A method and system for comparing and judging the aerodynamic performance of automobile exteriors

Abstract

This invention proposes a method and system for comparing and judging the drag performance of vehicle shapes, relating to the field of automotive aerodynamics analysis technology. The method includes acquiring three-dimensional point cloud data of a first and a second vehicle shape; preprocessing the two sets of point cloud data to obtain standardized first and second input point clouds; inputting the first and second input point clouds into a parameter-shared point cloud feature extraction network; and dynamically constructing local neighborhood relationships of the point clouds and performing layer-by-layer feature aggregation to obtain a first and second global feature vectors. This invention transforms the vehicle drag assessment problem into a relative comparison paradigm that determines which shape has a lower drag coefficient. This allows the model to directly learn about the drag advantages and disadvantages between the shapes, avoiding the problems of high training difficulty and insufficient prediction stability caused by direct regression prediction of absolute drag values ​​due to the highly nonlinear characteristics of aerodynamics.

Description

Technical Field

[0001] This invention relates to the field of automotive aerodynamics analysis technology, and in particular to a method and system for comparing and judging the drag performance of automotive shapes. Background Technology

[0002] Automotive aerodynamic performance is a crucial factor affecting overall vehicle energy consumption, high-speed stability, and noise control. Among these factors, the drag coefficient is a core indicator for measuring aerodynamic performance. A lower drag coefficient effectively reduces air resistance during high-speed driving, thereby reducing energy consumption and improving overall vehicle performance. Therefore, aerodynamic optimization has become a critical step in determining styling choices during the automotive exterior design phase. With the widespread application of 3D modeling technology, automotive exteriors can be accurately represented using 3D point cloud data. Point cloud data, composed of numerous 3D spatial coordinate points, can directly describe the geometry of an object's surface, offering advantages such as high accuracy and simple structure. In recent years, deep learning technology has made significant progress in the field of 3D point cloud analysis, providing a new technical approach for data-driven rapid evaluation of automotive aerodynamics.

[0003] Current technologies suffer from the following problems: First, at the task definition level, most existing methods attempt to directly predict the absolute value of the drag coefficient from the vehicle's geometric data, i.e., establishing a regression mapping model from the geometric shape to the specific drag value. However, aerodynamic processes are highly nonlinear, making such regression models difficult to train, resulting in insufficient predictive stability, and requiring a large amount of high-quality wind tunnel or computational fluid dynamics simulation data for annotation support. In actual vehicle design, engineers often focus more on the relative advantages and disadvantages of aerodynamic performance between different design schemes, rather than accurately obtaining the specific drag coefficient value of each scheme. There is a clear mismatch between the existing regression prediction paradigm and this practical need. Second, at the feature learning level, existing deep learning methods, when extracting vehicle geometric features from point clouds, typically do not consider the differences in the degree of influence of different shape regions on aerodynamic performance. Automotive aerodynamic studies show that aerodynamically sensitive areas such as the front area, the transition area between the windshield and the roof, and the rear area have a significantly greater impact on the overall vehicle drag performance than other areas. If the importance of these key areas is not highlighted during feature learning, the model may struggle to accurately capture the influence of shape changes on drag performance.

[0004] The aforementioned problems make it difficult for existing technologies to simultaneously balance evaluation efficiency and judgment accuracy when faced with a large number of automotive exterior design schemes for rapid screening. This makes it impossible to quickly select the better design scheme for wind tunnel testing or simulation verification, thus affecting the efficiency of implementing automotive exterior design schemes from design to testing. Summary of the Invention

[0005] In view of this, the present invention proposes a method and system for comparing and judging the drag performance of automobile shapes. By constructing a point cloud feature extraction network with shared parameters of two branches, two automobile shapes are mapped to the same feature space for expression. After weighting and enhancing the features with an aerodynamic sensitive area attention mechanism, differential feature vectors are constructed to realize the comparison and judgment of drag performance between shapes. The problem of automobile drag evaluation is transformed into a relative comparison paradigm of judging which shape has a lower drag coefficient. This allows the model to learn directly around the drag advantage and disadvantage relationship between shapes, avoiding the problems of high training difficulty and insufficient prediction stability caused by direct regression prediction of absolute drag values ​​due to the highly nonlinear characteristics of aerodynamics. At the same time, the introduction of the aerodynamic sensitive area attention mechanism enables the model to adaptively highlight key geometric areas that have a significant impact on drag performance, such as the front, roof transition and rear of the car.

[0006] The technical solution of this invention is implemented as follows: In a first aspect, the present invention provides a method for comparing and judging the aerodynamic drag performance of automobiles, including: Three-dimensional point cloud data of the first and second vehicle outlines are acquired. The two sets of point cloud data are preprocessed to obtain the standardized first and second input point clouds. The first and second input point clouds are respectively input into a point cloud feature extraction network with shared parameters. By dynamically constructing local neighborhood relationships of point clouds and performing layer-by-layer feature aggregation, the first and second global feature vectors are obtained. The first global feature vector and the second global feature vector are respectively input into the aerodynamic sensitive area attention network. The features of each point are weighted by spatial attention weights to obtain the first enhanced feature vector and the second enhanced feature vector. Based on the first enhanced feature vector and the second enhanced feature vector, a differential feature vector is constructed to characterize the geometric differences between the two car shapes. The differential feature vector is input into the contrast discrimination network, and after nonlinear mapping, the contrast discrimination result is output. The contrast discrimination result is used to indicate the car shape with a lower drag coefficient between the first car shape and the second car shape.

[0007] In one embodiment, the preprocessing of the two sets of point cloud data to obtain a normalized first input point cloud and a second input point cloud includes the following steps: A preset number of points are randomly sampled from two sets of 3D point cloud data to make the number of points in the two sets of point clouds the same, forming a first intermediate point cloud and a second intermediate point cloud. The first and second intermediate point clouds are centered to shift the overall mean of the point clouds to the origin, resulting in the normalized first and second input point clouds.

[0008] In one embodiment, the step of constructing local neighborhood relationships of point clouds and performing layer-by-layer feature aggregation to obtain a first global feature vector and a second global feature vector includes the following steps: In each feature transformation layer of the point cloud feature extraction network, based on the feature vector of each point in the current layer, the K-nearest neighbor algorithm is used to select a preset number of points in the feature space that are closest to each point to form a local neighborhood set of that point. For each point's local neighborhood set, construct the edge features between the point and its nearest neighbors, transform each edge feature using a shared convolutional mapping function, and apply max pooling to the transformed edge features to obtain the updated feature vector of the point. The stacked multi-layer local neighborhood construction and feature aggregation process concatenates the point feature vectors output by each layer along the feature dimension to obtain multi-scale point features; Perform global max pooling operation along the point dimension on the multi-scale point features to obtain the global feature vector corresponding to the car shape.

[0009] In one embodiment, constructing the edge features between the point and its nearest neighbors includes the following steps: For the center point and any of its nearest neighbors, calculate the difference between the feature vector of the nearest neighbor and the feature vector of the center point; The difference is concatenated with the feature vector of the center point itself to obtain the edge features between the center point and its nearest neighbors.

[0010] In one embodiment, the step of weighting the features of each point using spatial attention weights to obtain a first enhanced feature vector and a second enhanced feature vector includes the following steps: Using the aerodynamic sensitivity mapping function, an aerodynamic sensitivity score is generated for each point based on the feature vector of each point. The aerodynamic sensitivity scores of each point are mapped to spatial attention weights by a normalized exponential function, so that the sum of the spatial attention weights of all points is 1. The enhanced feature vector is obtained by multiplying the feature vector of each point with the corresponding spatial attention weight and then summing the results.

[0011] In one embodiment, after multiplying the feature vectors of each point with their corresponding spatial attention weights and summing the results to obtain the enhanced feature vectors, the following steps are further included: The enhanced feature vector is input into the attention mapping network, which is composed of a first weight matrix, a ReLU activation function, a second weight matrix, and a Sigmoid function in sequence. The attention weight vector of the feature dimension is calculated through the attention mapping network. The enhanced feature vector is multiplied element-wise with the feature dimension attention weight vector to obtain the target enhanced feature vector, and the target enhanced feature vector is used as the first enhanced feature vector or the second enhanced feature vector.

[0012] In one embodiment, the step of outputting the comparison and discrimination result after nonlinear mapping includes the following steps: The difference feature vector is input into a nonlinear mapping network consisting of at least one fully connected layer and an activation function to obtain the discriminant scalar value; The discriminant scalar value is mapped to a probability value using the Sigmoid function; When the probability value is greater than a preset threshold, the output indicates that the drag coefficient of the second car shape is less than that of the first car shape; otherwise, the output indicates that the drag coefficient of the first car shape is less than that of the second car shape.

[0013] Secondly, the present invention provides a vehicle exterior drag performance comparison and discrimination system for implementing the above-mentioned method, comprising: The preprocessing module is used to acquire three-dimensional point cloud data of the first and second vehicle shapes, and preprocess the two sets of point cloud data to obtain the standardized first and second input point clouds. The point cloud feature extraction module is used to input the first input point cloud and the second input point cloud into the parameter-shared point cloud feature extraction network, and obtain the first global feature vector and the second global feature vector by dynamically constructing the local neighborhood relationship of the point cloud and performing layer-by-layer feature aggregation. The aerodynamically sensitive region attention module is used to input the first global feature vector and the second global feature vector into the aerodynamically sensitive region attention network, respectively, and to weight the features of each point through spatial attention weights to obtain the first enhanced feature vector and the second enhanced feature vector. The differential feature construction module is used to construct differential feature vectors that characterize the geometric differences between two car shapes based on the first enhanced feature vector and the second enhanced feature vector. The comparison and discrimination module is used to input the differential feature vector into the comparison and discrimination network, and output the comparison and discrimination result after nonlinear mapping. The comparison and discrimination result is used to indicate the car shape with the lower drag coefficient between the first car shape and the second car shape.

[0014] Thirdly, the present invention provides an electronic device, including a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the above-mentioned method for comparing and judging the aerodynamic performance of automobile exterior shapes.

[0015] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method for comparing and judging the aerodynamic performance of automobile exteriors.

[0016] The method and system for comparing and judging the aerodynamic drag performance of automobiles in this invention have the following advantages over the prior art: 1. By constructing a point cloud feature extraction network with shared parameters in two branches, two car shapes are mapped to the same feature space for representation. After the features are weighted and enhanced by the aerodynamic sensitive region attention mechanism, differential feature vectors are constructed to compare and distinguish the drag performance between the shapes. The problem of car drag evaluation is transformed into a relative comparison paradigm to determine which of the two shapes has a lower drag coefficient. This allows the model to learn directly around the drag advantage and disadvantage relationship between the shapes, avoiding the problems of high training difficulty and insufficient prediction stability caused by direct regression prediction of absolute drag value due to the highly nonlinear characteristics of aerodynamics. At the same time, the introduction of the aerodynamic sensitive region attention mechanism enables the model to adaptively highlight key geometric areas that have a significant impact on drag performance, such as the front, roof transition and rear of the car. 2. By dynamically constructing local neighborhood relationships of points based on the current feature space in each feature transformation layer, and constructing edge features containing relative displacement and the center point's own features within the neighborhood, and extracting point cloud features layer by layer after convolutional mapping and max pooling aggregation, the network can adaptively capture local geometric structure changes in the car's shape that are sensitive to aerodynamic performance. At the same time, by stitching together multiple layers of features, a multi-scale geometric feature expression that takes into account both local details and overall contours is obtained, thereby improving the fine-grained representation ability of shape differences. 3. In the attention modeling of aerodynamic sensitive areas, an aerodynamic sensitivity mapping function composed of fully connected layers and activation functions is used to score the aerodynamic sensitivity of each point end-to-end. A normalized exponential function is used to generate spatial attention weights to weight and aggregate the features of each point. This mechanism does not rely on predefined physical prior formulas or manually labeled sensitive area templates. It can automatically learn the degree of differential influence of each shape region on aerodynamic performance during training, so that the geometric features of key areas such as the front, roof transition and rear of the car can obtain higher weights in the global expression, thereby effectively enhancing the accuracy and robustness of comparison and discrimination. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a schematic diagram illustrating the steps of the method for comparing and judging the aerodynamic drag performance of automobiles according to the present invention; Figure 2 This is a schematic diagram illustrating the sample binning matching principle of the vehicle exterior drag performance comparison and discrimination method of the present invention; Figure 3 This is a structural diagram of the point cloud feature extraction network in the vehicle shape drag performance comparison and discrimination method of the present invention; Figure 4 This is a schematic diagram of the aerodynamic sensitive area attention network in the vehicle shape drag performance comparison and discrimination method of the present invention. Detailed Implementation

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

[0020] like Figure 1-4 As shown, the vehicle shape drag performance comparison and discrimination method of the present invention uses paired vehicle shape point cloud data as input. By constructing a parameter-shared dual-branch deep neural network structure, features are extracted from the two sets of vehicle shape point clouds respectively, and a difference comparison relationship is established in the feature space, thereby outputting the discrimination result of the drag coefficient relationship between corresponding vehicle shapes. Unlike traditional drag coefficient regression prediction methods, the present invention transforms the vehicle aerodynamic performance evaluation problem into a relative comparison problem between shape schemes, thereby reducing the difficulty of model learning and improving the efficiency of rapid shape scheme selection during the design phase.

[0021] This invention comprises five main steps: point cloud data acquisition and preprocessing, sample binning and matching, feature extraction, attention feature fusion of aerodynamically sensitive areas, and wind resistance comparison and discrimination. These steps are interconnected and together constitute a complete comparative analysis process for the aerodynamic performance of automotive exteriors.

[0022] Overall, the method of this invention is implemented step by step according to the process of "point cloud acquisition and preprocessing - sample binning and matching - feature extraction - aerodynamic sensitive area attention feature fusion - drag comparison and discrimination". Starting from the original car shape point cloud input, geometric structure features are extracted layer by layer through deep neural networks, and the difference expression between different shapes is established in the feature space. Finally, the discrimination result of the drag coefficient relationship between the two cars is output, thus forming a complete end-to-end car aerodynamic performance comparison and analysis framework.

[0023] During model training, this invention utilizes the DriverNet vehicle shape dataset as a training sample source. This dataset contains approximately 8,000 vehicle shape samples, covering three typical vehicle structure types: sedans, hatchbacks, and SUVs. Each shape sample corresponds to a drag coefficient annotation obtained through aerodynamic simulation calculations, thus providing a reliable data foundation for the model to learn the relationship between vehicle shape geometry and aerodynamic performance. By pairwise combining different shape samples from this dataset, a large number of shape comparison samples can be constructed for training, effectively improving the model's learning ability and generalization performance.

[0024] Furthermore, this invention introduces an aerodynamically sensitive region attention mechanism during feature learning to address the issue of varying degrees of influence of different shaped regions on aerodynamic performance. In automotive aerodynamics, the front end, windshield transition area, roof area, and rear end are generally considered important aerodynamically sensitive regions affecting drag performance. Geometric changes in these regions often significantly alter airflow separation positions, boundary layer development processes, and wake structure characteristics, thus significantly impacting the overall vehicle drag coefficient. Therefore, by assigning different weights to features in different regions during feature learning through an attention mechanism, the model can focus more on key structural regions that have a greater impact on aerodynamic performance, thereby improving the accuracy of drag comparison and discrimination.

[0025] To provide a more detailed explanation, it is specifically divided into steps S100-S500.

[0026] Step S100: Obtain the three-dimensional point cloud data of the first and second vehicle outlines, and preprocess the two sets of point cloud data to obtain the standardized first and second input point clouds.

[0027] In this step, the first and second vehicle shapes represent two different vehicle design schemes for which aerodynamic performance will be compared. The 3D point cloud data for both shapes can be acquired using 3D scanning equipment or obtained from an existing 3D model database.

[0028] After acquiring the raw point cloud data, a uniform preprocessing operation is performed on each group of point clouds to ensure that the two groups of point clouds are consistent in terms of data scale and spatial distribution, thus providing standardized input data for subsequent feature extraction.

[0029] Step S200: Input the first input point cloud and the second input point cloud into the parameter-shared point cloud feature extraction network respectively. By dynamically constructing the local neighborhood relationship of the point cloud and performing layer-by-layer feature aggregation, the first global feature vector and the second global feature vector are obtained.

[0030] The point cloud feature extraction network adopts a dual-branch structure, with both branches having identical network structures and completely shared network parameters. The first input point cloud is fed into the first branch, and the second input point cloud is fed into the second branch; the two branches process in parallel. In each feature transformation layer, the network dynamically constructs local neighborhood relationships for each point in the feature space based on the feature vectors of each point in the current layer, enabling the neighborhood connectivity to adaptively adjust as the feature representation is updated. Based on this, convolutional mapping and aggregation operations are performed on the local geometric information within the neighborhood, extracting multi-scale geometric features from local details to the overall contour layer by layer.

[0031] Through this parameter-sharing dual-branch structure, the first and second vehicle shapes are mapped to the same feature space for representation, providing a consistent feature basis for subsequent difference comparison. Finally, each branch outputs a fixed-dimensional global feature vector corresponding to its respective input vehicle shape: the first global feature vector and the second global feature vector.

[0032] Step S300: Input the first global feature vector and the second global feature vector into the aerodynamic sensitive area attention network respectively, and weight the features of each point through spatial attention weights to obtain the first enhanced feature vector and the second enhanced feature vector.

[0033] The aerodynamically sensitive region attention network also employs a parameter-sharing dual-branch structure to process intermediate point-level features involved in the two global feature vectors. This attention network generates corresponding aerodynamic sensitivity scores for each point in the point cloud and calculates the spatial attention weights for each point based on these scores, enabling the model to adaptively assess the potential impact of different shaped regions on aerodynamic performance.

[0034] After obtaining the spatial attention weights of each point, the point-level features are weighted and aggregated to give features from aerodynamically sensitive areas such as the front, roof transition and rear of the vehicle a higher proportion in the global representation, thereby strengthening the influence of key structures on the overall feature representation and outputting the enhanced first and second enhanced feature vectors.

[0035] Step S400: Based on the first enhanced feature vector and the second enhanced feature vector, construct a differential feature vector to characterize the geometric differences between the two car shapes.

[0036] The two enhanced feature vectors output in step S300 are subjected to a difference operation in the feature space, and the dimension-wise difference between the first and second enhanced feature vectors is used as the difference feature vector. This difference feature vector describes the direction and magnitude of change in the geometric structural features of the second vehicle shape relative to the first vehicle shape, where each feature dimension reflects structural differences at different geometric semantic levels.

[0037] This differential modeling approach transforms the complex aerodynamic performance differences between two car shapes into a difference representation in the feature space, enabling the subsequent discrimination process to directly focus on the correlation between shape changes and drag performance changes.

[0038] To enhance the ability of differential features to express the complex nonlinear relationship between changes in vehicle shape geometry and changes in aerodynamic performance, after obtaining the differential feature vector d, a nonlinear transformation function can be further introduced to remap the differential features.

[0039] Specifically, the difference feature vector d obtained in step S400 is input into the nonlinear transformation function φ(·), which is implemented by a mapping network consisting of fully connected layers and activation functions:

[0040] Where d' is the difference feature vector after remapping. This nonlinear mapping network can further explore the complex nonlinear effects of vehicle shape geometry changes on aerodynamic performance in a high-dimensional feature space, enabling the fused features to more fully express the potential correlation between shape structure changes and drag performance changes.

[0041] After the nonlinear remapping in this step, the remapped difference feature vector d' is used as the input to the contrast discrimination network in step S500 for subsequent contrast discrimination output.

[0042] Step S500: Input the differential feature vector into the comparison discrimination network, and output the comparison discrimination result after nonlinear mapping. The comparison discrimination result is used to indicate the car shape with a lower drag coefficient between the first car shape and the second car shape.

[0043] The differential feature vector output by the comparison and discrimination network in step S400 is used as input. The shape difference information contained in the differential feature is gradually analyzed and mapped through multi-layer nonlinear transformation. The discrimination features related to the wind resistance performance are extracted, and the comparison and discrimination results are finally output.

[0044] The comparison result directly indicates which of the two car shapes has a lower drag coefficient, without needing to obtain the specific drag coefficient values ​​of each of the two car shapes. This achieves end-to-end processing from paired shape input to relative superiority / inferiority judgment.

[0045] The above steps S100-S500 together constitute the main process of this invention. This process transforms the traditional mode of evaluating the aerodynamic performance of automobiles, which relies on wind tunnel experiments or computational fluid dynamics simulations to obtain the absolute value of the drag coefficient of a single shape, into a mode of relative discrimination of the drag performance of pairs of shapes based on deep learning. It can quickly and accurately determine which of two automobile shape schemes has better drag performance without obtaining specific drag coefficient values.

[0046] In some embodiments, the preprocessing of the two sets of point cloud data in step S100 to obtain the normalized first input point cloud and second input point cloud can be achieved through sub-steps S101-S102.

[0047] Step S101: Randomly sample a preset number of points from the two sets of 3D point cloud data to make the number of points in the two sets of point clouds the same, forming a first intermediate point cloud and a second intermediate point cloud.

[0048] The original 3D point cloud data for the first and second vehicle outlines may originate from different acquisition devices or modeling tools, resulting in variations in point cloud size. To standardize the input data format, this step involves randomly sampling each group of point clouds, uniformly selecting a predetermined number of points from the original point clouds to form the first and second intermediate point clouds.

[0049] The number of points in the sampled point cloud remains consistent. The random sampling method can preserve the overall geometric distribution characteristics of the original point cloud to a certain extent, avoiding adverse effects on subsequent feature extraction due to differences in point cloud density.

[0050] Step S102: Center the first intermediate point cloud and the second intermediate point cloud to shift the overall mean of the point cloud to the origin of the coordinate system, and obtain the normalized first input point cloud and the second input point cloud.

[0051] For the first intermediate point cloud sampled in step S101, the average three-dimensional coordinates of all points in the point cloud are first calculated to obtain the overall mean coordinates of the point cloud. Then, the mean coordinates are subtracted from the coordinates of each point in the point cloud to translate the point cloud as a whole to the origin, resulting in the normalized first input point cloud. The same centering process is performed on the second intermediate point cloud to obtain the normalized second input point cloud.

[0052] Through the above-mentioned centralization process, the overall mean of the two sets of input point clouds is shifted to the origin of the coordinate system, which can eliminate the interference caused by the difference in absolute spatial position between different car shape models. This allows the subsequent feature extraction network to focus more on the geometric structural features of the car shape itself, rather than its absolute position in space, thereby improving the consistency of feature expression and the accuracy of comparison and discrimination.

[0053] In one specific embodiment, the model training phase also includes a sample binning and matching step to improve the model's learning performance on samples with different wind resistance levels.

[0054] Specifically, based on the range of drag coefficient values ​​corresponding to each car shape sample in the dataset, the samples are divided into multiple drag interval buckets. When constructing training sample pairs, competitor samples are generated using the bucket containing each car sample as a basis, combining three strategies: nearest neighbor sampling, cross-bucket sampling, and random sampling. Nearest neighbor sampling is used to construct comparative sample pairs between samples with similar drag values ​​to enhance the model's ability to discriminate subtle shape differences; cross-bucket sampling is used to construct sample pairs between samples with significant drag differences to highlight substantial performance differences; and random sampling is used to supplement sample diversity and improve the model's generalization ability.

[0055] By constructing multiple counterpart samples with different wind resistance levels for each sample, a large-scale and reasonably distributed set of sample pairs is obtained during training, thereby significantly improving the model's learning effect on wind resistance comparison patterns.

[0056] In some embodiments, the step S200 of constructing the local neighborhood relationship of the point cloud and performing layer-by-layer feature aggregation to obtain the first global feature vector and the second global feature vector can be implemented through the following sub-steps S201-S204.

[0057] Step S201: In each feature transformation layer of the point cloud feature extraction network, based on the feature vector of each point in the current layer, the K-nearest neighbor algorithm is used to select a preset number of points in the feature space that are closest to each point to form a local neighborhood set of that point.

[0058] Let the input car shape point cloud after preprocessing in step S100 be:

[0059] in Indicates the first The three-dimensional coordinates of the points To ensure a consistent number of sampled points, typical values ​​such as 1024 or 2048 can be used. Initially, the 3D coordinates of each point are used as input features for layer 0 of the network, denoted as:

[0060]

[0061] During feature extraction, for the first... Feature vector of any point in the layer ,pass The nearest neighbor search algorithm selects the feature space with the closest Euclidean distance. Each point forms its neighborhood set:

[0062] Since the neighborhood search is performed again after each layer of feature updates, the connections between points change dynamically with the feature representation. In shallow networks, the connections between points mainly reflect spatial proximity; as the network deepens and features become more abstract, points with similar semantic structures may be aggregated into the same neighborhood even if they are spatially far apart. This dynamic mapping mechanism enables the network to capture the potential geometric and semantic correlations in the point cloud layer by layer, which is beneficial for characterizing the key aerodynamically sensitive structural regions in the shape of a car.

[0063] Step S202: For the local neighborhood set of each point, construct the edge features between the point and each of its nearest neighbors, transform each edge feature through a shared convolutional mapping function, and apply max pooling operation to the transformed edge features to obtain the updated feature vector of the point.

[0064] For any neighborhood point The corresponding edge feature vector is constructed as follows:

[0065] By splicing the relative displacement vector with the center point's own features, the aforementioned edge features enable the network to perceive the changing trends of local geometry while retaining the semantic information of the center point within the overall structure, thus more comprehensively expressing the geometric features of local regions of the point cloud.

[0066] Subsequently, the edge features are input into a shared nonlinear mapping function for convolutional transformation to extract high-dimensional edge feature representations, specifically:

[0067] in, Typically, it consists of a one-dimensional convolution operator, a batch normalization layer, and an activation function, with parameters shared across all edge features within the same layer. Through this mapping process, the local geometric relationships between different point pairs can be uniformly modeled, enabling the network to learn local structural patterns that are discriminative in terms of aerodynamic performance.

[0068] To effectively aggregate information within the neighborhood to the center point, max pooling is used to symmetrically aggregate the mapping results of each point within its neighborhood, resulting in the updated feature representation of that point in the next layer:

[0069] in, Indicates the first Midpoint of layer Its neighboring points The corresponding edge feature mapping result; the max pooling operation is to take the maximum value of each edge feature vector in the neighborhood according to the feature dimension.

[0070] Through the above aggregation method, the features of each point can integrate the most significant local responses in its neighborhood, while ensuring the invariance of the order of neighborhood points, satisfying the requirement of the disorder of point cloud data itself, and enhancing the robustness of feature representation.

[0071] Step S203: Stacking multi-layer local neighborhood construction and feature aggregation process, concatenating the point feature vectors output by each layer in the feature dimension to obtain multi-scale point features.

[0072] By repeatedly executing the process of "dynamic nearest neighbor construction - edge feature construction - convolutional mapping - neighborhood aggregation" in the above steps, multiple layers of dynamic graph convolutional structures are stacked in the network to obtain feature representations of point clouds at different receptive field scales layer by layer, forming a multi-layer feature set:

[0073] in, Indicates the first The point feature matrix output by the layer. This indicates the number of points sampled in the point cloud. This indicates the feature dimension of the corresponding layer. Shallower features focus more on depicting the local geometric details of the car's exterior surface, such as curvature changes and edge transitions, while deeper features gradually express a larger range of structural contours and overall morphological information, such as the macroscopic geometric distribution of the front, roof, and rear areas, thus achieving hierarchical feature modeling from local to global.

[0074] To fully integrate the complementary information contained in features from different levels, the multi-level output features are concatenated along the feature dimension to construct a multi-scale fused feature representation:

[0075] Concat represents the splicing operation along the feature dimension. This multi-scale feature fusion method enables the network to retain both local detail information and overall contour information in the same feature representation, which is beneficial for subsequent modeling of the comprehensive aerodynamic characteristics of the car's shape.

[0076] Step S204: Perform global max pooling operation on the multi-scale point features along the point dimension to obtain the global feature vector corresponding to the car shape.

[0077] After obtaining multi-scale point features, a global symmetric pooling operation is applied to the point dimensions to aggregate the point cloud-level features into a fixed-dimensional global feature vector, which is used to characterize the overall geometric structure of the vehicle's exterior.

[0078] Where Pool represents the global max pooling operation. Represents the multi-scale fused feature matrix. The Middle The feature vectors corresponding to each point; the global max pooling operation is to take the maximum value along the point dimension for each feature dimension.

[0079] In the dual-branch network structure, the first input point cloud P1 and the second input point cloud P2 are mapped using the feature extraction process with completely shared parameters, respectively, to obtain the first global feature vector f1 and the second global feature vector f2. By sharing parameters, it is ensured that samples with different shapes are represented in the same feature space, so that the subsequent comparison process based on feature difference has consistent physical semantics and good comparability, providing a reliable feature basis for wind resistance coefficient comparison and discrimination.

[0080] In a specific embodiment, the construction of the edge features between the point and its nearest neighbors in step S202 can be further subdivided into S2021-S2022.

[0081] Step S2021: For the center point and any of its nearest neighbors, calculate the difference between the feature vector of the nearest neighbor and the feature vector of the center point.

[0082] The current feature transformation layer is the th Layer, the feature vector of the center point i in this layer is The eigenvector of its nearest neighbor j is , where j∈ The difference between the two eigenvectors is calculated, and the resulting difference vector reflects the relative positional relationship of the nearest neighbor points with respect to the center point in the feature space, thus characterizing the geometric change trend between points in the local neighborhood.

[0083] In shallow networks, since the features mainly encode the spatial coordinate information of the points, the difference vector directly describes the spatial offset direction and amount of the neighboring points relative to the center point. In deep networks, as the features are gradually abstracted into high-level semantic expressions, the difference vector reflects the differences in semantic features of the neighboring points, thereby adaptively capturing local structural change patterns that have distinguishing significance for aerodynamic performance.

[0084] Step S2022: Concatenate the difference with the feature vector of the center point itself to obtain the edge features between the center point and its nearest neighbors.

[0085] The difference vector calculated in step S321 is concatenated with the feature vector of the center point itself to construct the edge feature vector between the center point and its nearest neighbor. This edge feature vector contains two parts of information: first, the relative difference between the nearest neighbor and the center point, used to express the changes in local geometry; and second, the feature information of the center point itself, used to preserve the semantic location of the center point in the overall structure. Through this concatenation construction method, subsequent convolutional mapping operations can perceive the geometric change patterns in the local neighborhood without losing the semantic information of the center point itself in the global structure. This allows for a more comprehensive expression of the geometric features of the local region of the point cloud, which is beneficial for the network to learn local structural patterns that can distinguish the aerodynamic performance of the vehicle's shape.

[0086] In some embodiments, since the first enhanced feature vector and the second enhanced feature vector undergo the same processing through a parameter-shared attention network in step S300, the following description will take the point-level feature processing process corresponding to the first global feature vector as an example.

[0087] In automotive aerodynamics research, numerous wind tunnel tests and numerical simulations have shown that different body shapes contribute significantly to the overall aerodynamic performance of a vehicle. Specifically, the front region primarily influences the initial flow path and separation point of the incoming airflow; the transition area between the windshield and roof determines airflow attachment and boundary layer development; and the rear region directly affects the wake structure and pressure recovery process. These regions are often referred to as aerodynamically sensitive areas. If the importance of these key regions is not highlighted during feature learning, the model may struggle to accurately capture the impact of body shape changes on aerodynamic performance.

[0088] To address this issue, the aerodynamic sensitivity of each point is first estimated at the point cloud level, and spatial attention weights are generated accordingly to achieve adaptive enhancement of key aerodynamic features. This attention weighting process includes sub-steps S301-S303.

[0089] Step S301: Using the aerodynamic sensitivity mapping function, generate an aerodynamic sensitivity score for each point based on the feature vector of each point.

[0090] Let the point-level feature representation of the car shape point cloud obtained after passing through the feature extraction module be:

[0091] Indicates the first High-dimensional feature vectors of points This indicates the number of points contained in the point cloud. Indicates the feature dimension.

[0092] To characterize the potential impact of different points on aerodynamics, this invention first constructs an aerodynamic sensitivity estimation function:

[0093] in, This represents the aerodynamic sensitivity mapping function. Indicates the first The aerodynamic sensitivity score for each point reflects its potential importance within the aerodynamic structure of the vehicle's exterior. The aerodynamic sensitivity mapping function automatically learns the influence of each point on drag performance through an end-to-end training process, without relying on predefined physical prior formulas or manually labeled sensitive area templates.

[0094] Step S302: Map the aerodynamic sensitivity scores of each point to spatial attention weights using a normalized exponential function, so that the sum of the spatial attention weights of all points is 1.

[0095] By mapping the sensitivity scores using a normalization function, the spatial attention weights for the corresponding points are obtained:

[0096] in Indicates the first The spatial attention weights of each point are such that the sum of the weights of all points is 1. This normalization process allows the network to automatically allocate more weights to points corresponding to aerodynamically sensitive areas during the learning process. In automotive aerodynamics, the front area, the transition area between the windshield and the roof, and the rear area are generally considered important aerodynamically sensitive areas affecting drag performance. Changes in the geometry of these areas often significantly alter the airflow separation location, boundary layer development process, and wake structure characteristics.

[0097] Through this step of learning attention weights, the model can automatically discover and highlight the importance of these key regions during training.

[0098] Step S303: Multiply the feature vectors of each point by the corresponding spatial attention weights and sum them to obtain the enhanced feature vectors.

[0099] After obtaining the point-level attention weights, a weighted aggregation operation is used to construct an attention-enhanced global feature representation:

[0100] This weighted aggregation process allows features from aerodynamically sensitive regions to occupy a higher proportion in the global representation, thereby strengthening the influence of key structures such as the front, roof transition, and rear of the vehicle on the overall feature representation. In the dual-branch structure, the first and second global feature vectors are processed through the above steps to obtain the corresponding first and second enhanced feature vectors.

[0101] As an alternative or supplementary implementation of the point-level spatial attention mechanism in steps S301-S303, this embodiment can also use a feature-dimensional attention mechanism to enhance the global feature vector.

[0102] After obtaining the global feature vector f, the attention weights for each feature dimension are calculated using an attention mapping function. First, the global feature vector is linearly transformed using the first weight matrix, and then non-linearity is introduced through the ReLU activation function to obtain an intermediate representation. Subsequently, the intermediate representation is linearly transformed using the second weight matrix, and the output is mapped to the (0,1) interval using the Sigmoid function to obtain the attention weights for the corresponding feature dimensions.

[0103]

[0104] in and This represents the learnable weight matrix. Represents the ReLU activation function. This represents the Sigmoid function. This represents the attention weight for the corresponding feature dimension.

[0105] Through the above mapping process, the network can automatically learn the importance weights of each feature dimension based on the training samples, thereby highlighting the geometric structure information related to the aerodynamically sensitive region in the feature representation. Subsequently, the original features are weighted and enhanced through element-wise multiplication operations to obtain the attention-enhanced feature representation:

[0106]

[0107] in, This indicates element-wise multiplication.

[0108] Through this attention weighting process, the model can enhance its ability to express key aerodynamic structural features in the feature space, so that geometric changes that have a significant impact on aerodynamic performance, such as changes in front curvature, roof transition shape, and rear truncated structure, can obtain higher weights in the feature representation, thereby improving the accuracy of subsequent comparison and discrimination.

[0109] In some embodiments, the aerodynamic sensitivity mapping function It consists of at least one fully connected layer and an activation function. Specifically, the mapping function uses the feature vector x at each point. i As input, the input features are linearly transformed through a fully connected layer, and then nonlinear mapping capability is introduced through an activation function to output the aerodynamic sensitivity score s corresponding to that point. i .

[0110] When multiple fully connected layers are stacked, activation functions are set between each fully connected layer, thus forming a multilayer perceptron structure. This gives the mapping function a stronger nonlinear expressive power, enabling it to learn more complex mapping relationships between point-level features and aerodynamic sensitivity. The activation function can be the ReLU function or other commonly used nonlinear activation functions.

[0111] Aerodynamic sensitivity score s output by the aerodynamic sensitivity mapping function i The score reflects the potential importance of a point within the aerodynamic structure of a vehicle. A higher score indicates a greater potential impact of that point's location on the vehicle's aerodynamic performance; a lower score indicates a relatively smaller impact.

[0112] In automotive aerodynamics, the front area determines the initial separation position of the incoming airflow and the airflow path. The transition area between the windshield and the roof affects the airflow adhesion state and the boundary layer development process. The rear area directly affects the wake structure and pressure recovery process. Points in these aerodynamically sensitive areas are usually assigned a higher aerodynamic sensitivity score.

[0113] By performing end-to-end optimization of network parameters during training, the aerodynamic sensitivity mapping function can automatically learn and identify the aforementioned key aerodynamic regions without relying on predefined physical prior formulas or manually labeled sensitive region templates, thereby achieving adaptive aerodynamic sensitivity assessment for different car shape structures.

[0114] In some embodiments, the comparison and discrimination result output after nonlinear mapping in step S500 is implemented through sub-steps S501-S503.

[0115] Step S501: Input the difference feature vector into a nonlinear mapping network consisting of at least one fully connected layer and an activation function to obtain the discrimination scalar value.

[0116] The differential feature vector d output from step S400 is received as input. This differential feature vector characterizes the direction and magnitude of the geometric change of the second vehicle's shape relative to the first vehicle's shape in the feature space. The differential feature vector is then input into a nonlinear mapping network. In this network, the network consists of at least one fully connected layer and an activation function:

[0117]

[0118] in This represents a mapping function consisting of at least one layer of linear transformation, activation function, and regularization structure. The output is a real number, used to characterize the trend of wind resistance difference between the two shapes. When multiple fully connected layers are used, activation functions are set between each fully connected layer, which makes the network have stronger nonlinear expression capabilities and can fully explore the complex nonlinear effects of the geometric changes of the car shape on aerodynamic performance.

[0119] Step S502: Map the discrimination scalar value to a probability value using the Sigmoid function.

[0120] To transform the above discrimination scalar value into a probabilistic form with physical meaning, a sigmoid function is applied to the output to obtain the contrast discrimination probability value:

[0121] in This indicates the probability that the drag coefficient of the second car shape is less than that of the first car shape. When When the value is close to 1, it indicates that the drag coefficient of the second car's shape is very likely to be lower than that of the first car's shape; when... When it is close to 0, it means that the drag coefficient of the first car's shape is very likely to be lower than that of the second car's shape; when When the value is close to 0.5, it indicates that the difference in drag performance between the two car shapes is small.

[0122] Step S503: When the probability value is greater than the preset threshold, output the comparison judgment result indicating that the drag coefficient of the second car shape is less than that of the first car shape; otherwise, output the comparison judgment result indicating that the drag coefficient of the first car shape is less than that of the second car shape.

[0123] The obtained probability value Compare with a preset threshold. The preset threshold can be set to 0.5. When the value is greater than 0.5, the aerodynamic performance of the second vehicle shape is determined to be better than that of the first vehicle shape, and the output indicates that the drag coefficient of the second vehicle shape is less than that of the first vehicle shape; when... When the value is less than or equal to 0.5, the aerodynamic performance of the first vehicle shape is determined to be better than that of the second vehicle shape, and the comparison result indicating that the drag coefficient of the first vehicle shape is less than that of the second vehicle shape is output.

[0124] In addition, during the training phase, the actual drag coefficient relationship between sample pairs is used as a supervision signal. Let the true label be:

[0125]

[0126] in and These represent the actual drag coefficient values ​​corresponding to the first and second car exteriors, respectively. This scheme uses a binary cross-entropy loss function to measure the difference between the predicted probability and the true label:

[0127]

[0128] By backpropagating the aforementioned loss function, the network parameters of the feature extraction module, feature fusion module, and comparison and discrimination module are jointly optimized, enabling the entire network to gradually learn the mapping relationship between geometric differences in vehicle shape and wind resistance. After training convergence, the model can directly output the corresponding comparison and discrimination results for any two unknown vehicle shape point clouds as input during the inference phase.

[0129] The automotive aerodynamic drag performance comparison and discrimination system of the present invention is used to implement the above-mentioned method, including a preprocessing module, a point cloud feature extraction module, an aerodynamic sensitive area attention module, a differential feature construction module, and a comparison and discrimination module.

[0130] The preprocessing module is used to acquire three-dimensional point cloud data of the first and second vehicle shapes, and preprocesses the two sets of point cloud data to obtain the standardized first and second input point clouds.

[0131] The point cloud feature extraction module is used to input the first input point cloud and the second input point cloud into the parameter-shared point cloud feature extraction network, respectively. By dynamically constructing the local neighborhood relationship of the point cloud and performing layer-by-layer feature aggregation, the first global feature vector and the second global feature vector are obtained.

[0132] The aerodynamically sensitive region attention module is used to input the first global feature vector and the second global feature vector into the aerodynamically sensitive region attention network, respectively, and to weight the features of each point through spatial attention weights to obtain the first enhanced feature vector and the second enhanced feature vector.

[0133] The differential feature construction module is used to construct differential feature vectors that characterize the geometric differences between two car shapes based on the first enhanced feature vector and the second enhanced feature vector.

[0134] The comparison and discrimination module is used to input the differential feature vector into the comparison and discrimination network, and output the comparison and discrimination result after nonlinear mapping. The comparison and discrimination result is used to indicate the car shape with a lower drag coefficient between the first car shape and the second car shape.

[0135] In one specific embodiment, the 3D point cloud data of the vehicle's exterior is input into a preprocessing module. Random sampling ensures that the number of point clouds in each sample is consistent, and centralized processing is used to eliminate translational differences in the spatial location of the point clouds, thereby achieving data standardization. Subsequently, during the model training phase, representative pairs of vehicle exterior samples are constructed through sample binning and matching to enhance the model's ability to learn from samples with different wind resistance differences.

[0136] After constructing the sample pairs, the processed point cloud data is fed into the point cloud feature extraction module. This module employs a parameter-sharing dual-branch neural network structure to model the geometric features of the two car shape samples, thereby extracting the corresponding high-dimensional feature vectors f1 and f2. Because the two branches share network parameters, the extracted features reside in the same feature space, exhibiting good comparability.

[0137] Subsequently, these two global feature vectors are input into the aerodynamic sensitive area attention module. In the aerodynamic sensitive area attention module, the key geometric features related to aerodynamic performance are weighted and enhanced through the attention mechanism. Furthermore, the difference feature representation between the two shapes is constructed through the difference operation of the difference feature construction module.

[0138] Finally, the differential features are input into the comparison and discrimination module, and the discrimination probability of the drag coefficient relationship between the two car shapes is obtained through nonlinear mapping calculation. This enables a rapid comparison of the aerodynamic performance of different shape schemes, providing an efficient auxiliary decision-making basis for the overall vehicle shape design and optimization.

[0139] like Figure 2 As shown, this paper details the specific implementation logic of sample binning and matching during the training phase, which improves the model's discrimination accuracy through a refined sampling strategy. The scheme divides each car sample in the DrivarNet dataset into different drag range bins, as shown in bins 1 to 10 in the figure. When constructing training pairs, the core principle is "selecting the bin containing the sample," combining three sampling strategies: neighbor sampling (30%), which enhances the model's ability to discriminate minute shape details by selecting samples with similar drag; cross-bin sampling (60%), which highlights significant performance features by selecting samples with large drag differences; and random sampling (10%), which supplements sample diversity and improves generalization ability. This binning and matching method enables the model to learn the drag difference patterns at different levels.

[0140] like Figure 3 As shown, the dynamic graph convolutional network structure inside the point cloud feature extraction module is displayed, which is the core of realizing geometric feature modeling.

[0141] The point cloud feature extraction module first uses KNN calculation to dynamically construct the local neighborhood relationships of the point cloud in the feature space.

[0142] Subsequently, the network stacks multiple layers of edge convolutional structures. By splicing relative displacement vectors with center point features, the network can adaptively capture key areas in the car's shape that are sensitive to aerodynamic performance.

[0143] After multi-layer convolution mapping, the local features are aggregated into fixed-dimensional global features using the symmetric aggregation operation of max pooling, which effectively solves the problem of disorder in point cloud data.

[0144] Finally, the fully connected layer outputs the final global feature vector, providing a highly discriminative feature foundation for subsequent comparison tasks.

[0145] like Figure 4 As shown, the aerodynamic sensitive area attention module adopts a data processing structure from left to right and sets up two parallel input branches to represent the global feature vectors obtained after the two car shape point clouds pass through the feature extraction module.

[0146] Specifically, the upper input branch is the car shape feature vector f1, and the lower input branch is the car shape feature vector f2, which respectively represent the high-dimensional feature representations obtained after the two car shape point clouds are mapped by the feature extraction network. The input features first enter the aerodynamic sensitive area attention weight calculation module.

[0147] The aerodynamically sensitive region attention module performs a nonlinear transformation on the input feature vector through an attention mapping network to calculate the attention weights corresponding to each feature dimension. Specifically, the attention calculation network can be composed of a multilayer perceptron structure, combining the ReLU activation function and the sigmoid function to generate weight vectors. Through this process, the upper branch outputs attention weight W1, and the lower branch outputs attention weight W2, thus providing weight information for subsequent feature enhancement.

[0148] Subsequently, the input features and corresponding attention weights enter the feature weighting module, where element-wise multiplication is used to weight and enhance the features. Through this step, the network can assign different importance weights to different feature dimensions, thereby enhancing the representation of geometric features related to aerodynamically sensitive areas and giving higher weights to structural features that have a greater impact on aerodynamic performance in the feature representation.

[0149] After feature weighting, the two feature branches are fused in the differential feature calculation module. This module constructs differential feature vectors by performing differential operations on the enhanced features of the two shapes to characterize the direction and magnitude of geometric changes between the two car shapes, thereby transforming the complex aerodynamic performance differences between the shapes into differential representations in the feature space. Subsequently, the differential feature vectors are input to the nonlinear feature mapping module, where high-dimensional recoding of the differential features is performed through a multi-layer fully connected network. Through this process, the model can further analyze the complex nonlinear relationship between the geometric changes of the car shape and its aerodynamic performance.

[0150] Finally, the module outputs a fused feature vector, which integrates aerodynamic attention information and shape difference information. This feature can be used as the input feature for the subsequent drag comparison and discrimination module, thereby enabling the discrimination of the drag coefficient relationship between different car shape schemes.

[0151] The electronic device of the present invention includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the above-described method for comparing and judging the aerodynamic performance of automobile shapes.

[0152] The present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method for comparing and judging the aerodynamic performance of automobile exteriors.

[0153] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0154] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0155] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0156] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0157] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0158] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 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 of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for comparing and judging the aerodynamic drag performance of automobiles, characterized in that, include: Three-dimensional point cloud data of the first and second vehicle outlines are acquired. The two sets of point cloud data are preprocessed to obtain the standardized first and second input point clouds. The first and second input point clouds are respectively input into a point cloud feature extraction network with shared parameters. By dynamically constructing local neighborhood relationships of point clouds and performing layer-by-layer feature aggregation, the first and second global feature vectors are obtained. The first global feature vector and the second global feature vector are respectively input into the aerodynamic sensitive area attention network. The features of each point are weighted by spatial attention weights to obtain the first enhanced feature vector and the second enhanced feature vector. Based on the first enhanced feature vector and the second enhanced feature vector, a differential feature vector is constructed to characterize the geometric differences between the two car shapes. The difference feature vector is input into a nonlinear mapping network consisting of at least one fully connected layer and an activation function to obtain a discrimination scalar value. The discrimination scalar value is then mapped to a probability value using the Sigmoid function. When the probability value is greater than the preset threshold, the output indicates that the drag coefficient of the second car shape is less than that of the first car shape. Otherwise, the output indicates that the drag coefficient of the first car's shape is less than the comparison result of the second car's shape.

2. The method for comparing and judging the aerodynamic drag performance of automobiles as described in claim 1, characterized in that, The preprocessing of the two sets of point cloud data to obtain a standardized first input point cloud and second input point cloud includes the following steps: A preset number of points are randomly sampled from two sets of 3D point cloud data to make the number of points in the two sets of point clouds the same, forming a first intermediate point cloud and a second intermediate point cloud. The first and second intermediate point clouds are centered to shift the overall mean of the point clouds to the origin, resulting in the normalized first and second input point clouds.

3. The method for comparing and judging the aerodynamic performance of automobile exteriors as described in claim 1, characterized in that, The process of constructing local neighborhood relationships in the point cloud and performing layer-by-layer feature aggregation to obtain the first global feature vector and the second global feature vector includes the following steps: In each feature transformation layer of the point cloud feature extraction network, based on the feature vector of each point in the current layer, the K-nearest neighbor algorithm is used to select a preset number of points in the feature space that are closest to each point, forming a local neighborhood set for each point. For each point's local neighborhood set, construct the edge features between the corresponding point and its nearest neighbors, transform each edge feature through a shared convolutional mapping function, and apply max pooling operation to the transformed edge features to obtain the updated feature vector of the corresponding point. The stacked multi-layer local neighborhood construction and feature aggregation process concatenates the point feature vectors output by each layer along the feature dimension to obtain multi-scale point features; Perform global max pooling operation along the point dimension on the multi-scale point features to obtain the global feature vector corresponding to the car shape.

4. The method for comparing and judging the aerodynamic drag performance of automobiles as described in claim 3, characterized in that, The construction of edge features between a corresponding point and its nearest neighbors includes the following steps: For the center point and any of its nearest neighbors, calculate the difference between the feature vector of the nearest neighbor and the feature vector of the center point; The difference is concatenated with the feature vector of the center point itself to obtain the edge features between the center point and its nearest neighbors.

5. The method for comparing and judging the aerodynamic performance of automobile exteriors as described in claim 1, characterized in that, The step of weighting the features of each point using spatial attention weights to obtain the first enhanced feature vector and the second enhanced feature vector includes the following steps: Using the aerodynamic sensitivity mapping function, an aerodynamic sensitivity score is generated for each point based on the feature vector of each point. By mapping the aerodynamic sensitivity score of each point to a spatial attention weight using a normalized exponential function, the sum of the spatial attention weights of all points is equal to 1. The enhanced feature vector is obtained by multiplying the feature vector of each point with the corresponding spatial attention weight and then summing the results.

6. The method for comparing and judging the aerodynamic drag performance of automobiles as described in claim 5, characterized in that, After multiplying the feature vectors of each point by their corresponding spatial attention weights and summing the results to obtain the enhanced feature vectors, the following steps are also included: The enhanced feature vector is input into the attention mapping network, which is composed of a first weight matrix, a ReLU activation function, a second weight matrix, and a Sigmoid function in sequence. The attention weight vector of the feature dimension is calculated through the attention mapping network. The enhanced feature vector is multiplied element-wise with the feature dimension attention weight vector to obtain the target enhanced feature vector, and the target enhanced feature vector is used as the first enhanced feature vector or the second enhanced feature vector.

7. A system for comparing and judging the aerodynamic drag performance of automobiles, characterized in that, To implement the method of any one of claims 1-6, comprising: The preprocessing module is used to acquire three-dimensional point cloud data of the first and second vehicle shapes, and preprocess the two sets of point cloud data to obtain the standardized first and second input point clouds. The point cloud feature extraction module is used to input the first input point cloud and the second input point cloud into the parameter-shared point cloud feature extraction network, and obtain the first global feature vector and the second global feature vector by dynamically constructing the local neighborhood relationship of the point cloud and performing layer-by-layer feature aggregation. The aerodynamically sensitive region attention module is used to input the first global feature vector and the second global feature vector into the aerodynamically sensitive region attention network, respectively, and to weight the features of each point through spatial attention weights to obtain the first enhanced feature vector and the second enhanced feature vector. The differential feature construction module is used to construct differential feature vectors that characterize the geometric differences between two car shapes based on the first enhanced feature vector and the second enhanced feature vector. The comparison and discrimination module is used to input the differential feature vector into the comparison and discrimination network, and output the comparison and discrimination result after nonlinear mapping. The comparison and discrimination result is used to indicate the car shape with the lower drag coefficient between the first car shape and the second car shape.

8. An electronic device, characterized in that, It includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the method for comparing and judging the aerodynamic drag performance of an automobile as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for comparing and judging the aerodynamic drag performance of automobiles as described in any one of claims 1-6.

Images