Intelligent Generative Design Method for Anti-collision Beams

Abstract

This invention relates to the field of automotive design technology, and more particularly to an intelligent generative design method for crash beams. First, experimental analysis is conducted on the crash beam to obtain cross-sectional images and corresponding performance response data. Image recognition and text extraction are performed on the cross-sectional images. A multimodal, multi-layer fusion model is used to fuse the recognized images and extracted text to obtain multimodal design variables. The constructed conditional variational network is trained using the multimodal design variables and their corresponding performance response data to obtain a conditional variational model. The conditional variational model is used to generate design variables and predict the corresponding performance response data. The optimal performance response data is selected from the obtained performance response data, thereby deriving the optimal design scheme for the crash beam to be designed. This invention utilizes a conditional variational network to determine the optimal design scheme, effectively improving the efficiency and accuracy of optimization design.

Description

Technical Field

[0001] The present invention belongs to the technical field of automotive design, and particularly relates to an intelligent generative design method for a bumper beam. Background Art

[0002] The front bumper beam is an important safety component at the front end of an automotive body-in-white, and its performance directly affects the vehicle's collision safety. At present, the materials used for the front bumper beam are mainly high-strength steel, aluminum alloy, and carbon fiber reinforced composite (CFRP). CFRP and aluminum alloy front bumper beams are light in weight but high in price, and are mostly used in high-class vehicles. High-strength steel front bumper beams have a low cost but a large mass, and are commonly used in economy vehicles. Existing single-material front bumper beams have their own advantages and disadvantages, and their application ranges are limited. To meet the major technical requirements for the research and development of lightweight vehicle bodies in China, CFRP / steel hybrid structure front bumper beams have been proposed, which can give full play to the unique performance advantages of CFRP and steel, and can achieve low cost, high performance, and lightweight of the front bumper beam, with broad application prospects. At present, the research on the design scheme of CFRP / steel hybrid structure bumper beams is still in the exploration stage, and most inventors tentatively set the shapes of the CFRP layer and the metal layer (such as using common "mouth" shapes, "Ji" shapes, "Mu" shapes, etc.), and the connection types between the two are mostly stacked in a completely wrapped manner. It is difficult to effectively guarantee the performance of the bumper beam composed of the CFRP layer and the metal layer, and it is necessary to repeatedly trial and error and adjust to determine a reasonable structural shape and connection type. This design method has great blindness and low design efficiency. Therefore, in the existing field, there is a lack of an intelligent design method that can automatically determine the shapes and connection types of the CFRP layer and the metal layer according to the performance indicators of the CFRP / steel hybrid structure bumper beam (such as weight, collision peak force, collision intrusion amount, collision energy absorption). At present, there is a technical blank in the research of this field, which needs to be further overcome.

[0003] Currently, the traditional design method for automotive structures adopts the concept of "optimizing structural parameters → meeting performance requirements", and uses optimization algorithms to repeatedly iterate and update design variables. It mainly combines surrogate models (such as response surfaces, Kriging, radial basis neural networks, etc.) and intelligent optimization algorithms (genetic algorithms, particle swarm algorithms, etc.). Due to the complex characteristics of the design variables of the CFRP / steel hybrid structure, which are multimodal, high-dimensional, strongly coupled, and a mixture of discrete / continuous variables, and the performance response is highly nonlinear. It is difficult for traditional surrogate model methods to effectively construct the "multimodal high-dimensional variable-response" mapping relationship, and it is difficult to form a two-way mapping between the two. To obtain a set response value, only by repeatedly optimizing and iterating design variables, the calculation is complex and the high-dimensional variables often make the optimization algorithm difficult to converge. If the response conditions change, it is necessary to re-iterate and optimize, and it is difficult to constrain the design variables during optimization.

[0004] In recent years, deep generative models have been proposed in artificial intelligence, which can quickly create optimal designs based on objectives without iterative optimization processes and can fuse multivariate data, offering a solution to this problem. However, existing research uses deep generative models to generate a large number of design variables. To select the optimal design method from these variables, experiments or simulations are still needed to obtain performance index values, followed by a comprehensive decision based on the values ​​of the design variables and performance responses. This approach makes the computational steps of deep generative design complex and significantly inefficient. Therefore, it is necessary to improve traditional deep generative design methods to directly derive the optimal design variables and reduce intermediate solution steps. Currently, no research has reported on such a design method. Summary of the Invention

[0005] In view of this, the present invention aims to provide an intelligent generative design method for anti-collision beams. It innovatively applies a deep generative model of artificial intelligence to the integrated design of the structural shape and dimensions of carbon / steel composite anti-collision beams. By using the proposed conditional variational network, the optimal design scheme is determined. This solves the problems of low computational efficiency, poor accuracy of nonlinear solutions, and inability to directly map and solve for the optimal design variables based on response requirements in traditional design methods for optimizing design variables to obtain the best response quantity. It realizes one-step end-to-end intelligent design and provides a new approach and method for the design and development of automotive parts.

[0006] To achieve the above objectives, the technical solution created by this invention is implemented as follows:

[0007] A smart generative design method for crash beams includes:

[0008] S1: Conduct experimental analysis on the crash beam and obtain a dataset, which includes cross-sectional images of the crash beam and the corresponding performance response data.

[0009] S2: Perform image recognition and text extraction on the cross-sectional images in the dataset of step S1 to obtain the structural shape and structural dimensions respectively; fuse the structural shape and structural dimensions to obtain multimodal design variables;

[0010] S3: Using the multimodal design variables obtained in step S2 as input and the performance response data corresponding to the multimodal design variables as output, train the constructed conditional variational network to obtain the conditional variational model.

[0011] S4: Following steps S1~S2, extract multimodal design variables from the anti-collision beam to be designed, and input the obtained multimodal design variables into the conditional variational model obtained in step S3 to obtain the corresponding performance response data.

[0012] S5: Based on the multimodal design variables and performance response data generated in step S4, select the optimal performance response data, and derive the optimal design scheme for the anti-collision beam to be designed based on the optimal performance response data.

[0013] Furthermore, in step S1, the process of conducting experimental analysis on the anti-collision beam and obtaining the dataset includes: establishing a simulation model of the anti-collision beam and obtaining a simulation dataset; the simulation dataset includes cross-sectional images of the simulation model and its corresponding performance response; and conducting actual collision tests on the anti-collision beam to obtain a real dataset, which includes cross-sectional images of the anti-collision beam and its corresponding performance response values.

[0014] Furthermore, the process of establishing a simulation model of the crash barrier beam and obtaining the simulation dataset includes: establishing an implicit parameterized model that includes the structural shape and dimensions of the crash barrier beam; determining sample points with different structural shapes and dimensions in the crash barrier beam based on the Hammsley experimental design method; acquiring cross-sectional images of the sample points; performing low-speed collision finite element simulation analysis on the sample points, and extracting the performance response values ​​corresponding to each sample point from the simulation results; the simulation dataset includes cross-sectional images and corresponding performance response values.

[0015] Furthermore, in step S2, the process of extracting the structural shape and structural dimensions of the cross-sectional images in the simulation dataset includes: reading the structural dimensions and structural shape of the anti-collision beam cross-section from the Hamsley test design list obtained by the Hamsley test design method.

[0016] Furthermore, the process of conducting actual collision tests on the crash beam to obtain a real dataset includes: conducting multiple collision tests on the crash beam, extracting the experimentally measured performance response values, and extracting the cross-sectional images of the crash beam from the experimental image results; the real dataset includes the cross-sectional images and the corresponding performance response values.

[0017] Furthermore, in step S2, the process of extracting the structural shape and structural size of the cross-sectional image in the real dataset includes: using a convolutional neural model to identify the cross-sectional image in the real dataset to obtain the corresponding structural shape; after preprocessing the cross-sectional image in the real dataset by edge detection and contour extraction, the corresponding structural size is obtained by calibrating the scale and converting the length and thickness parameters.

[0018] Furthermore, in step S2, the process of fusing structural shape and structural dimensions to obtain multimodal design variables includes: integrating the structural shape of cross-sectional images in the real dataset and the structural shape of cross-sectional images in the simulation dataset to obtain a structural shape set; integrating the structural dimensions extracted from the real dataset and the structural dimensions extracted from the simulation dataset to obtain a structural dimension set; and inputting the structural shape set and the structural dimension set into a multimodal multilayer fusion model for fusion to obtain multimodal design variables.

[0019] Furthermore, in the conditional variational network of step S3: a multilayer perceptron-based encoder is used to encode the multimodal design variables, mapping the input multimodal design variables to the latent space to obtain the mean and log-variance of the latent space; the latent space is reparameterized based on the log-variance and the mean to obtain latent variables; the latent variables and the set conditional performance response are input into a hybrid density network-based decoder for decoding to obtain a Gaussian mixture parameter distribution; the Gaussian mixture parameter distribution is combined according to weights to obtain the generated design variables; a Gaussian process regression model is used to predict the performance of the generated design variables to obtain the corresponding performance response data.

[0020] Furthermore, in step S3, the conditional variational network is trained using a joint optimization loss function; the joint optimization loss function is:

[0021] ;

[0022] Where L represents the joint optimization loss function, Indicates the reconstruction loss. X represents the network parameters of the decoder, Y represents the multimodal design variables, Z represents the corresponding performance response data, and Z represents the latent variables. Denotes KL divergence, Represents the network parameters of the encoder. Indicates a negative feedback term. This represents the performance response data predicted by the conditional variational model obtained from the current training. This represents conditional performance response data. This represents the KL divergence proportionality coefficient. This represents the negative feedback proportional coefficient; the negative feedback proportional coefficient is obtained by the following formula:

[0023] ;

[0024] in, This represents the error between the predicted performance response data and the corresponding conditional performance response data. This represents the variance of the error.

[0025] Furthermore, step S5 includes the following process:

[0026] Principal component analysis was used to analyze the performance response data obtained from S4 to obtain the objective weight value of each performance response. Expert scoring and analytic hierarchy process (AHP) were used to obtain the subjective weight value of each performance response. The subjective and objective weight values ​​were then combined using a zero-sum game method to obtain the comprehensive weight value. Grey relational analysis was used to obtain the grey relational coefficient of each performance response scheme in the performance response data generated by S4. The grey relational coefficient was multiplied by the comprehensive weight value to obtain the grey relational degree of each performance response scheme. All performance response schemes were sorted according to the grey relational degree. The scheme with the largest grey relational degree value was the optimal performance response scheme, and the relevant parameters of the corresponding anti-collision beam were the optimal design scheme for the anti-collision beam.

[0027] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0028] (1) The intelligent generative design method for anti-collision beams created in this invention applies an artificial intelligence deep generative model to the integrated design of the structural shape and structural size of carbon / steel composite anti-collision beams. Based on the traditional conditional variational network architecture, a lightweight Gaussian process regression model and a multi-criteria decision model are introduced. The error feedback term is iterated into the network learning through back-propagation, which improves the generation accuracy. The proposed improved conditional variational network architecture can solve the optimal design scheme in one step, realizing direct end-to-end intelligent design from conditional performance response to optimal design variables, which greatly improves the design efficiency and provides a new technical approach and method for automotive structural design. It solves the problems of low computational efficiency, poor nonlinear solution accuracy, and inability to directly map and solve the optimal design variables according to the response requirements of the traditional design method for optimizing design variables to find the optimal response quantity. It provides a new approach and method for the design and development of automotive parts.

[0029] (2) The intelligent generative design method for the anti-collision beam created by the present invention is based on image recognition technology. It uses CNN to identify and predict the structural shape of the carbon / steel composite anti-collision beam, and uses edge detection, contour extraction, and proportional calibration and parameter conversion methods to identify and predict the structural dimensions of the carbon / steel composite anti-collision beam. This solves the technical problem of extracting and calibrating the complex and diverse cross-sections of the carbon / steel composite anti-collision beam, and greatly improves the design efficiency. In addition, the present invention integrates simulation and experimental data to enhance the accuracy of the model. It integrates the image variables of the structural shape and the text variables of the structural dimensions through multimodal information fusion, realizing the integrated collaborative design of structural shape and structural dimensions. This solves the problem that the connection shape and size of the CFRP layer and the metal layer of the carbon / steel composite anti-collision beam are difficult to determine directly and require repeated trial and error. This reduces the blindness of the design and improves the design efficiency.

[0030] (3) The intelligent generative design method for the anti-collision beam created in this invention introduces principal component analysis, expert scoring, analytic hierarchy process and zero-sum game method to obtain the comprehensive weight value for evaluating each performance response, which ensures the reliability and scientific nature of the weight and improves the accuracy of predicting the performance response. Attached Figure Description

[0031] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0032] Figure 1 A flowchart illustrating the intelligent generative design method for the anti-collision beam described in the embodiments of the present invention;

[0033] Figure 2 A schematic diagram illustrating the process of acquiring simulation datasets as described in an embodiment of the present invention;

[0034] Figure 3 A schematic diagram of the structural shape of the steel plate layer and CFRP layer of the anti-collision beam described in the embodiment of the present invention;

[0035] Figure 4 A schematic diagram illustrating the structural shape of the anti-collision beam described in an embodiment of the present invention;

[0036] Figure 5 A schematic diagram illustrating the process of acquiring the "multimodal design variable-performance response" dataset as described in the embodiments of the present invention;

[0037] Figure 6 This is a schematic diagram of the framework of the conditional variational network described in the embodiment of the present invention.

[0038] Explanation of reference numerals in the attached figures:

[0039] 1. Anti-collision beam; 101. Steel material layer; 102. CFRP layer. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof.

[0041] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0042] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0043] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0044] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0045] like Figure 1 As shown, the intelligent generative design method for crash beams includes:

[0046] S1: Conduct experimental analysis on the crash beam and obtain a dataset, which includes cross-sectional images of the crash beam and the corresponding performance response data.

[0047] In some embodiments, the process of experimentally analyzing the crash beam and obtaining a dataset includes: establishing a simulation model of the crash beam to obtain a simulation dataset, and conducting actual collision tests on the crash beam to obtain a real dataset. The simulation dataset includes cross-sectional images of the simulation model and their corresponding performance response values, while the real dataset includes cross-sectional images of the crash beam and their corresponding performance response values. The cross-sectional images in both the simulation and real datasets characterize the structural dimensions and shape of the crash beam.

[0048] Specifically, the process of establishing a simulation model of the crash barrier beam and obtaining the simulation dataset is as follows: Figure 2 Shown, including:

[0049] S101: Establish an implicit parametric model including the structural shape and structural dimensions of the bumper beam. In the embodiments of the present invention, a professional implicit parametric modeling software, such as SFE / Concept software, is used to establish a parametric model of the carbon / steel composite structural bumper beam for automobiles. The parameters include the structural dimensions and structural shape of the bumper beam.

[0050] S102: Determine sample points with different structural shapes and structural dimensions in the bumper beam based on the Hammersley experimental design method. In the embodiments of the present invention, multidisciplinary optimization software, such as Isight and Hyperstudy, etc., is adopted. Based on the experimental design module, the Hammersley experimental design method is used to generate a large number of data points, and a large number of sample points with different structural shapes and structural dimensions are generated through the bumper beam parametric model.

[0051] S103: Obtain the cross-sectional images of the sample points. The structural shapes of the steel plate layer and the CFRP (carbon fiber reinforced plastic) layer include "Ji" shape, "Kou" shape, "Ri" shape, "Mu" shape, half "Kou" shape, incomplete closed "Kou" shape, etc., as Figure 3 shown. At the same time, various structural shapes can be formed after the composite of the steel plate layer and the CFRP layer. For example, Figure 4 lists some possible structural shapes formed. Figure 4 It shows that in the structure of the bumper beam 1, there are various superposition methods between the steel material layer 101 and the CFRP layer 102. What the method provided by the present invention needs to solve is how to directly determine the superposition method between the steel material layer 101 and the CFRP layer 102 through end-to-end intelligent design, as well as the dimensions such as the thickness and length between the steel material layer 101 and the CFRP layer 102.

[0052] S104: Conduct low-speed collision finite element simulation analysis on the sample points, and extract the performance response values corresponding to each sample point from the simulation results. In the embodiments of the present invention, according to the requirements of national standards, low-speed impact simulation boundary conditions are given, and LS-DYNA software is called to conduct low-speed collision finite element simulation analysis on the sample points respectively, and a performance response data set of the total weight, peak force, energy absorption, and intrusion amount of the bumper beam is obtained.

[0053] S105: The simulation data set includes cross-sectional images and corresponding performance response values.

[0054] Specifically, the process of obtaining the real data set by conducting actual collision tests on the bumper beam includes:

[0055] S111: Multiple collision tests are conducted on the crash beam to extract the measured performance response values ​​and the cross-sectional image of the crash beam from the experimental image results. In this embodiment of the invention, based on a large number of low-speed collision tests accumulated during the historical research and development of CFRP / steel hybrid structure crash beams, the measured performance response values ​​are extracted from the tests, and the cross-sectional morphology of the crash beam is extracted from the experimental image results.

[0056] S112: The real dataset includes cross-sectional images and corresponding performance response values.

[0057] S2: Perform image recognition and text extraction on the cross-sectional images in the dataset from step S1 to obtain the structural shape and structural dimensions, respectively; fuse the structural shape and structural dimensions to obtain multimodal design variables. The multimodal design variables and their corresponding performance response data together constitute the "Multimodal Design Variables-Performance Response" dataset. The process of obtaining the "Multimodal Design Variables-Performance Response" dataset in step S2 is as follows: Figure 5 As shown.

[0058] It should be noted that since the cross-sectional images in the simulation dataset are automatically generated by software based on preset structural shapes and dimensions, image recognition of the simulation dataset is no longer required. Therefore, only text extraction of the simulation dataset is needed to obtain the structural dimensions and shapes. Specifically, in some embodiments, the text extraction process for the simulation dataset includes: reading the structural dimensions and shapes of the cross-sections of the crash beams from the Hammsley experimental design list obtained by the Hammsley experimental design method, thus completing the text extraction of the simulation dataset. The structural dimensions of the cross-sections include measurable dimensions such as the thickness and length of the structure.

[0059] In step S2 of this embodiment of the invention, the process of performing image recognition and text extraction on the cross-sectional images in the real dataset to obtain the structural shape and structural dimensions includes:

[0060] S211: A convolutional neural network model is used to identify cross-sectional images in a real dataset to obtain the corresponding structural shapes. In this embodiment, the convolutional neural network model used is a pre-trained LeNet network model. The LeNet network structure includes convolutional layers, pooling layers, and fully connected layers, and a cross-entropy loss function is specified. Gradient descent is used to optimize the parameters of the trained LeNet network, resulting in a pre-trained LeNet network model. The output of the LeNet network model is then compressed using an embedding method to obtain image feature vectors representing the structural shapes, thus completing the image recognition of the real dataset. During the training of the LeNet network, the cross-sectional images in the real dataset are normalized. Simultaneously, the image dataset is divided into training, validation, and test sets. The LeNet network is trained using the training set, and the LeNet network parameters are updated using the validation set via backpropagation to improve model accuracy. Finally, the model is evaluated and adjusted using the test set.

[0061] S212: After preprocessing the cross-sectional images in the real dataset by edge detection and contour extraction, the scaling ratio and length and thickness parameters are converted to obtain the structural dimensions of the cross-section of the crash beam, thus completing the text extraction from the real dataset. In this embodiment of the invention, edge detection and contour extraction can directly employ existing image recognition technologies, such as using the Canny algorithm for edge detection and the Sobel algorithm for edge extraction, to obtain the structural dimensions of the cross-section of the crash beam.

[0062] In some embodiments, the process of fusing structural shape and structural dimensions using a multimodal multilayer fusion model to obtain multimodal design variables includes: integrating the structural shape of cross-sectional images in a real dataset with the structural shape reflected in cross-sectional images in a simulation dataset to obtain a structural shape set; integrating the structural dimensions extracted from the real dataset with the structural dimensions extracted from the simulation dataset to obtain a structural dimension set; and inputting the structural shape set and structural dimension set into the multimodal multilayer fusion model for fusion to obtain multimodal design variables. In this embodiment of the invention, the multimodal multilayer fusion model adopts the deep neural network model disclosed in Chinese Patent Publication No. CN110674677A, published on January 10, 2020, entitled "A Multimodal Multilayer Fusion Deep Neural Network for Face Anti-Spoofing". During use, the structural shape set and structural dimension set are input into this deep neural network model, and multimodal design variables are output.

[0063] S3: Using the multimodal design variables obtained in step S2 as input and the performance response data corresponding to the multimodal design variables as output, train the constructed conditional variational network to obtain the conditional variational model.

[0064] This invention relates to a conditional variational network based on a conditional variational autoencoder (cVAE). A variational autoencoder (VAE) is a deep learning model consisting of an encoder and a decoder. The encoder is responsible for extracting key latent variables from the input data, mapping the input data to a probability distribution in the latent space. The core objective of the encoder is to construct the latent space of the input data through dimensionality reduction, preserving key information while identifying its hidden patterns. The decoder uses these latent variables to reconstruct the original input data, restoring the latent variables to the same dimension as the input data. Compared to the unsupervised learning of variational autoencoders (VAEs), conditional variational autoencoders (cVAEs) introduce supervised elements, enabling the model to learn labeled conditional information and enhancing the controllability of the decoder's output. Based on the trained neural network framework, the decoder obtains the required design variables by controlling the input of the supervised elements, i.e., the conditional performance response index, thus achieving inverse design. This invention then iteratively trains the conditional variational network until a performance index dataset with an accuracy of over 95% is generated, and the iteration converges. At this point, the training of the conditional variational network is complete, resulting in the conditional variational model.

[0065] In some embodiments, in the conditional variational network of step S3, a multilayer perceptron-based encoder is used to encode the multimodal design variables, mapping the input multimodal design variables to the latent space to obtain the probability distribution of the latent space, specifically obtaining the mean and log-variance of the latent space; the latent space is reparameterized based on the log-variance and mean to obtain latent variables; the latent variables and the set conditional performance response are input into a hybrid density network-based decoder for decoding to obtain a Gaussian mixture parameter distribution, where the conditional performance response refers to the performance response needing to meet certain conditions, which can be a definite value or a definite range of values, and the conditional performance response is adaptively selected and adjusted according to the actual situation; the Gaussian mixture parameter distribution is combined according to weights to obtain the generated design variables; the Gaussian process regression model (GPR) performs performance prediction on the generated design variables to obtain the corresponding performance response data. The GPR model here is a lightweight machine learning prediction model used to quickly predict performance response values ​​by inputting design variables into the model. In order to ensure that the accuracy of the GPR model meets the requirements, it needs to be pre-trained. 80% of the data is extracted from the multimodal design variables and performance response datasets obtained from simulation and experiment as the test set, and the remaining 20% ​​is used as the validation set. The model can only stop training and be used when the accuracy reaches 95% or above after verification.

[0066] The conditional variational network provided in the embodiments of the present invention is as follows: Figure 6 As shown, an encoder based on a multilayer perceptron (MLP) encodes multimodal design variables to obtain the mean of the multimodal design variables. and logarithmic variance , Represents the variance of multimodal design variables. The standard deviation of the multimodal design variables is represented; reparameterized sampling based on the logarithmic variance and mean yields the latent variable z, i.e. ,in, This represents a noise variable that follows a standard normal distribution, i.e. ; Compare the latent variable z with the given conditional performance response data The input is decoded in a hybrid density network (MDN) based decoder to obtain a Gaussian mixture parameter distribution, which includes the mixture weights. mean and variance The Gaussian mixture parameter distribution is selected based on the mixing weights, and the k-th Gaussian distribution is then drawn from the selected Gaussian distribution to obtain the generated design variables. Gaussian process regression model for generating design variables Perform performance prediction to obtain the performance response data predicted by the currently trained conditional variational model. .

[0067] In this embodiment of the invention, a multilayer perceptron-based encoder is constructed using five fully connected layers. Each fully connected layer has 256 neurons, the activation function is ReLU, and residual connections are introduced. Multimodal design variables are input into the encoder, and dimensionality is progressively reduced through the fully connected layers to learn the nonlinear relationship between the design space and performance indicators, outputting the mean of the latent space distribution. and logarithmic variance By reparameterizing the sampling to obtain the latent variable z, the non-differentiable random sampling process is transformed into a differentiable computation process, solving the problem of gradient backpropagation and facilitating optimization using gradient descent. Furthermore, this embodiment uses five fully connected layers to construct a decoder based on a hybrid density network. Each fully connected layer has 256 neurons and the activation function is Tanh. The latent variable and a preset conditional performance response are input into the decoder, which outputs a mixture of Gaussian parameter distributions. The decoder based on the hybrid density network dynamically learns the parameters (mixture weights, mean, and variance) of multiple basic probability distributions (such as Gaussian distributions), and linearly combines these distributions according to their weights, thereby constructing a multimodal probability model to accurately fit the complex distribution characteristics of the target data. Fully connected layers are the basic building blocks of multilayer perceptrons and hybrid density networks. The layers in a fully connected layer are fully connected, with the bottom layer being the input layer, the middle layers being hidden layers, and the last layer being the output layer. The input to a fully connected layer is flattened into a one-dimensional vector, and a linear transformation is achieved through weight matrix multiplication and bias term superposition.

[0068] To improve the matching accuracy between the performance response corresponding to the design variables generated by the conditional variational model and the conditional performance response, this invention proposes a backpropagation adjustment strategy. Specifically, in some embodiments, the generated design variables... The corresponding performance response value is obtained by predicting it using the GPR regression model. Combine it with conditional performance response error The decoder parameters are adjusted through backpropagation; specifically, the conditional variational network is trained by jointly optimizing the loss function. In this invention, a negative feedback term is introduced to construct the joint optimization loss function based on the original Maximum Evidence Lower Bound (ELBO). Specifically, the Maximum Evidence Lower Bound (ELBO) is as follows:

[0069] ;

[0070] Where l represents maximizing the lower bound of evidence, Indicates the network parameters of the decoder. Represents the network parameters of the encoder. Indicates the reconstruction loss. Let X represent the KL divergence, Y represent the multimodal design variables, Y represent the corresponding performance response data, and Z represent the latent variables.

[0071] The introduced negative feedback term is , The negative feedback proportional coefficient is obtained from the following formula:

[0072] ;

[0073] in, This represents the error variance. The joint optimization loss function is then obtained as:

[0074] ;

[0075] Where L represents the joint optimization loss function, This represents the KL divergence proportionality coefficient. A dynamic scheduling strategy is adopted (initial value 0.1, linearly increasing to 1.0), with a negative feedback proportional coefficient. Adaptive adjustment based on the error sample variance.

[0076] In this embodiment of the invention, a conditional variational network is trained using a multimodal design variable and performance response dataset. Specifically, 80% of the multimodal design variable and performance response dataset is divided into a training set and 20% into a validation set. The training set is used to train the conditional variational network, and the validation set is used to calculate the loss value of the joint optimization loss function. The training cutoff condition for the conditional variational network is that the loss value of the validation set in the multimodal design variable and performance response dataset decreases by less than 0.5% for 10 consecutive training epochs. In this embodiment of the invention, the Adam algorithm is used as the optimizer for training the conditional variational network, with default parameters... , The training batch size is 200, and the learning rate is 10. -8 .

[0077] S4: Following steps S1~S2, extract multimodal design variables from the anti-collision beam to be designed, and input the obtained multimodal design variables into the conditional variational model obtained in step S3 to obtain the corresponding performance response data.

[0078] S5: Based on the multimodal design variables and performance response data generated in step S4, the optimal design variable - performance response data is selected, and then the optimal design scheme of the anti-collision beam to be designed is obtained.

[0079] In this embodiment of the invention, step S5 includes: analyzing the performance response data obtained in S4 using principal component analysis to obtain the objective weight value of each performance response; obtaining the subjective weight value of each performance response using expert scoring and analytic hierarchy process; and solving for the comprehensive weight value by combining the subjective and objective weight values ​​through a zero-sum game method. This comprehensive weight value considers both the objective distribution characteristics of the data and the preference orientation of practical applications, avoiding the defects of maximum and minimum values ​​in a single subjective or objective weight method, thus ensuring the reliability and scientific nature of the weight. Using grey relational analysis, the grey relational coefficient of each group of performance response schemes in the performance response data generated in S4 is obtained, and the grey relational coefficient is multiplied by the comprehensive weight value to obtain the grey relational degree of each group of performance response schemes. All performance response schemes are sorted according to the size of the grey relational degree, and the scheme with the largest grey relational degree value is the optimal performance response scheme. The relevant parameters of the anti-collision beam to be designed corresponding to this scheme are the optimal design scheme for the anti-collision beam.

[0080] This invention incorporates a multi-criteria decision-making module combining principal component analysis (PCA) and grey relational analysis (GPR) with a GPR model. The aim is to achieve one-step, end-to-end intelligent design, where "end-to-end" refers to the process from input to output. The backpropagation adjustment strategy ensures the accuracy of the generated conditional variational model. Traditional conditional variational model architectures, using the conditional performance response as input, only generate design variables that meet performance indicators. However, the optimal design cannot be determined solely based on these variable values. Multiple sets of generated design variables need to be re-analyzed using simulation or experimentation to obtain performance indicators. These indicators are then compared with the conditional performance response to verify accuracy. If the accuracy is insufficient, the conditional variational model needs repeated debugging until the required precision is achieved. Finally, a comprehensive decision is made based on the performance indicators and design variables to select the optimal design, which is the output. Traditional conditional variational model architectures struggle to achieve one-step, end-to-end functionality, requiring multiple auxiliary functions. Furthermore, these auxiliary functions are not embedded into the conditional variational model through network learning, leading to repeated trial and error, resulting in low efficiency and accuracy. The purpose of this invention is to provide a one-step end-to-end design model based on a conditional variational model. It innovatively integrates a lightweight GPR prediction model at the back end of the conditional variational model, enabling rapid prediction of the corresponding performance response from generated design variables, eliminating the need for re-simulation or experimental analysis of multiple sets of generated design variables. Furthermore, this invention calculates the error between the GPR-predicted performance response and the conditional performance response, and incorporates the error term as a negative feedback term into the maximization of the lower bound of evidence, forming a joint optimization loss function. Through learning iteration, the accuracy of the generated conditional variational model is guaranteed to be above 95%, eliminating the need for repeated manual debugging of the conditional variational model parameters and greatly improving design efficiency. Furthermore, this invention uses a multi-criteria decision-making method to select the optimal design scheme from the generated design variables and the GPR-predicted performance response, directly outputting the optimal design scheme, eliminating the traditional steps of repeated comparison and manual selection of the optimal design variables, greatly improving design efficiency. Therefore, the technical solution provided by this invention realizes a one-step end-to-end design that automatically outputs the "optimal design scheme" upon inputting the "conditional performance response".

[0081] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0082] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A smart generative design method for anti-collision beams, characterized in that, include: S1: Conduct experimental analysis on the anti-collision beam to obtain a dataset, which includes cross-sectional images of the anti-collision beam and performance response data corresponding to the cross-sectional images; S2: Perform image recognition and text extraction on the cross-sectional images in the dataset from step S1 to obtain the structural shape and structural dimensions, respectively. By fusing structural shape and dimensions, multimodal design variables are obtained; S3: Using the multimodal design variables obtained in step S2 as input and the performance response data corresponding to the multimodal design variables as output, train the constructed conditional variational network to obtain the conditional variational model. In the conditional variational network of step S3: the multimodal design variables are encoded using an encoder based on a multilayer perceptron, and the input multimodal design variables are mapped to the latent space to obtain the mean and log-variance of the latent space; the latent space is reparameterized based on the log-variance and the mean to obtain the latent variables; The latent variables and the set conditional performance response are input into a decoder based on a hybrid density network for decoding to obtain a Gaussian mixture parameter distribution; the Gaussian mixture parameter distribution is combined according to weights to obtain the generated design variables; The Gaussian process regression model was used to predict the performance of the generated design variables, and the corresponding performance response data was obtained. S4: Following steps S1~S2, extract multimodal design variables from the anti-collision beam to be designed, and input the obtained multimodal design variables into the conditional variational model obtained in step S3 to obtain the corresponding performance response data. S5: Based on the multimodal design variables and performance response data generated in step S4, the optimal performance response data is selected, and the optimal design scheme of the anti-collision beam to be designed is obtained based on the optimal performance response data.

2. The intelligent generative design method for anti-collision beams according to claim 1, characterized in that, In step S1, the process of conducting experimental analysis on the crash beam and obtaining the dataset includes: A simulation model of the anti-collision beam is established to obtain a simulation dataset; the simulation dataset includes cross-sectional images of the simulation model and its corresponding performance response. In addition, actual collision tests were conducted on the crash beams to obtain a real dataset, which includes cross-sectional images of the crash beams and their corresponding performance response values.

3. The intelligent generative design method for anti-collision beams according to claim 2, characterized in that, The process of establishing a simulation model of the crash barrier beam and obtaining the simulation dataset includes: An implicit parametric model including the structural shape and dimensions of the anti-collision beam is established; The sample points for different structural shapes and dimensions in the anti-collision beam were determined based on the Hamsley experimental design method. Obtain the cross-sectional image of the sample point; Low-speed collision finite element simulation analysis was performed on the sample points, and the performance response values ​​corresponding to each sample point were extracted from the simulation results. The simulation dataset includes the cross-sectional images and the corresponding performance response values.

4. The intelligent generative design method for anti-collision beams according to claim 3, characterized in that, Step S2, the process of extracting the structural shape and size from the cross-sectional images of the simulation dataset, includes: The structural dimensions and shape of the anti-collision beam are read from the Hamsley test design list obtained from the Hamsley test design method.

5. The intelligent generative design method for anti-collision beams according to claim 2, characterized in that, The process of conducting actual crash tests on crash beams to obtain real-world datasets includes: Multiple collision tests were conducted on the anti-collision beam to extract the performance response values ​​measured in the tests, as well as the cross-sectional images of the anti-collision beam from the images captured in the tests. The real dataset includes cross-sectional images and corresponding performance response values.

6. The intelligent generative design method for anti-collision beams according to claim 2 or 5, characterized in that, Step S2, the process of extracting the structural shape and size from the cross-sectional images of the real dataset, includes: The cross-sectional images in the real dataset are identified using a convolutional neural model to obtain the corresponding structural shapes; After preprocessing the cross-sectional images in the real dataset by edge detection and contour extraction, the corresponding structural dimensions are obtained by calibrating the scale and converting the length and thickness parameters.

7. The intelligent generative design method for anti-collision beams according to claim 2, characterized in that, Step S2, the process of fusing structural shape and structural dimensions to obtain multimodal design variables, includes: The structural shapes of the cross-sectional images in the real dataset and the cross-sectional images in the simulation dataset are integrated to obtain a set of structural shapes; The structural dimensions extracted from the real dataset and the structural dimensions extracted from the simulation dataset are integrated to obtain a set of structural dimensions; The set of structural shapes and the set of structural dimensions are input into the multimodal multilayer fusion model for fusion to obtain the multimodal design variables.

8. The intelligent generative design method for anti-collision beams according to claim 1, characterized in that, In step S3, the conditional variational network is trained by jointly optimizing the loss function; The joint optimization loss function is: ； Where L represents the joint optimization loss function, Indicates the reconstruction loss. X represents the network parameters of the decoder, Y represents the multimodal design variables, Z represents the corresponding performance response data, and Z represents the latent variables. Denotes KL divergence, This represents the network parameters of the encoder. Indicates a negative feedback term. This represents the performance response data predicted by the conditional variational model obtained from the current training. This represents conditional performance response data. This represents the KL divergence proportionality coefficient. This represents the negative feedback proportionality coefficient; the negative feedback proportionality coefficient is obtained by the following formula: ； in, This represents the error between the predicted performance response data and the corresponding conditional performance response data. This represents the variance of the error.

9. The intelligent generative design method for anti-collision beams according to claim 1, characterized in that, Step S5 includes the following process: Principal component analysis was used to analyze the performance response data obtained from S4 and to determine the objective weight values ​​of each performance response. The subjective weight values ​​of each performance response are obtained by using expert scoring and analytic hierarchy process. The subjective and objective weight values ​​are then solved by a zero-sum game to obtain the comprehensive weight value. The grey relational analysis method is used to obtain the grey relational coefficient of each group of performance response schemes in the performance response data generated by S4, and the grey relational coefficient is multiplied by the comprehensive weight value to obtain the grey relational degree of each group of performance response schemes. All performance response schemes are sorted according to the magnitude of grey relational degree. The scheme with the largest grey relational degree value is the optimal performance response scheme, and the relevant parameters of the anti-collision beam to be designed are the optimal design scheme of the anti-collision beam.

Images