Generative model driven vehicle body beam structure section optimization method and system

Abstract

The invention relates to a generative model-driven vehicle body beam structure section optimization method and system, and the method comprises the steps: enabling a section shape formed by connecting control points after a structure section grayscale image is gridded to represent a structure section, constructing a section shape database in combination with process constraints, and providing high-quality training data and a standardized design space for a generative model; the generative model geometrically compresses a high-dimensional section into a low-dimensional submerged space, design dimension reduction and feature abstraction are achieved, and efficient sampling and shape generation are achieved; forming an input parameter by the generative model, and forming an initial training set of the proxy model in combination with finite element simulation marked mechanical properties to cover a global design space; the agent model replaces high-cost simulation, a newly-added sample is obtained in combination with an optimization point adding criterion so as to verify the iterative agent model, input parameters giving consideration to the structural section shape and the mechanical property are obtained, the final section shape is determined according to the input parameters, the method does not depend on experience of a designer and a test simulation result, and the process is simplified.

Description

Technical Field

[0001] The present application relates to the field of vehicle body structure optimization design, and in particular to a generative model-driven vehicle body beam structure section optimization method and system. Background Art

[0002] Lightweight design is a perennial theme in vehicle R&D. Research shows that the body structure influences over 60% of the vehicle's stiffness. Furthermore, the body accounts for the majority of the vehicle's weight, and its structural design is closely linked to the vehicle's structural form and overall performance. The conceptual design of the body accounts for approximately 70% of the vehicle's design cycle, typically using a thin-walled beam structure as the framework. The cross-sectional shape and thickness of the beam structure directly determine its weight, stiffness, and manufacturability.

[0003] However, traditional cross-section design is mainly based on reverse engineering, which mainly involves modifying existing models. This process is cumbersome, has a high number of iterations, and is highly dependent on the designer's experience and experimental simulation results. Summary of the Invention

[0004] The embodiments of the present application provide a generative model-driven vehicle body beam structure cross-section optimization method and system to address the problem in related technologies that traditional cross-section design is mainly based on reverse engineering, mainly modifying the existing model. This process is cumbersome, has a high number of iterations, and is highly dependent on the designer's experience and experimental simulation results.

[0005] In a first aspect, a generative model-driven vehicle body beam structure section optimization method is provided, comprising: The grayscale image of the structural section is gridded and constraints are set to sample control points; the control points are connected to obtain the cross-sectional shape; and then classified and stored to form a cross-sectional shape database; Training and forming a generative model based on the cross-sectional shape database; According to the cross-section design requirements, control points of the cross-section to be optimized are collected to obtain sampling data; the sampling data are reduced in dimension using the generative model to obtain input parameters; the input parameters are solved by simulation mechanics, and the calculated mechanical response is used as an output label; the input parameters and the output label are combined into a data set; The proxy model is constructed using the data set, and then new samples are obtained in combination with the optimization point addition criterion to verify the iterative proxy model and determine the final input parameters of the proxy model; and an optimized cross-sectional shape is formed according to the final input parameters.

[0006] In some embodiments, the size of the grayscale image is determined based on the degree of impact on the training time and computing power of the generative model, and the size of the structural section; The Latin hypercube design method, Monte Carlo sampling method or uniform design method is used to sample control points in a specified area of ​​the grayscale image after the size is determined. During the sampling process, the shortest line segment constraint and the stamping angle constraint are set to ensure that the cross-sectional shape obtained by connecting the control points meets the manufacturing process requirements.

[0007] In some embodiments, a generative model is built, and the ratio of convolutional layers / deconvolutional layers and fully connected layers in the encoder, decoder, and discriminator of the generative model, as well as the convolution kernel size, sampling step, dropout layer, activation function, batch processing, number of iterations, optimizer, and loss function value are set; The generative model is used to extract the geometric features of the cross-sectional shape database for training, and is continuously adjusted until the dimension of the latent space of the generative model, the balanced image feature extraction effect, and the latent space dimension all meet the optimization efficiency threshold.

[0008] In some embodiments, a Latin hypercube design method, a Monte Carlo sampling method, or a uniform design method is used to sample control points in a specified area of ​​a grayscale image of the section to be optimized to obtain the sampled data; during the sampling process, a shortest line segment constraint and a punching angle constraint are set to ensure that the cross-sectional shape obtained by connecting the control points meets the manufacturing process requirements; The generative model is used to reduce the dimensionality of the sampled data to the latent space dimension, and the latent space representation after dimensionality reduction is used as an input parameter.

[0009] In some embodiments, the finite element model is constructed by means of a parameterized script using Abaqus or Ansys finite element simulation software; The finite element model is used to solve the stiffness and stability of each control point in the input parameters to obtain the corresponding mechanical response; the mechanical response corresponding to each control point is used as a label to obtain the output label of the sampled data.

[0010] In some embodiments, a proxy model is constructed using the dataset, and then new samples are obtained in combination with an optimization point addition criterion to verify the iterative proxy model and determine the final input parameters of the proxy model; and an optimized cross-sectional shape is formed based on the final input parameters, specifically including the following steps: The input parameters are used as input and the output labels are used as output, and an approximate mapping relationship between the input and the output is constructed by combining a radial basis model, a kriging model, and a neural network model to form a proxy model; Obtaining new samples based on the data set using an optimized point addition criterion; Use the proxy model to take the new samples as input parameters to predict the output labels of the new samples; Update the proxy model based on the newly added samples and their output labels, and perform finite element verification until the iteration reaches convergence; The input parameters corresponding to the final proxy model are used as the optimization solution of the control points to form the optimized cross-sectional shape.

[0011] In some embodiments, the optimization point addition criterion includes one of an EI criterion, a PI criterion, and a RMSE criterion; The EI criteria are:

[0012] in, ; is the optimal true objective function value among all current control points; is a random variable with mean , the variance is Normal distribution; and represent the standard normal cumulative distribution function and the standard normal distribution probability density function respectively; The PI criteria are:

[0013] in, is a value smaller than the optimal value among all current control points; is a random variable with mean , the variance is Normal distribution; represents the standard normal cumulative distribution function.

[0014] In some embodiments, obtaining new samples based on the data set using an optimized point addition criterion further includes the following steps: For the cross-sectional shape corresponding to the newly added sample, if its geometric boundary is unclear or the cross-sectional shape cannot be constructed, the norm index between it and each cross-sectional shape in the cross-sectional shape database is calculated; The cross-sectional shape corresponding to the minimum norm index in the cross-sectional shape database is used as the target cross-sectional shape; The control points corresponding to the target section are processed by the generative model and then replaced with the newly added samples.

[0015] In some embodiments, after the optimized cross-sectional shape is obtained, the optimized cross-sectional shape is updated to the cross-sectional shape database.

[0016] In a second aspect, a generative model-driven vehicle body beam structure section optimization system is provided, which includes: The first module is used to grid the grayscale image of the structural section and set constraints to sample control points; connect the control points to obtain the cross-sectional shape; and then classify and store them to form a cross-sectional shape database; A second module is used to train and form a generative model based on the cross-sectional shape database; The third module is used to collect control points of the section to be optimized according to the section design requirements to obtain sampling data; use the generative model to reduce the dimension of the sampling data to obtain input parameters; perform simulation mechanical solution on the input parameters, and then use the calculated mechanical response as the output label; and form a data set with the input parameters and output labels; The fourth module is used to construct a proxy model using the data set, and then obtain new samples in combination with the optimization point addition criterion to verify the iterative proxy model and determine the input parameters of the final proxy model; and form an optimized cross-sectional shape based on the final input parameters.

[0017] The beneficial effects of the technical solution provided by this application include: The present invention provides a generative model-driven vehicle body beam structural section optimization method and system. The system uses a cross-sectional shape formed by connecting control points after meshing a grayscale image of the structural section to represent the structural section. Combined with process constraints, a cross-sectional shape database is constructed, providing high-quality training data for subsequent generative models and normalizing the design space. The generative model compresses high-dimensional cross-sectional geometry into a low-dimensional latent space, achieving design dimensionality reduction and feature abstraction, and enabling efficient sampling and shape generation. The generative model generates input parameters and, combined with finite element simulation to label mechanical properties, forms an initial training set for a proxy model, covering the global design space. The proxy model replaces high-cost simulations, optimizes to obtain new samples based on an optimized point addition criterion, and then verifies and iterates the proxy model to obtain input parameters that balance structural cross-sectional shape and mechanical properties. The structural cross-sectional shape is restored using control points corresponding to the input parameters to determine the final cross-sectional shape. This system does not rely on the designer's experience or experimental simulation results, simplifying the process, minimizing vehicle weight while meeting overall vehicle performance indicators, improving vehicle development and design efficiency, and providing more quality design space for other functional modules, thereby enhancing vehicle design flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0019] Figure 1 This is an architecture diagram of a generative model-driven vehicle body beam structure section optimization system provided in an embodiment of the present application; Figure 2 This is an example of generating some cross-section types in the cross-section shape database provided in the embodiment of the present application; Figure 3 A schematic diagram of constraints generated by cross-sectional shape sampling provided in an embodiment of the present application; Figure 4 A schematic diagram of the general variational autoencoder network structure provided in an embodiment of the present application; Figure 5 A schematic diagram of the proxy model optimization process provided in an embodiment of the present application; Figure 6 A schematic flow chart of the general process of the generative model-driven vehicle body beam structure section optimization method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0020] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0021] The purposes of this application are: Optimizing the cross-section design of vehicle body beam structures minimizes vehicle weight while meeting overall vehicle performance targets. This helps improve vehicle development efficiency, provides more room for quality design for other functional modules, and thus enhances overall vehicle design flexibility. Therefore, this paper proposes a generative model-driven cross-section optimization method for vehicle body beam structures, targeting key cross-sections such as pillars and transverse and longitudinal beams that impact vehicle performance. This method is applicable to the cross-section optimization design of thin-walled beam structures in various types of vehicles, including commercial and specialty vehicles.

[0022] The embodiments of the present application provide a generative model-driven vehicle body beam structure cross-section optimization method and system to address the problem in related technologies that traditional cross-section design is mainly based on reverse engineering, mainly modifying the existing model. This process is cumbersome, has a high number of iterations, and is highly dependent on the designer's experience and experimental simulation results.

[0023] See also Figure 1-Figure 2 , a generative model-driven vehicle body beam structure section optimization method, comprising: Step 100: grid the grayscale image of the structural section, set constraints to sample control points, connect the control points to obtain the cross-sectional shape, and then classify and store them to form a cross-sectional shape database; Step 200: training a generative model based on a cross-sectional shape database; Step 300: Collect control points of the section to be optimized according to the section design requirements to obtain sampling data; use the generative model to reduce the dimension of the sampling data to obtain input parameters; perform a simulation mechanical solution on the input parameters, and then use the calculated mechanical response as the output label; and form a data set with the input parameters and the output label; Step 400: Build a proxy model using the data set, then obtain new samples in combination with the optimization point addition criterion to verify the iterative proxy model and determine the final input parameters of the proxy model; and form an optimized cross-sectional shape based on the final input parameters.

[0024] The above method uses the cross-sectional shape formed by connecting the control points after gridding the grayscale image of the structural cross section to represent the structural cross section. Combined with process constraints, a cross-sectional shape database is constructed, providing high-quality training data for subsequent generative models and normalizing the design space. The generative model compresses the high-dimensional cross-sectional geometry into a low-dimensional latent space, achieving design dimensionality reduction and feature abstraction, and realizing efficient sampling and shape generation. The generative model forms input parameters and, combined with finite element simulation to label mechanical properties, forms an initial training set for the proxy model, covering the global design space. The proxy model replaces high-cost simulation and optimizes to obtain new samples based on the optimization point addition criterion. Then, the proxy model is verified and iterated to obtain input parameters that take into account both the structural cross-sectional shape and mechanical properties. The structural cross-sectional shape is restored using the control points corresponding to the input parameters to determine the final cross-sectional shape. This method does not rely on the designer's experience or experimental simulation results, simplifying the process, minimizing body weight while meeting overall body performance indicators, improving body development and design efficiency, and providing more quality design space for other functional modules, thereby enhancing vehicle design flexibility.

[0025] In some preferred embodiments, step 100 is specifically as follows: Determine the size of the grayscale image based on the impact on the training time and computing power of the generative model, as well as the size of the structural section; The Latin hypercube design method, Monte Carlo sampling method or uniform design method is used to sample control points in a specified area of ​​the grayscale image after the size is determined. During the sampling process, the shortest line segment constraint and the stamping angle constraint are set to ensure that the cross-sectional shape obtained by connecting the control points meets the manufacturing process requirements.

[0026] That is, the cross-sectional shape of the vehicle body beam structure is represented in the form of image data to achieve the purpose of standardizing the entire cross-sectional design space. The image type preferably selects a single-channel grayscale image, and the pixel values ​​are divided into 0 and 255 (0 corresponds to black, 255 corresponds to white). The image size is initially set to 64×64. The image size setting needs to consider whether the cross-sectional shape information is fully represented and the impact of the data scale on the generative model network training time and computing power. The cross-sectional shape database is classified according to the cross-sectional type, including open sections, single-chamber sections, double-chamber sections, and multi-chamber sections. The generation method is mainly to control the interconnection between parameterized grid control points. The control point position coordinates are sampled within the specified area, and the number of grids, the number of control points, and the corresponding specified area are determined according to the design requirements. Figure 2 The diagram shows the cross-sectional shape generation for an opening with six control points in a 4×4 grid, as well as for a single-chamber and double-chamber cross-section with eight control points. Experimental design is performed using methods such as Latin hypercube design, Monte Carlo sampling, or uniform design. A large number of control points are sampled in a specified area. Constraints are set during the sampling process to meet manufacturing process requirements. These constraints include: a. Shortest line segment constraint For stamping process, the minimum value of the cross-section segment should satisfy the following formula:

[0027] Where, is the minimum value of the allowed cross-section line segment; and is the coordinate of the endpoint of the line segment; is the modulus of the line segment.

[0028] b. Punch angle constraint The punching angle refers to the angle between the section line direction and the punching direction. The value range of the punching angle is .in, is the minimum punching angle.

[0029] Through the above steps, the cross-sectional shape database is generated and continuously updated with different design requirements. Figure 3 shown.

[0030] In some preferred embodiments, step 200 is specifically as follows: Build a generative model and set the ratio of convolutional / deconvolutional layers and fully connected layers in the encoder, decoder, and discriminator of the generative model, as well as the convolution kernel size, sampling step, dropout layer, activation function, batch processing, number of iterations, optimizer, and loss function value; The generative model is used to extract the geometric features of the cross-sectional shape database for training, and is continuously adjusted until the dimension of the latent space of the generative model, the balanced image feature extraction effect, and the latent space dimension all meet the optimization efficiency threshold.

[0031] That is, first build a generative model. The following takes the general variational autoencoder network as an example. Its network structure is shown as follows Figure 4 shown.

[0032] The variational autoencoder network assumes that both the input and the latent space feature vectors obey the distribution assumption, so that the latent space satisfies the characteristics of smoothness and continuity, so that random interpolation and sampling can be performed. The encoder, decoder, discriminator, the ratio of convolutional layers / deconvolution layers and fully connected layers, the size of the convolution kernel, the sampling step, the dropout layer, the activation function, etc.), batch processing, the number of iterations, the optimizer, and the loss function are set. The performance of the variational autoencoder network is evaluated by the evidence lower bound (ELBO) loss, which includes two parts: reconstruction loss and KL divergence loss. Among them, the reconstruction loss is used to measure the similarity between the data generated by the decoder and the original data; the KL divergence loss is used to measure the distance between the probability distribution of the latent space feature vector and the standard normal distribution. Among them, the reconstruction loss can be defined as:

[0033] Where m is the number of samples; and In addition, the reconstruction loss can also be defined as calculating the cross entropy, structural similarity, and peak signal-to-noise ratio between the true value and the predicted value.

[0034] Assume that the latent space feature vector is n-dimensional and its distribution mean is , the variance is Normal distribution, then the KL divergence loss is defined as:

[0035] The two loss terms are combined into the final ELBO loss function through the weighted coefficient.

[0036] Once the generative model is built, it is used to extract geometric features from the cross-sectional shape database. Data preprocessing and standardization (including standardization) are performed on the cross-sectional shape database. The required cross-sectional type dataset is selected as model input and the generative model is trained. The network model is continuously trained by adjusting the hyperparameters representing the latent space dimensionality to balance the effectiveness of image feature extraction (image reconstruction) with the impact of the latent space dimensionality on optimization efficiency.

[0037] In some preferred embodiments, step 300 is specifically as follows: ① Using the Latin hypercube design method, Monte Carlo sampling method, or uniform design method, control points are sampled in a specified area of ​​the grayscale image of the section to be optimized to obtain sampling data. During the sampling process, the shortest line segment constraint and the punching angle constraint are set to ensure that the cross-sectional shape obtained by connecting the control points meets the manufacturing process requirements. Use a generative model to reduce the dimensionality of the sampled data to the latent space dimension, and use the latent space representation after dimensionality reduction as the input parameter; Specifically, based on design requirements, a design experiment is performed for the desired cross-section type using methods such as Latin hypercube design. A small number of control points are sampled within a specified area. The sample points meet manufacturing process requirements, and the initial sampling is designed to maximize the information captured in the unknown design space. Using the encoder from a well-trained generative model, the sampled data is normalized and reduced to the latent space dimension, with the reduced latent space representation used as the input parameter.

[0038] ② Use Abaqus or Ansys finite element simulation software to construct finite element models through parametric scripts; The finite element model is used to solve the stiffness and stability of each control point in the input parameters to obtain the corresponding mechanical response; the mechanical response corresponding to each control point is used as a label to obtain the output label of the sampled data.

[0039] Specifically, for different mechanical problems, including stiffness, stability, and dynamics, finite element simulation software like Abaqus and Ansys, or custom code, is used to construct finite element models through parameterized scripts and solve the mechanical responses corresponding to different control points in the input parameters. The input parameters are labeled with single or multiple mechanical responses to obtain output labels. In other words, the latent space representation after dimensionality reduction serves as the input parameter for the sampled data, and the corresponding mechanical responses serve as the output labels for the sampled data, forming a dataset.

[0040] In some preferred embodiments, step 400 specifically includes: The proxy model is constructed using the data set, and then new samples are obtained in combination with the optimization point addition criterion to verify the iterative proxy model and determine the input parameters of the final proxy model; the optimized cross-sectional shape is formed according to the final input parameters, specifically including the following steps: The input parameters are used as input and the output labels are used as output, and a quasi-mapping relationship between input and output is constructed by combining a radial basis model, a kriging model, or a neural network model to form a proxy model; Obtaining new samples based on the data set using an optimized point addition criterion; Using the proxy model, the new samples are used as input parameters to predict the output labels of the new samples. Update the proxy model based on the newly added samples and their output labels, and perform finite element verification until the iteration reaches convergence; The input parameters corresponding to the final proxy model are used as the optimization solution of the control points to form the optimized cross-sectional shape.

[0041] After the step of obtaining new samples based on the data set using the optimized point addition criterion, and before the step of using the proxy model to use the new samples as input parameters to predict the output labels of the new samples, the following steps are also included: That is, using the optimized point addition criterion to obtain new samples based on the data set, further comprising the following steps: For the cross-sectional shape corresponding to the newly added sample, if its geometric boundary is unclear or the cross-sectional shape cannot be constructed, the norm index between it and each cross-sectional shape in the cross-sectional shape database is calculated; The cross-sectional shape corresponding to the minimum norm index in the cross-sectional shape database is used as the target cross-sectional shape; The control points corresponding to the target section are processed by the generative model and then replaced with the newly added samples.

[0042] The following is a detailed description: Using the composed dataset, we build proxy models based on programming languages ​​such as Matlab or Python, including radial basis function (RBF), Kriging, and artificial neural network (ANN). The Kriging model has the most extensive research and application. The Kriging model fits the input parameters and output labels in a given dataset through interpolation to obtain a proxy model. Then, the optimization point addition criterion is used to obtain new input parameters, and then the proxy model is used to give a response prediction for the new input parameters to obtain a new output label; the new output label and the new input parameter are used to update the proxy model.

[0043] The proxy model is:

[0044] in, It is Samples corresponding to the output response The weighting coefficient of The expression provides an unbiased estimate of unknown samples. Because the computational complexity of surrogate models for predicting any point is negligible, surrogate model-based optimization often employs traditional heuristic algorithms, such as genetic algorithms. During the iterative optimization process, the optimized point addition criterion determines new sampling points (additional samples). These new sampling points are then marked using the finite element method. The surrogate model is then updated and, based on the optimized point addition criterion and the updated surrogate model, new sampling point information is generated. This cycle repeats until optimization converges.

[0045] The optimization criteria used above mainly include: a. EI criterion (maximizing expected improvement)

[0046] in, ; is the optimal true objective function value among all current sample points; is a random variable with mean , the variance is Normal distribution; and represent the standard normal cumulative distribution function and the standard normal distribution probability density function, respectively.

[0047] b. PI criterion (maximizing the probability of improvement)

[0048] in, is a value smaller than the optimal value among all current sample points; is a random variable with mean , the variance is Normal distribution; represents the standard normal cumulative distribution function.

[0049] c. RMSE criterion (maximizing root mean square error)

[0050] in, The error at the unknown point predicted by the Kriging model.

[0051] d. MP criterion (minimizing the objective function)

[0052] in, The minimum value point predicted by the current surrogate model.

[0053] Furthermore, during the optimization process, due to the continuity of the latent space of the generative model, there is no guarantee that the new samples generated by the generator will have clear geometric boundaries and constructible cross-sectional shapes. Therefore, the norm indicators between the cross-sectional shapes corresponding to the newly added samples and the sample points in the database are considered. Norm indicators include: L1 norm, L2 norm, L∞ norm of the latent space feature vector; mean square error, cross entropy, structural similarity, and peak signal-to-noise ratio of the cross-sectional shape image.

[0054] The feasibility of the cross-sectional shape corresponding to the newly added sample is determined by the norm index. If it is not feasible, the cross-sectional shape with the closest norm index in the cross-sectional shape database is used as the target cross-sectional shape. The control points corresponding to the target cross-sectional shape are processed by the generative model and then replaced with the newly added sample. Then, modeling and accurate simulation analysis and calculation are carried out. Traditional heuristic algorithms such as genetic algorithms are used to carry out the optimal design. The schematic diagram of the agent model optimization is shown in the figure below. Figure 5 shown.

[0055] It should be noted that this method does not limit the generation method of the cross-sectional shape database (including regular geometric figures and irregular parametric figures); the higher the grayscale image pixels used to represent the cross-sectional shape, the more complex the cross-sectional shape that can be represented.

[0056] Through the above description, it should be understood that: Step 100: Provide a standardized design space to lay the data foundation for subsequent generative model training. Ensure the process feasibility of the cross-sectional shape through constraints.

[0057] Step 300: Initial sampling: LTD samples the latent space using methods such as Latin hypercube design to ensure that the control points reflect the unknown design space information, replacing the traditional high-dimensional parameter space and improving sampling efficiency. The generative model compresses the cross-sectional shape into a latent space vector, and the decoder generates new samples from the latent space. The continuity and smoothness of the latent space allow the generation of reasonable, unseen cross-sectional shapes, but feasibility must be verified in step 400. Finite element simulation software (such as Abaqus and Ansys) or custom code is used to calculate the mechanical responses (such as strength, buckling, and dynamic performance) of the initial samples. The mechanical responses serve as labels and, together with the latent space representation, form a dataset. This labeled dataset is provided for surrogate model training. Simulation verifies the mechanical properties of the control points to ensure design feasibility.

[0058] Step 400: Select new sampling points (control points) using the optimized point addition criterion. Simulation calculations are performed using the proxy model. The proxy model is updated based on the simulation results. The process continues iteratively until convergence. The proxy model significantly reduces the amount of simulation computation and improves optimization efficiency. During the optimization process, the optimized point addition criterion intelligently selects the next sampling point based on the proxy model's predictions, effectively approaching the optimal solution. The key is to balance exploration (unknown areas) with exploitation (the currently optimal area).

[0059] The generative model-driven approach reduces reliance on designer experience and shortens design cycles. It also achieves lightweight vehicle structures while meeting performance targets. It supports multi-section design to meet the diverse needs of commercial and specialty vehicles.

[0060] The following functions are achieved: Forward generation: cross-section shape database → generative model → input parameters obtained based on sampling data (multiple control points can form a cross-section).

[0061] Reverse verification: input parameters → simulation mark (solve the mechanical properties of the control point) → proxy model optimization → replace infeasible samples → update the cross-sectional shape database.

[0062] This application also provides a generative model-driven vehicle body beam structure section optimization system, which includes: The first module is used to grid the grayscale image of the structural section and set constraints to sample control points; connect the control points to obtain the cross-sectional shape; and then classify and store them to form a cross-sectional shape database; A second module is used to train and form a generative model based on the cross-sectional shape database; The third module is used to collect control points of the section to be optimized according to the section design requirements to obtain sampling data; use the generative model to reduce the dimension of the sampling data to obtain input parameters; perform simulation mechanical solution on the input parameters, and then use the calculated mechanical response as the output label; and form a data set with the input parameters and output labels; The fourth module is used to construct a proxy model using the data set, and then obtain new samples in combination with the optimization point addition criterion to verify the iterative proxy model and determine the input parameters of the final proxy model; and form an optimized cross-sectional shape based on the final input parameters.

[0063] The function of each module has been explained in the corresponding steps above and will not be repeated here.

[0064] In the description of this application, it should be noted that the terms "upper" and "lower" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on this application. Unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be internal communication between two elements. 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.

[0065] It should be noted that, in this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element.

[0066] The foregoing is merely a list of specific embodiments of the present application, intended to enable those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the broadest scope consistent with the principles and novel features of the present application.

Claims

1. A generative model-driven vehicle body beam structure section optimization method, characterized in that: It includes: Grid the grayscale image of the structural section and set constraints to sample control points; Connect the control points to obtain the cross-sectional shape; Then the cross-section shape database is formed by classification and storage; Training and forming a generative model based on the cross-sectional shape database; Collect control points of the section to be optimized according to the section design requirements to obtain sampling data; Using the generative model to reduce the dimension of the sampled data to obtain input parameters; Perform simulation mechanics solutions on the input parameters, and then use the calculated mechanical responses as output labels; Combine input parameters and output labels into a data set; The proxy model is constructed using the data set, and then new samples are obtained in combination with the optimization point addition criterion to verify the iterative proxy model and determine the final input parameters of the proxy model; and an optimized cross-sectional shape is formed according to the final input parameters.

2. The generative model-driven vehicle body beam structure cross-section optimization method according to claim 1, wherein: Determine the size of the grayscale image based on the impact on the training time and computing power of the generative model, as well as the size of the structural section; The control points are sampled in a designated area of ​​the grayscale image after the size is determined by using a Latin hypercube design method, a Monte Carlo sampling method or a uniform design method; During the sampling process, the shortest line segment constraint and the punching angle constraint are set to ensure that the cross-sectional shape obtained by connecting the control points meets the manufacturing process requirements.

3. The generative model-driven vehicle body beam structure section optimization method according to claim 1, characterized in that: The generative model is formed based on the cross-sectional shape database training, comprising the following steps: Build a generative model and set the ratio of convolutional / deconvolutional layers and fully connected layers in the encoder, decoder, and discriminator of the generative model, as well as the convolution kernel size, sampling step, dropout layer, activation function, batch processing, number of iterations, optimizer, and loss function value; The generative model is used to extract the geometric features of the cross-sectional shape database for training, and is continuously adjusted until the dimension of the latent space of the generative model, the balanced image feature extraction effect, and the latent space dimension all meet the optimization efficiency threshold.

4. The generative model-driven vehicle body beam structure section optimization method according to claim 1, wherein: Sampling control points in a designated area of ​​the grayscale image of the section to be optimized using a Latin hypercube design method, a Monte Carlo sampling method, or a uniform design method to obtain the sampling data; During the sampling process, the shortest line segment constraint and the punching angle constraint are set to ensure that the cross-sectional shape obtained by connecting the control points meets the manufacturing process requirements; The generative model is used to reduce the dimensionality of the sampled data to the latent space dimension, and the latent space representation after dimensionality reduction is used as an input parameter.

5. The generative model-driven vehicle body beam structure section optimization method according to claim 1, wherein: Use Abaqus or Ansys finite element simulation software to build finite element models through parametric scripts; The finite element model is used to solve the stiffness and stability of each control point in the input parameters to obtain the corresponding mechanical response; the mechanical response corresponding to each control point is used as a label to obtain the output label of the sampled data.

6. The generative model-driven vehicle body beam structure section optimization method according to claim 1, characterized in that: The proxy model is constructed using the data set, and then new samples are obtained in combination with the optimization point addition criterion to verify the iterative proxy model and determine the input parameters of the final proxy model; the optimized cross-sectional shape is formed according to the final input parameters, specifically including the following steps: The input parameters are used as input and the output labels are used as output, and an approximate mapping relationship between the input and the output is constructed by combining a radial basis model, a kriging model, and a neural network model to form a proxy model; Obtaining new samples based on the data set using an optimized point addition criterion; Use the proxy model to take the new samples as input parameters to predict the output labels of the new samples; Update the proxy model based on the newly added samples and their output labels, and perform finite element verification until the iteration reaches convergence; The input parameters corresponding to the final proxy model are used as the optimization solution of the control points to form the optimized cross-sectional shape.

7. The generative model-driven vehicle body beam structure section optimization method according to claim 6, characterized in that: The optimization point addition criterion includes one of the EI criterion and the PI criterion; The EI criteria are: in, ; is the optimal true objective function value among all current control points; is a random variable with mean , the variance is Normal distribution; and represent the standard normal cumulative distribution function and the standard normal distribution probability density function respectively; The PI criteria are: in, is a value smaller than the optimal value among all current control points; is a random variable with mean , the variance is Normal distribution; represents the standard normal cumulative distribution function.

8. The generative model-driven vehicle body beam structure section optimization method according to claim 6, wherein: Acquiring new samples based on the data set using the optimized point addition criterion also includes the following steps: For the cross-sectional shape corresponding to the newly added sample, if its geometric boundary is unclear or the cross-sectional shape cannot be constructed, the norm index between it and each cross-sectional shape in the cross-sectional shape database is calculated; The cross-sectional shape corresponding to the minimum norm index in the cross-sectional shape database is used as the target cross-sectional shape; The control points corresponding to the target section are processed by the generative model and then replaced with the newly added samples.

9. The generative model-driven vehicle body beam structure section optimization method according to claim 1, wherein: After the optimized cross-sectional shape is obtained, the optimized cross-sectional shape is updated to the cross-sectional shape database.

10. A generative model-driven vehicle body beam structure section optimization system, characterized in that: It includes: The first module is used to grid the grayscale image of the structural section and set constraints to sample control points; Connect the control points to obtain the cross-sectional shape; Then the cross-section shape database is formed by classification and storage; A second module is used to train and form a generative model based on the cross-sectional shape database; The third module is used to collect control points of the section to be optimized according to the section design requirements to obtain sampling data; Using the generative model to reduce the dimension of the sampled data to obtain input parameters; Perform a simulation mechanical solution on the input parameters, and then use the calculated mechanical response as the output label; the input parameters and output labels form a data set; The fourth module is used to construct a proxy model using the data set, and then obtain new samples in combination with the optimization point addition criterion to verify the iterative proxy model and determine the input parameters of the final proxy model; and form an optimized cross-sectional shape based on the final input parameters.

Images