Generative design method and system for vehicle body structure based on large model

Abstract

The invention belongs to the technical field of automobile structure design, and particularly relates to a large-model-based automobile body structure generative design method and system, and the method comprises the steps: building a multi-source database containing automobile body structure pictures, design variables and mechanical properties; training a vehicle body structure performance evaluation model and generating an evaluation model according to the database; on the basis of the generated evaluation model, the general text graph large model is finely adjusted, and a vehicle body structure generative design large model is obtained; generating a structure picture by using the obtained large model, and establishing a structure size evaluation model in combination with a database; inputting design requirements and constraint conditions, generating a vehicle body structure, evaluating the performance and size of the vehicle body structure, and outputting a final result. According to the method, the problems of low efficiency and strong dependence on experience of engineers in an existing automobile concept design stage are solved, the efficiency and creativity of automobile body structure design are improved, the design period is shortened, manual intervention is reduced, and automation and intelligence of automobile body structure design are promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobile structure design, and in particular, relates to a large model-based generative design method and system for a vehicle body structure. Background Art

[0002] With the rapid development of the automotive industry, consumers' demand for personalized and innovative vehicle designs continues to grow, prompting automakers to speed up the iteration of vehicle models. In order to maintain market competitiveness, automakers must shorten the development cycle, especially in the vehicle's conceptual design stage, which plays a vital role in determining the overall layout and functions of the vehicle.

[0003] The development process of existing vehicles usually includes multiple stages, and the conceptual design stage is the key to determining the overall vehicle solution. However, at this stage, engineers have limited product data, and many design decisions must rely on preset frameworks or personal experience, which increases uncertainty in the design process. Engineers need to continuously verify and optimize design solutions, which leads to longer design cycles and lower development efficiency. Especially in the context of rapidly changing market demands, this inefficient design process is difficult to meet industry requirements.

[0004] In addition, the existing conceptual design process relies heavily on the personal experience of engineers. Experienced engineers can complete complex design tasks more efficiently, but for inexperienced engineers, the decision-making process is more difficult when faced with multi-dimensional design requirements, further exacerbating the complexity and time cost of the development process. These problems are becoming more prominent under the current market's requirements for efficient and innovative design. There is an urgent need to improve design efficiency in the conceptual design stage and reduce reliance on personal experience. Therefore, how to improve the efficiency of conceptual design, optimize the decision-making process, and reduce the extension of the design cycle caused by repeated verification through new tools or methods has become a technical problem that needs to be solved in the current automotive design field. Summary of the invention

[0005] In view of the defects of the prior art, the present invention proposes a large-model-based generative design method and system for vehicle body structure. The method can learn from existing vehicle body structure data and mechanical performance information, generate a new vehicle body structure according to design requirements, and realize automated and efficient generative design of vehicle body structure.

[0006] In one aspect, the present invention provides a large model-based generative design method for a vehicle body structure, the method comprising:

[0007] Step 1: Establish a multi-source database containing vehicle body structure images, design variables, and mechanical properties;

[0008] Step 2: Train the vehicle body structure performance evaluation model based on the database in step 1;

[0009] Step 3: Generate an evaluation model by training the vehicle body structure based on the database in step 1;

[0010] Step 4: Based on the generated evaluation model in step 3, the back propagation method is used to fine-tune the training of the general Wensheng graph model to obtain the generative design model of the vehicle body structure;

[0011] Step 5: Generate a structural image using the generative design model of the vehicle body structure obtained in step 4, and establish a structural dimension assessment model in combination with the database in step 1;

[0012] Step 6: Using the generative design model of the body structure obtained in step 4, input the design instructions containing design requirements and constraints to generate the body structure, and use the structural performance evaluation model and structural dimension evaluation model of steps 2 and 5 to evaluate the performance and size, and output the body structure image and its performance and size.

[0013] In a preferred implementation, further, step 1 includes:

[0014] Step 1.1: Build a parametric model of the vehicle body structure and define the key design variables of each structure;

[0015] Step 1.2: According to the design space of the design variables, obtain multiple sets of design variable combinations through experimental design;

[0016] Step 1.3: Based on finite element simulation, obtain the mechanical performance data of different design variable combinations, and capture the corresponding structural images to establish a multi-source database containing body structure images, design variables and mechanical properties.

[0017] In a preferred implementation, further, step 2 includes:

[0018] Step 2.1: Extract the body structure pictures and their mechanical performance data in the database of step 1.3 to construct a training data set;

[0019] Step 2.2: Use the training data from step 2.1, take the body structure image as input and the mechanical performance data as output, train based on the ResNet residual network, and build a structural performance evaluation model.

[0020] In a preferred implementation, further, step 3 includes:

[0021] Step 3.1: Extract the body structure pictures and their mechanical performance data in the database of step 1.3, establish image-text pairs, and construct a training data set;

[0022] Step 3.2: Take the body structure picture and mechanical performance data text in step 3.1 as input, and the cosine similarity between the structure picture and the mechanical performance data text as output, train the VIT model, and build a generative evaluation model.

[0023] In a preferred implementation, further, step 4 includes:

[0024] Step 4.1: Extract the body structure pictures and their mechanical performance data in the database of step 1.3, establish image-text pairs, and construct a training data set;

[0025] Step 4.2: Based on the generation evaluation model obtained in step 3.2, construct a generation evaluation loss function;

[0026] Step 4.3: Based on the Stable Diffusion model, the design text instructions containing design requirements and constraints are used as input, the generation evaluation loss function of step 4.2 is introduced, and the model is trained based on back propagation to finally obtain a large model of generative design of the body structure.

[0027] In a preferred implementation, further, step 5 includes:

[0028] Step 5.1: Using the generative design large model of the vehicle body structure obtained in step 4.3, based on the structural dimensions and mechanical properties data in the database of step 1.3, construct text instructions, input them into the large model, and generate a structural image;

[0029] Step 5.2: Establish coordinate systems in the generated structural image and the body structure image in the database respectively, and use the local features in the structure as a reference to obtain the proportional relationship between the structure generated by the large model and the corresponding structural dimensions in the database, thereby constructing a structural dimension assessment model.

[0030] In a preferred implementation, further, step 6 includes:

[0031] Step 6.1: Using the generative design large model of the vehicle body structure obtained in step 4.3, input text instructions, where the text instructions include design requirements and mechanical performance constraint information, to generate a structural design solution;

[0032] Step 6.2: Evaluate the performance and size using the structural performance evaluation model and structural size evaluation model of step 2.2 and step 5.2, and output the structural image and its performance and size.

[0033] In a preferred implementation, further, step 2.2 includes:

[0034] Step 2.2.1: Input the vehicle body structure image in step 2.1 into the ResNet network, and use the ResNet network to extract the geometric features of the image;

[0035] Step 2.2.2: Map the extracted geometric features into predicted mechanical properties data through a fully connected layer;

[0036] Step 2.2.3: In order to measure the error between the mechanical properties predicted by the model and the actual mechanical properties, the mean square error is used as the loss function;

[0037] Step 2.2.4: The parameters of the model are gradually updated by minimizing the loss function, and the optimization process uses the Adam optimizer.

[0038] In another aspect, the present invention provides a large model-based generative design system for a vehicle body structure, comprising:

[0039] Database module, used to build a multi-source database containing vehicle body structure pictures, design variables and mechanical properties;

[0040] A structural performance evaluation model building module is used to train a vehicle body structural performance evaluation model based on a multi-source database;

[0041] Generate an evaluation model building module, which is used to train the vehicle body structure to generate an evaluation model based on a multi-source database;

[0042] The generative design big model fine-tuning module is used to fine-tune the training of the general Wensheng graph big model based on the generation evaluation model and the back propagation method to obtain the generative design big model of the vehicle body structure;

[0043] The structural dimension assessment model building module is used to generate structural images based on the generative design large model of the vehicle body structure and to establish a vehicle body structure dimension assessment model in combination with a multi-source database;

[0044] The body structure generation and evaluation module is used to input design instructions containing design requirements and constraints based on the body structure generative design large model, generate a body structure picture, and evaluate the performance and size of the generated body structure through the structural performance evaluation model and the structural size evaluation model, and output the body structure picture and its performance and size data.

[0045] The beneficial effects of the present invention are:

[0046] First, the generative design method of the body structure based on the large model of the present invention integrates multi-source databases, structural performance evaluation and size evaluation models to build a large model of generative design of the body structure, and realizes the creative generation of body structures that meet the design requirements. Compared with traditional design, the generative design method can quickly extract and learn the characteristics of existing structural data, automatically generate body structure pictures that meet performance and size requirements, greatly improve design efficiency, shorten the development cycle, reduce dependence on engineer experience, and ensure the accuracy and diversity of the design, and improve the automation and intelligence level of body structure design.

[0047] Second, in the preferred implementation, step 2 of the present invention can effectively improve the accuracy and efficiency of vehicle body structure performance evaluation by using vehicle body structure images and mechanical performance data combined with ResNet residual network for model training. Through the automatic feature extraction capability of the deep learning model, the complex mapping relationship between structural images and mechanical properties can be accurately captured to achieve rapid prediction of the mechanical properties of the vehicle body structure. Compared with traditional mechanical analysis methods, this image- and data-driven evaluation method greatly shortens the evaluation time and reduces the cost of complex mechanical simulation calculations, while providing more accurate and rapid feedback for structural design.

[0048] Third, in the preferred implementation, step 3 of the present invention uses the body structure picture and mechanical performance data, combined with the VIT model for training, to build a body structure generation evaluation model. Unlike traditional design, generative design can accurately capture the strong nonlinear mapping relationship between structure and performance by learning from existing data, thus realizing creative design. The cosine similarity between the structure picture and the mechanical performance data text is used to build a generation evaluation model to evaluate the generated solution, ensuring that the generated structure is highly matched with the input design requirements and constraints in terms of function and performance.

[0049] Fourth, in the preferred implementation, step 4 of the present invention realizes the automatic generation of the vehicle body structure by introducing the Stable Diffusion model and the structure generation evaluation loss function, combined with the back propagation method, which fully reflects the advantages of generative design. Compared with the traditional design process, this method can extract feature information from the existing vehicle body structure pictures and mechanical performance data, and creatively generate a vehicle body structure that meets the design requirements from 0 to 1. Through the structure generation evaluation loss function, the system automatically optimizes the generated design scheme to ensure that it not only meets the requirements in terms of geometric shape, but also meets the constraints of mechanical performance.

[0050] Fifth, in the preferred implementation, the structural dimension evaluation model in step 5 of the present invention effectively solves the problem of fuzzy and inaccurate dimensions in generative design. By establishing a coordinate system and calculating the dimension ratio based on local features, the newly generated structure can be accurately compared with the actual structural dimensions in the database to ensure that the design results meet the expected geometric requirements. At the same time, it can provide more accurate dimension feedback for subsequent modeling and design optimization.

[0051] Sixth, in the preferred implementation, step 6 of the present invention can realize the full process automation of body structure design by combining the body structure generative design large model and the performance and size evaluation model, and integrate the processing from design generation to performance and size evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of a large model-based generative design method for a vehicle body structure of the present invention;

[0053] Figure 2 is a schematic diagram of a large model-based generative design method for a vehicle body structure of the present invention;

[0054] Figure 3 is a schematic diagram of constructing a structural performance evaluation model of an embodiment of the present invention;

[0055] Figure 4 is a comparison diagram of the predicted value and the actual value of the torsional stiffness performance of the hood structure image by the structural performance evaluation model of the embodiment of the present invention;

[0056] Figure 5 is a comparison diagram of the predicted value and the actual value of the bending stiffness performance of the hood structure image by the structural performance evaluation model of the embodiment of the present invention;

[0057] Figure 6 is a schematic diagram of constructing a generation evaluation model according to an embodiment of the present invention;

[0058] Figure 7 is a schematic diagram of constructing a large model of a generative design of a vehicle body structure according to an embodiment of the present invention;

[0059] Figure 8 is a schematic diagram of constructing a structural dimension assessment model according to an embodiment of the present invention;

[0060] Fig. 9 is a schematic diagram of key dimensions of the hood structure of an embodiment of the present invention;

[0061] Fig.10 is a hood structure diagram generated by using a body structure generative design macro model according to an embodiment of the present invention;

[0062] Fig.11It is a schematic diagram of the large model-based generative design system for vehicle body structure of the present invention. DETAILED DESCRIPTION

[0063] In order to enable those skilled in the art to better understand the technical solution of the present application, the present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0064] In the description of this application, the terms "first", "second", and "third" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance; the term "plurality" refers to two or more, unless otherwise expressly defined. Terms such as "installed", "connected", "connected", and "fixed" should be understood in a broad sense. For example, "connected" can be a fixed connection, a detachable connection, or an integral connection; "connected" can be a direct connection or an indirect connection through an intermediate medium. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to the specific circumstances.

[0065] In the description of the present application, it should be understood that the directions or positional relationships indicated by the terms "up", "down", "left", "right", "front", "back", etc. are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or unit referred to must have a specific direction, be constructed and operated in a specific orientation, and therefore, should not be understood as a limitation on the present application.

[0066] In the description of this specification, the description of the terms "one embodiment / method", "some embodiments / methods", "specific embodiments / methods", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment / method or example are included in at least one embodiment / method or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment / method or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments / methods or examples in a suitable manner.

[0067] The purpose of the present invention is to provide a generative design method for body structure based on a large model. By establishing a multi-source database, training a structure performance evaluation model, training a structure generation evaluation model and fine-tuning a general generation model, the automatic generation and evaluation of body structure design schemes can be finally achieved. This method can not only greatly improve the design efficiency, but also reduce the dependence on experience while ensuring that the performance and size of the design scheme meet the requirements.

[0068] As the instruction manual Figure 1-2 , the method comprising:

[0069] Step 1: Establish a multi-source database containing vehicle body structure images, design variables, and mechanical properties.

[0070] Step 1: By building a multi-source database containing body structure images, design variables and mechanical properties, the system is ensured to have a variety of design samples and performance data. Specifically, it includes:

[0071] Step 1.1: Build a parametric model of the vehicle body structure and define the key design variables of each structure.

[0072] It should be noted that the vehicle body structure includes the hood, doors, body-in-white, etc. Each structure has unique geometric characteristics and mechanical requirements. By building parametric models of these structures, the foundation for subsequent simulation and design can be laid.

[0073] Specifically, a parametric modeling software (such as SFE-CONCEPT) is used to create a parametric model of the vehicle body structure. The parametric model allows the design variables of the structure to be flexibly modified. In the design of the vehicle body structure, the design variables usually include geometric features and material properties. Geometric design variables include thickness, length, width, radius of curvature, etc. These variables affect the shape and size of each structure of the vehicle body. Material design variables include material type (such as aluminum alloy, steel, carbon fiber, etc.), material density, elastic modulus, etc. The design variables of each structure are determined according to its function in the vehicle body, and these variables determine the mechanical properties and weight of the vehicle body structure.

[0074] For example: The main function of the hood is to protect the engine and absorb some of the energy in the event of a collision. Therefore, the key design variables of the hood may include thickness, material type (such as aluminum alloy or carbon fiber) and curvature, which directly affect its impact resistance and weight. The function of the door is to provide passengers with ingress and egress and provide side protection. Therefore, the key design variables of the door may include thickness, material strength and frame geometry, which affect its side impact resistance and rigidity. As the frame structure of the entire vehicle, the key design variables of the body in white may involve the strength, thickness and shape of the material to ensure that it has high stiffness and compression resistance.

[0075] Step 1.2: Based on the design space of design variables, obtain multiple sets of design variable combinations through experimental design.

[0076] Specifically, the experimental design methods include Latin square experimental design, full factorial experimental design or Monte Carlo method. Choose the appropriate experimental design method according to the number of design variables, the complexity of the interaction and the specific needs of the project. Latin square experimental design is suitable for use when there are many design variables, and can cover the entire design space with a smaller number of experiments. The values ​​of each design variable are designed to be arranged in a non-repeating manner, thereby effectively reducing the number of experiments. Full factorial experimental design is suitable for comprehensive exploration of the interactions between all design variables. This method generates all possible combinations of variables, including different values ​​of each design variable and all combinations of values ​​of other variables. The Monte Carlo method generates a large number of design variable combinations through random sampling, which is suitable for processing variables with large uncertainties or when multiple possibilities need to be simulated. The Monte Carlo method can obtain different combinations of design variables through multiple random simulations, which is especially suitable for complex and uncertain design spaces.

[0077] Taking the hood structure of the car body as an example, by changing the design variables such as thickness and material type, multiple combinations are generated, which are potential design solutions for the hood. Each combination will generate a unique design instance, that is, an independent 3D model is created for the variable combination, and these 3D models are used for simulation analysis in subsequent steps.

[0078] Step 1.3: Based on finite element simulation, obtain the mechanical performance data of different design variable combinations, and capture the corresponding structural images to establish a multi-source database containing body structure images, design variables and mechanical properties.

[0079] Specifically, finite element analysis tools (such as Optistruct, ANSYS, Abaqus) are used to simulate the mechanical properties of each design variable combination. The multi-source database of the vehicle body structure can be expressed as:

[0080] DB={(P i ,X i ,M i )|i=1,2,……,N} (1)

[0081] Where: DB represents the multi-source database of vehicle body structures; i represents the index of vehicle body structures, representing different structures (such as hood, door, body-in-white, etc.); N represents the number of all structures; P i Represents a picture of structure i (the picture is a front view or a three-dimensional structure view of each structure); X i represents the set of key design variables of structure i, X i ={x i1 ,x i2 ,……,x im}, where x ij represents the jth design variable of the structure, and the number of design variables is m; Mi represents the mechanical properties data set of structure i (such as stress distribution, deformation, mode, etc.), M i ={m i1 ,m i2 ,……,m ik}, where m ij Represents the jth mechanical performance data of the structure, and the number of mechanical performance indicators is k.

[0082] Taking the hood as an example, assuming that the design variables include thickness x1, material type x2, and the mechanical properties include stress distribution m1 and total deformation m2. Then the database of the hood can be expressed as:

[0083] (P 引擎盖 ,{x1,x2},{m1,m2}) (2)

[0084] Step 2: Train the vehicle body structure performance evaluation model based on the database in step 1.

[0085] It should be noted that the structural performance evaluation model evaluates the mechanical properties of the body-related structures. Given a body structure picture, it predicts and evaluates their mechanical performance data. This model focuses more on quantitative analysis and prediction of the mechanical properties of a given structure, aiming to determine the performance of the structure under specific mechanical conditions, such as stiffness, stress, strength, etc.

[0086] Step 2 uses the structural images and mechanical properties data in the database to train a structural performance evaluation model based on the ResNet residual network, extracting effective information from the existing design and performance data for evaluating the mechanical properties of the new design.

[0087] As the instruction manual Figure 3 , building a vehicle body structure performance evaluation model specifically includes:

[0088] Step 2.1: Extract the body structure images and their mechanical performance data from the database in step 1.3 to construct a training data set.

[0089] The structural image data and mechanical properties data are paired as the input and output of the model, and each pair of data will be used as a training sample for the model. The formula is expressed as:

[0090] {(P i ,M i )|i=1,2,……,N} (3)

[0091] Where: P i Indicates a structural image; M i Represents the mechanical properties data set corresponding to the structure.

[0092] Step 2.2: Use the training data from step 2.1, take the body structure image as input and the mechanical performance data as output, train based on the ResNet residual network, and build a structural performance evaluation model.

[0093] It should be noted that ResNet (Residual Network) is a deep learning model commonly used in image processing. It has a skip connections mechanism, which can effectively alleviate the gradient vanishing problem in deep neural networks and is suitable for processing complex image tasks.

[0094] The training process of the entire model can be expressed as a cyclic iterative process:

[0095] Step 2.2.1: The training sample pairs (P i ,M i ) in the picture P i Input ResNet network and use ResNet network to extract image P i The geometric characteristics of are as follows:

[0096] z i =f ResNet 9P i ) (4)

[0097] Where: z i represents the high-dimensional geometric features extracted from the image (such as the shape, curvature, thickness, etc. of the hood); f ResNet (P i ) represents the image P obtained through the ResNet residual network i The extracted feature vector z i .

[0098] Step 2.2.2: The extracted geometric features are mapped into predicted mechanical properties data through a fully connected layer. The formula is as follows:

[0099]

[0100] Where: Represents the mechanical properties predicted by the model, including stress distribution, deformation, modal frequency, etc.; W z represents the weight matrix of the fully connected layer; b z represents the bias term; z i Indicates that from the picture P i Extracted geometric features.

[0101] Step 2.2.3: To measure the mechanical properties predicted by the model and the real mechanical properties M iThe error between , using mean square error (MSE) as the loss function, the formula is as follows:

[0102]

[0103] Where: L(θ) represents the loss function of the model; Represents the mechanical property value predicted by the model; M i Represents real mechanical properties data.

[0104] Step 2.2.4: The parameters of the model are gradually updated by minimizing the loss function. The optimization process uses the Adam optimizer, and the formula is as follows:

[0105]

[0106] Where: θ t represents the model parameters at the tth iteration; η represents the learning rate; Represents the gradient of the loss function L(θ) with respect to the parameters.

[0107] After training, the structural performance evaluation model can predict the mechanical properties of the vehicle body based on the input vehicle body structure image. The core formula is:

[0108]

[0109] A specific embodiment includes:

[0110] Taking the hood structure as an example, first according to step 2.1, extract the front view or views from other angles of the hood, which are generated by the parametric modeling and simulation in step 1. This image will be used as the input of the model to learn the geometric features of the hood. Extract the mechanical properties data corresponding to the hood structure from the simulation results of step 1. The mechanical properties data include stress distribution, deformation, and modal analysis. These mechanical properties data will be used as the output of the model. Through training, the model will be able to predict the corresponding mechanical properties based on the input image. Then associate the hood structure image with its corresponding mechanical properties data to form a training sample. The formula for the training sample is (P 引擎盖 ,M 引擎盖 )=(hood structure picture,{stress distribution, deformation, modal analysis}).

[0111] Then according to step 2.2, (P 引擎盖 ,M 引擎盖 ) The training sample data is divided into a training set (80%) and a validation set (20%) according to the ratio. The training set is used to adjust the parameters of the ResNet model, and the validation set is used to evaluate the generalization ability of the model. The ResNet network structure includes an input layer, a convolutional layer, and a residual block. 引擎盖The input is sent to the ResNet residual network, which extracts geometric features from the image through a series of convolutional layers and residual blocks, including parameters such as the shape, curvature, and thickness of the hood. Using the regression method, the extracted features are mapped into mechanical performance data such as stress distribution, deformation, and modal frequency through the fully connected layer. Using the mean square error (MSE) as the loss function, the Adam optimizer is optimized to gradually reduce the error. After each training cycle, the model is tested using the validation set to evaluate the generalization ability of the model and prevent overfitting. Ultimately, the model is able to predict the mechanical properties of the hood based on the input hood structure image.

[0112] As the instruction manual Figure 4-5 , Figure 4 The prediction results of the torsional stiffness performance by the structural performance assessment model based on the hood structure picture are shown; Figure 5 The figure shows the model's prediction of the bending stiffness performance of the hood structure image. Figure 4 and Figure 5 It can be seen that the model prediction value is highly consistent with the actual value, which shows that the structural performance evaluation model has a high prediction accuracy. The high-precision prediction of the model helps to accurately evaluate the performance of the body structure generated by the generative design large model of the body structure, providing reliable support for further optimization of the design.

[0113] Step 3: Generate an evaluation model based on the vehicle body structure training database in step 1.

[0114] It should be noted that the generation evaluation model is used to evaluate the rationality of structure generation. The focus of this model is to determine the similarity between a certain structure image and mechanical property data through training, and then judge whether the structure is reasonable.

[0115] Step 3 uses the Vision Transformer (VIT) model to train a structure generation evaluation model to capture the complex relationship between the body structure image and mechanical properties, and uses the evaluation model to guide the generation process of the new design.

[0116] As the instruction manual Figure 6 ,Constructing the body structure generation evaluation model specifically includes:

[0117] Step 3.1: Extract the body structure images and their mechanical performance data from the database in step 1.3, establish image-text pairs, and construct a training data set.

[0118] It should be noted that the process of step 3.1 is similar to step 2.1 and will not be repeated here.

[0119] Step 3.2: Take the body structure picture and mechanical performance data text in step 3.1 as input, and the cosine similarity between the structure picture and the mechanical performance data text as output, train the VIT model, and build a generative evaluation model.

[0120] It should be noted that the VIT (Vision Transformer) model is an application of the Transformer architecture. It uses the self-attention mechanism to capture the global relationship of the input data. Specifically, in VIT, the input image is regarded as a sequence (the image is divided into a series of small blocks), and then the relationship between each block is learned through the self-attention mechanism, which can extract complex features from the image. The input image and mechanical properties text data are trained through the VIT model to learn the correspondence between the image and the text. Using cosine similarity as the loss function, the similarity between the structural image and the mechanical properties data is evaluated, and finally an evaluation model is trained to judge the rationality of the generated structure.

[0121] Specifically, the process of building a structure generation evaluation model includes:

[0122] Step 3.2.1: The VIT model receives the relevant structure image as input, divides the image into n fixed-size blocks, converts each block into a vector representation through linear mapping, adds position encoding, and inputs the feature vectors of all blocks into the Transformer encoder to extract the global features of the image.

[0123] The image is divided into n fixed-size blocks and can be expressed as:

[0124] P i →[p 1, p2,……,p n ] (9)

[0125] Where: P i Represents a structural image; each p k represents a block, and n represents the number of blocks.

[0126] Convert each block into a vector representation through a linear mapping:

[0127] z k =Linear(p k )+PositionEncoding(k) (10)

[0128] Where: z k Represents block p k The feature vector contains the characteristics and location information of the block.

[0129] The global features can be expressed as:

[0130] z i =f VIT ([z1,z2,……,z n ]) (11)

[0131] Where: z i Indicates the image P i The global feature vector contains the global relationship between each block in the image.

[0132] Step 3.2.2: In order to measure the similarity between the image features and the mechanical property vector, cosine similarity is used for comparison. The formula for cosine similarity is:

[0133]

[0134] Where: z i ·M i Represents the image feature vector z i and the mechanical property vector M i The dot product of ||z i || and ||M i || represents z i and M i The norm of .

[0135] Step 3.2.3: In order to train the model, the negative value of cosine similarity is minimized as the loss function. By minimizing this loss, the model can learn how to maximize the similarity between image features and mechanical properties, thereby establishing a strong association between them. During the training process, the optimization algorithm is used to optimize the parameters of the VIT model.

[0136] The loss function can be expressed as:

[0137] L=1-cosθ (13)

[0138] The optimization formula is:

[0139]

[0140] Where: θ t represents the model parameters at the tth iteration; η represents the learning rate; Represents the gradient of the loss function with respect to the model parameters.

[0141] A specific embodiment includes:

[0142] Taking the hood structure as an example, assume that a database contains a large number of hood structural images and corresponding mechanical properties data. The mechanical properties data include physical properties such as stress distribution and stiffness of the hood under different loads. Each set of data contains a 3D structural image of a hood and a text description of the mechanical properties of the structure, such as material strength, weight, deformation, etc. The structural images and mechanical properties data of each hood are extracted from the database and combined into image-text pairs. The image represents the 3D design model of the hood, while the text data contains the performance of the hood in various physical properties, such as stress analysis and strength calculation results. These image-text pairs will be used as training data for the model. The VIT model is used to process these data. The VIT model divides the structural image of the hood into 16×16 pixel blocks and extracts global features through a self-attention mechanism. At the same time, the mechanical properties data of the hood are input into the model. The VIT model establishes a mapping between images and texts by learning the relationship between them. During the model training process, the matching degree between the image and the text is evaluated by calculating the cosine similarity between the global feature vector of the image and the mechanical properties data vector. The higher the cosine similarity, the more the structural image matches the corresponding mechanical properties. Based on this, the model will generate a rationality evaluation of the hood structure. Finally, the trained VIT model is used as the hood structure generation evaluation model. New hood structure images and mechanical performance requirement text data can be input, and the rationality score of the hood structure solution can be given through the cosine similarity evaluation.

[0143] Step 4: Based on the structure generation evaluation model in step 3, the back propagation method is used to fine-tune the training of the general Wensheng graph model to obtain the body structure generative design model.

[0144] It should be noted that the backpropagation method is an important algorithm for training neural networks in deep learning. By calculating the gradient of the loss function, the parameters of the model are adjusted to minimize the error, making the prediction result closer to the target value.

[0145] The general text-to-image model is a large model for generating images from text. It is a generative model based on deep learning, which can convert text input (i.e. design requirements or descriptions) into corresponding image outputs. The model architecture of the general text-to-image model usually adopts diffusion models (such as Stable Diffusion) or generative adversarial networks (GAN) and other architectures, and gradually converts the input text information into images through deep learning algorithms.

[0146] Step 4 is based on the generated evaluation model obtained in step 3, uses the Stable Diffusion model for back propagation training, and combines the design requirements and constraints to generate a vehicle body structure image that meets the functional and performance requirements.

[0147] As the instruction manual Figure 7 , building a generative design model for the body structure specifically includes:

[0148] Step 4.1: Extract the body structure images and their mechanical performance data from the database in step 1.3, establish image-text pairs, and construct a training data set.

[0149] It should be noted that the process of step 4.1 is the same as step 2.1 and will not be repeated here.

[0150] Step 4.2: Based on the generation evaluation model obtained in step 3.2, construct a generation evaluation loss function.

[0151] This loss function is used to evaluate the difference between the generated structure and the mechanical performance design requirements and constraints, ensuring that the generated design meets the input mechanical performance requirements. The design of the loss function includes the mechanical performance difference loss. The mechanical performance difference loss is based on cosine similarity and measures the degree of match between the mechanical properties of the generated structure and the input mechanical performance description. The generation evaluation loss function is expressed as follows:

[0152]

[0153] Where: p represents the number of mechanical properties characteristics; θ i Represents the cosine similarity between the generated structure and the input mechanical property description on the i-th mechanical property feature.

[0154] Step 4.3: Based on the Stable Diffusion model, the design text instructions containing design requirements and constraints are used as input, the generation evaluation loss function of step 4.2 is introduced, and the model is trained based on back propagation to finally obtain a large model of generative design of the body structure.

[0155] It should be noted that the Stable Diffusion model is a generative model based on diffusion models. It generates images that meet the target from random noise through a step-by-step denoising method. The StableDiffusion model uses version V1.5, which is an improved stable diffusion model that is good at generating high-quality, detailed images based on the input text description. Constraints include size, material, performance requirements, etc. These constraints are introduced into the Stable Diffusion model through feedback from the loss function to adjust the generated details.

[0156] The process of model training based on back propagation includes: inputting the design text instructions of any structure of the vehicle body into the Stable Diffusion model, and generating a set of vehicle body structure images through the Stable Diffusion model. Use the generation evaluation loss function to evaluate the difference between the generated image and the text description. According to the loss function, back-propagate the gradient information and update the model parameters so that the next generated image is more in line with the design requirements and mechanical performance requirements of the text description. After multiple iterations, the model is gradually optimized, so that the accuracy and rationality of the generated structure are continuously improved. Finally, the trained Stable Diffusion model can be used as a large model for structural generative design, which can automatically generate vehicle body structure images that meet the design requirements based on the input design text.

[0157] Step 5: Generate an image using the generative design model of the vehicle body structure obtained in step 4, and build a structural dimension assessment model based on the database in step 1.

[0158] Step 5 establishes a size ratio relationship between the generated structure and the existing structure in the database, and obtains the size data of the generated structure through the size assessment model.

[0159] As the instruction manual Figure 8 , building a vehicle body structure size assessment model specifically includes:

[0160] Step 5.1: Using the generative design large model of the body structure obtained in step 4.3, based on the structural dimensions and mechanical properties data in the database in step 1.3, construct text instructions, input them into the large model, and generate a structural image.

[0161] The text instructions are the design requirements input into the body structure generative design big model, including the structural dimensions and mechanical properties data in the database of step 1.3. Taking the hood as an example, the text instructions are constructed as follows: "Generate a hood with a size of 1200x700mm, made of aluminum alloy, 2mm thick, torsional stiffness greater than 80N·m / deg, bending stiffness greater than 139N / mm, and mass less than 13kg." These instructions tell the body structure generative design big model what kind of hood structure needs to be generated, especially the requirements for the geometric dimensions and mechanical properties of the hood. After entering the text instructions, the body structure generative design big model generates a hood structure picture that meets the design requirements based on these requirements, and these generated pictures will be used for subsequent dimensional evaluation.

[0162] Step 5.2: Establish coordinate systems in the generated structural image and the body structure image in the database respectively, and use the local features in the structure as a reference to obtain the proportional relationship between the structure generated by the large model and the corresponding structural dimensions in the database, thereby constructing a structural dimension assessment model.

[0163] Specifically, the coordinate system established in step 5.2 is used to calibrate the geometric features and relative positions in the image. The coordinate system can be a Cartesian coordinate system on a two-dimensional plane, assuming that the origin is a fixed reference point on the hood, such as the lower left corner or center position of the hood. In the generated model structure image and the real model structure image in the database, the coordinate system is used to determine the geometric dimensions of the local features, which can be the edge, aperture, curved area or other specific geometric dimensions of the hood. The model takes the hood as an example, such as: the total length and width of the hood, the curve length of the edge of the hood, the diameter of the aperture on the hood or the position of the hole.

[0164] By accurately calibrating local features, the size ratio relationship between the generated image and the real image can be obtained. The local features of the generated image and the real image are compared one by one, and the size ratio of each feature is calculated. Then, by weighted averaging multiple proportional relationships of local features, the overall size ratio relationship can be obtained to evaluate whether the generated structure meets the actual requirements in terms of size. The proportional relationship of a local feature size between the large model generated structure and the corresponding structure in the database is expressed by the following formula:

[0165]

[0166] Where: R feature Indicates the size ratio of a local feature; S generated represents the size of the feature in the generated image (e.g. the diameter of a hole); S real Represents the size of the corresponding feature of the real image in the database.

[0167] During the training process of the evaluation model, a size loss function can be designed based on the size ratio difference between the generated structure and the real structure:

[0168]

[0169] Where: represents the size ratio of the i-th local feature; m represents the total number of local features.

[0170] It should be noted that, ideally, the proportional relationship is 1, that is, the size of the generated structure and the real structure are exactly the same, and the loss is 0.

[0171] As the instruction manual Fig. 9 , Fig. 9 The key dimensions of the hood structure are shown in Figure 1. In order to evaluate the hood structure dimensions, the structural dimension evaluation model is used to evaluate and analyze the hood structure diagram. Table 1 lists the comparison data between the actual dimensions and the evaluated dimensions.

[0172] Table 1

[0173] Key Dimensions Actual size (mm) Evaluation size (mm) error A 136 134.2 1.32％ B 158 158.9 0.57％ C 244 246.1 0.86％ D 84 84.3 0.36％ E 80 78.9 1.37％ F 58 58.9 1.55％

[0174] From the data in Table 1, it can be seen that the error between the evaluation size and the actual size is within a reasonable range, and the maximum error does not exceed 1.55%. This shows that the accuracy of the structural size evaluation model is high, which is conducive to the accurate size evaluation of the body structure generated by the large model of the body structure generative design.

[0175] Step 6: Using the generative design model of the body structure obtained in step 4, input the design instructions containing design requirements and constraints to generate the body structure, and use the structural performance evaluation model and structural dimension evaluation model of steps 2 and 5 to evaluate the performance and size, and output the body structure image and its performance and size.

[0176] Step 6: By inputting the design requirements and mechanical performance constraint information, the generated structure is automatically evaluated through the performance and size evaluation model, and the structure image and its performance and size are output. Specifically, it includes:

[0177] Step 6.1: Using the generative design large model of the vehicle body structure obtained in step 4.3, input text instructions, where the text instructions include design requirements and mechanical performance constraint information, to generate a structural design solution.

[0178] Step 6.2: Evaluate the performance and size using the structural performance evaluation model and structural size evaluation model of step 2.2 and step 5.2, and output the structural image and its performance and size.

[0179] As the instruction manual Fig.10 Taking the hood as an example, the generative design model of the hood structure obtained based on the hood database is input with design instructions containing design requirements and constraints. The design instructions are: "Generate a bilaterally symmetrical hood inner panel structure diagram, including 3 longitudinal stiffeners and 4 transverse stiffeners. The hood mass is less than 13kg, the torsional stiffness is greater than 83N·m / deg, and the bending stiffness is greater than 138N / mm", and then the hood structure picture is generated. Fig.10 It can be seen that the generated hood structure meets the design requirements, the layout of each part is reasonable, the structural design is clear and regular, and it is easy to produce and manufacture.

[0180] The hood structure image is evaluated for both performance and size through the structural performance evaluation model and the structural size evaluation model to ensure that the generated design meets expectations. During performance evaluation, the hood structure performance evaluation model is a model trained based on the ResNet residual network, which can predict its mechanical performance data based on the input structure image. During size evaluation, the hood structure size evaluation model is a model based on the comparison of the coordinate system and local features, which evaluates whether the geometric dimensions of the generated structure meet expectations. The performance evaluation results and the size evaluation results will be integrated into a final report to evaluate whether the generated hood design meets the design requirements in terms of both size and performance. The performance evaluation report includes stress distribution diagrams, stiffness data, strength analysis, etc. The size evaluation report includes the size deviation of each local feature, the percentage of total size deviation, etc.

[0181] The evaluation results are shown in Table 2.

[0182] Table 2

[0183]

[0184] As can be seen from Table 2, the total mass of the hood is 12.88kg. This mass is below the 13kg required by the design directive, indicating that the generated hood meets the quality control requirements. The evaluated torsional stiffness is 83.76N·m / deg, which is slightly higher than the 83N·m / deg required in the design directive. Higher torsional stiffness means that the hood has better resistance to torsion when subjected to force. The evaluated bending stiffness is 140.19N / mm, which exceeds the 138N / mm required in the design directive. This shows that the generated hood also has good bending resistance when subjected to bending force. The evaluation values ​​of the dimensions of key components such as the rear crossbeam, front crossbeam, side longitudinal beam, and side crossbeam are all within a reasonable range and meet the design requirements. The dimensions of key positions can be used to establish a three-dimensional model.

[0185] The present invention realizes the full process automation of vehicle body structure design by integrating multi-source database, generative large model, structural performance evaluation model and structural size evaluation model. This method can not only automatically generate vehicle body structure images that meet design requirements, but also ensure the reliability and readability of the design scheme through performance and size evaluation, greatly improve design efficiency, shorten design cycle, reduce dependence on manual experience, and improve the intelligence and accuracy of design, which is suitable for rapid iteration and innovation of complex structures.

[0186] As the instruction manual Fig.11 Another object of the present invention is to provide a large-scale model-based generative design system for a vehicle body structure, the system comprising:

[0187] Database module, used to build a multi-source database containing vehicle body structure pictures, design variables and mechanical properties;

[0188] A structural performance evaluation model building module is used to train a vehicle body structural performance evaluation model based on a multi-source database;

[0189] Generate an evaluation model building module, which is used to train the vehicle body structure to generate an evaluation model based on a multi-source database;

[0190] The generative design big model fine-tuning module is used to train and fine-tune the general Wensheng graph big model based on the generation and evaluation model using the back propagation method to obtain the generative design big model of the vehicle body structure;

[0191] The structural dimension assessment model building module is used to generate structural images based on the generative design large model of the vehicle body structure and to establish a vehicle body structure dimension assessment model in combination with a multi-source database;

[0192] The body structure generation and evaluation module is used to input design instructions containing design requirements and constraints based on the body structure generative design large model, generate a body structure picture, and evaluate the performance and size of the generated body structure through the structural performance evaluation model and the structural size evaluation model, and output the body structure picture and its performance and size data.

[0193] The above is only an embodiment of the present invention, and the common sense such as the known specific structure and characteristics in the scheme is not described in detail here. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present invention is defined by the attached claims rather than the above description, so it is intended to include all changes within the meaning and scope of the equivalent elements of the claims. Any figure mark in the claims should not be regarded as limiting the claims involved.

Claims

1. A generative design method for a vehicle body structure based on a large model, characterized in that: Methods include: Step 1: Establish a multi-source database containing vehicle body structure images, design variables, and mechanical properties; Step 2: Train the vehicle body structure performance evaluation model based on the database in step 1; Step 3: Generate an evaluation model by training the vehicle body structure based on the database in step 1; Step 4: Based on the generated evaluation model in step 3, the back propagation method is used to fine-tune the training of the general Wensheng graph model to obtain the generative design model of the vehicle body structure; Step 5: Generate a structural image using the generative design model of the vehicle body structure obtained in step 4, and establish a structural dimension assessment model in combination with the database in step 1; Step 6: Using the generative design large model of the body structure obtained in step 4, input the design instructions containing design requirements and constraints to generate the body structure, and use the structural performance evaluation model and structural dimension evaluation model of steps 2 and 5 to evaluate the performance and size, and output the body structure image and its performance and size.

2. The large model-based generative design method for vehicle body structure according to claim 1, characterized in that: Step 1 includes: Step 1.1: Build a parametric model of the vehicle body structure and define the key design variables of each structure; Step 1.2: According to the design space of the design variables, obtain multiple sets of design variable combinations through experimental design; Step 1.3: Based on finite element simulation, obtain the mechanical performance data of different design variable combinations, and capture the corresponding structural images to establish a multi-source database containing body structure images, design variables and mechanical properties.

3. The large model-based generative design method for vehicle body structure according to claim 1, characterized in that: Step 2 includes: Step 2.1: Extract the body structure pictures and their mechanical performance data in the database of step 1.3 to construct a training data set; Step 2.2: Use the training data from step 2.1, take the body structure image as input and the mechanical performance data as output, train based on the ResNet residual network, and build a structural performance evaluation model.

4. The large model-based generative design method for vehicle body structure according to claim 1, characterized in that: Step 3 includes: Step 3.1: Extract the body structure pictures and their mechanical performance data in the database of step 1.3, establish image-text pairs, and construct a training data set; Step 3.2: Take the body structure picture and mechanical performance data text in step 3.1 as input, and the cosine similarity between the structure picture and the mechanical performance data text as output, train the VIT model, and build a generative evaluation model.

5. The large model-based generative design method for vehicle body structure according to claim 1, characterized in that: Step 4 includes: Step 4.1: Extract the body structure pictures and their mechanical performance data in the database of step 1.3, establish image-text pairs, and construct a training data set; Step 4.2: Based on the generation evaluation model obtained in step 3.2, construct a generation evaluation loss function; Step 4.3: Based on the Stable Diffusion model, the design text instructions containing design requirements and constraints are used as input, the generation evaluation loss function of step 4.2 is introduced, and the model is trained based on back propagation to finally obtain a large model of generative design of the body structure.

6. The large model-based generative design method for vehicle body structure according to claim 1, characterized in that: Step 5 includes: Step 5.1: Using the generative design large model of the vehicle body structure obtained in step 4.3, based on the structural dimensions and mechanical properties data in the database in step 1.3, construct text instructions, input them into the large model, and generate a structural image; Step 5.2: Establish coordinate systems in the generated structural image and the body structure image in the database respectively, and use the local features in the structure as a reference to obtain the proportional relationship between the structure generated by the large model and the corresponding structural dimensions in the database, thereby constructing a structural dimension assessment model.

7. The large model-based generative design method for vehicle body structure according to claim 1, characterized in that: Step 6 includes: Step 6.1: Using the generative design large model of the vehicle body structure obtained in step 4.3, input text instructions, where the text instructions include design requirements and mechanical performance constraint information, to generate a structural design solution; Step 6.2: Evaluate the performance and size using the structural performance evaluation model and structural size evaluation model of step 2.2 and step 5.2, and output the structural image and its performance and size.

8. The large model-based generative design method for vehicle body structure according to claim 2, characterized in that: Step 2.2 includes: Step 2.2.1: Input the vehicle body structure image in step 2.1 into the ResNet network, and use the ResNet network to extract the geometric features of the image; Step 2.2.2: Map the extracted geometric features into predicted mechanical properties data through a fully connected layer; Step 2.2.3: In order to measure the error between the mechanical properties predicted by the model and the actual mechanical properties, the mean square error is used as the loss function; Step 2.2.4: The parameters of the model are gradually updated by minimizing the loss function, and the optimization process uses the Adam optimizer.

9. A generative design system for vehicle body structure based on a large model, characterized in that: include: Database module, used to build a multi-source database containing vehicle body structure pictures, design variables and mechanical properties; A structural performance evaluation model building module is used to train a vehicle body structural performance evaluation model based on a multi-source database; Generate an evaluation model building module, which is used to train the vehicle body structure to generate an evaluation model based on a multi-source database; The generative design big model fine-tuning module is used to fine-tune the training of the general Wensheng graph big model based on the generation evaluation model and the back propagation method to obtain the generative design big model of the vehicle body structure; The structural dimension assessment model building module is used to generate structural images based on the generative design large model of the vehicle body structure and to establish a vehicle body structure dimension assessment model in combination with a multi-source database; The body structure generation and evaluation module is used to input design instructions containing design requirements and constraints based on the body structure generative design large model, generate a body structure picture, and evaluate the performance and size of the generated body structure through the structural performance evaluation model and the structural size evaluation model, and output the body structure picture and its performance and size data.

Images