Automobile wind resistance coefficient prediction method and device based on multivariate conditional encoding

Abstract

This invention provides a method and apparatus for predicting the drag coefficient of a vehicle based on multivariate conditional coding, relating to the field of artificial intelligence. The method includes: acquiring three-dimensional surface geometry data of the vehicle and various operating condition parameter vectors of the vehicle's external aerodynamic performance; determining surface point cloud data, surface normal vectors, and area elements based on the three-dimensional surface geometry data; performing multivariate conditional coding on the various operating condition parameter vectors to generate a unified condition vector; inputting the surface point cloud data and the unified condition vector into a prediction network for condition-driven geometric feature learning and physical quantity prediction processing, outputting the vehicle's surface physical field; determining the aerodynamic direction based on the operating condition parameters; and using the aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vectors, and area elements to perform integral output processing based on the surface physical field to obtain the target drag coefficient of the vehicle. This invention can significantly improve prediction accuracy and reduce model maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to a method and apparatus for predicting the drag coefficient of a vehicle based on multivariate conditional coding. Background Technology

[0002] Currently, computational fluid dynamics (CFD) simulations are commonly used in automotive R&D to obtain various aerodynamic parameters of vehicles. However, traditional CFD simulation processes have long computation cycles and consume significant resources, making it difficult to meet the demands of rapid iterative design. Related technologies propose using data-driven surrogate models to shorten prediction time; however, this approach primarily models single operating conditions. When multiple heterogeneous parameters change simultaneously, multiple models need to be trained separately, leading to an increase in the number of models, higher maintenance costs, and limitations in model deployment due to mesh generation constraints. Furthermore, the derivation of the drag coefficient does not consider the contribution of wall shear stress, thus affecting the accuracy of drag coefficient prediction. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a method and apparatus for predicting the drag coefficient of a vehicle based on multivariate conditional coding, which can significantly improve the prediction accuracy and reduce the model maintenance cost.

[0004] In a first aspect, embodiments of the present invention provide a method for predicting the drag coefficient of a vehicle based on multivariate conditional coding. The method includes: acquiring three-dimensional surface geometry data of the vehicle and various operating condition parameter vectors of the vehicle's external flow field aerodynamic performance; determining surface point cloud data, surface normal vectors, and area elements based on the three-dimensional surface geometry data; performing multivariate conditional coding on the various operating condition parameter vectors to generate a unified condition vector; inputting the surface point cloud data and the unified condition vector into a prediction network for condition-driven geometric feature learning and physical quantity prediction processing; outputting the surface physical field of the vehicle, wherein the surface physical field includes surface pressure field data and wall shear stress field data; determining the aerodynamic direction based on the operating condition parameters; and using the aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vectors, and area elements to perform integral output processing based on the surface physical field to obtain the target drag coefficient of the vehicle.

[0005] In one implementation, the step of determining surface point cloud data, surface normal vectors, and area elements based on three-dimensional surface geometric data includes: performing surface point cloud extraction processing on the three-dimensional surface geometric data to obtain surface point cloud data; determining the area values ​​of each triangular facet in the three-dimensional surface geometric data as area elements, and determining the normal vectors of the triangular facets as surface normal vectors.

[0006] In one implementation, the step of performing multivariate conditional encoding on various working condition parameter vectors to generate a unified condition vector includes: performing parameter identification processing on each working condition parameter vector to obtain angular parameters and continuous parameters corresponding to each working condition parameter vector, and encoding the angular parameters and continuous parameters respectively to obtain angular parameter encoding results and continuous parameter encoding results; concatenating the angular parameter encoding results and continuous parameter encoding results to obtain a concatenated result, and performing feature fusion processing on the concatenated result through a conditional mapping network to generate a unified condition vector.

[0007] In one embodiment, the steps of encoding angular parameters and continuous parameters respectively to obtain angular parameter encoding results and continuous parameter encoding results include: performing periodic sensing encoding on the angular parameters to map the angular parameters into periodic continuous feature representations to eliminate discontinuities at angular boundaries, thereby obtaining angular parameter encoding results; and performing frequency domain embedding on the continuous parameters to map the continuous parameters into feature representations at multiple different frequency scales to enhance the representation ability of parameters with different dimensions, thereby obtaining continuous parameter encoding results.

[0008] In one embodiment, the prediction network includes a geometric feature extraction unit and a physical quantity prediction unit. The steps of inputting surface point cloud data and a unified conditional vector into the prediction network, performing condition-driven geometric feature learning and physical quantity prediction processing, and outputting the surface physical field of the vehicle include: inputting surface point cloud data into the geometric feature extraction unit for feature extraction processing to obtain the potential geometric features of the vehicle surface; and performing physical quantity prediction processing on the potential geometric features and the unified conditional vector through the physical quantity prediction unit to obtain surface pressure field data and wall shear stress field data.

[0009] In one embodiment, the step of obtaining the target drag coefficient of a vehicle by performing integral output processing based on the surface physical field using aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vector, and area element includes: determining the unit vector of the drag direction based on the yaw angle in the operating parameters; performing integral processing along the drag direction on each sampling unit of the vehicle surface according to the surface pressure field data, wall shear stress field data, surface normal vector, and area element to obtain the force situation of each sampling unit in the drag direction; and summing and normalizing the force situation corresponding to each sampling unit to obtain the target drag coefficient.

[0010] In one implementation, after obtaining the target drag coefficient, the process includes: constructing a training sample set and obtaining the true drag coefficient values ​​corresponding to the training sample set, wherein the training sample set includes: three-dimensional surface geometric data, working condition parameter vectors, surface pressure field data, wall shear stress field data, and the target drag coefficient; and training the prediction network using a joint loss function based on the true drag coefficient values ​​and the training sample set to obtain an updated target prediction network.

[0011] Secondly, embodiments of the present invention also provide a vehicle drag coefficient prediction device based on multivariate conditional coding. The device includes: a data processing module, which acquires three-dimensional surface geometry data of the vehicle and various operating condition parameter vectors of the vehicle's external flow field aerodynamic performance, and determines surface point cloud data, surface normal vectors, and area elements based on the three-dimensional surface geometry data; a multivariate conditional coding module, which performs multivariate conditional coding processing on the various operating condition parameter vectors to generate a unified conditional vector, and inputs the surface point cloud data and the unified conditional vector into a prediction network for condition-driven geometric feature learning and physical quantity prediction processing, outputting the vehicle's surface physical field, wherein the surface physical field includes: surface pressure field data and wall shear stress field data; and an aerodynamic coefficient integration module, which determines the aerodynamic direction according to the operating condition parameters, and uses the aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vectors, and area elements to perform integration output processing based on the surface physical field to obtain the vehicle's target drag coefficient.

[0012] Thirdly, embodiments of the present invention also provide a server, including a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement any of the methods provided in the first aspect.

[0013] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement any of the methods provided in the first aspect.

[0014] The embodiments of the present invention bring the following beneficial effects: This invention provides a method and apparatus for predicting the drag coefficient of a vehicle based on multivariate conditional coding. The method first acquires the three-dimensional surface geometry data of the vehicle and various operating condition parameter vectors of the vehicle's external aerodynamic performance. Based on the three-dimensional surface geometry data, it determines surface point cloud data, surface normal vectors, and area elements. Then, it performs multivariate conditional coding on the various operating condition parameter vectors to generate a unified condition vector. The surface point cloud data and the unified condition vector are input into a prediction network for condition-driven geometric feature learning and physical quantity prediction, outputting the vehicle's surface physical field. Finally, the aerodynamic direction is determined based on the operating condition parameters, and the aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vectors, and area elements are used to perform integral output processing based on the surface physical field to obtain the target drag coefficient of the vehicle. This invention can significantly improve prediction accuracy and reduce model maintenance costs.

[0015] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for predicting the drag coefficient of a vehicle based on multivariate conditional coding, provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating the specific process of a vehicle drag coefficient prediction method based on multivariate conditional coding provided in an embodiment of the present invention; Figure 3 A schematic diagram of a multivariate conditional encoding provided in an embodiment of the present invention; Figure 4 A schematic diagram of an aerodynamic coefficient integration path provided in an embodiment of the present invention; Figure 5 A schematic diagram of a vehicle drag coefficient prediction device based on multivariate conditional coding provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a server provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Currently, in the automotive R&D process, aerodynamic parameters such as drag coefficient, lift coefficient, and surface pressure distribution are usually obtained through computational fluid dynamics simulation. Related technologies suggest that the traditional computational fluid dynamics simulation process includes steps such as geometry cleaning, volume mesh generation, boundary condition setting, and numerical solution. However, when dealing with combinations of different styling schemes, different yaw angles, and different vehicle speeds, this existing simulation process needs to repeat the above process, resulting in a long overall calculation cycle and high resource consumption. Therefore, it is difficult to meet the needs of rapid iterative design.

[0021] While existing data-driven proxy models can shorten prediction time, they still have the following shortcomings: 1. Insufficient multi-condition modeling capability. Most existing solutions model for a single condition or a small number of scalar conditions. When multiple heterogeneous parameters such as yaw angle, incoming flow velocity, Reynolds number, and ground clearance change simultaneously, multiple models often need to be trained separately, which increases the number of models, raises maintenance costs, and reduces the ability to generalize to unseen combinations of conditions.

[0022] 2. Reliance on CFD mesh input. Existing high-precision methods typically rely on CFD simulation meshes, volume mesh nodes, or discrete topologies associated with the simulation mesh as input. When the new geometry to be evaluated has not yet generated a simulation mesh, such methods are difficult to deploy directly and are still limited by the mesh generation process.

[0023] 3. Lack of consistency between drag coefficient and physical field. In existing methods, the drag coefficient is often obtained by direct regression after global pooling, or by approximate derivation based solely on surface pressure, without simultaneously considering the contribution of wall shear stress. This results in a lack of consistency between the global aerodynamic coefficient and the local surface physical field.

[0024] Based on this, the vehicle drag coefficient prediction method and apparatus based on multivariate conditional coding provided by this invention can solve the problems in the prior art that require training multiple models separately when the heterogeneous parameters change under multiple operating conditions, the prediction process relies on CFD volume mesh leading to complex deployment, and the lack of consistency between the drag coefficient and the surface pressure field and wall shear stress field. By directly using the vehicle geometric point cloud and operating parameters, it outputs at least one physical quantity among the surface pressure field, wall shear stress field and velocity field, and obtains the drag coefficient and lift coefficient by integrating the surface physical field, thereby significantly improving the prediction accuracy and reducing the model maintenance cost.

[0025] See Figure 1 The diagram shows a flowchart of a vehicle drag coefficient prediction method based on multivariate conditional coding. The method mainly includes the following steps S102 to S106: Step S102: Obtain the three-dimensional surface geometry data of the vehicle and the various working condition parameter vectors of the aerodynamic performance of the vehicle's external flow field, and determine the surface point cloud data, surface normal vector, and area element based on the three-dimensional surface geometry data.

[0026] In one implementation, three-dimensional surface geometry data refers to a computer-aided design model or mesh file that can characterize the shape of the outer surface of a car. This geometry data describes the surface structure of the outer surface of the car, which is composed of a large number of triangular facets. Each triangular facet contains the spatial coordinate information of three vertices, the facet normal vector, and the facet area. By reading such files, a complete description of the car's shape can be obtained.

[0027] Simultaneously, the aerodynamic parameter vectors of various operating conditions for the vehicle's external flow field are obtained. These operating parameter vectors refer to a set of multiple external condition parameters that affect the aerodynamic performance of the vehicle during driving. Specifically, these may include: yaw angle, incoming flow velocity, Reynolds number, and ground clearance. The yaw angle is the angle between the incoming flow direction and the vehicle's driving direction, used to characterize crosswind conditions; the incoming flow velocity refers to the vehicle's speed or wind speed; the Reynolds number is a dimensionless number used to characterize the fluid flow state, determined by fluid density, incoming flow velocity, and characteristic length; and the ground clearance is the vertical distance between the vehicle chassis and the ground, used to characterize the impact of the vehicle's ground clearance on the underside airflow.

[0028] After acquiring the aforementioned geometric data and operating parameters, surface point cloud data, surface normal vectors, and area elements are determined based on the three-dimensional surface geometric data. Specifically, surface point cloud extraction processing is performed on the three-dimensional surface geometric data. This surface point cloud extraction processing refers to the process of extracting a representative set of discrete point coordinates from the triangular facets constituting the outer surface of the vehicle. In one embodiment, the coordinates of the three vertices of each triangular facet can be extracted to form a point cloud; as another optional embodiment, the coordinates of the center point of each triangular facet can be extracted to form a point cloud, or a more uniformly distributed point cloud data can be obtained through uniform resampling. Through this processing, surface point cloud data is obtained, which is a set of discrete points with three-dimensional coordinates, where each point represents a sampling position on the outer surface of the vehicle.

[0029] Furthermore, the area values ​​of each triangular facet in the 3D surface geometry data are defined as area elements. Here, the area element refers to the size of the area corresponding to each triangular facet, used to characterize the weight of that sampling unit in subsequent integration calculations. Simultaneously, the normal vector of the triangular facet is defined as the surface normal vector. Here, the surface normal vector is a unit vector perpendicular to the plane containing the triangular facet and pointing outwards from the vehicle, used to characterize the orientation of the surface at that location. Through the above processing, each sampling point is associated with a corresponding surface normal vector and area element, providing the necessary basic data for subsequent physics field prediction and aerodynamic coefficient integration.

[0030] Step S104: Perform multivariate conditional encoding on the vectors of various working conditions to generate a unified conditional vector. Input the surface point cloud data and the unified conditional vector into the prediction network to perform condition-driven geometric feature learning and physical quantity prediction processing, and output the physical field data of the vehicle surface. The physical field data includes at least surface pressure field data and wall shear stress field data.

[0031] In one implementation, multivariate conditional encoding can be performed on the vectors of various operating parameters to generate a unified conditional vector. This multivariate conditional encoding refers to the process of converting operating parameters with different properties and dimensions into a unified feature vector that can be processed by a neural network. Since the operating parameters include both angular and continuous parameters, which have different properties, they need to be processed separately. Specifically, parameter identification processing is performed on each operating parameter vector to distinguish between angular and continuous parameters. For angular parameters, such as yaw angle, periodic-aware encoding is performed. Angles are periodic; for example, 0 degrees and 360 degrees represent the same physical direction. Directly inputting the original angle value would lead to numerical discontinuities. Periodic-aware encoding maps the angle to a combination of sine and cosine function values, maintaining continuity at the angle boundaries. For continuous parameters, such as incoming flow velocity, Reynolds number, and ground clearance, frequency domain embedding is performed. Since the value range of continuous parameters varies greatly, frequency domain embedding maps the parameters to sine and cosine function values ​​at multiple different frequency scales, forming a rich feature representation in a high-dimensional space. The angle-type parameter encoding results and the continuous parameter encoding results are concatenated to obtain the concatenated result. Then, the concatenated result is subjected to feature fusion processing through a conditional mapping network to generate a unified conditional vector.

[0032] Furthermore, the surface point cloud data and the unified conditional vector are input into the prediction network for condition-driven geometric feature learning and physical quantity prediction processing, outputting the physical field data of the vehicle surface. This physical field data includes at least surface pressure field data and wall shear stress field data. The surface pressure field refers to the pressure distribution of air acting on various points on the vehicle's outer surface, while the wall shear stress field refers to the frictional force distribution of air acting on the surface along the tangent direction. The prediction network includes a geometric feature extraction unit and a physical quantity prediction unit. Specifically, the surface point cloud data is input into the geometric feature extraction unit for feature extraction processing to obtain the potential geometric features of the vehicle surface. These potential geometric features contain key geometric information such as the overall shape of the vehicle, local protrusions, and depressions. Then, the physical quantity prediction unit performs physical quantity prediction processing on the potential geometric features and the unified conditional vector. The unified conditional vector is injected into the prediction network as a conditional control signal, informing the network of the current operating conditions, enabling the network to output the corresponding physical field according to different operating conditions. Through this processing, the surface pressure field data and wall shear stress field data are finally obtained.

[0033] Step S106: Determine the aerodynamic direction based on the operating parameters, and use the aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vector and area element to perform integral output processing based on the surface physical field to obtain the target drag coefficient of the vehicle.

[0034] In one implementation, the model must first output the surface pressure field and the wall shear stress field, and then integrate these two physical fields to obtain the drag coefficient or lift coefficient. During retraining, the error of the drag coefficient will propagate back, forcing the model to calibrate the surface physical fields as well. The drag coefficient represents the degree to which air hinders the car's movement, and is integrated along the drag direction. The lift coefficient represents the degree to which air lifts or presses the car down, and is integrated along the lift direction. Except for the different integration paths, the calculation steps for the two are the same. The lift coefficient can be calculated according to actual usage requirements and has no impact on the calculation of the drag coefficient.

[0035] Specifically, the aerodynamic direction is determined based on the operating parameters. Then, using the aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vector, and area elements, an integral output based on the surface physics field is performed to obtain the target drag coefficient of the vehicle. First, the unit vector of the drag direction is determined based on the yaw angle in the operating parameters. Here, the aerodynamic direction refers to the direction of the resultant force generated by air acting on the vehicle surface. The drag direction is a major component of the aerodynamic direction, specifically referring to the opposite direction of the incoming flow, i.e., the direction in which air impedes the vehicle's forward movement. Since the yaw angle describes the angle between the incoming flow direction and the vehicle's direction of travel, the unit vector of the drag direction can be calculated using trigonometric functions based on the yaw angle. This unit vector is used in subsequent calculations to project the force on each sampling unit onto the drag direction.

[0036] Furthermore, based on the surface pressure field data, wall shear stress field data, surface normal vector, and area element, the force on each sampling unit of the vehicle surface is integrated along the resistance direction to obtain the force situation of each sampling unit in the resistance direction. Here, the integration process refers to projecting the pressure and friction forces onto the resistance direction separately for each sampling unit and then summing them. Specifically, for each sampling unit, its surface pressure value is first multiplied by the dot product of the surface normal vector and the unit vector of the resistance direction to obtain the contribution of pressure in the resistance direction; then, its wall shear stress vector is multiplied by the unit vector of the resistance direction to obtain the contribution of friction in the resistance direction; the two are then added together to obtain the total force on that sampling unit in the resistance direction. Since each sampling unit corresponds to an area element, this force value already implicitly includes the influence of area weighting.

[0037] Finally, the forces acting on each sampling unit are summed, and the summation result is normalized to obtain the target drag coefficient. The summation here refers to adding the forces acting on all sampling units in the drag direction to obtain the total air resistance experienced by the vehicle. The normalization process involves dividing the total resistance by the product of the reference dynamic pressure and the reference area, where the reference dynamic pressure is equal to half multiplied by the fluid density and then multiplied by the square of the incoming flow velocity. This normalization process yields a dimensionless drag coefficient, which is the aerodynamic drag performance index of the vehicle under given operating conditions.

[0038] The vehicle drag coefficient prediction method based on multivariate conditional coding provided in this invention uses surface point clouds extracted from the vehicle's three-dimensional curved surface geometry data as geometric input and operating parameters such as yaw angle, incoming flow velocity or Reynolds number, and ground clearance as conditional input. First, different types of operating parameters are encoded and mapped into a unified conditional vector. Then, the unified conditional vector drives the geometric feature learning and physical quantity prediction process, outputting at least one physical quantity from the surface pressure field, wall shear stress field, and velocity field. Subsequently, based on the predicted surface pressure field and wall shear stress field, combined with the surface normal vector, area element, and aerodynamic direction determined by the operating parameters, area-weighted integration is performed on the vehicle surface to obtain the drag coefficient and lift coefficient. This scheme does not require the generation of a CFD volume mesh and can perform rapid prediction from a single model for new geometries and multiple operating condition combinations, while enhancing the consistency between global aerodynamic coefficients and local surface physical fields.

[0039] See Figure 2 The diagram shows a specific flowchart of a method for predicting the drag coefficient of a car based on multivariate conditional coding. This embodiment of the invention also provides an implementation method for predicting the drag coefficient of a car based on multivariate conditional coding, as detailed in (1) to (4) below: (1) Geometric point cloud extraction and preprocessing: Obtain vehicle geometric data and perform geometric point cloud extraction and preprocessing on the vehicle geometric data. Specifically, the input is three-dimensional surface geometric data of the vehicle. The geometric data can be .stl, .obj, .step or other geometric files that can represent the outer surface of the vehicle body. First, surface point clouds are extracted from the three-dimensional surface geometric data of the vehicle. Calculate or read the external normal vector corresponding to each surface sampling point. and area elements Then construct the query point set. The query point set includes surface query points, and may also include volume query points when flow field prediction is required. Finally, the surface point cloud and query point coordinates are positionally encoded to obtain the embedded geometric point cloud. and embedded query points .

[0040] In one implementation, surface point cloud extraction processing can be performed on the three-dimensional surface geometric data to obtain surface point cloud data. The area values ​​of each triangular facet in the three-dimensional surface geometric data are determined as area elements, and the normal vectors of the triangular facets are determined as surface normal vectors. The surface point cloud can be composed of triangular facet vertices, facet center points, or uniform point sets obtained through resampling. Position encoding refers to the process of coordinate normalization, geometric feature construction, or feature embedding. Position encoding can adopt sine / cosine encoding, multi-frequency position encoding, or other encoding methods that can enhance the coordinate representation capability.

[0041] (2) Multivariate condition coding: Obtaining operating condition parameters Multivariate conditional encoding is performed on the operating parameters to obtain a unified condition vector, where the operating parameters are... This may include yaw angle Incoming flow velocity Reynolds number Ground gap And other parameters that affect the aerodynamic external flow field.

[0042] Specifically, the input is a vector of operating condition parameters. Since operating parameters may include both angular and continuous parameters, this invention first encodes the different types of parameters separately, and then maps them into a unified condition vector: angular parameters are periodically sensed and encoded to eliminate discontinuities caused by angular period boundaries; continuous parameters are frequency-domain embedded or subjected to other nonlinear mappings to enhance the representation of parameters with different dimensions; the encoded results of different parameters are concatenated and input into a condition mapping network to obtain the unified condition vector. The unified condition vector is input into the prediction network as conditional information to control geometric feature learning and physical quantity prediction.

[0043] In one implementation, parameter identification processing can be performed on the various working condition parameter vectors to obtain the corresponding angle-type parameters and continuous parameters. The angle-type parameters and continuous parameters are then encoded separately to obtain angle-type parameter encoding results and continuous parameter encoding results. These results are then concatenated to obtain a concatenated result. A conditional mapping network is used to perform feature fusion processing on the concatenated result to generate a unified conditional vector. Specifically, periodic-aware encoding processing is performed on the angle-type parameters, mapping them to periodically continuous feature representations to eliminate discontinuities at angle boundaries, resulting in angle-type parameter encoding results. Frequency-domain embedding processing is performed on the continuous parameters, mapping them to feature representations at multiple different frequency scales to enhance the representation ability of parameters with different dimensions, resulting in continuous parameter encoding results.

[0044] In one implementation, see Figure 3 The diagram shown illustrates a multivariate conditional coding scheme for yaw angle. The encoding can be represented as:

[0045] in, and First-order periodic information used to characterize the yaw angle; and The second-order periodic information used to characterize the yaw angle enhances the model's ability to express angle change patterns; the joint representation of sine and cosine terms avoids numerical discontinuities near periodic boundaries such as 0 or 360 degrees.

[0046] For continuous parameters The encoding can be represented as:

[0047] Where x represents the input continuous operating condition parameter, which can be the incoming flow velocity, Reynolds number, ground clearance or other continuous variables; Indicates preset One frequency scale parameter, Indicates the number of frequencies used in frequency domain embedding; , This represents the variation characteristics of the characterization parameter x at the Kth frequency scale.

[0048] By jointly encoding multiple frequency scales, the model's ability to represent operating parameters with different dimensions and varying ranges can be enhanced. For a unified representation of multivariable operating parameters, this invention concatenates the encoding results of different types of parameters and inputs them into a conditional mapping network to obtain a unified conditional vector, the expression of which is:

[0049] in, Indicates a splicing operation; This represents a unified condition vector, used to characterize the comprehensive condition information of the current working condition combination; The conditional mapping network is preferably a multilayer perceptron. Encoding results for different parameters; This indicates that the encoding results of other operating condition parameters can be further combined.

[0050] Furthermore, the unified conditional vector can be used to generate feature injection parameters, gating weights, or other control signals to adjust intermediate features in one or more hidden layers of the prediction network.

[0051] (3) Condition-driven geometric feature learning and physical quantity prediction: The preprocessed geometric point cloud and unified condition vector are input into the prediction network to perform condition-driven geometric feature learning and physical quantity prediction, and output at least one of the surface pressure field, wall shear stress field and velocity field.

[0052] Specifically, the input is the embedded geometric point cloud. Query point representation and unified condition vector The prediction network jointly models the vehicle geometry and operating conditions and outputs the physical quantities to be determined. The prediction network may include a geometric feature extraction unit and a physical quantity prediction unit, and performs at least the following operations: first, extract the potential geometric feature representation based on the surface point cloud; then, combine the unified condition vector with the geometric feature extraction process and the query point physical quantity prediction process; finally, based on the geometric feature representation, the query point location, and the operating conditions, output at least one physical quantity among the surface pressure field, the wall shear stress field, and the velocity field.

[0053] In one embodiment, surface point cloud data can be input into a geometric feature extraction unit for feature extraction processing to obtain potential geometric features of the vehicle surface. Then, a physical quantity prediction unit can be used to perform physical quantity prediction processing on the potential geometric features and the unified condition vector to obtain surface pressure field data and wall shear stress field data. The prediction network includes a geometric feature extraction unit and a physical quantity prediction unit.

[0054] Furthermore, the relationship between the latent geometric feature representation and the physical quantity output can be expressed as:

[0055]

[0056] in, For surface pressure field, For the wall shear stress field, It represents the velocity field; the output can be at least one of the above physical quantities.

[0057] The prediction network can adopt a Geometry-Informed Neural Operator (GINO) style architecture, including: a geometric encoding unit, a latent global operator processing unit, and a geometric decoding unit. The geometric encoding unit maps the input surface point cloud to a latent regular grid or latent regular representation; the latent global operator processing unit performs global operator operations on the latent regular representation, preferably using a Fourier Neural Operator-type module; and the geometric decoding unit maps latent features to arbitrary query point locations, outputting the surface pressure field, wall shear stress field, and optional velocity field. A unified condition vector can be used as conditional information injection or a control signal in one or more stages of the encoding, latent global operator, and decoding phases. and It can be implemented using the encoder-processor-decoder structure in a GINO-style architecture.

[0058] In some implementations, conditional information injection can be achieved through feature-wise linear transformation, which can be expressed as:

[0059] in, Intermediate features representing the conditional information to be injected. and These represent the scaling factor and offset factor generated from the unified condition vector, respectively. This indicates element-wise multiplication.

[0060] (4) Aerodynamic coefficient integral output based on surface physical field: Based on the surface pressure field and wall shear stress field, combined with the surface normal vector, the area weight corresponding to the surface sampling point or surface sampling unit, and the aerodynamic direction determined by the working condition parameters, the integral output based on the surface physical field is performed to obtain the drag coefficient and / or lift coefficient. That is to say, the present invention does not directly output the drag coefficient through the regression head, but first predicts the surface pressure field and wall shear stress field, and then performs area-weighted integration on the surface physical field to obtain the global aerodynamic coefficient.

[0061] In one implementation, the unit vector of the drag direction can be determined based on the yaw angle in the operating parameters. Then, based on the surface pressure field data, wall shear stress field data, surface normal vector, and area element, each sampling unit on the vehicle surface is integrated along the drag direction to obtain the force situation of each sampling unit in the drag direction. The force situation corresponding to each sampling unit is then summed and normalized to obtain the target drag coefficient.

[0062] In one implementation, see Figure 4The diagram shows an integral path for aerodynamic coefficients, with drag coefficients... It can be represented as:

[0063] in, Indicates fluid density; Indicates the incoming flow velocity; Indicates the reference area; Indicates the first The external normal vector of each surface sampling unit; Indicates the first Area elements of each surface sampling unit; This represents the unit vector of drag direction determined by the yaw angle.

[0064] In one embodiment, the unit vector in the direction of drag can be expressed as:

[0065] Lift coefficient It can be obtained by integrating along the direction of lift.

[0066] Furthermore, the prediction network can be pre-trained, or the results of each training iteration can be used to update the prediction network for the next round of wind resistance prediction. When updating using the prediction results, a training sample set is first constructed, and the ground truth values ​​of the wind resistance coefficients corresponding to the training sample set are obtained. Then, using a joint loss function, the prediction network is trained based on the ground truth values ​​of the wind resistance coefficients and the training sample set to obtain the updated target prediction network. The training sample set includes: three-dimensional surface geometric data, operating condition parameter vectors, surface pressure field data, wall shear stress field data, and the target wind resistance coefficient.

[0067] In one implementation, to obtain a prediction model or update module, the network can be trained based on offline CFD samples. The training samples may include true values ​​of the surface pressure field, the wall shear stress field, the aerodynamic coefficients, and optionally, the velocity field.

[0068] A joint loss function can be used during training:

[0069] in, Used to constrain surface pressure field prediction errors; Used to constrain the prediction error of wall shear stress field; Used to constrain the error between the aerodynamic coefficients obtained by integration and the true values; This is an optional setting used to constrain velocity field prediction errors. This is an optional parameter used to constrain the residuals of continuity equations or other physical residuals.

[0070] Since the aerodynamic coefficients are obtained by integrating the surface physical field, the gradient of the coefficient loss can be backpropagated along the integration path to the surface physical field prediction branch, thereby improving the consistency between global indices and local physical quantities.

[0071] In practical applications, please refer to Examples 1 to 3 below: Example 1: Model Training and Deployment.

[0072] S101: Sample Data Construction. 300 sets of 3D automotive surface geometry samples with different geometric parameters were constructed. These parameters included length, width, height, rear tilt angle, and corner radius. CFD calculations were performed on each geometry under 16 operating conditions, including 4 incoming flow velocity levels and 4 yaw angle levels, resulting in a total of 4,800 samples.

[0073] In this embodiment: the incoming flow velocity is 10 m / s, 30 m / s, 50 m / s, and 70 m / s; the yaw angle is... , , and The CFD-annotated data includes surface pressure field, wall shear stress field, aerodynamic coefficient labels, and optional velocity field labels. The dataset is divided into training set (80%), validation set (10%), and test set (10%) according to geometric dimensions to avoid different working conditions of the same geometry appearing in the training set and test set at the same time.

[0074] S102: Geometry and Operating Condition Preprocessing. For each set of 3D automotive surface geometry, surface point clouds are extracted from its triangular patch data, and the normal vector and area element corresponding to each sampling unit are calculated. Regarding operating condition parameters, periodic sensing encoding is used for the yaw angle, and continuous parameter embedding encoding is used for the incoming flow velocity or Reynolds number. The encoded results of each parameter are then concatenated and mapped to a unified condition vector.

[0075] In this embodiment, approximately 100,000 surface sampling points can be extracted from a single geometric sample; the conditional vector dimension can be set to 256.

[0076] S103: Network Construction and Training. A prediction network, preferably employing a GINO-style architecture, is constructed. The prediction network includes a GNO (Graph Neural Operator) or a similar geometric encoder, a global operator module located on a latent regular grid, and a GNO geometric decoder. The geometric encoder maps the vehicle surface point cloud to a latent regular grid or latent regular representation; the global operator module performs global feature propagation on the latent regular representation, preferably using a Fourier NeuralOperator-type module; and the geometric decoder maps latent features to surface query points or volume query points and outputs the corresponding physical quantities. During training, a unified condition vector can be injected as conditional information into one or more stages of the encoding, latent global operator, and decoding stages to achieve unified modeling for different operating conditions. To adapt to memory constraints, random subsampling of query points is performed during training.

[0077] In this embodiment, the example configuration shown in Table 1 can be used. The number of feature extraction layers and the number of prediction layers in Table 1 are exemplary network depth indicators used to characterize the network size of the above preferred architecture. It is not limited to using a unique combination of layers.

[0078] Table 1 Example Configuration Table

[0079] During training, a joint loss is used to jointly constrain the surface pressure field, wall shear stress field, aerodynamic coefficients, and optional velocity field. For example, the following can be configured:

[0080] S104: Online Inference. For the new vehicle geometry to be evaluated, simply extract its surface point cloud, normal vector, and area elements, and input the target operating parameters to output at least one of the following results: surface pressure distribution, wall shear stress distribution, velocity field, as well as drag coefficient and lift coefficient.

[0081] In this embodiment, the total time for geometric preprocessing, condition coding, and model forward inference is approximately 10 seconds.

[0082] Example 2: Unified Modeling of Multiple Operating Conditions Based on Multivariate Conditional Encoding. In this example, yaw angle, incoming flow velocity, and ground clearance are input as joint conditions into the same prediction network. By encoding angular and continuous parameters and mapping them to a unified conditional vector, a single model can cover multiple combinations of operating conditions without the need to train independent models for different operating conditions.

[0083] The unified condition vector can also serve as a condition control signal in the aforementioned preferred architecture during the encoding, potential global operator, and decoding stages.

[0084] Compared to approaches that model only a single condition, this implementation reduces the number of models and maintenance costs, and improves the ability to generalize to unseen combinations of operating conditions.

[0085] Example 3: Consistency Constraints for Aerodynamic Coefficients Based on Integral Paths. In this example, instead of setting a separate regression head for the direct regression drag coefficient, the model first outputs the surface pressure and wall shear stress, and then integrates the drag coefficient based on the surface normal vector, area element, and drag direction determined by the yaw angle.

[0086] Compared to the approach of directly regressing aerodynamic coefficients, this implementation allows the aerodynamic coefficient monitoring signal to act on the surface physics prediction branch, thereby enhancing the consistency between global indices and local physics.

[0087] In summary, this invention can solve the problems of requiring multiple models for multi-condition prediction, prediction relying on CFD volumetric meshes, and inconsistencies between drag coefficient and surface physical field: First, this invention identifies operating parameters such as yaw angle and incoming flow velocity, distinguishes between angular and continuous parameters, and then performs periodic sensing encoding and frequency domain embedding encoding respectively. The encoding results are then concatenated and fused into a unified conditional vector. Through this multivariate conditional encoding, a single prediction network can simultaneously accept combined inputs of multiple operating parameters, eliminating the need to train multiple models separately for different operating conditions.

[0088] Secondly, this invention only needs to extract surface point cloud data from three-dimensional curved surface geometry data, and determine the area value and normal vector of the triangular facet as area elements and surface normal vector. The entire prediction process only depends on the geometric information of the vehicle surface, without the need to generate a CFD volume mesh for each new geometry, which greatly reduces the deployment complexity.

[0089] Finally, this invention first predicts the surface pressure field and wall shear stress field output by the network, then determines the drag direction based on the yaw angle, integrates each sampling unit along the drag direction, and finally obtains the drag coefficient through summation and normalization. Since the drag coefficient is derived from the integral of the predicted physical field rather than a direct regression output, the supervisory signal for the drag coefficient can propagate back to the physical field prediction branch, forcing the model to simultaneously learn the accurate surface physical field, thereby enhancing the consistency between the global coefficients and the local physical field.

[0090] Regarding the vehicle drag coefficient prediction method based on multivariate conditional coding provided in the foregoing embodiments, this invention provides a vehicle drag coefficient prediction device based on multivariate conditional coding. (See attached image) Figure 5 The diagram shows a structural schematic of a vehicle drag coefficient prediction device based on multivariate conditional coding. The device includes the following parts: The data processing module 502 acquires the three-dimensional curved surface geometry data of the vehicle and the various working condition parameter vectors of the aerodynamic performance of the vehicle's external flow field, and determines the surface point cloud data, surface normal vector and area element based on the three-dimensional curved surface geometry data. The multivariate conditional encoding module 504 performs multivariate conditional encoding on the vectors of various working conditions to generate a unified conditional vector. It then inputs the surface point cloud data and the unified conditional vector into the prediction network to perform condition-driven geometric feature learning and physical quantity prediction processing, and outputs the surface physical field of the vehicle. The surface physical field includes surface pressure field data and wall shear stress field data. The aerodynamic coefficient integration module 506 determines the aerodynamic direction based on the operating parameters, and uses the aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vector and area element to perform integral output processing based on the surface physical field to obtain the target drag coefficient of the vehicle.

[0091] The vehicle drag coefficient prediction device based on multivariate conditional coding provided in this application embodiment can significantly improve prediction accuracy and reduce model maintenance costs.

[0092] In one embodiment, when performing the step of determining surface point cloud data, surface normal vectors, and area elements based on three-dimensional surface geometric data, the data processing module 502 is further configured to: perform surface point cloud extraction processing on the three-dimensional surface geometric data to obtain surface point cloud data; determine the area values ​​of each triangular facet in the three-dimensional surface geometric data as area elements, and determine the normal vectors of the triangular facets as surface normal vectors.

[0093] In one embodiment, when performing multivariate conditional encoding processing on various working condition parameter vectors to generate a unified condition vector, the multivariate conditional encoding module 504 is further configured to: perform parameter identification processing on various working condition parameter vectors to obtain angular parameters and continuous parameters corresponding to each working condition parameter vector, and encode the angular parameters and continuous parameters respectively to obtain angular parameter encoding results and continuous parameter encoding results; concatenate the angular parameter encoding results and continuous parameter encoding results to obtain a concatenated result, and perform feature fusion processing on the concatenated result through a conditional mapping network to generate a unified condition vector.

[0094] In one embodiment, when performing the steps of encoding the angular parameters and continuous parameters respectively to obtain the angular parameter encoding result and the continuous parameter encoding result, the multivariate conditional encoding module 504 is further configured to: perform periodic sensing encoding processing on the angular parameters, mapping the angular parameters to periodic continuous feature representations to eliminate the discontinuity of the angular boundary, and obtain the angular parameter encoding result; and perform frequency domain embedding processing on the continuous parameters, mapping the continuous parameters to feature representations at multiple different frequency scales to enhance the representation ability of parameters with different dimensions, and obtain the continuous parameter encoding result.

[0095] In one embodiment, when the prediction network includes a geometric feature extraction unit and a physical quantity prediction unit, and the surface point cloud data and a unified conditional vector are input into the prediction network to perform condition-driven geometric feature learning and physical quantity prediction processing, and output the surface physical field of the vehicle, the multivariate conditional encoding module 504 is further used to: input the surface point cloud data into the geometric feature extraction unit for feature extraction processing to obtain the potential geometric features of the vehicle surface; and perform physical quantity prediction processing on the potential geometric features and the unified conditional vector through the physical quantity prediction unit to obtain surface pressure field data and wall shear stress field data.

[0096] In one embodiment, when performing the step of integrating the aerodynamic coefficient of a vehicle based on the surface physical field using aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vector, and area element to obtain the target drag coefficient, the aerodynamic coefficient integration module 506 is further configured to: determine the unit vector of the drag direction based on the yaw angle in the operating parameters; integrate each sampling unit on the vehicle surface along the drag direction based on the surface pressure field data, wall shear stress field data, surface normal vector, and area element to obtain the force situation of each sampling unit in the drag direction; and sum and normalize the force situation corresponding to each sampling unit to obtain the target drag coefficient.

[0097] In one embodiment, after obtaining the target drag coefficient, the aerodynamic coefficient integration module 506 is further configured to: construct a training sample set and obtain the true drag coefficient value corresponding to the training sample set, wherein the training sample set includes: three-dimensional surface geometric data, working condition parameter vector, surface pressure field data, wall shear stress field data and target drag coefficient; and perform model training on the prediction network based on the true drag coefficient value and the training sample set through a joint loss function to obtain an updated target prediction network.

[0098] The device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0099] This invention provides a server, specifically, the server includes a processor and a storage device; the storage device stores a computer program, which, when run by the processor, executes the method described in any of the above embodiments.

[0100] Figure 6 This is a schematic diagram of the structure of a server provided in an embodiment of the present invention. The server 100 includes: a processor 60, a memory 61, a bus 62, and a communication interface 63. The processor 60, the communication interface 63, and the memory 61 are connected through the bus 62. The processor 60 is used to execute executable modules, such as computer programs, stored in the memory 61.

[0101] The memory 61 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 63 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0102] Bus 62 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0103] The memory 61 is used to store programs. After receiving an execution instruction, the processor 60 executes the program. The method executed by the device for defining the flow process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor 60 or implemented by the processor 60.

[0104] Processor 60 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 60 or by instructions in software form. Processor 60 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 61. Processor 60 reads the information in memory 61 and, in conjunction with its hardware, completes the steps of the above method.

[0105] The computer program product of the readable storage medium provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.

[0106] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0107] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the drag coefficient of a vehicle based on multivariate conditional coding, characterized in that, The method includes: The three-dimensional surface geometry data of the vehicle and the various working condition parameter vectors of the aerodynamic performance of the vehicle's external flow field are obtained, and the surface point cloud data, surface normal vector and area element are determined based on the three-dimensional surface geometry data. Multivariate conditional encoding is performed on the vectors of the various working conditions to generate a unified conditional vector. The surface point cloud data and the unified conditional vector are then input into a prediction network for condition-driven geometric feature learning and physical quantity prediction to output the surface physical field of the vehicle. The surface physical field includes surface pressure field data and wall shear stress field data. The aerodynamic direction is determined based on the operating parameters, and the aerodynamic direction, surface pressure field data, wall shear stress field data, surface normal vector, and area element are used to perform integral output processing based on the surface physical field to obtain the target drag coefficient of the vehicle.

2. The method for predicting vehicle drag coefficient based on multivariate conditional coding according to claim 1, characterized in that, The step of determining surface point cloud data, surface normal vector, and area elements based on the three-dimensional surface geometry data includes: The surface point cloud data is obtained by performing surface point cloud extraction processing on the three-dimensional surface geometric data; The area values ​​of each triangular facet in the three-dimensional surface geometry data are determined as the area elements, and the normal vectors of the triangular facets are determined as the surface normal vectors.

3. The method for predicting vehicle drag coefficient based on multivariate conditional coding according to claim 1, characterized in that, The step of performing multivariate conditional encoding on each of the aforementioned working condition parameter vectors to generate a unified condition vector includes: For each of the aforementioned working condition parameter vectors, parameter identification processing is performed to obtain the angle-type parameters and continuous parameters corresponding to each of the aforementioned working condition parameter vectors. The angle-type parameters and the continuous parameters are then encoded to obtain the angle-type parameter encoding results and the continuous parameter encoding results, respectively. The angle-type parameter encoding result and the continuous-type parameter encoding result are concatenated to obtain a concatenated result. The concatenated result is then subjected to feature fusion processing through a conditional mapping network to generate the unified conditional vector.

4. The method for predicting vehicle drag coefficient based on multivariate conditional coding according to claim 3, characterized in that, The step of encoding the angle-type parameter and the continuous-type parameter respectively to obtain the angle-type parameter encoding result and the continuous-type parameter encoding result includes: The angle-type parameters are subjected to periodic sensing encoding processing, and the angle-type parameters are mapped to periodic continuous feature representations to eliminate the discontinuity of the angle boundary, thereby obtaining the encoding result of the angle-type parameters; The continuous parameters are subjected to frequency domain embedding processing, which maps the continuous parameters to feature representations at multiple different frequency scales to enhance the representation ability of parameters with different dimensions, thereby obtaining the encoding result of the continuous parameters.

5. The method for predicting vehicle drag coefficient based on multivariate conditional coding according to claim 1, characterized in that, The prediction network includes a geometric feature extraction unit and a physical quantity prediction unit. The step of inputting the surface point cloud data and the unified conditional vector into the prediction network, performing condition-driven geometric feature learning and physical quantity prediction processing, and outputting the surface physical field of the vehicle includes: The surface point cloud data is input into the geometric feature extraction unit for feature extraction processing to obtain the potential geometric features of the car surface. The physical quantity prediction unit performs physical quantity prediction processing on the potential geometric features and the unified condition vector to obtain the surface pressure field data and the wall shear stress field data.

6. The method for predicting vehicle drag coefficient based on multivariate conditional coding according to claim 1, characterized in that, The step of using the aerodynamic direction, the surface pressure field data, the wall shear stress field data, the surface normal vector, and the area element to perform integral output processing based on the surface physical field to obtain the target drag coefficient of the vehicle includes: Based on the yaw angle in the operating parameters, determine the unit vector of the resistance direction; Based on the surface pressure field data, the wall shear stress field data, the surface normal vector, and the area element, each sampling unit on the vehicle surface is integrated along the resistance direction to obtain the force situation of each sampling unit in the resistance direction. The target drag coefficient is obtained by summing and normalizing the force conditions corresponding to each sampling unit.

7. The method for predicting vehicle drag coefficient based on multivariate conditional coding according to claim 6, characterized in that, After the step of obtaining the target drag coefficient, the following is included: Construct a training sample set and obtain the true value of the drag coefficient corresponding to the training sample set, wherein the training sample set includes: the three-dimensional surface geometric data, the working condition parameter vector, the surface pressure field data, the wall shear stress field data, and the target drag coefficient; By using a joint loss function, the prediction network is trained based on the true value of the drag coefficient and the training sample set to obtain an updated target prediction network.

8. A vehicle drag coefficient prediction device based on multivariate conditional coding, characterized in that, The device includes: The data processing module acquires the three-dimensional curved surface geometry data of the vehicle and the various working condition parameter vectors of the aerodynamic performance of the vehicle's external flow field, and determines the surface point cloud data, surface normal vector, and area element based on the three-dimensional curved surface geometry data. The multivariate conditional encoding module performs multivariate conditional encoding on each of the working condition parameter vectors to generate a unified conditional vector. The surface point cloud data and the unified conditional vector are then input into the prediction network for condition-driven geometric feature learning and physical quantity prediction. The output is the surface physical field of the vehicle, wherein the surface physical field includes surface pressure field data and wall shear stress field data. The aerodynamic coefficient integration module determines the aerodynamic direction based on the operating parameters, and uses the aerodynamic direction, the surface pressure field data, the wall shear stress field data, the surface normal vector, and the area element to perform integration output processing based on the surface physical field to obtain the target drag coefficient of the vehicle.

9. A server, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 7.

Images