Three-dimensional vehicle pressure prediction method and system based on hor net

Abstract

This invention proposes a 3D vehicle pressure prediction method and system based on HorNet, belonging to the technical field of computer deep learning. First, a 3D vehicle CAD model dataset is acquired and divided into training and testing sets. Simulation calculations of vehicle pressure are performed on each 3D vehicle CAD model, and the simulation results are used as real data labels. The 3D vehicle CAD models in the dataset are stored as an octree data structure. The input data for training and testing in the octree data structure is determined. Then, a HorNet convolutional neural network model is constructed. The HorNet convolutional neural network model is trained using data from the training set in the input data, and the loss is compared with the real data labels to obtain a trained neural network model. This model is used for vehicle aerodynamic pressure prediction, testing the pressure state of different parts of the vehicle. It features fast calculation speed and high prediction accuracy.

Description

Technical Field

[0001] This invention relates to the technical field of computer deep learning, and more specifically, to a three-dimensional vehicle pressure prediction method and system based on HorNet. Background Technology

[0002] Aerodynamic layout design is a crucial aspect of automotive design. By reducing the stress experienced by a moving vehicle, its aerodynamic performance can be significantly improved, which is of great importance for energy conservation and emission reduction in gasoline vehicles and for increasing the driving range of electric vehicles.

[0003] However, the calculation of key aerodynamic parameters for automobiles still relies on traditional CFD (Computational Fluid Dynamics). As the number of meshes and degrees of freedom increase, the computation time grows exponentially, which cannot meet the requirements of rapid automotive design. In CFD simulations, partial differential equation solvers are typically used to simulate the physical properties of a given geometry. However, running physical simulations in the finite element method is a computationally expensive, memory-intensive, and time-consuming process. These drawbacks limit opportunities for design space exploration and prohibit interactive design. Fluid dynamics problems, due to their high dimensionality, strong nonlinearity, and large data volume, are precisely an important area where deep learning excels and can bring about paradigm innovation in research.

[0004] To alleviate the many limitations of traditional CFD, deep learning has recently been introduced into CFD, similar to many other engineering fields. Deep learning uses a data-driven approach to learn the similarity relationships between data, reducing errors caused by mesh partitioning and parameter settings during simulation calculations and improving the accuracy of CFD simulations. Convolutional Neural Networks (CNNs) are neural networks specifically designed to process mesh-like data. Using the mathematical operation of convolution, they are one of the most representative algorithms in deep learning. Due to their inherent excellent characteristics, CNNs are naturally suitable for many computer vision tasks. For example, a method for optimizing device shape using a convolutional neural network is disclosed in the prior art. This method includes discretizing the shape of the device into multiple points along the device's contour or into a surface mesh, and inputting the multiple points or surface mesh into a convolutional neural network to calculate the predicted value of at least one hydrodynamic parameter. However, the convolution operation in standard CNNs does not explicitly consider the spatial interactions, thus failing to significantly increase the model's expressive power.

[0005] Currently, extensive experimental applications of HorNet, a network based on recursive and gated convolutions, in ImageNet classification, COCO object detection, and ADE20K semantic segmentation demonstrate that, with similar overall architecture and training configurations, HorNet significantly outperforms Swin Transformers and ConvNeXt. HorNet also exhibits good scalability to more training data and larger model sizes. HorNet is efficient, scalable, and translation-invariant. It not only meets the requirements of dynamic weights and long-distance modeling but also achieves high-order spatial interactions. HorNet's effectiveness in common visual recognition tasks demonstrates higher accuracy and better performance compared to other CNN networks, significantly increasing the model's expressive power. Therefore, applying HorNet convolutional network technology to 3D vehicle pressure prediction to address the computationally intensive and slow computational problems inherent in traditional CFD simulations that use high-order differential equations or discretized algebraic equations for physical simulation is of great significance. Summary of the Invention

[0006] To address the issues of high computational load and slow speed in traditional CFD simulation for calculating key aerodynamic parameters of automobiles to predict vehicle pressure, this invention proposes a three-dimensional vehicle pressure prediction method and system based on HorNet. By using the HorNet convolutional neural network, the method predicts vehicle body pressure and tests the pressure state of different parts of the vehicle, achieving fast computation speed and high prediction accuracy.

[0007] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows:

[0008] A three-dimensional vehicle pressure prediction method based on HorNet, the method comprising the following steps:

[0009] S1. Obtain the 3D automotive CAD model dataset and divide it into a training set and a test set;

[0010] S2. Perform vehicle pressure simulation calculations for each 3D vehicle CAD model;

[0011] S3. Construct an octree to store the three-dimensional car CAD model in the car dataset as an octree data structure, determine the input data for training and testing in the octree data structure, and use the pressure field generated on the car surface obtained by simulation calculation in S2 as the label of the input data.

[0012] S4. Construct the HorNet convolutional neural network model;

[0013] S5. Input the data belonging to the training set into the HorNet convolutional neural network model to learn and predict the output pressure parameters. Compare the output pressure parameters with the label loss and continuously backpropagate gradients to update and train the HorNet convolutional neural network model until the HorNet convolutional neural network model is trained.

[0014] S6. Input the data belonging to the test set from the input data into the trained HorNet convolutional neural network model to predict the pressure on each surface of the 3D car.

[0015] This technical solution stores the centralized 3D automotive CAD model in an octree data structure, determines the input data for training and testing in the octree data structure, and then constructs a HorNet convolutional neural network model. The HorNet convolutional neural network model is trained using the data in the input data that belong to the training set. The trained neural network model can be used for automotive aerodynamic pressure prediction and testing the pressure state of different parts of the vehicle. It has fast calculation speed and high prediction accuracy.

[0016] Preferably, in step S1, the three-dimensional automotive CAD model dataset is represented as: D = {D1, D2, ..., D...} n}, using 80% of the 3D automotive CAD model data as the training set F = {F1, F2, ..., F} m}, 20% of the 3D automotive CAD model data is used as the test set T = {T1, T2, ..., T} h}

[0017] Preferably, in step S2, the vehicle pressure is simulated and calculated for each three-dimensional vehicle CAD model using XFlow. The process is as follows:

[0018] S21. Determine the length, width, and height of the simulation computation domain, calculate the blocking ratio, and confirm that the blocking ratio meets the requirements. The calculation formula is:

[0019]

[0020] Where ω represents the blockage ratio, and S represents the projected area of ​​the car model on the cross-section of the air duct. N This indicates the cross-sectional area of ​​the wind tunnel, specifically the nozzle area of ​​the wind tunnel test section.

[0021] S22. Select XFlow's solver as the 3D kernel to simulate the external flow field of a car under isothermal conditions. Set the boundary condition to velocity. Set three levels of velocity for each 3D car CAD model and perform numerical simulation calculations. Set the virtual wind tunnel floor to a stationary wall. The simulation time is 0.1s after the air flows through the entire virtual wind tunnel.

[0022] S23. The simulation results of each three-dimensional car model under three speed levels are denoted as Y1, Y2, and Y3, respectively. Y1 represents the pressure field under wind speed level 1, and Y2 and Y3 represent the pressure fields of the car under wind speed levels 2 and 3, respectively.

[0023] Preferably, in step S3, when constructing the octree, the number of levels of the octree is first determined, i.e., the maximum depth to be partitioned, and then the dataset D = {D1, D2, ..., D...} is calculated. n The 3D automotive CAD model is scaled and placed into an axis-aligned unit cube;

[0024] Using the unit cube as the root node, evenly divide the cube into 8 identical smaller cubes. Traverse all non-empty octets at the current depth, i.e., traverse nodes containing the shape of a 3D car CAD model. In the next layer, evenly divide the non-empty octets into eight identical octets again, while octets without the shape of a 3D car CAD model are no longer divided. Repeat this process until the specified number of layers is reached.

[0025] Preferably, the input data is the average normal vector of the shape of the octet of the three-dimensional automotive CAD model, where the deepest leaf node in the octree data structure is located.

[0026] Preferably, the HorNet convolutional neural network model constructed in step S4 includes sequentially connected HorBlock modules and multilayer feedforward neural networks (FFN). The HorBlock module contains recursively connected Layer Norn modules and gated convolutions (g). n Conv, the FFN module contains recursively connected Layer Norn modules and MLP units; the gated convolution g n Conv is used to implement long-range modeling and high-order spatial interactions. It is constructed from standard convolution, linear projection, and element-wise multiplication. High-order spatial interactions are divided into first-order spatial interactions and higher-order spatial interactions. First-order spatial interactions are as follows:

[0027] Let the input features of the input data of S3 be x∈R N×C Gated convolution g n The output of Conv is:

[0028]

[0029] p1=g(q0)⊙p0∈R N×C y = Φ out (p0)∈R N×2C

[0030] Where, Φ in , Φout All are linear projection operations to complete the information exchange at the channel dimension; g()b represents depth-wise convolution, and ⊙ represents the element-wise multiplication of vectors; p1=g(q0)⊙p0∈R N×C y = Φ out (p0)∈R N×2C express and its surrounding features First-order interactions;

[0031]

[0032] Where, δ i This represents a local window of the Depth-wise Conv, with center coordinate i, and w represents the weight of the Depth-wise Conv.

[0033] Higher-order spatial interactions are:

[0034] First, through Φ in We obtain a series of projective features p0 and {q0, q1, ..., q}. k}, where k = 0, 1, ..., n-1;

[0035]

[0036] Then perform gated convolution recursively:

[0037] p k =g k (q k )⊙h k (p k ) / α, k = 0, 1, ..., n-1

[0038] In this process, each recursive step is divided by α to stabilize the training, {g k} is a series of depth-wise convolution operations, denoted by {h k The number of channels of the matched feature is counted in each recursive process;

[0039]

[0040] Finally, output q from the last recursion. n Input to projection layer Φ out , to obtain g n The result of Conv is to realize the nth-order interaction between a certain feature and its surrounding features.

[0041] Here, the g-based recursive and gated convolutions in the HorNet convolutional network nConv enables convolutional neural networks to perform high-order spatial interactions, making them well-suited for handling high-dimensional problems in CAE physics simulations.

[0042] Preferably, in step S5, adjacent octets are selected at the same octree depth, the ReLU activation function is used, BatchNorm normalization is used to accelerate neural network training, max pooling is used as the pooling layer, and the convolution operation Φ is written in unfolded form. c :

[0043]

[0044] Among them, O ijk T(·) represents the octet adjacent to O, and T(·) represents the octet adjacent to O. ijk The relevant feature vector, T (n) (·) represents the nth channel of the eigenvector. These are the weights of the convolution operation; if O ijk It does not exist in the octree, so T(O) ijk The zero vector is set; in this way, the convolution operation is transformed into a matrix multiplication, which can be efficiently computed on the GPU.

[0045] Let the training set F = {F1, F2, ..., F...} m The average normal vector of the 3D car CAD model shape in the octet containing the deepest leaf node of the octree is used as the input data of the HorNet convolutional neural network model. The pressure field generated on the car surface obtained by S2 simulation is used as the labels Y1, Y2, and Y3 of the input data. Through convolution operation Φ c Pooling operations are used to output the pressure field predictions for the corresponding wind speed conditions using the HorNet convolutional neural network model. Where i takes values ​​of 1, 2, and 3; the pressure field prediction quantity The loss is compared with the corresponding labels Y1, Y2 and Y3. The parameters of the HorNet convolutional neural network model are updated by continuously backpropagating gradients until the HorNet convolutional neural network model is trained.

[0046] Preferably, the mean squared error (MSE) is used as the evaluation index for the regression model:

[0047]

[0048] Where m is the total number of predictions y 人The term "predicted value" refers to the total predicted value, while "y" refers to the total true label value of the pressure field generated on the car surface obtained from the simulation calculation in step S2. The mean square error (MSE) is the average of the sum of squares of the errors between corresponding points in the predicted data and the true label value data. The smaller the MSE, the closer it is to 0, the better the selection and fitting effect of the HorNet convolutional neural network model, the more accurate the data prediction, and the smoother the data obtained from the MSE is compared with the mean absolute error.

[0049] Preferably, in step S6, the input data belongs to the test set T = {T1, T2, ..., T...} h The data in the octree structure is the average normal vector of the shape of the 3D car CAD model in the octet containing the deepest leaf node in the octree structure.

[0050] A three-dimensional vehicle pressure prediction system based on HorNet, the system comprising:

[0051] The dataset acquisition unit is used to acquire a 3D automotive CAD model dataset and divide the 3D automotive CAD model dataset into a training set and a test set.

[0052] The simulation calculation unit performs simulation calculations of vehicle pressure for each three-dimensional automotive CAD model;

[0053] Octree data structure building unit is used to construct octrees, store the three-dimensional car CAD model in the car dataset as an octree data structure, determine the input data for training and testing in the octree data structure, and use the pressure field generated on the car surface obtained by S2 simulation calculation as the label of the input data.

[0054] Model design building blocks are used to build the HorNet convolutional neural network model.

[0055] The model training unit inputs data belonging to the training set into the HorNet convolutional neural network model to learn and predict output stress parameters. It compares the output stress parameters with the label loss and continuously backpropagates gradients to update and train the HorNet convolutional neural network model until the HorNet convolutional neural network model training is completed.

[0056] The prediction unit inputs the data belonging to the test set from the input data into the trained HorNet convolutional neural network model to predict the pressure on each surface of the 3D car.

[0057] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0058] This invention proposes a 3D vehicle pressure prediction method and system based on HorNet. First, a 3D vehicle CAD model dataset is acquired and divided into training and testing sets. Simulation calculations of vehicle pressure are performed on each 3D vehicle CAD model, and the simulation results are used as real data labels. The 3D vehicle CAD models in the dataset are stored as an octree data structure. The input data for training and testing in the octree data structure is determined. Then, a HorNet convolutional neural network model is constructed. The HorNet convolutional neural network model is trained using data from the training set in the input data, and the loss is compared with the real data labels to obtain a trained neural network model. This model is used for vehicle aerodynamic pressure prediction, testing the pressure state of different parts of the vehicle. It features fast calculation speed and high prediction accuracy. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the three-dimensional vehicle pressure prediction method based on HorNet proposed in Embodiment 1 of the present invention.

[0060] Figure 2 This diagram shows the overall structure of the HorNet convolutional neural network model proposed in Embodiment 2 of the present invention.

[0061] Figure 3 This refers to the gated convolution g in the HorNet convolutional neural network model proposed in Embodiment 2 of the present invention. n Structural diagram of Conv;

[0062] Figure 4 This represents the convergence curve of the loss function during the training and testing processes proposed in Embodiment 2 of the present invention;

[0063] Figure 5 This diagram illustrates the three-dimensional vehicle pressure prediction system based on HorNet proposed in Embodiment 3 of the present invention. Detailed Implementation

[0064] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0065] To better illustrate this embodiment, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions;

[0066] It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings.

[0067] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0068] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0069] Example 1

[0070] In this embodiment, a three-dimensional vehicle pressure prediction method based on HorNet is proposed. The flowchart of this method is shown below. Figure 1 As shown, see Figure 1 The method includes the following steps:

[0071] S1. Obtain the 3D automotive CAD model dataset and divide it into a training set and a test set;

[0072] S2. Perform vehicle pressure simulation calculations for each 3D vehicle CAD model;

[0073] S3. Construct an octree to store the three-dimensional car CAD model in the car dataset as an octree data structure, determine the input data for training and testing in the octree data structure, and use the pressure field generated on the car surface obtained by simulation calculation in S2 as the label of the input data.

[0074] S4. Construct the HorNet convolutional neural network model;

[0075] S5. Input the data belonging to the training set into the HorNet convolutional neural network model to learn and predict the output pressure parameters. Compare the output pressure parameters with the label loss and continuously backpropagate gradients to update and train the HorNet convolutional neural network model until the HorNet convolutional neural network model is trained.

[0076] S6. Input the data belonging to the test set from the input data into the trained HorNet convolutional neural network model to predict the pressure on each surface of the 3D car.

[0077] Example 2

[0078] In this embodiment, 400 3D model files of the automobile subclass under the large-scale 3D model dataset ShapeNet with rich annotations were selected as the obtained 3D automobile CAD model dataset, and these 400 3D automobile CAD model datasets are represented as: D = {D1, D2, ..., D...} n}, using 80% of the 3D automotive CAD model data as the training set F = {F1, F2, ..., F} m}, 20% of the 3D automotive CAD model data is used as the test set T = {T1, T2, ..., T} h}

[0079] Because the blocking ratio has a non-negligible impact on the experimental results, the smaller the blocking ratio, the more accurate the experimental results. When the blocking ratio is below 2%, the interference caused by the blocking ratio on the experimental results can be disregarded. In step S2, the vehicle pressure is simulated and calculated for each 3D vehicle CAD model using XFlow. The process is as follows:

[0080] S21. To accommodate vehicles of varying sizes in the 3D automotive CAD model dataset, the length, width, and height of the simulation computation domain are determined, and the blocking ratio is calculated. In this embodiment, the simulation computation domain is determined to be 55 meters long, 20 meters wide, and 9 meters high. Considering this numerical simulation experiment, the blocking ratio calculation formula is:

[0081]

[0082] Where ω represents the blockage ratio, and S represents the projected area of ​​the car model on the cross-section of the air duct. N This represents the cross-sectional area of ​​the wind tunnel, i.e., the nozzle area of ​​the wind tunnel test section; at this point, the blockage ratio is less than 1.5%, which meets the requirements.

[0083] S22. Select XFlow's solver as the 3D kernel to simulate the external flow field of a car under isothermal conditions. Set the boundary condition as velocity. To expand the data volume of the dataset, set three levels of velocity for each 3D car CAD model for numerical simulation calculation. In this embodiment, the velocities in the x-direction are 20m / s, 30m / s, and 40m / s, respectively. The velocities basically cover the approximate speed of the car during actual driving. The virtual wind tunnel ground is set as a stationary wall. The simulation time is 0.1s after the air flows through the entire virtual wind tunnel to ensure that the flow is fully developed.

[0084] S23. The simulation results of each three-dimensional car model under three speed levels are denoted as Y1, Y2, and Y3, respectively. Y1 represents the pressure field under wind speed level 1 (20 m / s), and Y2 and Y3 represent the pressure fields of the car under wind speed levels 2 (30 m / s) and 3 (40 m / s), respectively.

[0085] In this embodiment, when constructing the octree, the first step is to determine the number of levels, i.e., the maximum depth required for partitioning, which determines the resolution of the subsequent model. Generally, the more levels, the more accurate the model, but the computational cost also increases exponentially. Here, a balance is struck between accuracy and computational cost, and a 6-level octree is chosen for construction. Then, the dataset D = {D1, D2, ..., D...} is... n The 3D automotive CAD model is scaled and placed into an axis-aligned unit cube;

[0086] Using the unit cube as the root node, evenly divide the cube into 8 identical smaller cubes. Traverse all non-empty octets at the current depth, i.e., traverse nodes containing the shape of a 3D car CAD model. In the next layer, evenly divide the non-empty octets into eight identical octets again, while octets without the shape of a 3D car CAD model are no longer divided. Repeat this process until the specified number of layers is reached.

[0087] In this embodiment, the input data is the average normal vector of the shape of the octet of the three-dimensional car CAD model, where the deepest leaf node in the octree data structure is located.

[0088] HorNet convolutional network based on recursive and gated convolutions g n Conv enables convolutional neural networks to perform high-order spatial interactions, making them well-suited for handling high-dimensional problems in CAE physics simulations. For example... Figure 2 As shown, the HorNet convolutional neural network model constructed in this embodiment includes a HorBlock module connected in sequence and a multi-layer feedforward neural network FFN. The HorBlock module contains a recursively connected Layer Nom module and a gated convolution g. n Conv, the FFN module contains a recursively connected Layer Norn module and an MLP unit; Figure 3 For gated convolution g n The specific structure diagram of Conv, gated convolution g n Conv is used to implement long-range modeling and high-order spatial interactions. It is constructed from standard convolution, linear projection, and element-wise multiplication. High-order spatial interactions are divided into first-order spatial interactions and higher-order spatial interactions. First-order spatial interactions are as follows:

[0089] Let the input features of the input data of S3 be x∈R N×C Gated convolution g n The output of Conv is:

[0090]

[0091] p1=g(q0)⊙p0∈R N×C y = Φ out (p0)∈R N×2C

[0092] Where, Φ in , Φ out All are linear projection operations to complete the information exchange at the channel dimension; g()b represents depth-wise convolution, and ⊙ represents the element-wise multiplication of vectors; p1=g(q0)⊙p0∈R N×C y = Φ out(p0)∈R N×2C express and its surrounding features First-order interactions;

[0093]

[0094] Where, δ i This represents a local window of the Depth-wise Conv, with center coordinate i, and w represents the weight of the Depth-wise Conv.

[0095] Higher-order spatial interactions are:

[0096] First, through Φ in We obtain a series of projective features p0 and {q0, q1, ..., q}. k}, where k = 0, 1, ..., n-1;

[0097]

[0098] Then perform gated convolution recursively:

[0099] p k =g k (q k )⊙h k (p k ) / α, k = 0, 1, ..., n-1

[0100] In this process, each recursive step is divided by α to stabilize the training, {g k} is a series of depth-wise convolution operations, denoted by {h k The number of channels of the matched feature is counted in each recursive process;

[0101]

[0102] Finally, output q from the last recursion. n Input to projection layer Φ out , to obtain g n The result of Conv is to realize the nth-order interaction between a certain feature and its surrounding features.

[0103] Adjacent octets at the same octree depth are selected. The ReLU activation function is used, and BatchNorm normalization is used to accelerate neural network training. Max pooling is used as the pooling layer. The convolution operation Φ is written in unfolded form. c :

[0104]

[0105] Among them, O ijkT(·) represents the octet adjacent to O, and T(·) represents the octet adjacent to O. ijk The relevant feature vector, T (n) (·) represents the nth channel of the eigenvector. These are the weights of the convolution operation; if O ijk It does not exist in the octree, so T(O) ijk The zero vector is set; in this way, the convolution operation is transformed into a matrix multiplication, which can be efficiently computed on the GPU.

[0106] Let the training set F = {F1, F2, ..., F...} m The average normal vector of the 3D car CAD model shape in the octet containing the deepest leaf node of the octree is used as the input data of the HorNet convolutional neural network model. The pressure field generated on the car surface obtained by S2 simulation is used as the labels Y1, Y2, and Y3 of the input data. Through convolution operation Φ c Pooling operations are used to output the pressure field predictions for the corresponding wind speed conditions using the HorNet convolutional neural network model. Where i takes values ​​of 1, 2, 3; the pressure field prediction quantity The loss is compared with the corresponding labels Y1, Y2 and Y3. The parameters of the HorNet convolutional neural network model are updated by continuously backpropagating gradients until the HorNet convolutional neural network model is trained.

[0107] Finally, the mean squared error (MSE) is used as an evaluation metric for the regression model:

[0108]

[0109] Where m is the total number of predictions y 人 The term "predicted value" refers to the total predicted value, while "y" refers to the total true label value of the pressure field generated on the car surface obtained from the simulation calculation in step S2. The mean square error (MSE) is the average of the sum of squares of the errors between corresponding points in the predicted data and the true label value data. The smaller the MSE, the closer it is to 0, the better the selection and fitting effect of the HorNet convolutional neural network model, the more accurate the data prediction, and the smoother the data obtained from the MSE is compared with the mean absolute error.

[0110] In this embodiment, the input data belongs to the test set T = {T1, T2, ..., T...} h The data in the octree structure is the average normal vector of the shape of the 3D car CAD model in the eighth of the octree structure, where the deepest leaf node is located. The model performs a series of operations and finally outputs the predicted pressure on each face of the 3D car. Figure 4 The convergence curves of the loss function are shown during the training and testing processes. Figure 4 The dashed line represents the training loss, and the dotted line represents the test loss. The experimental results show that the HorNet convolutional neural network model has good performance in this experiment, can quickly achieve convergence of the loss function, has good generalization ability, and shows high accuracy on the test set.

[0111] Example 3

[0112] like Figure 5 As shown, this embodiment proposes a three-dimensional vehicle pressure prediction system based on HorNet. See [link to relevant documentation]. Figure 5 The system includes:

[0113] The dataset acquisition unit is used to acquire a 3D automotive CAD model dataset and divide the 3D automotive CAD model dataset into a training set and a test set.

[0114] The simulation calculation unit performs simulation calculations of vehicle pressure for each three-dimensional automotive CAD model;

[0115] Octree data structure building unit is used to construct octrees, store the three-dimensional car CAD model in the car dataset as an octree data structure, determine the input data for training and testing in the octree data structure, and use the pressure field generated on the car surface obtained by S2 simulation calculation as the label of the input data.

[0116] Model design building blocks are used to build the HorNet convolutional neural network model.

[0117] The model training unit inputs data belonging to the training set into the HorNet convolutional neural network model to learn and predict output stress parameters. It compares the output stress parameters with the label loss and continuously backpropagates gradients to update and train the HorNet convolutional neural network model until the HorNet convolutional neural network model training is completed.

[0118] The prediction unit inputs the data belonging to the test set from the input data into the trained HorNet convolutional neural network model to predict the pressure on each surface of the 3D car.

[0119] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A three-dimensional vehicle pressure prediction method based on HorNet, characterized in that, The method includes the following steps: S1. Obtain the 3D automotive CAD model dataset and divide it into a training set and a test set; S2. Perform vehicle pressure simulation calculations for each 3D vehicle CAD model; S3. Construct an octree to store the three-dimensional car CAD model in the car dataset as an octree data structure, determine the input data for training and testing in the octree data structure, and use the pressure field generated on the car surface obtained by simulation calculation in S2 as the label of the input data. S4. Construct the HorNet convolutional neural network model; Step S4 describes the construction of the HorNet convolutional neural network model, which includes sequentially connected HorBlock modules and multi-layer feedforward neural networks (FFN). The HorBlock module contains recursively connected Layer Norn modules and gated convolutions (g). n Conv, the FFN module contains recursively connected Layer Norn modules and MLP units; the gated convolution g n Conv is used to implement long-range modeling and high-order spatial interactions. It is constructed from standard convolution, linear projection, and element-wise multiplication. High-order spatial interactions are divided into first-order spatial interactions and higher-order spatial interactions. First-order spatial interactions are as follows: Let the input features of the input data of S3 be x∈R N×C Gated convolution g n Conv's output is: ， , in, , All are linear projection operations to complete the information exchange at the channel dimension; g() represents depth-wise convolution, and ⊙ represents the element-wise multiplication of vectors. =g( ) ∈R N×C ,y= ( )∈R N×2C express and its surrounding features First-order interactions; = in, This represents a local window of the Depth-wise Conv, with its center coordinates being... i , w Indicates the weights of the Depth-wise Conv; Higher-order spatial interactions are: Firstly, through A series of projection features were obtained and{ ， ，...， }, where k = 0, 1, ..., n-1; Then perform gated convolution recursively: = ( ) ( ) a, k=0, 1,..., n-1 In this process, each recursive step is divided by α to stabilize the training. } is a series of depth-wise convolution operations, denoted by { The number of channels of the matched feature is counted in each recursive process; Finally, output the last recursive call. Input to projection layer ,get g n The result of Conv is to realize the nth-order interaction between a certain feature and its surrounding features; S5. Input the data belonging to the training set into the HorNet convolutional neural network model to learn and predict the output pressure parameters. Compare the output pressure parameters with the label loss and continuously backpropagate gradients to update and train the HorNet convolutional neural network model until the HorNet convolutional neural network model is trained. S6. Input the data belonging to the test set from the input data into the trained HorNet convolutional neural network model to predict the pressure on each surface of the 3D car.

2. The HorNet-based three-dimensional vehicle pressure prediction method according to claim 1, characterized in that, In step S1, let the 3D automotive CAD model dataset be represented as: D = { , ,......, },Will The 3D automotive CAD model data is used as the training set F={ , ,......, }, The 3D automotive CAD model data is used as the test set T={ , ,......, } 3. The three-dimensional vehicle pressure prediction method based on HorNet according to claim 1, characterized in that, In step S2, XFlow is used to simulate and calculate the vehicle pressure for each 3D vehicle CAD model. The process is as follows: S21. Determine the length, width, and height of the simulation computation domain, calculate the blocking ratio, and confirm that the blocking ratio meets the requirements. The calculation formula is: in, This represents the blockage ratio, and S represents the projected area of ​​the car model on the cross-section of the air duct. This indicates the cross-sectional area of ​​the wind tunnel, specifically the nozzle area of ​​the wind tunnel test section. S22. Select XFlow's solver as the 3D kernel to simulate the external flow field of a car under isothermal conditions. Set the boundary condition to velocity. Set three levels of velocity for each 3D car CAD model and perform numerical simulation calculations. Set the virtual wind tunnel floor to a stationary wall. The simulation time is 0.1s after the air flows through the entire virtual wind tunnel. S23. Record the simulation results of each 3D car model at three speed levels as follows: , , ,in, This represents the pressure field under wind speed level 1. , These represent the pressure fields of the car under wind speeds of level 2 and level 3, respectively.

4. The HorNet-based three-dimensional vehicle pressure prediction method according to claim 3, characterized in that, In step S3, when constructing the octree, first determine the number of levels in the octree, i.e., the maximum depth to be partitioned, and then calculate the dataset D={ , ,......, The 3D automotive CAD model is scaled and placed into an axis-aligned unit cube; Using the unit cube as the root node, evenly divide the cube into 8 identical smaller cubes. Traverse all non-empty octets at the current depth, i.e., traverse nodes containing the shape of a 3D car CAD model. In the next layer, evenly divide the non-empty octets into eight identical octets again, while octets without the shape of a 3D car CAD model are no longer divided. Repeat this process until the specified number of layers is reached.

5. The three-dimensional vehicle pressure prediction method based on HorNet according to claim 4, characterized in that, The input data is the average normal vector of the shape of the octet of the 3D automotive CAD model, which contains the deepest leaf node in the octet data structure.

6. The three-dimensional vehicle pressure prediction method based on HorNet according to claim 1, characterized in that, In step S5, adjacent octets are selected at the same octree depth. The ReLU activation function is used, and BatchNorm normalization is used to accelerate neural network training. Max pooling is used as the pooling layer, and the convolution operation is written in unfolded form. : in, express The adjacent octets, Indicates and Related feature vectors This represents the nth channel of the eigenvector. These are the weights of the convolution operation; if It does not exist in an octree. Set it to a zero vector; in this way, the convolution operation is transformed into a matrix multiplication, which can be efficiently computed on the GPU; Let the training set F={ , ,......, The average normal vector of the 3D car CAD model shape in the octet containing the deepest leaf node of the octree is used as the input data for the HorNet convolutional neural network model, and the pressure field generated on the car surface calculated by S2 simulation is used as the label of the input data. , and Through convolution operations Pooling operations are used to output the pressure field predictions for the corresponding wind speed conditions using the HorNet convolutional neural network model. ,in, i The values ​​are 1, 2, and 3; the pressure field prediction quantity is... and corresponding tags , and By comparing the loss, the parameters of the HorNet convolutional neural network model are updated through continuous backpropagation of gradients until the HorNet convolutional neural network model is trained.

7. The HorNet-based three-dimensional vehicle pressure prediction method according to claim 6, characterized in that, Using the mean squared error (MSE) as an evaluation metric for the regression model: Where m is the total number of predictions performed. The term "predicted value" refers to the total predicted value, while "y" refers to the total true label value of the pressure field generated on the car surface obtained from the simulation calculation in step S2. The mean square error (MSE) is the average of the sum of squares of the errors between corresponding points in the predicted data and the true label value data. The smaller the MSE, the closer it is to 0, the better the selection and fitting effect of the HorNet convolutional neural network model, the more accurate the data prediction, and the smoother the data obtained from the MSE is compared with the mean absolute error.

8. The three-dimensional vehicle pressure prediction method based on HorNet according to claim 1, characterized in that, In step S6, the input data belonging to the test set T={ , ,......, The data in the octree structure is the average normal vector of the shape of the 3D car CAD model in the octet containing the deepest leaf node in the octree structure.

9. A three-dimensional vehicle pressure prediction system based on HorNet, characterized in that, The system is used to implement the HorNet-based three-dimensional vehicle pressure prediction method as described in claim 1, including: The dataset acquisition unit is used to acquire a 3D automotive CAD model dataset and divide the 3D automotive CAD model dataset into a training set and a test set. The simulation calculation unit performs simulation calculations of vehicle pressure for each three-dimensional automotive CAD model; Octree data structure building unit is used to construct octrees, store the three-dimensional car CAD model in the car dataset as an octree data structure, determine the input data for training and testing in the octree data structure, and use the pressure field generated on the car surface obtained by S2 simulation calculation as the label of the input data. Model design building blocks are used to build the HorNet convolutional neural network model. The model training unit inputs data belonging to the training set into the HorNet convolutional neural network model to learn and predict output stress parameters. It compares the output stress parameters with the label loss and continuously backpropagates gradients to update and train the HorNet convolutional neural network model until the HorNet convolutional neural network model training is completed. The prediction unit inputs the data belonging to the test set from the input data into the trained HorNet convolutional neural network model to predict the pressure on each surface of the 3D car.

Images