A method and system for predicting the drag contribution of a car rear wing design

Abstract

This invention proposes a method and system for predicting the drag contribution of automotive rear wing designs, belonging to the fields of automotive aerodynamics and artificial intelligence. The method includes: acquiring and preprocessing raw CFD flow field images to generate standardized binary images; extracting flow field depth and tilt angle features from the binary images, using them together as multi-source features characterizing the wake field structure; constructing a deep learning model, using the multi-source features as input, training the deep learning model, and outputting scalar parameters; inputting the flow field image of the automotive rear wing design to be predicted into the trained deep learning model, extracting multi-source features, and combining them with changes in target design parameters to quickly predict the drag contribution using a quadratic function. This invention significantly improves prediction efficiency and has strong generalization ability, providing accurate and efficient decision-making basis for automotive rear wing drag optimization design, and is suitable for rapid iteration and parameter scanning in automotive rear wing design.

Description

Technical Field

[0001] This invention belongs to the fields of automotive aerodynamics and artificial intelligence, and particularly relates to a method and system for predicting the drag contribution of automotive rear wing designs. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In automotive research and development, aerodynamic performance directly affects a vehicle's top speed and energy consumption. The drag coefficient is a core evaluation indicator, and the rear wing design is a key aspect of optimizing drag.

[0004] Currently, the industry mainly relies on high-fidelity computational fluid dynamics (CFD) numerical simulations to evaluate the impact of tail fin design on wind resistance. However, a complete simulation takes several hours to several days, resulting in high computational costs and making it difficult to support the rapid iteration and scanning of a large number of parameters during the design phase. Existing AI-based accelerated prediction methods also have significant shortcomings, as they are mostly purely data-driven, lack the embedding of aerodynamic physics principles, and have poor generalization capabilities.

[0005] Furthermore, existing models target a single scenario and typically predict drag coefficient values ​​directly, rather than the drag difference between two shapes. They cannot accurately calculate the drag increment caused by changes in design parameters, nor can they extract physical features from CFD flow field images to guide the design, making it difficult to meet the actual R&D needs of automotive rear wing drag optimization. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method and system for predicting the drag contribution of automotive rear wing designs, which addresses the technical problems of low efficiency in traditional CFD simulation, poor generalization ability of existing models, lack of physical interpretability, and insufficient design guidance.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for predicting the drag contribution of a car rear wing design; A method for predicting the drag contribution of a car rear wing design includes: The original CFD flow field image is acquired and preprocessed to generate a standardized binary image; For each binary image, a convolutional neural network is used to automatically extract the flow field depth features. At the same time, geometric analysis is used to extract the tilt angle features of the top edge of the main wake region. The depth features and the tilt angle features are used together as multi-source features to characterize the wake field structure. Construct a deep learning model, take the multi-source features as model input, train the deep learning model, and output a scalar parameter of a quadratic function; The flow field image of the car rear wing design to be predicted is input into the trained deep learning model, its multi-source features are extracted, and combined with the change of the target design parameters, the drag contribution is quickly predicted through the quadratic function.

[0008] As a further technical solution, the original CFD flow field image is acquired and preprocessed, including: Automatically detect and remove irrelevant information and color bars at the bottom of the image; Identify the vehicle body region and crop out the region containing the complete wake using the right boundary of the vehicle body as a reference; generate a binary image that highlights the vortex structure by enhancing the blue channel and thresholding. The corresponding aerodynamic parameters are automatically parsed based on the image file name information, and training samples are generated through systematic pairing.

[0009] As a further technical solution, the tilt angle characteristics of the top edge of the main wake region are extracted through geometric analysis, including: Identify the largest connected region in the binary image as the main wake region, traverse its right three-fifths along the horizontal direction, record the vertical coordinates of the first foreground pixel in each column, and form an edge point set. A linear fit is performed on the set of edge points, the inclination angle of the fitted line is calculated, and it is normalized to the interval [-1, 1] as a feature.

[0010] As a further technical solution, the structure of the constructed deep learning model includes: The convolutional neural network part consists of multiple convolutional layers and pooling layers, which are used to extract image depth features and obtain image feature vectors through global pooling. The feature fusion layer is used to concatenate the image feature vector with the tilt angle feature of the top edge of the main wake region into a comprehensive feature vector. A fully connected network is used to process the comprehensive feature vector and output the scalar parameters of the quadratic function.

[0011] As a further technical solution, the training process of the deep learning model also includes the construction of training samples, which include: Using flow field images of different rear wing designs of the same vehicle model as a pair, the difference in rear wing height between them is calculated as one of the input features, and the difference in drag coefficient is used as the prediction label; symmetrical samples are generated by swapping the order of the image pairs to expand the training dataset.

[0012] As a further technical solution, the quadratic function is used to describe the physical relationship between the change in tail fin design parameters and the drag contribution:

[0013] in, Contribution to wind resistance; This refers to the distance the tail fin deflects downwards; , It is a scalar parameter.

[0014] The second aspect of this invention provides a system for predicting the drag contribution of a car rear wing design.

[0015] A system for predicting the drag contribution of a car rear wing design includes: The data preprocessing module is configured to: acquire the original CFD flow field image and preprocess it to generate a standardized binary image; The feature extraction module is configured to: automatically extract flow field depth features using a convolutional neural network for each binary image, and simultaneously extract the tilt angle features of the top edge of the main wake region through geometric analysis, and use the depth features and the tilt angle features together as multi-source features characterizing the wake field structure. The model building and training module is configured to: build a deep learning model, use the multi-source features as model input, train the deep learning model, and output a scalar parameter of a quadratic function; The prediction application module is configured to: input the flow field image of the car rear wing design to be predicted into the trained deep learning model, extract its multi-source features, and combine them with the change in the target design parameters to quickly predict the wind resistance contribution through the quadratic function.

[0016] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the method for predicting the drag contribution of a car rear wing scheme as described in the first aspect of the present invention.

[0017] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the method for predicting the drag contribution of a car rear wing scheme as described in the first aspect of the present invention.

[0018] The fifth aspect of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the method for predicting the drag contribution of a car rear wing scheme as described in the first aspect of the present invention.

[0019] The above one or more technical solutions have the following beneficial effects: (1) This method abandons the traditional time-consuming CFD numerical simulation method and relies on the trained deep learning model to quickly predict the contribution of wind resistance. Only one forward propagation of the neural network is needed to obtain the result, which can greatly shorten the time for wind resistance evaluation after the tail wing design parameters are adjusted. It can efficiently support the rapid scanning and iterative testing of a large number of variables in the design stage, while saving a lot of computing resources and reducing the manpower and time costs related to CFD simulation, allowing designers to focus more on the optimization of the scheme itself.

[0020] (2) The model used in this invention does not directly output the drag value, but learns the quadratic function parameters that characterize the relationship between tail wing design parameters and drag contribution. It integrates the basic principles of aerodynamics into the modeling process, so that the prediction results naturally conform to the physical change trend. It effectively avoids the problem that pure data-driven models are prone to violating physical laws outside the data distribution, and improves the generalization ability of the model under different design schemes. No matter how the baseline design is adjusted, it can reliably predict the drag increment caused by parameter changes, and is more practical for guiding actual design.

[0021] (3) By fusing the depth features of the flow field image with the geometric boundary features with clear physical meaning for prediction, it can not only quantitatively assess the changes in wind resistance, providing engineers with objective and accurate decision support for deciding whether to implement optimization schemes, thus getting rid of excessive reliance on human experience judgment and reducing subjective errors; it can also extract core aerodynamic features such as wake vortex structure and airflow separation state from the flow field image, allowing designers to clearly grasp the influence of design parameter changes on the wake field, providing multiple reference bases for the accurate and scientific optimization design of the car rear wing, and helping to improve the aerodynamic performance of the whole vehicle.

[0022] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0024] Figure 1 This is a flowchart of the method in the first embodiment.

[0025] Figure 2 This is a system structure diagram of the second embodiment. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] The overall approach of this invention addresses the problems of low efficiency in drag assessment and lack of physical interpretability in existing models during automotive rear wing design. By integrating automotive aerodynamics, computer vision, and machine learning techniques, it extracts aerodynamically significant geometric features from CFD flow field images and combines these with deep features automatically learned by convolutional neural networks to jointly characterize the wake field structure. A deep learning model is constructed, using multi-source features as input, and the model outputs a quadratic function whose parameters describe the physical relationship between changes in rear wing design parameters and drag contribution. Finally, this quadratic function is used to quickly predict the drag impact of arbitrary design changes. This method embeds physical laws into the model design, achieving efficient, accurate, and interpretable drag prediction.

[0030] Example 1 This embodiment discloses a method for predicting the drag contribution of a car rear wing design. The method involves preprocessing CFD flow field images and constructing a paired sample set. Then, depth features and geometric boundary features of the main wake region are extracted from the images. A deep learning model that integrates multi-source features is used to learn a quadratic function parameter representing the relationship between the tail wing's downforce and drag variation. Finally, based on this quadratic function, the flow field characteristics of the new design scheme and the tail wing's downforce parameters are input to quickly output the corresponding drag contribution, achieving an efficient and accurate assessment of the drag impact of the car rear wing design.

[0031] Specifically, such as Figure 1 As shown, a method for predicting the drag contribution of a car rear wing design includes: Step S1: Obtain the original CFD flow field image and perform preprocessing to generate a standardized binary image.

[0032] Raw CFD flow field images contain a large amount of irrelevant information and suffer from inconsistent sizes and formats, as well as low visual contrast of flow field features, making them unsuitable for direct use in training deep learning models. Furthermore, image data from a single design state cannot demonstrate the correlation between changes in design parameters and wind resistance. Therefore, a streamlined image preprocessing procedure is employed to generate standardized, feature-rich binary flow field images. Based on a strategy of pairing different tail wing design states for the same vehicle model, a standard training dataset is constructed using wind resistance changes as a supervisory signal, providing a high-quality, highly correlated data foundation for subsequent model training.

[0033] Five consecutive operations are performed on the original CFD flow field image: irrelevant information removal, vehicle body region identification and wake region cropping, size normalization, blue channel enhancement, and threshold segmentation to generate a binary image. The specific operations are as follows: The bottom of the original CFD flow field image typically contains information unrelated to the wake flow structure, such as color bars, simulation parameter annotations, and coordinate scales. This information can interfere with the accuracy of subsequent feature extraction. Therefore, we first analyze the image pixel values ​​to automatically detect pixel regions containing irrelevant information and set the pixel values ​​of these regions to the background color, thus completely removing the irrelevant information. In this embodiment, for the original CFD flow field image with a resolution of 3840×2160, the region at the bottom of the image with a y-coordinate ≥1341 is identified as an irrelevant information region and is directly cropped and removed.

[0034] Based on the pixel value distribution and contour features of the image, the Canny edge detection algorithm combined with a contour fitting algorithm is used to intelligently identify the car body region in the image and determine the right boundary pixel coordinates of the car body region. Using the right boundary of the car body as a reference, a rectangular region containing the complete wake field structure is cropped to the right of the image. This ensures that the cropped image only contains wake field information related to the rear wing's drag, eliminating irrelevant areas such as the front and sides of the car body. In this embodiment, the coordinates of the rectangular region cropped for the wake field are (x1=2494, y1=540, x2=3744, y2=1340), which completely covers the wake vortex structure and airflow separation area at the rear of the car.

[0035] Different CFD simulation projects and different vehicle models generate flow field images of varying sizes. To meet the input requirements of deep learning models, the cropped wake region images are uniformly scaled to a standard size of 256×256 pixels. The scaling process uses a bilinear interpolation algorithm to ensure that the geometric and texture features of the flow field images are not distorted, thus ensuring the comparability of features between different images.

[0036] In CFD flow field images, the vortex structure and airflow separation boundary in the wake region exhibit more pronounced pixel value differences in the blue channel. To enhance the visual contrast of the flow field features, single-channel enhancement processing is performed on the standardized RGB image. Specifically, the blue channel of the image is extracted, gamma correction is applied to the pixel values ​​of the blue channel to increase the brightness of the dark areas, and contrast stretching is applied to the pixel values, mapping them to the full range of [0, 255], making the difference between the vortex structure and the background more significant.

[0037] For the enhanced blue channel image, the OTSU method is used to automatically calculate the optimal segmentation threshold, dividing the image into foreground and background to generate a binary flow field image. The foreground represents the critical wake region, while the background represents other regions. The OTSU method automatically determines the optimal threshold based on the image's grayscale distribution, avoiding the subjectivity of manually setting thresholds and ensuring consistent segmentation standards across different images. The generated binary image accurately highlights the core geometric structure of the wake field.

[0038] After the above five operations, the original CFD flow field image is converted into a standardized 256×256 pixel binary flow field image. This image removes all irrelevant information, and highlights the wake vortex structure and airflow separation boundary features, which can be directly used for subsequent feature extraction operations.

[0039] Furthermore, the correlation between tail fin design parameter changes and drag changes is established. By systematically pairing labeled CFD flow field images, training samples are generated with drag coefficient differences as supervision signals and tail fin design parameter changes, binary flow field images, and features as inputs. The symmetry of the samples is used to expand the dataset size, and finally a standard and efficient training sample set is constructed.

[0040] In this embodiment, the filenames of all original CFD flow field images adopt standardized naming rules and include key information such as vehicle model number, tail wing downforce distance, and drag coefficient Cd value. Through string parsing algorithm, the aerodynamic parameters such as vehicle model number, tail wing downforce distance x_raw, and drag coefficient Cd corresponding to each design state are automatically extracted from the image filenames and stored as structured data (Pandas DataFrame) to facilitate subsequent sample pairing.

[0041] Different car models have different rear wing design benchmarks, and the range of values ​​for the rear wing deflection distance varies. In order to eliminate the scale differences between car models, the extracted original rear wing deflection distance x_raw is linearly normalized and mapped to the interval [-1,1] to obtain the normalized rear wing design parameter change x.

[0042] Based on vehicle model number, binary flow field images of different rear wing design states for the same vehicle model are paired up, with each pair corresponding to a baseline design state and a modified design state. For each pair of images (denoted as image A and image B), the corresponding normalized rear wing downforce is extracted. , and drag coefficient , The changes in tail fin design parameters Δx and the difference in drag coefficient ΔCd are calculated and used as supervision signals for model training.

[0043] For each pair of paired images, a training sample is generated using the binary flow field image of the baseline design state, the features of that image, and the change in tail fin design parameters Δx as model inputs, and the drag coefficient difference ΔCd as the label. The input of the sample corresponds one-to-one with the label, ensuring that the model can learn the correlation between changes in design parameters and changes in drag.

[0044] Because image pairing is symmetrical, exchanging the baseline and changed states of each pair of images generates new training samples. Expanding the training sample set through symmetrical samples doubles its size, fully leveraging the value of existing data. Simultaneously, it allows the model to learn the wind resistance variation patterns corresponding to positive and negative changes in design parameters, thus improving the model's generalization ability.

[0045] All generated training samples were randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used for iterative parameter updates of the deep learning model; the validation set was used for performance evaluation during model training, monitoring for overfitting, and adjusting model hyperparameters based on the validation set's performance; the test set was used to objectively evaluate the model's final prediction accuracy and generalization ability after training was complete. All datasets were stored in TFRecord format to improve data retrieval efficiency during model training.

[0046] After the above operations, the constructed standard training sample set has the characteristics of sufficient sample quantity, close correlation between input and label, standardized data format, and wide coverage of vehicle models. It can provide high-quality data support for the training of subsequent deep learning models and ensure that the model learns the real physical relationship between the changes in tail wing design parameters and the contribution of wind resistance.

[0047] Step S2: For each binary image, a convolutional neural network is used to automatically extract the flow field depth features. At the same time, the tilt angle features of the top edge of the main wake region are extracted through geometric analysis. The depth features and the tilt angle features are used together as multi-source features to characterize the wake field structure.

[0048] The core of this step is to extract multi-source features that accurately characterize the wake field structure from standardized binary flow field images. These features include wake field depth features based on convolutional neural networks (CNNs) and the tilt angle features of the top edge of the main wake region based on geometric analysis. These two types of features complement each other: depth features capture the overall structure and complex texture of the wake field, while geometric boundary features have clear aerodynamic physical significance, directly reflecting changes in airflow separation and wake vortex structure. Combining these features as model input significantly improves the model's prediction accuracy and physical interpretability. This step performs multi-source feature extraction on each binary flow field image generated in step 1. The extracted features correspond one-to-one with the image and are stored as feature vectors. Specifically, this process is divided into two sub-steps: depth feature extraction and geometric boundary feature extraction.

[0049] Step S21: A Convolutional Neural Network (CNN) is used to automatically learn the deep features in the binary flow field image. CNN has powerful local feature extraction and feature fusion capabilities, and can extract features from low-dimensional to high-dimensional layers of the image, such as edge and contour features at the bottom layer, vortex structure features in the middle layer, and overall wake field structure features at the top layer. These deep features can comprehensively and accurately characterize the complex structure of the wake field, which cannot be replaced by manual features. In this embodiment, ResNet18 is selected as the basic CNN network. ResNet18 effectively solves the gradient vanishing problem of deep networks through its residual connection structure, ensuring the effectiveness of feature extraction. At the same time, its network structure is simple, computationally efficient, and suitable for industrial applications. The specific extraction process is as follows: The standardized 256×256 pixel binary flow field image is converted into a tensor format with a shape of [1, 256, 256], and the pixel values ​​are normalized to map the pixel values ​​from [0, 255] to the [0, 1] interval to meet the input requirements of the ResNet18 network.

[0050] The preprocessed image tensor is input into the ResNet18 network, and features are extracted sequentially through convolutional layers, batch normalization layers, ReLU activation functions, and pooling layers. The convolutional layers use 3×3 kernels with a stride of 1 and padding of 1 to ensure the feature map size remains unchanged. The pooling layers use 2×2 max-pooling kernels with a stride of 2 to downsample the feature map and reduce computation. The ResNet18 contains four residual blocks, each consisting of two convolutional layers. After layer-by-layer processing, the binary flow field image is converted into a high-dimensional feature map with a shape of [512, 8, 8].

[0051] Global average pooling (GAP) is performed on the high-dimensional feature map output by the CNN, transforming the 8×8 feature map into 1×1 feature values, ultimately generating a one-dimensional depth feature vector of length 512. Global average pooling effectively preserves the global information of the feature map, avoids the overfitting problem of fully connected layers, and generates a feature vector with a moderate dimension, which is convenient for subsequent feature fusion and model training.

[0052] After the above process, each binary flow field image corresponds to a 512-dimensional depth feature vector. This vector contains core information such as the overall structure of the wake field, vortex distribution, and airflow separation boundary, and can comprehensively characterize the depth features of the wake field.

[0053] Step S22: Using geometric analysis, the tilt angle of the top edge of the main wake region is extracted from the binary flow field image. This feature has clear aerodynamic physical significance: the tilt angle of the top edge of the main wake region directly reflects the separation state of the airflow at the rear of the vehicle. The larger the angle, the more severe the airflow separation, the more unstable the wake vortex structure, and the higher the corresponding drag coefficient; conversely, the smaller the angle, the less severe the airflow separation, the more stable the wake field, and the lower the drag coefficient. This feature quantifies the aerodynamic characteristics of the wake field into specific numerical features, effectively improving the physical interpretability of the model. The specific extraction process is as follows: In a binary flow field image, the foreground region (pixel value 0, black) represents the wake region, which includes the main wake region and secondary vortex regions. A connected component analysis algorithm is used to detect connected components in the foreground region of the binary flow field image, calculating the area, circumscribed rectangle, center coordinates, and other geometric parameters of each connected component. The connected component with the largest area is identified as the main wake region, because the main wake region is the most significant airflow separation area at the rear of the vehicle, and its structural changes have the greatest impact on wind resistance; the influence of secondary vortex regions is negligible.

[0054] The main wake region is defined by using its circumscribed rectangle as the boundary. The right three-fifths of the main wake region is selected as the feature extraction region (this region is the core area of ​​the wake structure, where airflow separation is most pronounced). Within this feature extraction region, each pixel column is traversed horizontally (x-axis direction). For each column, the vertical coordinate (y-coordinate) of the first foreground pixel (pixel value 0) is recorded from the top (minimum y-axis value) to the bottom (maximum y-axis value). This pixel is the top edge point of the main wake region in that column. After the traversal, a set of top edge points containing multiple vertical coordinates is obtained, which accurately represents the geometry of the top edge of the main wake region.

[0055] For the extracted top edge point set, a least squares method is used for linear fitting to obtain a fitted straight line that characterizes the overall tilt trend of the top edge of the main wake region. The equation of the fitted line is: y = kx + c, where k is the slope of the fitted line and c is the intercept. The slope k directly reflects the degree of edge tilt. The tilt angle θ (in radians) of the fitted line is calculated based on the slope k using the formula: θ = arctan(k). The range of the tilt angle θ at the top edge of the main wake region varies in different binary flow field images. To meet the model input requirements, the tilt angle θ is linearly normalized and mapped to the interval [-1,1] to obtain the normalized tilt angle feature.

[0056] After the above process, each binary flow field image corresponds to a normalized tilt angle feature θ_norm (within the range of [-1,1]). This feature is a one-dimensional scalar feature with clear aerodynamic physical meaning and can directly reflect the airflow separation state in the main wake region.

[0057] Step S23 involves concatenating the one-dimensional tilt angle features extracted from the 512-dimensional depth feature vector to generate a multi-source feature vector of length 513. This vector simultaneously contains the depth features of the wake field and physically meaningful geometric boundary features, comprehensively and accurately characterizing the wake field structure and serving as the core input feature for subsequent deep learning models. The multi-source feature vector is stored using a floating-point array with a precision of 32-bit floating-point numbers to ensure the numerical accuracy of the features.

[0058] Step S3: Construct a deep learning model, use the multi-source features as model input, train the deep learning model, and output a scalar parameter of a quadratic function.

[0059] This step is the core of the invention, aiming to construct a deep learning model that integrates multi-source features and embeds quadratic function constraints. Through model training, the model learns the mapping relationship between multi-source features of the wake field, changes in tail fin design parameters, and quadratic function parameters. Ultimately, it outputs quadratic function scalar parameters a and b that accurately describe the relationship between changes in tail fin design parameters and drag contribution. Unlike existing purely data-driven AI models, the model of this invention does not directly predict the drag coefficient Cd value, but instead learns the parameters of the quadratic function. The drag contribution is calculated through the quadratic function, making the prediction results naturally conform to the physical laws of aerodynamics, significantly improving the model's physical interpretability and generalization ability. This step is specifically divided into three sub-steps: deep learning model structure design, model training hyperparameter setting, and model training and convergence judgment.

[0060] Step S31: The deep learning model constructed in this invention is an end-to-end model that integrates multiple input sources. The model structure is functionally divided into three parts: a CNN feature extraction layer, a feature fusion layer, and a fully connected network layer. The model's inputs are a standardized binary flow field image, multi-source feature vectors, and normalized tail fin design parameter variation Δx. The output is two scalar parameters a and b of a quadratic function, where the quadratic function is:

[0061] in, Contribution to wind resistance; This refers to the distance the tail fin deflects downwards; , It is a scalar parameter.

[0062] This quadratic function accurately describes the physical relationship between the tail fin's downward pressure and its contribution to wind resistance, conforming to the fundamental principle in aerodynamics that design parameters and drag coefficients exhibit a nonlinear quadratic relationship. The overall model structure is built using the PyTorch framework, and the specific design of each part is as follows: The CNN feature extraction layer uses the same ResNet18 network as in step S2 to extract deep features from the input standardized binary flow field image, outputting a 512-dimensional depth feature vector. This layer shares the network structure with the deep feature extraction network in step S2, but its parameters are updated independently during model training to ensure that the model can adaptively adjust the feature extraction parameters according to the characteristics of the training samples, thereby improving the accuracy of feature extraction.

[0063] The core function of the feature fusion layer is to concatenate the 512-dimensional depth feature vector output by the CNN feature extraction layer, the 1-dimensional tilt angle feature extracted by geometric analysis, and the normalized tail fin design parameter variation Δx to generate a comprehensive feature vector. In this embodiment, the normalized tail fin design parameter variation Δx is a one-dimensional scalar feature; therefore, it is concatenated with the tilt angle feature, and then with the depth feature vector to finally generate a comprehensive feature vector of length 514. The feature fusion layer uses a simple concatenation operation, which can retain the original information of each feature to the maximum extent, avoid information loss during the feature fusion process, and has high computational efficiency, making it suitable for industrial applications.

[0064] The fully connected network layer serves as the model's prediction layer. Its core function is to perform a nonlinear transformation on the 514-dimensional comprehensive feature vector output by the feature fusion layer, ultimately outputting two scalar parameters a and b of a quadratic function. The fully connected network layer adopts a three-layer fully connected structure, with the design of each layer as follows: The first fully connected layer has an input dimension of 514 and an output dimension of 256. The activation function is ReLU, which realizes non-linear transformation of features and improves the model's fitting ability. The second fully connected layer has an input dimension of 256 and an output dimension of 128. The activation function is ReLU, which further performs non-linear transformation on the features to explore the deep correlation between features. The third fully connected layer (output layer): The input dimension is 128, the output dimension is 2, there is no activation function, and it directly outputs two scalar values, which correspond to the parameters a and b of the quadratic function respectively.

[0065] Each layer of the fully connected network adds a Dropout layer and a batch normalization layer. The Dropout layer can randomly drop some neurons, effectively preventing the model from overfitting; the batch normalization layer can standardize the input of each layer, accelerating the convergence speed of the model and improving the training stability of the model.

[0066] The overall forward propagation process of the model includes inputting a standardized binary flow field image into a CNN feature extraction layer, which outputs a depth feature vector; the depth feature vector, along with the tilt angle feature and the change in tail fin design parameters, is concatenated into a comprehensive feature vector through a feature fusion layer; the comprehensive feature vector is input into a fully connected network layer, and after nonlinear transformation, it outputs quadratic function parameters a and b; finally, the drag contribution is calculated based on the quadratic function.

[0067] The hyperparameters used in model training directly affect the model's training efficiency, convergence speed, and final prediction accuracy. This implementation method determined the optimal combination of hyperparameters through extensive comparative experiments. All hyperparameters are visualized and monitored using TensorBoard for easy dynamic adjustment during training. The specific hyperparameter settings are as follows: The Adam optimizer was chosen as the parameter optimizer for the model. The Adam optimizer combines the advantages of momentum and adaptive learning rate methods, and can adaptively adjust the learning rate of each parameter. It has fast convergence speed, high training stability, and is suitable for end-to-end training of deep learning models.

[0068] The training objective of the model is to minimize the error between the predicted wind resistance contribution calculated from the quadratic function parameters a and b output by the model and the actual value. Therefore, mean squared error (MSE) is chosen as the model's loss function. MSE effectively measures the squared error between the predicted and actual values ​​and is a commonly used loss function for regression problems. The model is trained using mini-batch stochastic gradient descent (SGD) with a batch size of 32. This means that 32 samples are randomly selected from the training set each time to update the model parameters. Mini-batch training improves the model's generalization ability while ensuring training efficiency. The maximum number of iterations (epochs) is set to 200, meaning that all samples in the training set are traversed by the model 200 times, ensuring that the model can fully learn the correlation between features and labels.

[0069] To avoid oscillations in the later stages of model training that prevent convergence to the optimal solution, a step-decay learning rate strategy is adopted. When the model's loss value on the validation set no longer decreases for 10 consecutive epochs, the learning rate is multiplied by a decay factor of 0.5 until the learning rate drops to 1e-6, at which point the learning rate decay stops.

[0070] To prevent overfitting, an early stopping strategy is adopted. When the model's loss value on the validation set no longer decreases for 20 consecutive epochs, the model training is stopped immediately, and the model weights at this point are saved as the optimal model weights to avoid overfitting caused by continued model training.

[0071] Based on the model structure and hyperparameter settings described above, the model training process is executed. The training process consists of four stages: model initialization, iterative training on the training set, performance evaluation on the validation set, convergence determination, and model saving. All training processes are performed on GPUs, utilizing multi-GPU parallel computing to improve training efficiency. The specific process is as follows: All parameters of the constructed deep learning model are randomly initialized (using the Xavier initialization method) to ensure that the initial values ​​of the parameters follow a normal distribution, thus avoiding the vanishing or exploding gradient problem in the early stages of model training. The model is then set to training mode (model.train()) and the model parameters are loaded into the GPU.

[0072] The training set is divided into multiple training batches of size 32. For each training batch, five operations are performed sequentially: data loading, forward propagation, loss calculation, back propagation, and parameter update. A batch of samples is loaded from the training set, including standardized binary flow field images, multi-source feature vectors, normalized tail fin design parameter changes Δx, and the true value of drag contribution ΔCd_true, and the data is loaded into the GPU. The samples are input into the model, passing through the CNN feature extraction layer, feature fusion layer, and fully connected network layer, outputting quadratic function parameters a and b. These parameters are then substituted into the quadratic function to calculate the predicted drag contribution value. The mean squared error between the predicted and true values ​​is calculated according to the loss function formula. Backpropagation is performed on the loss value to calculate the gradient values ​​of all model parameters. The Adam optimizer iteratively updates the model parameters based on the calculated gradient values, thus training the model. After each epoch of training, the model is set to evaluation mode, and model inference is performed on the validation set. The mean squared error and mean absolute error of the validation set are calculated as evaluation metrics for model performance. During evaluation, the training mode of the Dropout layer and batch normalization layer is disabled to avoid the influence of random factors on the evaluation results. Backpropagation and parameter update operations are not performed; only forward propagation inference is performed.

[0073] Finally, based on the performance metrics of the validation set, and combining the learning rate decay strategy and the early stopping strategy, the convergence of the model is determined: if the loss value of the validation set does not decrease for 10 consecutive epochs, the learning rate is decayed by multiplying the learning rate by 0.5; if the loss value of the validation set does not decrease for 20 consecutive epochs, the early stopping strategy is triggered, and model training is stopped; if the model training reaches the maximum number of iterations (200), training is automatically stopped. During model training, the model weights with the smallest loss value on the validation set in each epoch are saved as the optimal model weights. After model training is completed, the optimal model weights are loaded into the model, and the final performance evaluation is performed on the test set to verify the model's prediction accuracy and generalization ability.

[0074] Step S4: Input the flow field image of the car rear wing design to be predicted into the trained deep learning model, extract its multi-source features, and combine them with the change in the target design parameters to quickly predict the wind resistance contribution through the quadratic function.

[0075] Step S3 completes the training of the deep learning model and saves the optimal model weights. This step is the engineering application stage of the model. For the new car rear wing design, standardized image preprocessing and multi-source feature extraction are performed. The extracted features are input into the trained deep learning model, which quickly outputs quadratic function parameters a and b. Combined with the changes in the target rear wing design parameters, the drag contribution of the design can be quickly calculated using the quadratic function. The entire prediction process does not require CFD simulation and takes only seconds, significantly improving the evaluation efficiency of the rear wing design. The operation process of this step is highly consistent with the previous steps, ensuring the consistency and accuracy of the prediction results. The specific operation is as follows: For the new car rear wing design (determining the target rear wing downforce), CFD simulation software was used to generate CFD flow field images of the wake field of the design. The image format and resolution were kept consistent with the original images in the training sample set to ensure compatibility with subsequent preprocessing operations.

[0076] For the original CFD flow field image of the new design scheme, perform the same image preprocessing operations as in step 1, including irrelevant information removal, vehicle body area identification and wake region cropping, size normalization, blue channel enhancement, and threshold segmentation to generate a normalized binary flow field image of 256×256 pixels.

[0077] For the generated standardized binary flow field image, perform the same multi-source feature extraction operation as in step S2, including ResNet18-based depth feature extraction and geometric analysis-based top edge tilt angle feature extraction of the main wake region, and stitch them together to generate a 513-dimensional multi-source feature vector.

[0078] Extract the target tail fin downforce x_target_raw of the new design scheme, and map it to the interval [-1,1] using the same normalization formula as in step S1 to obtain the normalized change in the target design parameters.

[0079] The trained deep learning model is set to evaluation mode. The standardized binary flow field image, multi-source feature vector, and normalized target design parameter changes are input into the model. Forward propagation inference is performed to quickly output the two scalar parameters a and b of the quadratic function. The entire inference process is performed on the GPU and takes only milliseconds.

[0080] By substituting the parameters a and b from the model output and the changes in the normalized target design parameters into a quadratic function, the drag contribution of the new car rear wing design can be directly calculated.

[0081] The calculated drag contribution ΔCd is output as the prediction result, along with the corresponding quadratic function curve, which visually demonstrates the relationship between the tail fin's downward pressure distance and the drag contribution. Engineers can use the prediction results to determine whether the design achieves drag optimization, providing an objective and accurate basis for selecting and optimizing tail fin design options.

[0082] The core advantage of this step lies in its speed and accuracy. The entire prediction process, from image preprocessing to result output, takes only seconds, while traditional CFD simulation methods take about 2 hours per attempt, improving efficiency by more than a thousand times. At the same time, because the model learns the physical relationship between the tail wing design parameters and the contribution of wind resistance during the training process, the prediction results have extremely high accuracy and can effectively replace traditional CFD simulation methods, supporting rapid iteration and parameter scanning in the automotive tail wing design stage.

[0083] Example 2 This embodiment discloses a system for predicting the drag contribution of a car rear wing design; like Figure 2 As shown, a system for predicting the drag contribution of a car rear wing includes: The data preprocessing module 201 is configured to: acquire the original CFD flow field image and perform preprocessing to generate a standardized binary image; The feature extraction module 202 is configured to: for each binary image, use a convolutional neural network to automatically extract the flow field depth features, and at the same time extract the tilt angle features of the top edge of the main wake region through geometric analysis, and use the depth features and the tilt angle features together as multi-source features characterizing the wake field structure. The model building and training module 203 is configured to: build a deep learning model, use the multi-source features as model input, train the deep learning model, and output a scalar parameter of a quadratic function; The prediction application module 204 is configured to: input the flow field image of the car rear wing scheme to be predicted into the trained deep learning model, extract its multi-source features, and combine them with the change of the target design parameters to quickly predict the wind resistance contribution through the quadratic function.

[0084] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0085] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the method for predicting the drag contribution of a car rear wing scheme as described in Example 1.

[0086] Example 4 The purpose of this embodiment is to provide an electronic device.

[0087] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the method for predicting the drag contribution of a car rear wing scheme as described in Embodiment 1.

[0088] Example 5 Embodiment 5 of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the method for predicting the drag contribution of a car rear wing scheme as described in Embodiment 1.

[0089] The steps and methods involved in the apparatuses of Embodiments 2, 3, 4, and 5 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0090] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0091] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting the drag contribution of a car rear wing design, characterized in that, include: The original CFD flow field image is acquired and preprocessed to generate a standardized binary image; For each binary image, a convolutional neural network is used to automatically extract the flow field depth features. At the same time, geometric analysis is used to extract the tilt angle features of the top edge of the main wake region. The depth features and the tilt angle features are used together as multi-source features to characterize the wake field structure. Construct a deep learning model, take the multi-source features as model input, train the deep learning model, and output a scalar parameter of a quadratic function; The flow field image of the car rear wing design to be predicted is input into the trained deep learning model, its multi-source features are extracted, and combined with the change of the target design parameters, the drag contribution is quickly predicted through the quadratic function.

2. The method for predicting the drag contribution of a car rear wing as described in claim 1, characterized in that, Acquire the raw CFD flow field image and perform preprocessing, including: Automatically detect and remove irrelevant information and color bars at the bottom of the image; Identify the vehicle body region and crop out the region containing the complete wake using the right boundary of the vehicle body as a reference; generate a binary image that highlights the vortex structure by enhancing the blue channel and thresholding. The corresponding aerodynamic parameters are automatically parsed based on the image file name information, and training samples are generated through systematic pairing.

3. The method for predicting the drag contribution of a car rear wing design as described in claim 1, characterized in that, The tilt angle features of the top edge of the main wake region were extracted through geometric analysis, including: Identify the largest connected region in the binary image as the main wake region, traverse its right three-fifths along the horizontal direction, record the vertical coordinates of the first foreground pixel in each column, and form an edge point set. A linear fit is performed on the set of edge points, the inclination angle of the fitted line is calculated, and it is normalized to the interval [-1, 1] as a feature.

4. The method for predicting the drag contribution of a car rear wing as described in claim 1, characterized in that, The structure of the constructed deep learning model includes: The convolutional neural network part consists of multiple convolutional layers and pooling layers, which are used to extract image depth features and obtain image feature vectors through global pooling. The feature fusion layer is used to concatenate the image feature vector with the tilt angle feature of the top edge of the main wake region into a comprehensive feature vector. A fully connected network is used to process the comprehensive feature vector and output the scalar parameters of the quadratic function.

5. The method for predicting the drag contribution of a car rear wing as described in claim 1, characterized in that, The training process of the deep learning model also includes the construction of training samples, which include: Using flow field images of different rear wing designs of the same vehicle model as a pair, the difference in rear wing height between them is calculated as one of the input features, and the difference in drag coefficient is used as the prediction label; symmetrical samples are generated by swapping the order of the image pairs to expand the training dataset.

6. The method for predicting the drag contribution of a car rear wing as described in claim 1, characterized in that, Quadratic functions are used to describe the physical relationship between changes in tail fin design parameters and their contribution to wind resistance: in, Contribution to wind resistance; This refers to the distance the tail fin deflects downwards; , It is a scalar parameter.

7. A system for predicting the drag contribution of a car rear wing design, characterized in that, include: The data preprocessing module is configured to: acquire the original CFD flow field image and preprocess it to generate a standardized binary image; The feature extraction module is configured to: automatically extract flow field depth features using a convolutional neural network for each binary image, and simultaneously extract the tilt angle features of the top edge of the main wake region through geometric analysis, and use the depth features and the tilt angle features together as multi-source features characterizing the wake field structure. The model building and training module is configured to: build a deep learning model, use the multi-source features as model input, train the deep learning model, and output a scalar parameter of a quadratic function; The prediction application module is configured to: input the flow field image of the car rear wing design to be predicted into the trained deep learning model, extract its multi-source features, and combine them with the change in the target design parameters to quickly predict the wind resistance contribution through the quadratic function.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for predicting the drag contribution of a car rear wing scheme as described in any one of claims 1-7.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for predicting the drag contribution of a car rear wing scheme as described in any one of claims 1-7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps in the method for predicting the drag contribution of a car rear wing scheme as described in any one of claims 1-6.

Images