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

Abstract

The application provides a whole vehicle point cloud wind resistance coefficient prediction method and system based on multi-scale learning and convolution, relates to the wind resistance coefficient prediction technical field, and specifically comprises the following steps: acquiring point cloud data of a whole vehicle model and performing pretreatment; the pretreated point cloud data is sampled by a farthest point sampling method to retain geometric key points; a convolutional neural network model is constructed, the point cloud data is input into the network, twice multi-scale convolution operations are sequentially performed, local details and global structure features are extracted through different scale convolution kernels, the features are mapped from low dimension to 512 dimensions and expanded to 1024 dimensions, and feature aggregation is completed; the feature expression capability is further enhanced through two layers of convolution; global features are aggregated by a maximum pooling layer, the features are reduced in dimension by a multilayer perceptron; a physical guided attention mechanism is introduced to physically constrain the spatial weight of the features; a mapping relationship between the point cloud features and the wind resistance coefficient is established through a full connection layer, and a wind resistance coefficient prediction result is output. The application can improve the accuracy and efficiency of wind resistance coefficient prediction.

Description

TECHNICAL FIELD

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

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

[0003] At present, the common wind resistance coefficient prediction method often fails to fully exploit the potential information of point cloud data in the multi-scale level, easily ignores the synergistic effect of local and global features, and is difficult to comprehensively consider the complex factors affecting the wind resistance coefficient. In addition, the existing point cloud data processing means generally has the problem of insufficient precision, which is difficult to effectively retain key geometric information, and is easy to cause feature dimension redundancy or information loss, thereby affecting the overall performance and precision of the prediction model, and failing to meet the application requirements of high precision and rapid prediction of the automobile industry. SUMMARY

[0004] The purpose of the present application is to provide a vehicle point cloud wind resistance coefficient prediction method and system based on multi-scale learning and convolution, in order to solve the problem of insufficient consideration of complex factors affecting the wind resistance coefficient and limited point cloud feature processing capability in the prior art, resulting in low wind resistance coefficient prediction accuracy.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a vehicle point cloud wind resistance coefficient prediction method based on multi-scale learning and convolution, comprising the following steps: obtaining point cloud data of a vehicle model, preprocessing the point cloud data; sampling the preprocessed point cloud data by a farthest point sampling method; establishing a neural network model, inputting the sampled point cloud data into the neural network model, performing two multi-scale convolution operations by using different scale convolution kernels, mapping the sampled point cloud data from low-dimensional features to high-dimensional features, extracting local detail features and global structure features of the point cloud data after multi-scale convolution operation and performing feature aggregation to obtain target point cloud features; introducing a physical guided attention mechanism to physically constrain the attention weight of the neural network model, and gradually reducing the dimension of the target point cloud features by a multi-layer perception machine, then establishing a mapping relationship between the target point cloud features and the wind resistance coefficient through a fully connected layer, and finally outputting the wind resistance coefficient prediction result.

[0006] Optionally, the point cloud data of the whole vehicle model is acquired and preprocessed; the farthest point sampling method is used to sample the preprocessed point cloud data to retain geometric key points, enhance the information representativeness of the point cloud data and maintain the integrity of the vehicle structure; a convolution-based neural network model is constructed, and the point cloud data is input into the neural network model; first, local feature data is extracted through two multi-scale convolution operations (1x3 and 3x3 convolution kernels are used respectively), the local detail features of the point cloud data in different local fields are extracted, and the feature dimension is gradually increased to 1024 dimensions to form a fusion feature representation; then, two single-scale convolutions are used to further enhance the feature expression ability, and the feature dimension is maintained at 1024; next, a max-pooling operation is used to realize global feature aggregation, and a multi-layer perceptron (MLP) is used to gradually reduce the dimension of the features; at the same time, a physically guided attention mechanism is introduced to apply physical constraints to the spatial feature weights in the neural network, guiding the model to focus on key structural areas that significantly affect the drag coefficient; finally, a full connection layer is used to establish a mapping relationship between the point cloud features and the drag coefficient, and a drag coefficient prediction result is output.

[0007] Optionally, the preprocessing step includes normalizing the point cloud data to scale 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 preprocessed point cloud data, constructing a sampling strategy by iteratively selecting the point farthest from the selected point set, thereby obtaining a key point cloud set that maintains the geometric details of the vehicle surface and has uniform spatial distribution, and the calculation formula is: wherein, is the next point, is the maximum value, is the minimum value, is the candidate point, is the selected point, is the candidate point set, is the selected point set, is the Euclidean distance between the candidate point and the selected point.

[0009] Optionally, the step of using different scale convolution kernels to perform two multi-scale convolution operations to map the sampled point cloud data from low-dimensional features to high-dimensional features includes: using 1x3 and 3x3 convolution kernels to perform multi-scale convolution operations on the sampled point cloud data to map the features from low-dimensional to 512-dimensional and then to 1024-dimensional; wherein the calculation formula for using 1x3 and 3x3 convolution kernels to perform multi-scale convolution operations on the point cloud data is: wherein, is the scale of a convolution kernel, is a convolution result, is point cloud data.

[0010] Optionally, the step of performing two multi-scale convolution operations further comprises: performing batch normalization processing on the point cloud data after multi-scale convolution operation, and the calculation formula is: , wherein y is a batch normalization result, x is the point cloud data after multi-scale convolution operation, is the mean of the batch data, is the variance of the batch data, and is a learnable parameter, is a small constant; a nonlinear conversion operation is introduced through a ReLU activation function, and the calculation formula is: , wherein is a linear rectifier activation function, x is the point cloud data after multi-scale convolution operation, is the maximum value of 0 and x; a part of neurons is randomly inactivated, and the calculation formula is: , wherein y is a random inactivation result, x is the point cloud data after multi-scale convolution operation, and p is a probability parameter.

[0011] Optionally, the step of simultaneously extracting local detail features and global structure features of the point cloud data after multi-scale convolution operation and performing feature aggregation comprises: simultaneously extracting local detail features and global structure features and splicing them into a new feature matrix, and the calculation formula is: , wherein is a multi-scale feature matrix, is a splicing operation, is a feature matrix; wherein the output feature dimension of the first multi-scale convolution is 512, and the feature matrix is , the output feature dimension after two multi-scale convolutions is 1024, and the feature matrix is ; the multi-scale feature matrix is subjected to a maximum pooling operation to extract global features, and the output dimension after pooling is 1024, and the calculation formula is: , wherein is a feature value in a local region, is a global feature value.

[0012] Optionally, the step of gradually reducing the dimension of the target point cloud feature through a multi-layer perception machine comprises: the multi-layer perception machine comprises 5 fully connected layers, each fully connected layer sequentially performs batch normalization, ReLU activation function nonlinear conversion operation and random inactivation, and gradually maps the dimension of the global feature vector from 1024 to 512, 256, 128 and 64, and finally outputs a 1-dimensional wind resistance coefficient value, and the calculation formula is: , wherein, is an output feature of the multi-layer perceptron, is an activation function, is a weight matrix, is a global feature, is a bias term.

[0013] Optionally, the step of establishing the mapping relationship between the target point cloud feature and the wind resistance coefficient through the fully connected layer specifically comprises: adopting a fully connected mapping strategy, mapping the target point cloud feature to a wind resistance coefficient value in aerodynamics through a fully connected layer, and the calculation formula is: , wherein, is a wind resistance coefficient, is a weight matrix, is an output feature of the multi-layer perceptron, is a bias term.

[0014] Optionally, the step of introducing the physical guided attention mechanism specifically comprises: obtaining the whole vehicle point cloud feature information extracted by the neural network model; based on the geometric position relationship of each point in the point cloud, calculating an original spatial attention map to depict the preliminary attention degree of each region in the model; introducing physical prior information constructed based on the simplified form of the Navier-Stokes equation, wherein the physical prior information includes local pressure gradient and velocity gradient on the surface of the whole vehicle to reflect the possible aerodynamic influence of each region of the vehicle under the action of air flow; constructing a physical guiding function to weight and correct the original spatial attention map, so that the regions concerned by the model are more consistent with the principle of aerodynamics, and focus on the key structures such as the front, roof and tail that have a significant impact on the wind resistance coefficient; output the attention map after the physical guiding correction, and fuse it with the target point cloud feature to further enhance the feature expression ability and be used for the subsequent wind resistance coefficient regression prediction task.

[0015] Optionally, the step of introducing the physical guided attention mechanism to physically constrain the attention weight of the neural network model specifically comprises: introducing the simplified form of the Navier-Stokes equation as a soft constraint, and the calculation formula is: , wherein, is a correction term, and are weight coefficients, is a redundant parameter, is a Reynolds number, is a wind resistance coefficient, is a far-field fluid density, is a local fluid density, is a fluid velocity.

[0016] In another aspect, the present application also provides a whole vehicle point cloud wind resistance coefficient prediction system based on multi-scale learning and convolution, which is used to realize the whole vehicle point cloud wind resistance coefficient prediction method based on multi-scale learning and convolution, and comprises: a point cloud acquisition module, which is used to acquire point cloud data of a whole vehicle model and pre-process the point cloud data; a farthest point sampling module, which is used to sample the pre-processed point cloud data by a farthest point sampling method to retain geometric key points, enhance information representativeness of the point cloud data and maintain integrity of the vehicle structure; a neural network module, which is used to establish a neural network model, input the sampled point cloud data into the neural network model, perform two multi-scale convolution operations by using different scale convolution kernels, map the sampled point cloud data from low-dimensional features to high-dimensional features, extract local detail features and global structure features of the point cloud data after the multi-scale convolution operations and perform feature aggregation to obtain target point cloud features; and a wind resistance coefficient prediction module, which is used to physically constrain attention weights of the neural network model by introducing a physical guided attention mechanism, gradually reduce dimensions of the target point cloud features by a multi-layer perception machine, establish a mapping relationship between the target point cloud features and wind resistance coefficients by a fully connected layer and finally output a wind resistance coefficient prediction result.

[0017] Compared with the prior art, the present application has the following beneficial effects:

[0018] (1) After the point cloud data of the whole vehicle model is acquired and pre-processed, the pre-processed point cloud data is sampled by the farthest point sampling method, so that the sampling points are as evenly distributed as possible to cover the entire point cloud space, to retain geometric key points, enhance data information representativeness and ensure integrity of the vehicle geometric structure, so as to provide basic feature representation for subsequent multi-scale convolution.

[0019] (2) The neural network model is established, and the point cloud data is input into the neural network model for wind resistance coefficient prediction, which can quickly process the task of unordered point cloud.

[0020] (3) The point cloud data is mapped from low-dimensional features to 512 dimensions and expanded to 1024 dimensions by two multi-scale convolution operations, which greatly improves the expression ability of the vehicle point cloud data features, provides richer feature representation for subsequent tasks, improves the accuracy of model prediction, extracts local detail features and global structure features of the point cloud data and performs feature aggregation to obtain target point cloud features, solves the problem that the current wind resistance prediction method cannot consider both local detail features and global structure features, and enhances the diversity and robustness of feature representation.

[0021] (4) The target point cloud feature is gradually reduced in dimension by a multi-layer perception machine, the most significant feature information of a model is reserved, a physical guided attention mechanism is introduced to physically constrain the attention weight of the neural network model, the reliability of the model prediction is improved, then a mapping relationship between the target point cloud feature and the wind resistance coefficient is established through a full connection layer, and finally a wind resistance coefficient prediction result is output, so that end-to-end prediction is realized. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 It is a schematic diagram of the method steps of the application.

[0023] Figure 2 It is a schematic diagram of the intelligent prediction structure of the wind resistance coefficient of the system of the application.

[0024] Figure 3 It is a schematic diagram of the method framework of the application.

[0025] Figure 4 It is a schematic diagram of the physical guided attention mechanism module of the application.

[0026] Figure 5 It is a schematic diagram of the multi-layer perception machine module of the application.

[0027] Figure 6 It is a schematic diagram of the feature dimension compression structure of the application.

[0028] Figure 7 It is a schematic diagram of the system structure of the application.

[0029] In the figure: 10-point cloud acquisition module, 20-farthest point sampling module, 30-neural network module, 40-wind resistance coefficient prediction module. DETAILED DESCRIPTION

[0030] The schemes of the application will be clearly and completely described below by combining the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments.

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

[0032] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this 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 groups thereof. 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 the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0033] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0034] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0035] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

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

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

[0038] Specifically, this application preprocesses the point cloud data of the whole vehicle model after obtaining it, converting the raw point cloud data into a standardized format that can be processed by the network model.

[0039] S200. The preprocessed point cloud data is sampled 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.

[0040] Specifically, the farthest point sampling method is used to sample the preprocessed point cloud data, so that the sampling points can uniformly cover the entire point cloud space as much as possible, so as to retain geometric key points, enhance data information representation, and ensure the integrity of the automobile geometric structure, thereby providing a basic feature representation for subsequent multi-scale convolution.

[0041] S300, a neural network model is established, the sampled point cloud data is input into the neural network model, two multi-scale convolution operations are performed by using different scale convolution kernels, the sampled point cloud data is mapped from low-dimensional features to high-dimensional features, and local detail features and global structure features of the point cloud data after the multi-scale convolution operation are extracted and aggregated to obtain target point cloud features.

[0042] Specifically, the neural network model is established, and the point cloud data is input into the neural network model for wind resistance coefficient prediction, which can quickly process the task of unordered point cloud. The point cloud data is mapped from low-dimensional features to 512-dimensional features and expanded to 1024-dimensional features through two multi-scale convolution operations, which greatly improves the expression ability of the automobile point cloud data features, provides richer feature representation for subsequent tasks, improves the accuracy of model prediction, extracts local detail features and global structure features of the point cloud data and aggregates the features to obtain target point cloud features, solves the problem that the current wind resistance prediction method cannot consider local detail features and global structure features, and enhances the diversity and robustness of feature representation.

[0043] S400, the attention weight of the neural network model is physically constrained by introducing a physical guided attention mechanism, and the target point cloud features are gradually reduced in dimension by a multi-layer perceptron, a mapping relationship between the target point cloud features and the wind resistance coefficient is established through a fully connected layer, and finally a wind resistance coefficient prediction result is output.

[0044] Specifically, the attention weight of the neural network model is physically constrained by introducing a physical guided attention mechanism, which guides the network to focus on the key areas that affect the wind resistance coefficient, improves the reliability of the model prediction, and then gradually reduces the target point cloud features in dimension by a multi-layer perceptron, retains the most significant feature information of the model, establishes a mapping relationship between the target point cloud features and the wind resistance coefficient, and finally outputs a wind resistance coefficient prediction result, realizing end-to-end prediction.

[0045] Optionally, the point cloud data of the whole vehicle model is obtained and preprocessed; the farthest point sampling method is used to sample the preprocessed point cloud data, so as to retain geometric key points, enhance the representativeness of point cloud data information, and maintain the integrity of the automobile structure; a convolution-based neural network model is constructed, and the point cloud data is input into the neural network model:

[0046] Firstly, local feature data is extracted by two multi-scale convolution operations, (using 1x3 and 3x3 scale convolution kernels respectively), to extract local detail features of point cloud data in different local fields and gradually increase the feature dimension to 1024 dimensions to form a fusion feature representation; then, two single-scale convolutions are used to further enhance the feature expression ability, and the feature dimension is maintained at 1024; then, a max-pooling operation is used to realize global feature aggregation, and a multi-layer perception (MLP) is used to gradually reduce the dimension of the features; at the same time, a physically guided attention mechanism is introduced to impose physical constraints on the spatial feature weights in the neural network, guiding the model to focus on key structural areas that significantly affect the drag coefficient; finally, a fully connected layer is used to establish a mapping relationship between the point cloud features and the drag coefficient, and the drag coefficient prediction result is output.

[0047] It needs to be understood that, since the shape and structure of the car have multi-scale features when predicting the car's drag coefficient, for example, the overall streamlined profile of the car is a macro-scale feature, i.e., a global structural feature; while the tiny bumps and dents on the surface of the car body, rearview mirrors, door handles, etc. are micro-scale features, i.e., local detail features. For existing drag prediction models using neural networks, it is difficult to capture these different scale features at the same time. It may focus on the overall shape and ignore the impact of local details on drag, or vice versa, resulting in inaccurate prediction of the drag coefficient. At the same time, in the prediction of car drag, the change of local airflow is often closely related to the local geometric features of the car surface, for example, the local shape near the front bumper and air intake grille of the car will significantly affect the airflow in that area, and thus affect the drag. Existing drag prediction models using neural networks do not have a mechanism specifically 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 that integrates macroscopic car body streamline features and microscopic surface detail features, there are deficiencies in dealing with complex relationships between car shape and drag. Therefore, based on the neural network model architecture and using different scale convolution kernels to perform convolution operations on the input point cloud feature data, the application can achieve multi-level feature extraction, capturing both local details and global structures to enhance the richness of feature representation and improve feature relevance and prediction accuracy. At the same time, the introduction of feature visualization and physical constraints enhances the model's interpretability.

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

[0049] Specifically, the preprocessing of the point cloud data is mainly used for standardizing the point cloud data, converting the point cloud data into a standardized format that can be processed by the model, to ensure the stability and convergence of the model training. The core function mainly includes: 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 difference of different vehicle point cloud data, thereby accelerating the model convergence, making the optimization process smoother, reducing the training time, and avoiding gradient explosion or disappearance problem, improving the model robustness and generalization ability, to enhance the adaptability of the model to the scale change of the input data, and reduce the influence of noise and abnormal values.

[0050] In some embodiments, the step of sampling the preprocessed point cloud data by the farthest point sampling method specifically includes: performing farthest point sampling on the preprocessed point cloud data, constructing a sampling strategy by iteratively selecting the point farthest from the selected point set, thereby obtaining a key point cloud set that maintains the geometric details of the vehicle surface and has uniform spatial distribution, and the calculation formula is: wherein, is the next point, is the maximum value, is the minimum value, is the candidate point, is the selected point, is the candidate point set, is the selected point set, is the Euclidean distance between the candidate point and the selected point.

[0051] Specifically, by the farthest point sampling method, the preprocessed point cloud data is sampled, which preferentially captures the geometric significant regions such as edges and corners of the vehicle surface, such as the contour line of the vehicle body, the edge of the vehicle lamp, etc., which can retain the key information of the geometric features of the vehicle surface, in order to provide basic feature representation for subsequent multi-scale convolution; at the same time, it has good spatial distribution uniformity, ensuring that the sampling points are uniformly distributed in the entire point cloud space, retaining local features and covering global shapes.

[0052] In some embodiments, the step of mapping the sampled point cloud data from low-dimensional features to high-dimensional features by using different scale convolution kernels includes: simultaneously performing multi-scale convolution operation on the sampled point cloud data by using 1x3 and 3x3 convolution kernels, to expand the neural network width, map the features from low-dimensional to 512-dimensional and further expand to 1024-dimensional, then further enhance the feature expression ability by two layers of convolution, map the point cloud data from low-dimensional features to 512-dimensional, then continuously use multi-scale convolution operation to expand the features to 1024-dimensional; wherein the calculation formula for simultaneously performing multi-scale convolution operation on the point cloud data by using 1x3 and 3x3 convolution kernels is: , wherein is a convolution kernel with a size of is a convolution result, is point cloud data.

[0053] Specifically, the 1x3 convolution kernel is used to extract the fine structure of the vehicle surface in the point cloud data, and the 3x3 convolution kernel is used to cover a wider neighborhood to capture the overall shape of the vehicle and the relationship between the components; the point cloud data is mapped from a low-dimensional feature to 512 dimensions, and then the feature is expanded to 1024 dimensions by using a multi-scale convolution operation in succession, which can better capture local details and global structure, improve the richness of the feature, greatly enhance the expression ability of the point cloud data feature, and improve the prediction accuracy of the model.

[0054] In some embodiments, the step of performing two multi-scale convolution operations further comprises: performing batch normalization processing on the point cloud data after the multi-scale convolution operation, respectively, and the calculation formula is: , wherein y is the batch normalization result, x is the point cloud data after the multi-scale convolution operation, is the mean of the batch data, is the variance of the batch data, and are learnable parameters, is a small constant; a nonlinear conversion operation is introduced by using a ReLU activation function, and the calculation formula is: , wherein is a linear rectifier activation function, x is the point cloud data after the multi-scale convolution operation, is the maximum value of 0 and x; a part of neurons is randomly inactivated, and the calculation formula is: , wherein y is the random inactivation result, x is the point cloud data after the multi-scale convolution operation, and p is a probability parameter.

[0055] ​Specifically, in addition to the multi-scale convolution operation, in order to enhance the expression ability of the model, accelerate the convergence and prevent overfitting, batch normalization, ReLU activation function nonlinear conversion and dropout operation are added after the convolution operation; the batch normalization operation can standardize the input of each layer during training, ensure that the input features are more balanced in each layer, thereby reducing the "internal covariate shift" in training, thereby accelerating the training, improving the model performance and stability. At the same time, by using the batch normalization operation, the dependence of the model on specific samples can be reduced to avoid overfitting as much as possible and make the model more stable during training and reduce the problem of gradient vanishing or explosion. The nonlinear feature conversion operation, that is, the ReLU (Rectified Linear Unit) activation function, captures the complex patterns in the image and introduces nonlinear features, which is conducive to learning more complex features and preventing gradient vanishing problems, improving the training speed and performance of the network. Random inactivation operation is used to randomly inactivate part of the neurons to prevent overfitting. Therefore, the feature data after the above four operations can make the model have faster convergence speed, more stable training process, higher accuracy and generalization ability, and effectively avoid the problems of gradient vanishing and explosion, and enhance the expression ability of the model.

[0056] In some embodiments, the step of simultaneously extracting the local detail features and the global structure features of the point cloud data after the multi-scale convolution operation and performing feature aggregation specifically comprises: simultaneously extracting the local detail features and the global structure features and splicing them into a new feature matrix, and the calculation formula is: , wherein, is a multi-scale feature matrix, is a splicing operation, is a feature matrix; wherein the output feature dimension of the first multi-scale convolution is 512, and the feature matrix is , the feature dimension output after two multi-scale convolutions is 1024, and the feature matrix is ; the multi-scale feature matrix is subjected to a max-pooling operation to extract global features, and the output dimension after pooling is 1024, and the calculation formula is: , wherein, is a feature value in a local region, is a global feature value.

[0057] Specifically, the local detail features and the global structure features are simultaneously extracted and spliced into a new feature matrix, which can take into account the local detail information and the global structure information to form a more comprehensive feature representation, so that the model can pay attention to details and take into account the overall structure, avoiding the limitations of single features.

[0058] In some embodiments, the step of gradually reducing the dimensionality of the target point cloud feature by the multi-layer perception machine specifically comprises: the multi-layer perception machine comprises 5 fully connected layers, each fully connected layer sequentially performs batch normalization, ReLU activation function nonlinear conversion operation and random inactivation, gradually maps the dimensionality of the global feature vector from 1024 dimensions to 512 dimensions, 256 dimensions, 128 dimensions, 64 dimensions, and finally outputs a 1-dimensional wind resistance coefficient value, and the calculation formula is: , wherein, is the output feature of the multi-layer perception machine, is an activation function, is a weight matrix, is a global feature, is a bias term.

[0059] Specifically, the multi-layer perception machine MLP is responsible for gradually reducing the dimensionality of high-dimensional features and mapping to the final wind resistance coefficient value. Through multiple nonlinear transformations, it extracts high-order features in the point cloud and enhances the expression ability of the model. The 1024-dimensional high-dimensional feature is gradually reduced to a 64-dimensional low-dimensional space through five multi-layer perception machine MLP operations, and the most significant feature information of the model is retained.

[0060] In some embodiments, the step of establishing a mapping relationship between the target point cloud feature and the wind resistance coefficient by the fully connected layer specifically comprises: adopting a fully connected mapping strategy, mapping the target point cloud feature to the wind resistance coefficient value in aerodynamics through the fully connected layer, and the calculation formula is: , wherein, is the wind resistance coefficient, is a weight matrix, is the output feature of the multi-layer perception machine, is a bias term.

[0061] Specifically, the fully connected layer maps the target point cloud feature to the wind resistance coefficient Cd value, which can realize end-to-end prediction.

[0062] The present application can effectively improve the prediction accuracy and enhance the robustness of the model. In terms of economic effect, it has high optimization efficiency, can effectively reduce the optimization period, shorten the listing time, and reduce the time, economic and other costs of research and development optimization. At the same time, compared with traditional CFD, it can effectively reduce the hardware cost and improve the economic benefit. In terms of application effect, it effectively predicts the wind resistance coefficient of the automobile, which can effectively improve the fuel economy and other performances of the automobile.

[0063] In some embodiments, the step of introducing a physical guided attention mechanism to physically constrain the attention weight of the neural network model specifically comprises: introducing a simplified form of Navier-Stokes equation as a soft constraint, and the calculation formula is: , wherein, for the correction term, and for the weight coefficient, for the redundancy parameter, for the Reynolds number, for the wind resistance coefficient, for the far-field fluid density, for the local fluid density, for the fluid velocity.

[0064] Specifically, the step of introducing the physical guided attention mechanism specifically includes: obtaining the whole vehicle point cloud feature information extracted by the neural network model; based on the geometric position relationship of each point in the point cloud, calculating an original spatial attention map to depict the preliminary attention degree of each region in the model; introducing physical prior information constructed based on a simplified form of Navier-Stokes equation, wherein the physical prior information includes local pressure gradient and velocity gradient on the surface of the whole vehicle to reflect the possible aerodynamic influence of each region of the vehicle under the action of air flow; constructing a physical guiding function to weight and correct the original spatial attention map, so that the regions concerned by the model are more consistent with the aerodynamics principle, and focus on the key structures such as the vehicle head, roof and tail that have significant influence on the wind resistance coefficient; output the attention map corrected by the physical guidance, and fuse it with the target point cloud feature to further enhance the feature expression ability and use it for the subsequent wind resistance coefficient regression prediction task. By introducing the simplified form of Navier-Stokes equation as a soft constraint, the model can be guided to focus on the regions that conform to the physical law, for example, in the wind resistance prediction task, the model will automatically focus on the high-curvature regions and windward area on the surface of the vehicle body, and ignore irrelevant regions; associating the attention weight with the physical quantity can make the regions concerned by the model have clear physical meaning and enhance the interpretability of the model.

[0065] In another aspect, the present application also provides a whole vehicle point cloud wind resistance coefficient prediction system based on multi-scale learning and convolution, which is used to implement the whole vehicle point cloud wind resistance coefficient prediction method based on multi-scale learning and convolution, and comprises: a point cloud acquisition module, which is used to acquire point cloud data of a whole vehicle model and pre-process the point cloud data; a farthest point sampling module, which is used to sample the pre-processed point cloud data by a farthest point sampling method to retain geometric key points, enhance information representativeness of the point cloud data and maintain integrity of the vehicle structure; a neural network module, which is used to establish a neural network model, input the sampled point cloud data into the neural network model, perform two multi-scale convolution operations by using different scale convolution kernels, map the sampled point cloud data from low-dimensional features to high-dimensional features, extract local detail features and global structure features of the point cloud data after the multi-scale convolution operations and perform feature aggregation to obtain target point cloud features; and a wind resistance coefficient prediction module, which is used to physically constrain attention weights of the neural network model by introducing a physical guided attention mechanism and gradually reduce dimensions of the target point cloud features by a multi-layer perception machine, establish a mapping relationship between the target point cloud features and the wind resistance coefficient by a fully connected layer and finally output a wind resistance coefficient prediction result.

[0066] The integrated unit, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such an understanding, the technical solutions of the present application, essentially or in part, or all or part of the technical solutions 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 can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various other media that can store program codes.

[0067] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing relevant hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the computer program can include the processes of the above-mentioned embodiment methods. Any reference to memory, storage, database or other medium used in the embodiments of the present application can include non-volatile and / or volatile memory. The non-volatile memory can include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. The volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory 10 bus dynamic RAM (RDRAM), etc.

[0068] The above merely describes the embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent transformation or direct or indirect application in the related technical field using the content of the present application specification and drawings is also included in the patent protection scope of the present application.

Claims

1. A whole vehicle point cloud wind resistance coefficient prediction method based on multi-scale learning and convolution, characterized by the steps of The method comprises the following steps: acquiring point cloud data of a whole vehicle model and preprocessing the point cloud data; sampling the preprocessed point cloud data by a farthest point sampling method; establishing a neural network model, inputting the sampled point cloud data into the neural network model, performing two multi-scale convolution operations by using different scale convolution kernels, mapping the sampled point cloud data from low-dimensional features to high-dimensional features, simultaneously extracting local detail features and global structure features of the point cloud data after the multi-scale convolution operation and performing feature aggregation to obtain target point cloud features; the step of simultaneously extracting local detail features and global structure features and splicing them into a new feature matrix comprises the following steps: simultaneously extracting local detail features and global structure features and splicing them into a new feature matrix; wherein the output feature dimension of the first multi-scale convolution is 512 dimensions, and the feature dimension output after two multi-scale convolutions is 1024 dimensions; performing a maximum pooling operation on the multi-scale feature matrix to extract a global feature vector, and the output dimension after the pooling is 1024 dimensions; introducing a physical guided attention mechanism to physically constrain the attention weight of the neural network model, and the step of introducing the physical guided attention mechanism comprises the following steps: acquiring the whole vehicle target point cloud features extracted by the neural network model; calculating an original spatial attention map based on the geometric position relationship of each point in the point cloud; introducing physical prior information constructed based on a simplified form of the Navier-Stokes equation, wherein the physical prior information comprises local pressure gradient and velocity gradient on the surface of the whole vehicle; constructing a physical guide function to weight and correct the original spatial attention map; outputting the attention map after the physical guide correction, fusing it with the target point cloud features, and using it for wind resistance coefficient regression prediction task; and gradually reducing the dimension of the target point cloud features by a multilayer perception machine, and then establishing a mapping relationship between the target point cloud features and the wind resistance coefficient by a fully connected layer, and the step of establishing the mapping relationship between the target point cloud features and the wind resistance coefficient by the fully connected layer comprises the following steps: adopting a fully connected mapping strategy to map the target point cloud features to the wind resistance coefficient value in aerodynamics through the fully connected layer, and finally outputting the wind resistance 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 comprises: normalizing the point cloud data to scale the coordinates to the range of [0, 1].

3. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, characterized in that, The step of sampling the preprocessed point cloud data by the farthest point sampling method comprises the following steps: performing farthest point sampling on the preprocessed point cloud data, constructing a sampling strategy by iteratively selecting the point farthest from the selected point set, thereby obtaining a key point cloud set that maintains the geometric details of the vehicle surface and has uniform spatial distribution.

4. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, characterized in that, The step of using different scale convolution kernels to perform two multi-scale convolution operations to map the sampled point cloud data from low-dimensional features to high-dimensional features comprises the following steps: adopting 1x3 and 3x3 convolution kernels to simultaneously perform multi-scale convolution operations on the sampled point cloud data, mapping the features from low-dimensional to 512-dimensional and expanding to 1024-dimensional.

5. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, characterized in that, The step of performing two multi-scale convolution operations further comprises: The point cloud data after multi-scale convolution operation is respectively subjected to batch normalization processing, and the calculation formula is: where y is the batch normalization result, x is the point cloud data after the multi-scale convolution operation, is the mean of the batch data, is the variance of the batch data, and are learnable parameters, is a small constant; Nonlinear conversion operation is introduced through ReLU activation function, and the calculation formula is: , wherein, is a linear rectified activation function, x is the point cloud data after the multi-scale convolution operation, is the maximum value of 0 and x; Part of neurons is randomly inactivated, and the calculation formula is: , where y is the result of random inactivation, x is the point cloud data after the multi-scale convolution operation, and p is a probability parameter.

6. The vehicle point cloud drag coefficient prediction method based on multi-scale learning and convolution according to claim 1, characterized in that, The step of gradually reducing the dimension of the target point cloud feature through the multi-layer perception comprises: The multi-layer perception comprises five fully connected layers, each fully connected layer sequentially performs batch normalization, ReLU activation function nonlinear conversion operation and random inactivation, and gradually maps the dimension of the global feature vector from 1024 dimensions to 512 dimensions, 256 dimensions, 128 dimensions and 64 dimensions, and finally outputs a 1-dimensional wind resistance coefficient value.

7. A multi-scale learning and convolution based whole vehicle point cloud drag coefficient prediction system for implementing the multi-scale learning and convolution based whole vehicle point cloud drag coefficient prediction method according to any one of claims 1-6, characterized in that, Comprise: The point cloud acquisition module is used for acquiring point cloud data of a whole vehicle model and pre-processing the point cloud data; The farthest point sampling module is used for sampling the pre-processed point cloud data through a farthest point sampling method; The neural network module is used for establishing a neural network model, inputting the sampled point cloud data into the neural network model, performing two multi-scale convolution operations through different scale convolution kernels, mapping the sampled point cloud data from low-dimensional features to high-dimensional features, extracting local detail features and global structure features of the point cloud data after multi-scale convolution operation and performing feature aggregation, and obtaining a target point cloud feature; The wind resistance coefficient prediction module is used for physically constraining attention weights of the neural network model by introducing a physical guided attention mechanism, gradually reducing the dimension of the target point cloud feature through a multi-layer perception, establishing a mapping relationship between the target point cloud feature and the wind resistance coefficient through a fully connected layer, and finally outputting a wind resistance coefficient prediction result.

Images