteen design samples were generated using a sampling method, and simulation calculations were performed to extract the intrusion displacement data of the middle part of the anti-collision beam during the collision process. The constructed dataset was dimensionality-reduced to establish a surrogate model reflecting the relationship between design parameters and intrusion amount. After the model accuracy verification was passed, the optimization objective was set as minimizing the intrusion displacement at a specific time point in the early stages of the collision, while constraining its maximum deformation within a certain range to ensure occupant survival space. A set of optimal cross-sectional dimensions and material parameters was obtained by solving the problem using a sequential least squares programming algorithm. The anti-collision beam design updated based on these results, verified by whole-vehicle-level side-impact simulation, showed a significant reduction in intrusion speed, effectively improving side-impact safety.

[0087] Example 4: Impact-resistant design at the bottom of the battery pack casing

[0088] Given the impact resistance requirements of the bottom of the electric vehicle battery pack, this embodiment optimizes its shell structure. Engineers constructed a battery pack shell model including mounting points and simulated bottom scratches and impacts in the simulation. The shell bottom plate thickness, stiffener spacing, and material density were selected as key design parameters. Following a sampling strategy, 12 sample points were selected for finite element analysis to collect stress distribution data in key areas of the shell bottom. Due to the extremely high dimensionality of the stress cloud map data, singular value decomposition (SVD) was used for intrinsic orthogonal decomposition, reducing the data from thousands of nodes to a dozen or so principal modes, greatly improving modeling efficiency. The surrogate model trained based on the dimensionality-reduced data can quickly predict stress concentration under different design parameters. The optimization process primarily aimed to reduce overall weight, but strictly limited the maximum stress value to below the allowable stress of the material. The final optimized scheme successfully reduced the shell weight by approximately 5% while meeting impact resistance requirements, contributing to an increase in the vehicle's driving range.

[0089] Example 5: Co-optimization of fatigue life and weight of control arm

[0090] This embodiment focuses on the lower control arm of an SUV model, aiming to achieve lightweight design while ensuring fatigue life. During the geometric model preparation phase, the main force transmission path characteristics were preserved. In the finite element analysis, in addition to static stiffness analysis, fatigue life simulation was performed based on the load spectrum to obtain the number of life cycles corresponding to each sample design. Design variables included the plate thickness of the control arm and the material thickness at different locations. Ten sample points were generated using a sampling method, and the displacement response and minimum fatigue life value at the connection points between the control arm and the body and wheels were extracted after simulation. Proxy models were constructed for the displacement and life data respectively. In the multi-objective optimization setting, the weighted sum of weight and displacement was used as the objective function, while fatigue life was a necessary constraint. The optimization algorithm was used to quickly explore the "virtual design space" composed of the proxy models, finding a series of Pareto optimal solutions. Engineers selected the design scheme with the lightest weight while meeting the design requirement of 200,000 kilometers of fatigue life, successfully achieving a balance between performance and lightweight design.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based optimization design method for automotive parts, characterized in that, Includes the following steps: S1. Construct geometric models of automotive parts and simplify the models; S2. Based on the simplified geometric model, a finite element model of the automotive component is established using finite element analysis software. Material parameters, load conditions, and constraint conditions are set, and a collision simulation analysis is performed to obtain stress cloud map and deformation cloud map data of the automotive component during the collision process. S3. Using the Latin hypercube sampling method, multiple sample points are selected within the design parameter space of the automotive parts; S4. For each sample point selected in step S3, perform the collision simulation analysis described in step S2, extract the displacement data of the preset nodes, and construct a dataset for machine learning. S5. The dataset obtained in step S4 is subjected to dimensionality reduction using the intrinsic orthogonal decomposition (POD) method. The POD method is implemented through singular value decomposition (SVD), specifically including: Construct the response matrix Q; Perform singular value decomposition on Q: Q = U × Σ × V T , where U is the left singular matrix, V is the right singular matrix, and is the singular value matrix; The order of reduction, k, is determined according to the formula, where For eigenvalues, For singular values, p ≥ 0.95 is required; S6. Divide the dimensionality-reduced dataset from step S5 into a training set and a test set. Use the training set to construct a Kriging surrogate model. The expression of the Kriging surrogate model is as follows: Y(x) = β0 + Z(x), where β0 is the deterministic value of the global trend model, and Z(x) is a value with a mean of 0 and a variance of 0. static stochastic process S7. Use the test set to verify the accuracy of the Kriging agent model established in step S6, and calculate the decision coefficients R. 2 and root mean square error (RMSE), the R 2 The calculation formula is: The formula for calculating RMSE is: S8. Based on the Kriging surrogate model verified in step S6, with the objective function being the minimization of the deformation displacement of the automotive component in a specified direction, and the constraint being the maximum stress of the automotive component, the Sequential Least Squares Programming (SLSQP) algorithm is used to optimize the solution and obtain the optimal combination of design parameters; the optimization model is expressed as: Min F(x) Where F(x) is the objective function, representing the displacement in the specified direction; Smax is the maximum stress borne by the automotive component; Smin is the minimum stress borne by the automotive component; x1 is the thickness design parameter; x2 is the elastic modulus design parameter; S9. Update the design of the automotive parts based on the optimal design parameter combination obtained in step S8.

2. The machine learning-based automotive component optimization design method according to claim 1, characterized in that: In step S1, the simplification process includes removing bolts and welded nuts from the automotive parts.

3. The machine learning-based automotive component optimization design method according to claim 1, characterized in that: In step S2, the automotive component is an automotive energy-absorbing box, and the material parameters include an elastic modulus of 200 GPa and a density of 7.22 g / cm³. 3 The load condition is set such that the energy-absorbing box impacts the planar rigid wall at a speed of 8.94 m / s.

4. The machine learning-based automotive component optimization design method according to claim 1, characterized in that: In step S3, the number of sample points is 9.

5. The machine learning-based automotive component optimization design method according to claim 1, characterized in that: In step S4, the preset node is node 259, and the extracted data is the displacement data of this node in the Z-axis direction.

6. The machine learning-based automotive component optimization design method according to claim 1, characterized in that: In step S6, the covariance of the random variable Z(x) at different spatial locations in the Kriging proxy model is expressed as: cov(Z(x), Z(x′))=σ 2 H(x, x′), where the correlation function The parameter θ is determined by the maximum likelihood estimation method.

7. The machine learning-based automotive component optimization design method according to claim 1, characterized in that: In step S7, the decision coefficient R 2 The requirement is greater than 96%.

8. A system for implementing the machine learning-based automotive component optimization design method according to any one of claims 1 to 7, characterized in that, include: The geometric modeling and simplification module is used to perform step S1; The finite element simulation analysis module is used to perform collision simulation analysis in steps S2 and S4. The sample sampling module is used to execute step S3; The data dimensionality reduction module is used to execute step S5; The proxy model construction and verification module is used to execute steps S6 and S7; The optimization solution module is used to execute step S8; The design output module is used to execute step S9.

9. The machine learning-based automotive component optimization design method according to claim 1, characterized in that: When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 7.

10. The machine learning-based automotive component optimization design method according to claim 1, characterized in that: When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images