Vehicle point cloud wind resistance coefficient prediction method and system based on multi-scale learning and convolution

Abstract

The invention provides a whole vehicle point cloud windage coefficient prediction method and system based on multi-scale learning and convolution, and relates to the technical field of windage coefficient prediction, and the method specifically comprises the steps: obtaining point cloud data of a whole vehicle model, and carrying out the preprocessing of the point cloud data; sampling the preprocessed point cloud data through a farthest point sampling method so as to reserve geometric key points; constructing a convolutional neural network model, inputting the point cloud data into the network, sequentially carrying out two times of multi-scale convolution operations, extracting local details and global structure features through convolution kernels of different scales, mapping the features from low dimensions to 512 dimensions and expanding the features to 1024 dimensions, and completing feature aggregation; the feature expression capability is further enhanced through two-layer convolution; utilizing a maximum pooling layer to aggregate global features, and performing dimension reduction on the features through a multi-layer perceptron; introducing a physical guidance attention mechanism to physically constrain the spatial weight of the features; and establishing a mapping relation between the point cloud features and the wind resistance coefficient through a full connection layer, and outputting a wind resistance coefficient prediction result. According to the invention, the accuracy and efficiency of wind resistance coefficient prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind resistance coefficient prediction, and particularly to a method and system for predicting the wind resistance coefficient of a whole vehicle point cloud based on multi-scale learning and convolution. Background Art

[0002] In the design, research and development, and manufacturing processes of modern automobiles, the wind resistance coefficient is an important parameter for measuring the aerodynamic performance of vehicles. A lower wind resistance coefficient not only helps to reduce the air resistance of the vehicle during driving, thus significantly reducing energy consumption, but also improves the stability and handling performance of the vehicle under high-speed driving conditions, enhancing the safety and ride comfort of the whole vehicle.

[0003] Currently, common wind resistance coefficient prediction methods often fail to fully exploit the potential information of point cloud data at multiple scales, easily overlook the synergistic effect of local and global features, and are difficult to comprehensively consider the complex factors affecting the wind resistance coefficient. In addition, existing point cloud data processing means generally have the problem of insufficient accuracy, are difficult to effectively retain key geometric information, and are prone to feature dimension redundancy or information loss, thus affecting the overall performance and accuracy of the prediction model and unable to meet the application requirements of the automotive industry for high-precision and rapid prediction. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for predicting the wind resistance coefficient of a whole vehicle point cloud based on multi-scale learning and convolution, so as to solve the problem in the prior art that the complex factors affecting the wind resistance coefficient are not comprehensively considered and the point cloud feature processing ability is limited, resulting in low accuracy of wind resistance coefficient prediction.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A method for predicting the wind resistance coefficient of a whole vehicle point cloud based on multi-scale learning and convolution, the steps include: obtaining the point cloud data of the whole vehicle model, and preprocessing the point cloud data; sampling the preprocessed point cloud data by the farthest point sampling method to retain geometric key points. Establish a neural network model, input the point cloud data into the neural network model, perform two multi-scale convolution operations using different scale convolution kernels, map the point cloud data from low-dimensional features to high-dimensional features, and at the same time extract the local detail features and global structure features of the point cloud data and perform feature aggregation to obtain target point cloud features; introduce a physical-guided attention mechanism to physically constrain the attention weights of the neural network model, and gradually reduce the dimension of the target point cloud features through a multi-layer perceptron, then establish a mapping relationship between the target point cloud features and the wind resistance coefficient, and finally output the wind resistance coefficient prediction result.

[0006] Optionally, obtain the point cloud data of the vehicle model and perform preprocessing; sample the preprocessed point cloud data using the farthest point sampling method to retain geometric key points, enhance the representativeness of the point cloud data information, and maintain the integrity of the vehicle structure; construct a convolutional neural network model and input the point cloud data into the neural network model: First, extract local feature data through two multi-scale convolutional operations (using convolutional kernels with scales of 1×3 and 3×3 respectively), extract the local detail features of the point cloud data in different local fields of view, and gradually increase the feature dimension to 1024 dimensions to form a fused feature representation; Subsequently, further enhance the feature expression ability through two layers of single-scale convolutions, keeping the feature dimension at 1024 dimensions; Then, use max pooling operation to achieve global feature aggregation, and perform step-by-step dimensionality reduction processing on the features through a multi-layer perceptron (MLP); At the same time, introduce a physical-guided attention mechanism to impose physical constraints on the spatial feature weights in the neural network, guiding the model to focus on the key structural areas that have a significant impact on the drag coefficient; Finally, establish a mapping relationship between the point cloud features and the drag coefficient through a fully connected layer, and output the predicted result of the drag coefficient.

[0007] Optionally, the preprocessing step includes: performing normalization processing on the point cloud data and scaling the coordinates to the range [0, 1].

[0008] Optionally, the step of sampling the preprocessed point cloud data by the farthest point sampling method specifically includes: performing farthest point sampling on the point cloud data, and constructing a sampling strategy by iteratively selecting the point farthest from the selected point set, so as to obtain a key point cloud set that not only maintains the geometric details of the vehicle surface but also has uniform spatial distribution. Its calculation formula is: In the formula, next_point is the next point, max is the maximum value, min is the minimum value, p is the candidate point, q is the selected point, points is the set of candidate points, selected is the set of selected points, and ||p - q||2 is the Euclidean distance between the candidate point and the selected point.

[0009] Optionally, the step of using convolutional kernels of different scales to perform two multi-scale convolutional operations to map the point cloud data from low-dimensional features to high-dimensional features includes: using convolutional kernels of 1×3 and 3×3 to perform multi-scale convolutional operations on the point cloud data simultaneously to expand the breadth of the neural network, map the features from low-dimensional to 512 dimensions and further expand to 1024 dimensions, and then further enhance the feature expression ability through two layers of convolutions; Among them, the calculation formula for using convolutional kernels of 1×3 and 3×3 to perform multi-scale convolutional operations on the point cloud data simultaneously is: F k×k = Conv(P norm , K k×k ), where K k×k is the convolutional kernel with a scale of k×k, and Fk×k is the convolution result, P norm is the point cloud data.

[0010] Optionally, after the step of performing the two multi-scale convolution operations, the method further includes: performing batch normalization processing on the point cloud data after the multi-scale convolution operation, and its calculation formula is: In the formula, y is the batch normalization result, x is the point cloud data, μ is the mean of the batch data, σ 2 is the batch data variance, γ and β are learnable parameters, and ε is a small constant; introducing a non-linear transformation operation through the ReLU activation function, and its calculation formula is: RELU(x) = max(0, x), where RELU is the rectified linear activation function, x is the point cloud data, and max(0, x) is the maximum value of 0 and x; randomly inactivating some neurons, and its calculation formula is: In the formula, y is the random inactivation result, x is the point cloud data, and p is the probability parameter.

[0011] Optionally, the step of simultaneously extracting local detail features and global structure features and performing feature aggregation specifically includes: simultaneously extracting local detail features and global structure features and splicing them into a new feature matrix, and its calculation formula is: F multi-scale = Concat(F k×k , F k×k ), where F multi-scale is the multi-scale feature matrix, Concat is the splicing operation, and F k×k is the feature matrix; among them, the output feature dimension of the first multi-scale convolution is 512 dimensions, the feature matrix is n×512, and the output feature dimension after two multi-scale convolutions is 1024 dimensions, and the feature matrix is n×1024; performing a max pooling operation on the multi-scale feature matrix to extract global features, and the output dimension after pooling is 1024 dimensions, and its calculation formula is: F pooled = max(F multi-scalel ), where F multi-scalel is the eigenvalue within the local area, and F pooled is the global eigenvalue.

[0012] Optionally, the step of gradually reducing the dimension of the target point cloud feature through the multi-layer perceptron specifically includes: the multi-layer perceptron includes 5 fully connected layers, and each fully connected layer performs batch normalization, ReLU activation function non-linear transformation operation, and random inactivation in sequence, gradually mapping the dimension of the global feature vector from 1024 dimensions to 512 dimensions, 256 dimensions, 128 dimensions, 64 dimensions, and finally outputting a wind resistance coefficient value of 1 dimension, and its calculation formula is: F MLP = ReLU(W.F pooled + b), where F MLPis the output feature of the multi-layer perceptron, RELU is the activation function, W is the weight matrix, and F pooled is the global feature, and b is the bias term.

[0013] Optionally, the step of establishing the mapping relationship between the target point cloud feature and the drag coefficient through the fully connected layer specifically includes: adopting a fully connected mapping strategy, and mapping the target point cloud feature to the drag coefficient value in aerodynamics through the fully connected layer. The calculation formula is: C d = W out .F MLP + b out , where C d is the drag coefficient, W out is the weight matrix, F MLP is the output feature of the multi-layer perceptron, and b out is the bias term.

[0014] Optionally, the step of introducing the physically-guided attention mechanism specifically includes: obtaining the vehicle point cloud feature information extracted by the neural network module; calculating the original spatial attention map based on the geometric position relationship of each point in the point cloud to depict the initial attention of each region in the model; introducing the physical prior information constructed based on the simplified form of the Navier-Stokes equation, where the physical prior information includes the local pressure gradient and velocity gradient on the vehicle surface to reflect the possible aerodynamic effects of each region under the action of air flow; constructing a physically-guided function to weight and correct the original spatial attention map, so that the regions focused by the model are more in line with the principles of aerodynamics, and the key structures such as the front of the vehicle, the roof, and the tail that have a significant impact on the drag coefficient are focused on; outputting the physically-guided corrected attention map, and fusing it with the target point cloud feature to further enhance the feature expression ability and be used for subsequent drag coefficient regression prediction tasks.

[0015] Optionally, the step of introducing the physically-guided attention mechanism to physically constrain the attention weights of the neural network model specifically includes: introducing the simplified form of the Navier-Stokes equation as a soft constraint, and its calculation formula is: where L phy is the correction term, λ1 and λ2 are weight coefficients, α is a redundant parameter, is the Reynolds number, C d is the drag coefficient, ρ∞ is the far-field fluid density, ρ is the local fluid density, and v is the fluid velocity.

[0016] On the other hand, the present invention also provides a vehicle point cloud drag coefficient prediction system based on multi-scale learning and convolution, including: a point cloud acquisition module for acquiring point cloud data of a vehicle model and preprocessing the point cloud data; a farthest point sampling module for sampling the preprocessed point cloud data by the farthest point sampling method to retain geometric key points, enhance the representativeness of point cloud data information and maintain the integrity of the vehicle structure; a neural network module for establishing a neural network model, inputting the point cloud data processed by the farthest point sampling module into the neural network model, performing two multi-scale convolution operations using different scale convolution kernels, mapping the point cloud data from low-dimensional features to high-dimensional features, and at the same time extracting the local detail features and global structure features of the point cloud data and performing feature aggregation to obtain target point cloud features; a drag coefficient prediction module for physically constraining the attention weights of the neural network model by introducing a physical-guided attention mechanism and gradually reducing the dimension of the target point cloud features through a multi-layer perceptron, establishing a mapping relationship between the target point cloud features and the drag coefficient, and finally outputting the drag coefficient prediction result.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] After acquiring the point cloud data of the vehicle model in this application, preprocessing is carried out, and further sampling of the preprocessed point cloud data is performed by the farthest point sampling method, so that the sampling points can cover the entire point cloud space as evenly as possible to retain geometric key points, enhance the representativeness of data information, and at the same time ensure the integrity of the vehicle geometric structure, so as to provide a basic feature representation for subsequent multi-scale convolution.

[0019] A neural network model is established, and the point cloud data is input into the neural network model for drag coefficient prediction, which can quickly process the task of unordered point cloud.

[0020] Through two multi-scale convolution operations, the point cloud data is mapped from low-dimensional features to 512 dimensions and extended to 1024 dimensions, which greatly improves the expression ability of the features of vehicle point cloud data, provides a richer feature representation for subsequent tasks, improves the accuracy of model prediction, and at the same time extracts the local detail features and global structure features of the point cloud data and performs feature aggregation to obtain target point cloud features, solving the problem that the current drag prediction methods are difficult to take into account both local detail features and global structure features, and enhancing the diversity and robustness of feature representation.

[0021] The target point cloud features are gradually reduced in dimension through a multi-layer perceptron, and the most significant feature information of the model is retained. A physical-guided attention mechanism is introduced to physically constrain the attention weights of the neural network model, improving the reliability of model prediction. Then, a mapping relationship between the target point cloud features and the drag coefficient is established through a fully connected layer, and finally, the drag coefficient prediction result is output to achieve end-to-end prediction. Description of the Drawings

[0022] Figure 1 This is a schematic diagram of the method step flow of the present invention.

[0023] Figure 2 This is a schematic diagram of the intelligent prediction structure of the drag coefficient of the system of the present invention.

[0024] Figure 3 This is a schematic diagram of the method framework flow of the present invention.

[0025] Figure 4 This is a schematic diagram of the structure of the physical-guided attention mechanism module of the present invention.

[0026] Figure 5 This is a schematic diagram of the structure of the multi-layer perceptron module of the present invention.

[0027] Figure 6 This is a schematic diagram of the feature dimension compression structure of the present invention.

[0028] Figure 7 This is a schematic diagram of the system structure of the present invention.

[0029] In the figure: 10 - point cloud acquisition module, 20 - farthest point sampling module, 30 - neural network module, 40 - drag coefficient prediction module. Detailed Embodiments

[0030] Next, the solutions of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0031] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so as to describe the embodiments of this application here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0032] Those skilled in the art of the present technology can understand that, unless specifically stated, the singular forms "a", "an", "the", and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present application means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or their groups. It should be understood that when we say an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling. The phrase "and / or" used herein includes all or any unit and all combinations of one or more related listed items.

[0033] Those skilled in the art of the present technology can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the art to which the present application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless specifically defined as here.

[0034] It should be understood that the sequence numbers and magnitudes of the steps in this embodiment do not imply the order of execution. The order of execution of each process is determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0035] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0036] Please refer to Figures 1-7 , the present invention provides a method for predicting the aerodynamic drag coefficient of a vehicle body point cloud based on multi-scale learning and convolution, and the steps include: The steps include:

[0037] S100. Obtain the point cloud data of the vehicle body model, and preprocess the point cloud data.

[0038] Specifically, after obtaining the point cloud data of the vehicle body model in the present application, preprocessing is performed to convert the original point cloud data into a standardized format that can be processed by the network model.

[0039] S200. Sample the preprocessed point cloud data by the farthest point sampling method to retain geometric key points, enhance the representativeness of the point cloud data information, and maintain the integrity of the vehicle structure.

[0040] Specifically, the preprocessed point cloud data is sampled by the farthest point sampling method, so that the sampling points can cover the entire point cloud space as evenly as possible, retain geometric key points, enhance the representativeness of data information, and ensure the integrity of the automotive geometric structure, so as to provide a basic feature representation for subsequent multi-scale convolution.

[0041] S300. Establish a neural network model, input the point cloud data into the neural network model, perform two multi-scale convolution operations using different scale convolution kernels, map the point cloud data from low-dimensional features to high-dimensional features, extract the local detail features and global structure features of the point cloud data and perform feature aggregation to obtain the target point cloud features.

[0042] Specifically, establish a neural network model, input the point cloud data into the neural network model for drag coefficient prediction, and can quickly process the task of unordered point cloud. Through two multi-scale convolution operations, the point cloud data is mapped from low-dimensional features to 512 dimensions and extended to 1024 dimensions, greatly improving the expression ability of the features of automotive point cloud data, providing a richer feature representation for subsequent tasks, improving the accuracy of model prediction, and extracting the local detail features and global structure features of the point cloud data and performing feature aggregation to obtain the target point cloud features, solving the problem that the current drag prediction methods are difficult to take into account both local detail features and global structure features, and enhancing the diversity and robustness of feature representation.

[0043] S400. Physically constrain the attention weights of the neural network model by introducing a physically-guided attention mechanism, and gradually reduce the dimension of the target point cloud features through a multi-layer perceptron, establish a mapping relationship between the target point cloud features and the drag coefficient, and finally output the drag coefficient prediction result.

[0044] Specifically, introduce a physically-guided attention mechanism to physically constrain the attention weights of the neural network model, guide the network to focus on the key regions affecting the drag coefficient, improve the reliability of model prediction, then gradually reduce the dimension of the target point cloud features through a multi-layer perceptron, retain the most significant feature information of the model and establish a mapping relationship between the target point cloud features and the drag coefficient, and finally output the drag coefficient prediction result to achieve end-to-end prediction.

[0045] Optionally, obtain the point cloud data of the whole vehicle model and perform preprocessing; sample the preprocessed point cloud data by the farthest point sampling method to retain geometric key points, enhance the representativeness of point cloud data information and maintain the integrity of the automotive structure; construct a neural network model based on convolution, and input the point cloud data into the neural network model:

[0046] First, local feature data is extracted through two multi-scale convolution operations (using convolution kernels of scales 1×3 and 3×3 respectively), extracting local detailed features of point cloud data in different local fields of view and gradually increasing the feature dimension to 1024 dimensions to form a fused feature representation; subsequently, the feature expression ability is further enhanced through two layers of single-scale convolution, maintaining the feature dimension at 1024 dimensions; then, max pooling operation is used to achieve global feature aggregation, and the feature is gradually dimension-reduced through a multi-layer perceptron (MLP); at the same time, a physical-guided attention mechanism is introduced to impose physical constraints on the spatial feature weights in the neural network, guiding the model to focus on the key structural regions that significantly affect the drag coefficient; finally, a mapping relationship between the point cloud features and the drag coefficient is established through a fully connected layer, and the drag coefficient prediction result is output.

[0047] It should be understood that when predicting the drag coefficient of a car, the shape and structure of the car have multi-scale features. For example, the overall streamlined contour of the car is a macroscopic-scale feature, that is, a global structural feature; while the tiny bumps and depressions on the body surface, rearview mirrors, door handles, etc. are microscopic-scale features, that is, local detailed features. For existing drag prediction models such as neural networks, it is difficult to capture these different-scale features simultaneously. It may focus on the overall shape and ignore the influence of local details on drag, or vice versa, resulting in inaccurate prediction of the drag coefficient. At the same time, in car drag prediction, the change of local air flow is often closely related to the local geometric features of the car surface. For example, the local shape near the front bumper and intake grille of the car will significantly affect the air flow in this area, thereby affecting the drag. Existing drag prediction models such as neural networks do not have a dedicated mechanism for local feature extraction and may not be able to accurately obtain these local features related to drag. Due to the lack of a fusion mechanism for integrating macroscopic body streamlined features and microscopic surface detail features, there are deficiencies in dealing with the complex relationship between car shape and drag. Therefore, based on the neural network model architecture, this application can perform multi-level feature extraction on the input point cloud feature data by using convolution kernels of different scales, can capture local details and global structures simultaneously, enhance the richness of feature representation, and improve feature correlation and prediction accuracy. At the same time, the introduction of its feature visualization and physical constraints enhances the interpretability of the model.

[0048] In some embodiments, the preprocessing step includes: normalizing the point cloud data and scaling the coordinates to the range [0, 1].

[0049] Specifically, the preprocessing of the point cloud data is mainly used to standardize the point cloud data and convert it into a standardized format that can be processed by the model to ensure the stability and convergence of model training. Its core functions mainly include: mapping the point cloud coordinates and feature values to a unified numerical range, specifically including scaling the coordinates to the range of [0, 1] to eliminate the scale differences of point cloud data of different vehicles, thereby accelerating model convergence, making the optimization process smoother, and reducing training time. At the same time, it avoids the problems of gradient explosion or disappearance, improves the robustness and generalization ability of the model, enhances the adaptability of the model to scale changes of input data, and reduces the influence of noise and outliers.

[0050] In some embodiments, the step of uniformly sampling the preprocessed point cloud data by the farthest point sampling method specifically includes: performing farthest point sampling on the point cloud data, and constructing a sampling strategy by iteratively selecting the point farthest from the selected point set, so as to obtain a key point cloud set that not only maintains the geometric details of the vehicle surface but also has uniform spatial distribution. Its calculation formula is: In the formula, next_point is the next point, max is the maximum value, min is the minimum value, p is the candidate point, q is the selected point, points is the candidate point set, selected is the selected point set, and ||p - q||2 is the Euclidean distance between the candidate point and the selected point. Specifically, sampling the preprocessed point cloud data by the farthest point sampling method can preferentially capture geometrically significant regions such as the edges and corners of the vehicle surface, such as the contour line of the vehicle body and the edge of the headlight, etc., and can retain the key information of the geometric features of the vehicle surface to provide a basic feature representation for subsequent multi-scale convolution. At the same time, it has good uniform spatial distribution, ensuring that the sampling points are evenly distributed in the entire point cloud space, retaining local features while covering the global shape.

[0051] In some embodiments, the step of performing two multi-scale convolution operations on the point cloud data using different scale convolution kernels to map the point cloud data from low-dimensional features to high-dimensional features includes: simultaneously performing multi-scale convolution operations on the point cloud data using 1×3 and 3×3 convolution kernels to expand the breadth of the neural network, mapping the features from low-dimensional to 512 dimensions and further expanding to 1024 dimensions. Subsequently, the feature expression ability is further enhanced through two layers of convolution, mapping the point cloud data from low-dimensional features to 512 dimensions, and then continuously using multi-scale convolution operations to expand the features to 1024 dimensions. Among them, the calculation formula for simultaneously performing multi-scale convolution operations on the point cloud data using 1×3 and 3×3 convolution kernels is: F k×k =Conv(P norm , K k×k ), where K k×k is the convolution kernel with a scale of k×k, F k×k is the convolution result, and Pnorm is point cloud data.

[0052] Specifically, a 1×3 convolutional kernel is used to extract the fine structure of the vehicle surface in the point cloud data, and a 3×3 convolutional kernel is used to cover a wider neighborhood to capture the overall shape of the vehicle and the relationships between its components; mapping the point cloud data from low-dimensional features to 512 dimensions, and then continuously using multi-scale convolutional operations to expand the features to 1024 dimensions can better capture local details and global structures, enhance the richness of the features, greatly enhance the expressive ability of the point cloud data features, and improve the accuracy of model prediction.

[0053] In some embodiments, after the step of performing two multi-scale convolutional operations, it further includes: performing batch normalization processing on the point cloud data after the multi-scale convolutional operations, and its calculation formula is: In the formula, y is the batch normalization result, x is the point cloud data, μ is the mean of the batch data, σ 2 is the variance of the batch data, γ and β are learnable parameters, and ε is a small constant; introducing a non-linear transformation operation through the ReLU activation function, and its calculation formula is: RELU(x) = max(0, x), where RELU is the rectified linear unit activation function, x is the point cloud data, and max(0, x) is to take the maximum value between 0 and x; randomly deactivating some neurons, and its calculation formula is: In the formula, y is the result of random inactivation, x is the point cloud data, and p is the probability parameter.

[0054] Specifically, in addition to the multi-scale convolution operation, in order to enhance the model's expressive power, accelerate convergence, and prevent overfitting, batch normalization, ReLU activation function non-linear transformation, and dropout operations are added after the convolution operation. The batch normalization operation can standardize the input of each layer during training, ensuring that the input features are more balanced in each layer, thereby reducing the "internal covariate shift" during training, accelerating training, and improving the model's performance and stability. At the same time, by using the batch normalization operation, the model's dependence on specific samples can be reduced to avoid overfitting as much as possible, and the model can be more stable during training, reducing the problems of gradient disappearance or explosion. The non-linear feature transformation operation uses the ReLU (Rectified Linear Unit) activation function to capture complex patterns in the image and introduce non-linear features, which is beneficial for learning more complex features and preventing the problem of gradient disappearance, improving the training speed and performance of the network. The dropout operation is used to randomly deactivate some neurons to prevent overfitting. Therefore, the feature data after the above four operations can enable the model to have a faster convergence speed, a more stable training process, higher accuracy and generalization ability, effectively avoid the problems of gradient disappearance and explosion, and enhance the model's expressive power.

[0055] In some embodiments, the step of simultaneously extracting local detail features and global structure features and performing feature aggregation specifically includes: simultaneously extracting local detail features and global structure features and splicing them into a new feature matrix, and its calculation formula is: F multi-scale = Concat(F k×k , F k×k ), where in the formula, F multi-scale is the multi-scale feature matrix, Concat is the splicing operation, and F k×k is the feature matrix; among them, the output feature dimension of the first multi-scale convolution is 512 dimensions, the feature matrix is n×512, and the output feature dimension after two multi-scale convolutions is 1024 dimensions, and the feature matrix is n×1024; perform a max pooling operation on the multi-scale feature matrix to extract global features, and the output dimension after pooling is 1024 dimensions, and its calculation formula is: F pooled = max(F multi-scalel ), where in the formula, F multi-scalel is the eigenvalue in the local area, and F pooled is the global eigenvalue.

[0056] Specifically, simultaneously extracting local detail features and global structure features and splicing them into a new feature matrix can take into account both local detail information and global structure information to form a more comprehensive feature representation, enabling the model to pay attention to details while also taking into account the overall structure and avoiding the limitations of single features.

[0057] In some embodiments, the step of gradually reducing the dimensionality of the target point cloud features through a multi-layer perceptron specifically includes: the multi-layer perceptron contains 5 fully connected layers, and each fully connected layer sequentially performs batch normalization, non-linear transformation operation with the ReLU activation function, and dropout, gradually mapping the dimensionality of the global feature vector from 1024 dimensions to 512 dimensions, 256 dimensions, 128 dimensions, 64 dimensions, and finally outputting a drag coefficient value of 1 dimension. Its calculation formula is: F MLP = RELU(W.F pooled + b), where F MLP is the output feature of the multi-layer perceptron, RELU is the activation function, W is the weight matrix, F pooled is the global feature, and b is the bias term.

[0058] Specifically, the multi-layer perceptron MLP is responsible for gradually reducing the dimensionality of the high-dimensional features and mapping them to the final drag coefficient value. Through multi-layer non-linear transformations, it extracts high-order features in the point cloud and enhances the expressive power of the model. The 1024-dimensional high-dimensional features are gradually reduced to a 64-dimensional low-dimensional space through five operations of the multi-layer perceptron MLP, retaining the most significant feature information of the model.

[0059] In some embodiments, the step of establishing the mapping relationship between the target point cloud features and the drag coefficient through a fully connected layer specifically includes: adopting a fully connected mapping strategy, and mapping the target point cloud features to the drag coefficient value in aerodynamics through a fully connected layer. Its calculation formula is: C d = W out ·F MLP + b out , where C d is the drag coefficient, W out is the weight matrix, F MLP is the output feature of the multi-layer perceptron, and b out is the bias term.

[0060] Specifically, the fully connected layer maps the target point cloud features to the drag coefficient Cd value, enabling end-to-end prediction.

[0061] This application can effectively improve the prediction accuracy and enhance the robustness of the model. In terms of economic effects, it has a high optimization efficiency, can effectively reduce the optimization cycle, shorten the time to market, and reduce the time, economic and other costs of R & D optimization. At the same time, compared with traditional CFD, it can effectively reduce the hardware cost and improve the economic benefits. In terms of application effects, its effective prediction of the vehicle drag coefficient can effectively improve the performance of the vehicle fuel economy and so on.

[0062] In some embodiments, the step of introducing a physical guiding attention mechanism to impose physical constraints on the attention weights of the neural network model specifically includes: introducing a simplified form of the Navier-Stokes equation as a soft constraint, and its calculation formula is: In the formula, L phy is a correction term, λ1 and λ2 are weight coefficients, α is a redundant parameter, is the Reynolds number, C d is the drag coefficient, ρ∞ is the far-field fluid density, ρ is the local fluid density, and v is the fluid velocity.

[0063] Specifically, the step of introducing a physical guiding attention mechanism specifically includes: obtaining the vehicle point cloud feature information extracted by the neural network module; calculating an original spatial attention map based on the geometric position relationship of each point in the point cloud to depict the initial attention degree of each region in the model; introducing physical prior information constructed based on a simplified form of the Navier-Stokes equation, where the physical prior information includes the local pressure gradient and velocity gradient on the vehicle surface to reflect the possible aerodynamic effects of each region under the action of airflow; constructing a physical guiding function to perform weighted correction on the original spatial attention map, so that the regions concerned by the model are more in line with the principles of aerodynamics, and focus on key structures such as the front of the vehicle, the roof, and the tail that have a significant impact on the drag coefficient; outputting the attention map corrected by physical guidance, fusing it with the target point cloud features to further enhance the feature expression ability, and using it for subsequent drag coefficient regression prediction tasks. By introducing a simplified form of the Navier-Stokes equation as a soft constraint, the model can be guided to focus on regions that conform to physical laws. For example, in the drag prediction task, the model will automatically focus on the high-curvature regions and the windward area of the vehicle body and ignore irrelevant regions; associating the attention weights with physical quantities can make the regions concerned by the model have clear physical meanings and enhance the interpretability of the model.

[0064] On the other hand, the present invention also provides a vehicle point cloud drag coefficient prediction system based on multi-scale learning and convolution, including: a point cloud acquisition module, configured to acquire point cloud data of a vehicle model and preprocess the point cloud data; a farthest point sampling module, configured to sample the preprocessed point cloud data by the farthest point sampling method to retain geometric key points, enhance the representativeness of the point cloud data information and maintain the integrity of the vehicle structure; a neural network module, configured to establish a neural network model, input the point cloud data into the neural network model, perform two multi-scale convolution operations using different scale convolution kernels, map the point cloud data from low-dimensional features to high-dimensional features, and at the same time extract local detail features and global structure features of the point cloud data and perform feature aggregation to obtain target point cloud features; a drag coefficient prediction module, configured to physically constrain the attention weights of the neural network model by introducing a physical-guided attention mechanism and gradually reduce the dimension of the target point cloud features through a multi-layer perceptron, establish a mapping relationship between the target point cloud features and the drag coefficient, and finally output a drag coefficient prediction result.

[0065] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, 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. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present invention. The aforementioned storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc that can store program codes.

[0066] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0067] The above are only embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for predicting the aerodynamic drag coefficient of a vehicle's point cloud based on multi-scale learning and convolution, characterized in that the steps Including: Obtain the point cloud data of the whole vehicle model and preprocess the point cloud data; Sample the preprocessed point cloud data by the farthest point sampling method to retain geometric key points. Build a neural network model, input the point cloud data into the neural network model, perform two multi-scale convolution operations using different scale convolution kernels, map the point cloud data from low-dimensional features to high-dimensional features, and at the same time extract the local detail features and global structure features of the point cloud data and perform feature aggregation to obtain the target point cloud features; Physically constrain the attention weights of the neural network model by introducing a physical-guided attention mechanism, gradually reduce the dimension of the target point cloud features through a multi-layer perceptron, then establish a mapping relationship between the target point cloud features and the drag coefficient, and finally output the drag coefficient prediction result.

2. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, characterized in that The preprocessing step includes: Normalize the point cloud data and scale the coordinates to the range of [0, 1].

3. The vehicle point cloud aerodynamic drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, wherein The step of uniformly sampling the preprocessed point cloud data by the farthest point sampling method specifically includes: Perform farthest point sampling on the point cloud data, and construct a sampling strategy by iteratively selecting the point farthest from the selected point set, so as to obtain a key point cloud set that not only maintains the geometric details of the vehicle surface but also has uniform spatial distribution.

4. The vehicle point cloud aerodynamic drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, wherein, The step of performing two multi-scale convolution operations using different scale convolution kernels to map the point cloud data from low-dimensional features to high-dimensional features includes: Simultaneously perform multi-scale convolution operations on the point cloud data using 1×3 and 3×3 convolution kernels to expand the breadth of the neural network, map the features from low-dimensional to 512 dimensions and further expand to 1024 dimensions. Subsequently, enhance the feature expression ability through two layers of convolution.

5. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, wherein, After the step of performing two multi-scale convolution operations, it also includes: Perform batch normalization processing on the point cloud data after multi-scale convolution operations respectively, and its calculation formula is: Where y is the result of batch normalization, x is the point cloud data, μ is the mean of the batch data, and σ 2 is the variance of the batch data, γ and β are learnable parameters, and ε is a small constant; Introduce a non-linear transformation operation through the ReLU activation function, and its calculation formula is: RELU(x) = max(0, x), where RELU is the rectified linear unit activation function, x is the point cloud data, and max(0, x) is to take the maximum value between 0 and x; Randomly deactivate some neurons, and its calculation formula is: Where y is the result of random inactivation, x is the point cloud data, and p is the probability parameter.

6. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, wherein The step of simultaneously extracting local detail features and global structure features and performing feature aggregation specifically includes: Simultaneously extract local detail features and global structure features and splice them into a new feature matrix; Among them, the output feature dimension of the first multi-scale convolution is 512 dimensions, and the output feature dimension after two multi-scale convolutions is 1024 dimensions; Perform a max pooling operation on the multi-scale feature matrix to extract global features, and the output dimension after pooling is 1024 dimensions.

7. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 6, characterized in that The step of gradually reducing the dimension of the target point cloud features through a multi-layer perceptron specifically includes: The multi-layer perceptron contains 5 fully connected layers. Each fully connected layer sequentially performs batch normalization, non-linear transformation operation of the ReLU activation function, and random inactivation, and gradually maps the dimension of the global feature vector from 1024 dimensions to 512 dimensions, 256 dimensions, 128 dimensions, 64 dimensions, and finally outputs a drag coefficient value of 1 dimension.

8. The vehicle point cloud aerodynamic drag coefficient prediction method based on multi-scale learning and convolution according to claim 7, characterized in that, The steps of establishing the mapping relationship between the target point cloud feature and the drag coefficient through the fully connected layer specifically include: Adopt a fully connected mapping strategy, and map the target point cloud feature to the drag coefficient value in aerodynamics through the fully connected layer.

9. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, wherein, The steps of introducing the physical-guided attention mechanism specifically include: Obtain the vehicle point cloud feature information extracted by the neural network module; Based on the geometric position relationship of each point in the point cloud, calculate the original spatial attention map to depict the initial attention degree of each region in the model; Introduce the physical prior information constructed based on the simplified form of the Navier-Stokes equation, where the physical prior information includes the local pressure gradient and velocity gradient on the vehicle surface; Construct a physical-guided function to weight and correct the original spatial attention map; Output the attention map corrected by physical guidance, fuse it with the target point cloud feature to further enhance the feature expression ability, and use it for the subsequent drag coefficient regression prediction task.

10. A vehicle point cloud aerodynamic drag coefficient prediction system based on multi-scale learning and convolution, characterized in that, It includes: A point cloud acquisition module, which is used to acquire the point cloud data of the vehicle model and preprocess the point cloud data; A farthest point sampling module, which is used to sample the preprocessed point cloud data by the farthest point sampling method to retain geometric key points; A neural network module, which is used to establish a neural network model, input the point cloud data into the neural network model, perform two multi-scale convolution operations using different scale convolution kernels, map the point cloud data from low-dimensional features to high-dimensional features, and at the same time extract the local detail features and global structure features of the point cloud data and perform feature aggregation to obtain the target point cloud feature; A drag coefficient prediction module, which is used to physically constrain the attention weights of the neural network model by introducing a physical-guided attention mechanism and gradually reduce the dimension of the target point cloud feature through a multi-layer perceptron, establish the mapping relationship between the target point cloud feature and the drag coefficient, and finally output the drag coefficient prediction result.

Images