Combined prediction method and system for automobile wind resistance coefficient and surface pressure field

Abstract

The application provides a combined prediction method and system for a wind resistance coefficient and a surface pressure field, and relates to the technical field of automobile aerodynamics. The combined prediction method predicts the pressure field and the wind resistance coefficient in parallel in the same network, and calibrates the wind resistance coefficient derived from the pressure field obtained by pressure field integration and the wind resistance coefficient directly regressed by the network with physical consistency constraints, so that the model has physical credibility and interpretability while ensuring numerical accuracy, thereby being capable of locating a key stress area and serving an automobile modeling iteration process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automotive aerodynamics, in particular to a combined prediction method and system for the drag coefficient and surface pressure field of a vehicle. BACKGROUND

[0002] The aerodynamic performance of the overall vehicle shape directly affects the energy consumption and stability of the vehicle. In engineering, both the drag coefficient of the vehicle (Cd) and the surface pressure field are needed to evaluate the pros and cons of the scheme, and the surface pressure field is needed to locate the contribution of the "windward area" to the drag for styling optimization and local modification.

[0003] The existing prediction methods for the drag coefficient and surface pressure field of a vehicle mainly include two categories. The first category is the traditional drag and pressure field prediction method based on CFD / RANS. This method usually takes CAD and STL geometry as input, performs meshing, sets up turbulence model, configures boundary conditions and numerical parameters, solves RANS using finite volume method, obtains surface pressure and shear stress field, and then integrates to obtain the drag coefficient. The interrelation of the process chain is: geometry → mesh → control equation and turbulence closure → numerical solution → field variable → aerodynamic force post-processing. The advantages of this route are strong physical interpretation and complete pressure field output; the disadvantages are long calculation period, high cost, sensitivity to mesh and model selection, and difficulty in supporting high-frequency iteration and screening in the early stage of vehicle styling in complex shape and multiple scheme scenarios. The second category is the existing common drag and pressure field prediction method based on deep learning proxy model. This method takes the deep learning model to directly predict the vehicle drag coefficient and pressure field from the vehicle point cloud, such as point cloud direct regression, representing the shape with 3D point cloud, using PointNet / PointNet++ network for feature coding, and regressing with multi-layer perception and other aerodynamic indicators. Although this method can support high-frequency iteration and screening in the early stage of vehicle styling.

[0004] However, the existing prediction methods based on proxy models focus on single indicator or single field prediction, and there is a lack of collaborative modeling of targets in the same framework, making it difficult to provide both the overall vehicle drag coefficient and the surface pressure field prediction while maintaining their physical consistency. Even if multiple targets are predicted simultaneously, it is difficult to ensure the mechanism of physical consistency between them, and it is easy to appear numerically accurate but inconsistent with each other, which weakens the usability of engineering. SUMMARY

[0005] (I) Technical problems solved

[0006] ​​In view of the deficiencies of the prior art, the present application provides a combined prediction method and system for the wind resistance coefficient and surface pressure field of an automobile, which solves the technical problem of the mechanism that the existing prediction method through a proxy model is difficult to guarantee the physical consistency of the wind resistance coefficient and surface pressure field of the whole vehicle.

[0007] (II) Technical solutions

[0008] To achieve the above object, the present application is implemented by the following technical solutions:

[0009] In a first aspect, the present application provides a combined prediction method for the wind resistance coefficient and surface pressure field of an automobile, comprising:

[0010] Obtaining a training data set comprising automobile point cloud data and label data, and preprocessing the automobile point cloud data in the training data set to obtain a sampling point matrix and an appearance parameter vector; and uniformly sampling the sampling point matrix to obtain a uniform point matrix;

[0011] Training an initial combined prediction model through the sampling point matrix, the appearance parameter vector and the uniform point matrix, optimizing the loss value between the output of the combined prediction model and the label data, and obtaining the combined prediction model;

[0012] The initial combined prediction model comprises a whole vehicle wind resistance prediction sub-network, an automobile surface pressure field prediction sub-network, a multi-layer perception machine, an image segmentation network and a physical consistency constraint calculation module.

[0013] The physical consistency constraint calculation module is used to integrate the predicted surface pressure field to obtain the wind resistance coefficient derived from the pressure field.

[0014] The whole vehicle wind resistance prediction sub-network is used to process the uniform point matrix and the appearance parameter vector to obtain wind resistance features; the automobile surface pressure field prediction sub-network is used to process the sampling point matrix to obtain pressure field features; the multi-layer perception machine is used to process the wind resistance features and the wind resistance coefficient derived from the pressure field to obtain the wind resistance coefficient of the automobile; and the image segmentation network is used to process the pressure field features and the wind resistance coefficient derived from the pressure field to obtain the surface pressure field of the automobile.

[0015] Preferably, the whole vehicle wind resistance prediction sub-network comprises a PointNet++ module, an MLP module and a feature fusion module, wherein the PointNet++ module is used to process the uniform point matrix to obtain point cloud structure features, the MLP module is used to encode the appearance parameters to obtain appearance features, and the feature fusion module is used to align and fuse the point cloud structure features and the appearance features.

[0016] Preferably, the process of aligning and fusing the point cloud structural features and shape features by the feature fusion module includes:

[0017] The linear mapping of the feature fusion module is Then cross attention is The formula for the cross-attention feature fusion module is:

[0018]

[0019] in, p =0.1 indicates the dropout rate of Dropout; , representing the structural features of the point cloud; , represents the shape feature; C represents the number of feature channels, B Indicates the batch size.

[0020] Preferably, the vehicle surface pressure field prediction subnetwork includes a first normalization layer, a physical attention layer, a second normalization layer, and a feedforward layer. After the first normalization layer processes the sampling point matrix, the output data passes through the physical attention layer. The output of the physical attention layer and the sampling point matrix are concatenated with residuals to obtain residual features. The residual features are processed by the second normalization layer and then input to the feedforward layer. The output of the feedforward layer and the residual features are concatenated with residuals to obtain pressure field features.

[0021] Preferably, the physical attention layer includes a first linear layer, a multi-head attention module, and a second linear layer;

[0022] The multi-head attention module includes a slice layer, a label-specific attention layer, and a reverse slice layer.

[0023] Preferably, the step of integrating the predicted surface pressure field to obtain the drag coefficient derived from the pressure field includes:

[0024] Given the unit vector of the incoming flow direction air density ,speed Reference windward area A Dynamic pressure ; Pressure subnet outputs surface pressure field Then, discrete integration yields the drag coefficient derived from the pressure field. :

[0025]

[0026] in, Indicates the correction factor; This represents the total number of points on the surface of the car's point cloud; Indicates the first i The external normal vector of each point; represents the area of the i th point.

[0027] Preferably, the loss value between the output of the joint prediction model and the label data is calculated according to the following formula:

[0028]

[0029] wherein the first term is the wind resistance supervision, the second term is the pressure supervision, and the third term is the physical consistency constraint; 、 、 is a weight coefficient; is the wind resistance coefficient derived from the pressure field obtained by pressure field integration; is the wind resistance coefficient predicted by the joint prediction model; represents the true value of the wind resistance coefficient; represents the pressure field on the surface of the i th point predicted by the joint prediction model; represents the true value of the pressure field on the surface of the i th point.

[0030] In a second aspect, the present application provides a joint prediction system for vehicle wind resistance coefficient and surface pressure field, comprising:

[0031] a data acquisition and preprocessing module for acquiring a training data set comprising vehicle point cloud data and label data, and preprocessing the vehicle point cloud data in the training data set to obtain a sampling point matrix and an appearance parameter vector; and uniformly sampling the sampling point matrix to obtain a uniform point matrix;

[0032] a model training module for training an initial joint prediction model through the sampling point matrix, the appearance parameter vector and the uniform point matrix, optimizing the loss value between the model output and the label data, and obtaining a joint prediction model; the joint prediction model is used for processing the sampling point matrix, the appearance parameter vector and the uniform point matrix of a vehicle to be evaluated to obtain the vehicle wind resistance coefficient and the vehicle surface pressure field;

[0033] wherein the initial joint prediction model comprises a whole vehicle wind resistance prediction sub-network, a vehicle surface pressure field prediction sub-network, a multi-layer perception machine, an image segmentation network and a physical consistency constraint calculation module;

[0034] wherein the physical consistency constraint calculation module is used for integrating the predicted surface pressure field to obtain the wind resistance coefficient derived from the pressure field;

[0035] The whole vehicle drag prediction subnetwork is used to process the uniform point matrix and the shape parameter vector to obtain the drag characteristics; the vehicle surface pressure field prediction subnetwork is used to process the sampling point matrix to obtain the pressure field characteristics; the multilayer perceptron is used to process the drag characteristics and the drag coefficient derived from the pressure field to obtain the vehicle drag coefficient; the image segmentation network is used to process the pressure field characteristics and the drag coefficient derived from the pressure field to obtain the vehicle surface pressure field.

[0036] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for the joint prediction of vehicle drag coefficient and surface pressure field, wherein the computer program causes a computer to execute the joint prediction method of vehicle drag coefficient and surface pressure field as described above.

[0037] Fourthly, the present invention provides an electronic device, comprising:

[0038] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing a joint prediction method for vehicle drag coefficient and surface pressure field as described above.

[0039] (III) Beneficial Effects

[0040] This invention provides a method and system for jointly predicting the drag coefficient and surface pressure field of an automobile. Compared with the prior art, it has the following advantages:

[0041] The joint prediction method of this invention predicts pressure fields in parallel within the same network. With drag coefficient The drag coefficient derived from the pressure field and obtained by integrating the pressure field is constrained by physical consistency. drag coefficient directly regressed from the network Calibration is performed to ensure that the model has physical credibility and interpretability while maintaining numerical accuracy, thereby enabling the location of key stress areas and serving the automotive styling iteration process. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A block diagram of a method for jointly predicting the drag coefficient and surface pressure field of a vehicle;

[0044] Figure 2 A specific block diagram of the automobile surface pressure field prediction sub-network. DETAILED DESCRIPTION

[0045] For the purpose, technical solutions and advantages of the embodiments of the present application, the technical solutions in the embodiments of the present application are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] The embodiments of the present application provide a method and system for jointly predicting automobile drag coefficient and surface pressure field, solve the technical problem that the existing prediction method through the proxy model is difficult to guarantee the physical consistency mechanism of the whole vehicle drag coefficient and surface pressure field, realize parallel prediction of pressure field and drag coefficient in the same network, and calibrate the drag coefficient derived from the pressure field by pressure field integration and the drag coefficient directly regressed by the network with the physical consistency constraint, so that the model has physical credibility and interpretability while ensuring numerical accuracy, thereby being able to locate the key stress area and serve the automobile modeling iteration process.

[0047] The technical solutions in the embodiments of the present application are as follows to solve the above technical problems:

[0048] The existing prediction method through the proxy model of drag and pressure field based on deep learning focuses on single index or single field prediction, lacks collaborative modeling among targets in the same framework, and is difficult to simultaneously provide whole vehicle drag coefficient and surface pressure field prediction and maintain the physical consistency of the two. When predicting simultaneously, the drag coefficient and the pressure field are simultaneously regressed in a parallel minimum error manner, although two types of outputs can be given numerically, but the closed physical relationship between pressure field integration and drag coefficient is not explicitly modeled. That is, even if the two parameters are predicted simultaneously, the physical relationship that "the drag coefficient derived from the pressure field by pressure field integration should be consistent with the predicted value of the drag coefficient" is not explicitly included in the constraint, and the situation that each is accurate numerically but inconsistent with each other occurs, which weakens the usability of engineering. Most of the existing methods adopt a pure data-driven strategy, lack of "physical features" and "physical losses" consistent with external flow mechanics. At the same time, point cloud and grid representation are sensitive to sampling density and distribution, lack adaptive sampling strategies for key areas such as high curvature, resulting in the loss of some details; the evaluation system focuses on error such as MSE / MAE, lacks indicators and benchmarks to measure the consistency of "predicted drag coefficient and drag coefficient obtained by pressure field integration", and is difficult to integrate with the existing automobile modeling iteration process and guide local optimization.

[0049] To overcome the defects of the prior art, the embodiment of the present application proposes a combined prediction method for the automobile wind resistance coefficient and the surface pressure field, which predicts the pressure field and the wind resistance coefficient in the same network in parallel , and calibrates the wind resistance coefficient derived from the pressure field by integrating the pressure field with the physically consistent constraint and the wind resistance coefficient directly regressed by the network to make the model have physical credibility and explainability while ensuring numerical accuracy, so as to locate the key stress area and serve the automobile modeling iteration process.

[0050] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings in the specification and specific embodiments.

[0051] The embodiment of the present application provides a combined prediction method for the automobile wind resistance coefficient and the surface pressure field, as shown in Figure 1 , which comprises:

[0052] S1, obtaining a training data set comprising automobile point cloud data and label data, and preprocessing the automobile point cloud data in the training data set to obtain a sampling point matrix and an appearance parameter vector; and uniformly sampling the sampling point matrix to obtain a uniform point matrix;

[0053] S2, training an initial combined prediction model by using the sampling point matrix, the appearance parameter vector and the uniform point matrix, optimizing the loss value between the model output and the label data, and obtaining a combined prediction model; the combined prediction model is used for processing the sampling point matrix, the appearance parameter vector and the uniform point matrix of an automobile to be evaluated to obtain the automobile wind resistance coefficient and the automobile surface pressure field;

[0054] The initial combined prediction model comprises a whole vehicle wind resistance prediction sub-network, an automobile surface pressure field prediction sub-network, a multi-layer perception machine, an image segmentation network and a physical consistency constraint calculation module.

[0055] The physical consistency constraint calculation module is used for integrating the predicted surface pressure field to obtain the wind resistance coefficient derived from the pressure field.

[0056] The whole vehicle wind resistance prediction sub-network is used for processing the uniform point matrix and the appearance parameter vector to obtain a wind resistance feature; the automobile surface pressure field prediction sub-network is used for processing the sampling point matrix to obtain a pressure field feature; the multi-layer perception machine is used for processing the wind resistance feature and the wind resistance coefficient derived from the pressure field to obtain the automobile wind resistance coefficient; and the image segmentation network is used for processing the pressure field feature and the wind resistance coefficient derived from the pressure field to obtain the automobile surface pressure field.

[0057] The joint prediction method of this invention predicts pressure fields in parallel within the same network. With drag coefficient The drag coefficient derived from the pressure field and obtained by integrating the pressure field is constrained by physical consistency. drag coefficient directly regressed from the network Calibration is performed to ensure that the model has physical credibility and interpretability while maintaining numerical accuracy, thereby enabling the location of key stress areas and serving the automotive styling iteration process.

[0058] It should be noted that after the joint prediction model is trained, the physical consistency constraint calculation module no longer functions during actual use.

[0059] The structure and training process of the joint prediction model are described in detail below:

[0060] In this embodiment of the invention, in order to make the training of the joint prediction model faster and more stable, the vehicle drag prediction sub-network is initialized in advance. The initialization process of the vehicle drag prediction sub-network is as follows:

[0061] The system uses a 3D point cloud (STL) of the vehicle as input and employs curvature-based adaptive sampling. Given a fixed number of sampling points, it retains aerodynamically sensitive areas with high curvature based on curvature indicators, reduces redundant points in flat areas, and ensures high-fidelity representation of key surface details. Simultaneously, it automatically extracts global shape parameters (vehicle length L, vehicle width W, and vehicle height H) from the point cloud. The sampled point cloud is used to extract local-point cloud structure features using PointNet++ in the vehicle drag prediction subnetwork. The shape parameters are encoded using MLP in the same subnetwork. These two feature sets are then fused by the feature fusion module in the vehicle drag prediction subnetwork before being output as the drag coefficient via a regression head. The training phase optimizes using MSE. With truth value ,get Rapid prediction results provide a stable initialization for subsequent joint training.

[0062] The curvature sampling process is as follows:

[0063] Given a car point cloud, , The goal is to prioritize high curvature regions when extracting M samples, for each point... Constructing the covariance matrix ,Pick eigenvalues ,by As a point The curvature index, and the covariance matrix are:

[0064]

[0065] wherein, is a point of the neighborhood set, is a neighborhood mean. represents the total number of points of the point cloud surface of the car; is a point in the neighborhood set .

[0066] The fusion process of the feature fusion module can be described as:

[0067] For a batch of samples, the point cloud structure feature is obtained by the PointNet+ module The global dimensions (L, W, H) are upgraded by the MLP and copied to the shape feature of each point , assuming that the linear mapping of the feature fusion module is , then the cross attention is The formula of the cross attention feature fusion module is:

[0068]

[0069] wherein, p =0.1 represents the dropout rate of Dropout; C represents the number of feature channels, B represents the batch size.

[0070] It should be noted that the whole vehicle wind resistance prediction subnetwork includes PointNet++, MLP, feature fusion module, and the initialization process of the network, plus the regression head used in the initialization process, can be used as a wind resistance prediction model alone to quickly predict the wind resistance coefficient.

[0071] After initializing the whole vehicle wind resistance prediction subnetwork, the VTK file containing the three-dimensional point cloud data of the car can be used to train the joint prediction model, as follows:

[0072] The three-dimensional point cloud data of the car in the VTK file is preprocessed, and the preprocessing process is as follows:

[0073] The point set is obtained by simplifying and sampling the three-dimensional point cloud data of the car based on curvature, and the point set is uniformly sampled, and the global shape information is automatically extracted from the point cloud. The specific process is consistent with the sampling and shape parameter information extraction in the whole vehicle wind resistance prediction subnetwork process, and will not be repeated here.

[0074] The whole vehicle wind resistance prediction sub-network uses the fusion regression of "uniform sampling point cloud + shape parameters", and the preliminary prediction of the car surface pressure field prediction sub-network is obtained by using the Transolver point-level regression structure. To ensure the physical consistency of the outputs of the two, the physical relationship of "pressure field integration-wind resistance coefficient" is introduced as a constraint in the joint training.

[0075] The wind resistance coefficient derived from the pressure field by pressure field integration is:

[0076] Given the incoming flow direction unit vector , air density , velocity , reference windward area A , dynamic pressure . The point-level pressure field output by the pressure sub-network is After discrete integration, we get:

[0077]

[0078] wherein represents the correction coefficient; represents the total number of points on the surface of the car point cloud; represents the outward normal vector of the i th point; represents the area of the i th point.

[0079] The loss function of the joint prediction model in the training process is as follows:

[0080]

[0081] wherein the first term is the wind resistance supervision, the second term is the pressure supervision, and the third term is the physical consistency constraint; , , are weight coefficients; is the wind resistance coefficient derived from the pressure field by pressure field integration; is the wind resistance coefficient predicted by the joint prediction model; represents the true value of the wind resistance coefficient; represents the car surface pressure field predicted by the joint prediction model at the i th point; represents the true value of the car surface pressure field at the i th point.

[0082] The car surface pressure field prediction sub-network is as follows: Figure 2As shown, the pressure field feature is obtained by performing residual connection between the output of the first normalization layer and the sample point matrix after the sample point matrix is processed by the first normalization layer, performing residual connection between the output of the physical attention layer and the sample point matrix after the data output by the first normalization layer is processed by the physical attention layer, and performing residual connection between the output of the feedforward layer and the residual feature after the residual feature is processed by the second normalization layer and input into the feedforward layer.

[0083] The physical attention layer comprises a first linear layer, a multi-head attention module, and a second linear layer.

[0084] The following verifies the message of the embodiment of the application through specific data sets:

[0085] Verification experiment 1:

[0086] In order to verify the effectiveness of the method disclosed in the embodiment of the application, the DrivAerNet++ data set is used in the verification experiment, the DrivAerNet++ data set comprises 8000 automobile designs, each design is subjected to high-fidelity CFD simulation, and typical body categories such as fastback, notchback, and estateback are covered.

[0087] (1) In the wind resistance prediction experiment, for each sample, the curvature-based downsampling is performed to 2w points; the model is trained for 300 epochs, and the batch size = 1. The data is divided into: 5600 vehicles in the training set, 800 vehicles in the verification set, and 1600 vehicles in the test set. The results show that the wind resistance prediction technology proposed in the application has a significant advantage in the DrivAerNet++ data set. Specifically, under the condition of sampling 20,000 points, the method of the embodiment of the application is compared with the traditional PointNet, GCNN and other methods in the verification process, and the results are shown in Table 1:

[0088] Table 1 Comparison results of the method of the embodiment of the application and the traditional PointNet, GCNN and other methods

[0089]

[0090] (2) In the wind resistance-pressure field combined prediction experiment, the combined training prediction was carried out on the public data set DrivAerNet++ data set, the curvature of the automobile 3D grid was sampled to 10w points according to the curvature, and the simplified grid of the automobile was generated on the basis, and the overall shape parameters of the vehicle were extracted; the simplified grid is uniformly sampled, and the overall shape parameters are input into the whole vehicle wind resistance prediction sub-network, and at the same time, the simplified grid is fully sampled, and the surface pressure field prediction sub-network is input. The experimental results show that under the condition of fluid density 1.184 kg / m^3, flow velocity 30 m / s, and flow direction along the x axis of the DrivAerNet++ data set, the MRE of the wind resistance coefficient and the pressure field is , the MSE is , the MAE is , the consistency coefficient is 0.9607, the good consistency ratio is 96.2%, and the index is used as the physical consistency benchmark of the public data set DrivAerNet++. It is found that the C d and the surface pressure field output by the combined prediction network have good physical consistency.

[0091] Verification experiment 2:

[0092] The data set of a certain brand of automobile is used in this experiment, and the data set of a certain brand of automobile is much richer than the vehicle structure contained in the public data set, including but not limited to cars, SUVs, off-road vehicles, MPVs, pick-ups and other vehicle types. These vehicle types have significant differences in geometric structure, wind resistance source, vortex distribution, chassis design and spoiler layout. The results show that the wind resistance prediction technology proposed in the embodiment of the application has excellent generalization ability for diversified vehicle types on this data set, and the ablation experiment of different sampling point numbers shows that the prediction effect is best when the sampling point number is 5w.

[0093] Table 2 Performance of the method of the embodiment of the application at different sampling point numbers

[0094]

[0095] The embodiment of the application also provides a combined prediction system for the wind resistance coefficient and the surface pressure field of an automobile, comprising:

[0096] A data acquisition and preprocessing module is used to acquire a training data set comprising automobile point cloud data and label data, and to preprocess the automobile point cloud data in the training data set to obtain a sampling point matrix and an appearance parameter vector; and uniformly sample the sampling point matrix to obtain a uniform point matrix;

[0097] The model training module is configured to train an initial joint prediction model by using the sampling point matrix, the shape parameter vector and the uniform point matrix, optimize a loss value between a model output and label data, and obtain a joint prediction model.

[0098] The initial joint prediction model comprises a whole vehicle wind resistance prediction sub-network, an automobile surface pressure field prediction sub-network, a multi-layer perception machine, an image segmentation network and a physical consistency constraint calculation module.

[0099] The physical consistency constraint calculation module is configured to perform integration according to the predicted surface pressure field to obtain a wind resistance coefficient derived from the pressure field.

[0100] The whole vehicle wind resistance prediction sub-network is configured to process the uniform point matrix and the shape parameter vector to obtain a wind resistance feature; the automobile surface pressure field prediction sub-network is configured to process the sampling point matrix to obtain a pressure field feature; the multi-layer perception machine is configured to process the wind resistance feature and the wind resistance coefficient derived from the pressure field to obtain the automobile wind resistance coefficient; and the image segmentation network is configured to process the pressure field feature and the wind resistance coefficient derived from the pressure field to obtain the automobile surface pressure field.

[0101] It can be understood that the joint prediction system of the automobile wind resistance coefficient and the surface pressure field provided by the embodiment of the present application corresponds to the joint prediction method of the automobile wind resistance coefficient and the surface pressure field, and the explanation, examples, beneficial effects and the like of the related content can refer to the corresponding content in the joint prediction method of the automobile wind resistance coefficient and the surface pressure field, which will not be repeated here.

[0102] The embodiment of the present application further provides a computer readable storage medium storing a computer program for joint prediction of an automobile wind resistance coefficient and a surface pressure field, wherein the computer program causes a computer to execute the joint prediction method of the automobile wind resistance coefficient and the surface pressure field as described above.

[0103] The embodiment of the present application further provides an electronic device comprising one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the program comprises a program for executing the joint prediction method of the automobile wind resistance coefficient and the surface pressure field as described above.

[0104] Compared with the prior art, the embodiment of the present application has the following beneficial effects:

[0105] 1. The joint prediction method of the embodiment of the present application predicts the pressure field in parallel in the same network. With and the physical consistency constraint is used to constrain the pressure field integral Direct regression with network Calibration is performed to make the model physically plausible and interpretable while ensuring numerical accuracy, thereby enabling the localization of key stress areas and serving the automotive styling iteration process.

[0106] 2、The wind resistance coefficient prediction of the embodiment of the present application does not require CFD mesh partitioning and numerical solving, and the tedious geometry repair, turbulence model and boundary condition configuration are omitted; it can be directly used for repeated iteration and scheme screening in the early stage of styling, greatly reducing the manpower, computing power and time cost. At the same time, the double-channel representation of "local geometry + global shape" is adopted, the vehicle shape parameters (length, width, height) are automatically extracted from the point cloud, and the local-point cloud structure features of the curvature sampling point cloud are encoded respectively, then cross-attention fusion is performed in the network, which takes into account the details and contour prior, and improves the cross-model generalization of wind resistance prediction.

[0107] It should be noted that, in this text, relational terms such as first and second are used merely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such a process, method, article or apparatus. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus including the element.

[0108] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit it; although the foregoing embodiments of the present application have been described in detail, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for combined prediction of the drag coefficient and the surface pressure field of a car, characterized in that, The method comprises the following steps: acquire a training data set comprising car point cloud data and label data, and preprocess the car point cloud data in the training data set to obtain a sampling point matrix and an outline parameter vector; uniformly sample the sampling point matrix to obtain a uniform point matrix; train an initial joint prediction model through the sampling point matrix, the outline parameter vector and the uniform point matrix, optimize the loss value between the output of the joint prediction model and the label data, and obtain the joint prediction model; wherein the initial joint prediction model comprises a whole vehicle wind resistance prediction subnetwork, a car surface pressure field prediction subnetwork, a multi-layer perception machine, an image segmentation network and a physical consistency constraint calculation module; wherein the physical consistency constraint calculation module is used to integrate the predicted surface pressure field to obtain a wind resistance coefficient derived from the pressure field; the whole vehicle wind resistance prediction subnetwork is used to process the uniform point matrix and the outline parameter vector to obtain a wind resistance feature; the car surface pressure field prediction subnetwork is used to process the sampling point matrix to obtain a pressure field feature; the multi-layer perception machine is used to process the wind resistance feature and the wind resistance coefficient derived from the pressure field to obtain a car wind resistance coefficient; and the image segmentation network is used to process the pressure field feature and the wind resistance coefficient derived from the pressure field to obtain a car surface pressure field.

2. The method of claim 1, wherein the method further comprises: The whole vehicle wind resistance prediction subnetwork comprises a PointNet++ module, an MLP module and a feature fusion module, wherein the PointNet++ module is used to process the uniform point matrix to obtain a point cloud structure feature, and the MLP module is used to encode the outline parameter to obtain an outline feature; and the feature fusion module is used to align and fuse the point cloud structure feature and the outline feature.

3. The method of claim 2, wherein the step of calculating the pressure field is performed by solving the following equation: ###0001### where p is the pressure field, U is the velocity field, and n is the unit normal vector of the surface. The process of aligning and fusing the point cloud structure feature and the outline feature by the feature fusion module comprises: The linear mapping of the feature fusion module is The cross attention is The formula of the cross attention feature fusion module is wherein, p = 0.1 represents the dropout rate of Dropout; represents the point cloud structure feature; represents the shape feature; C represents the number of feature channels, B represents the batch size.

4. The method of claim 1, wherein the method further comprises: The car surface pressure field prediction subnetwork comprises a first normalization layer, a physical attention layer, a second normalization layer and a feedforward layer; the data output by the first normalization layer after processing the sampling point matrix is input into the physical attention layer, the output of the physical attention layer and the sampling point matrix are connected in residual to obtain residual features, the residual features are processed by the second normalization layer and then input into the feedforward layer, and the output of the feedforward layer and the residual features are connected in residual to obtain the pressure field feature.

5. The method of claim 4, wherein the step of calculating the pressure field is performed by solving the following equation: ###0002### where p is the pressure field, U is the velocity field, and n is the normal vector of the surface. The physical attention layer comprises a first linear layer, a multi-head attention module and a second linear layer; wherein the multi-head attention module comprises a slicing layer, an attention layer for labels and an inverse slicing layer.

6. The method of claim 1 to 5, wherein the method is characterized by, The integral of the predicted surface pressure field to obtain the wind resistance coefficient derived from the pressure field comprises: Given incoming flow direction unit vector , air density , velocity , reference frontal area A , dynamic pressure ; pressure sub-net outputs surface pressure field , after which a discrete integration yields the drag coefficient derived from the pressure field : wherein, denotes a correction factor; denotes the total number of points of the car point cloud surface; denotes the outward normal vector of the i point; denotes the area of the i point.

7. The method of claim 1 to 5, wherein the method is characterized by, The loss value between the output of the joint prediction model and the label data is calculated according to the following formula: where the first term is the wind drag supervision, the second term is the pressure supervision, and the third term is the physical consistency constraint; , , is a weight coefficient; is a wind drag coefficient derived from the pressure field by pressure field integration; is a wind drag coefficient predicted by the joint prediction model; denotes the true value of the wind drag coefficient; denotes the pressure field on the surface of the car at the i point predicted by the joint prediction model; denotes the pressure field on the surface of the car at the i point predicted by the joint prediction model; 8. A system for combined prediction of the drag coefficient and the surface pressure field of a car, characterized in that, The method comprises the following steps: a data acquisition and preprocessing module is used to acquire a training data set comprising car point cloud data and label data, and preprocess the car point cloud data in the training data set to obtain a sampling point matrix and an outline parameter vector; and the sampling point matrix is uniformly sampled to obtain a uniform point matrix; A model training module is configured to train an initial joint prediction model by using a sampling point matrix, an outline parameter vector, and a uniform point matrix, optimize a loss value between a model output and label data, and obtain a joint prediction model; the joint prediction model is configured to process a sampling point matrix, an outline parameter vector, and a uniform point matrix of a vehicle to be evaluated, and obtain an automobile wind resistance coefficient and an automobile surface pressure field. The initial joint prediction model includes a whole vehicle wind resistance prediction sub-network, an automobile surface pressure field prediction sub-network, a multi-layer perception machine, an image segmentation network, and a physical consistency constraint calculation module. The physical consistency constraint calculation module is configured to integrate a predicted surface pressure field to obtain a wind resistance coefficient derived from the pressure field. The whole vehicle wind resistance prediction sub-network is configured to process a uniform point matrix and an outline parameter vector to obtain a wind resistance feature; the automobile surface pressure field prediction sub-network is configured to process a sampling point matrix to obtain a pressure field feature; the multi-layer perception machine is configured to process the wind resistance feature and the wind resistance coefficient derived from the pressure field to obtain an automobile wind resistance coefficient; and the image segmentation network is configured to process the pressure field feature and the wind resistance coefficient derived from the pressure field to obtain an automobile surface pressure field.

9. A computer readable storage medium, characterized in that, A computer program for joint prediction of an automobile wind resistance coefficient and a surface pressure field is stored, wherein the computer program causes a computer to execute the joint prediction method of the automobile wind resistance coefficient and the surface pressure field according to any one of claims 1-7.

10. An electronic device, comprising: It includes: One or more processors, memories, and one or more programs, wherein the one or more programs are stored in the memories and configured to be executed by the one or more processors, and the programs include a method for executing the joint prediction method of the automobile wind resistance coefficient and the surface pressure field according to any one of claims 1-7.

Images