Aerodynamic parameter prediction-oriented interpretable appearance feature learning and quantitative representation method

Abstract

The invention discloses an explainable appearance feature learning and quantitative representation method for aerodynamic parameter prediction, and belongs to the technical field of aerodynamics and artificial intelligence crossing. The method comprises the following steps: constructing a collaborative network architecture comprising an aerodynamic prediction module, an airfoil concept learning module and a quantitative distillation agent module; while high-precision and high-efficiency aerodynamic parameter prediction is realized, a prediction result can be decomposed into the sum of quantitative contributions of different airfoil profile concepts, so that a direct explainable basis is provided for the prediction result, and expert users are assisted in understanding and verifying the prediction process of the model.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of aerodynamics and artificial intelligence, and in particular to an interpretable shape feature learning and quantification method for aerodynamic parameter prediction. Background Technology

[0002] Accurately obtaining the aerodynamic parameters of airfoils is a crucial foundation for aircraft design and optimization. Traditionally, wind tunnel testing and computational fluid dynamics (CFD) have been two main techniques. Wind tunnel testing offers high reliability, but its process design heavily relies on expert experience, and the testing cycle is long and costly. CFD methods analyze the interaction between fluids and specific boundary conditions by numerically solving fluid dynamic governing equations (such as the Navier-Stokes equations). While CFD can provide high-precision calculation results, its solution process involves a large number of iterative calculations, consuming enormous computational resources and taking a long time, making it difficult to meet the urgent need for rapid iterative evaluation of schemes in modern aircraft design.

[0003] In recent years, with the rapid development of artificial intelligence technology, data-driven intelligent prediction methods have been widely used in aerodynamics research. These methods utilize the powerful nonlinear learning and function approximation capabilities of neural networks to construct the mapping relationship between airfoil geometry, inflow conditions, and aerodynamic parameters. Compared to traditional CFD calculations, intelligent prediction methods can reduce the calculation time for a single airfoil flow field from hours to seconds, significantly improving design efficiency.

[0004] However, existing intelligent prediction methods generally suffer from a key flaw: they are typically "black box models." These models only output predictions without explaining their internal decision-making mechanisms or providing interpretable evidence from input to output. This makes it difficult for domain experts to understand how the model arrives at specific conclusions based on airfoil shape characteristics, thus hindering their trust in and acceptance of the model's predictions and limiting its application in critical engineering tasks.

[0005] Currently, the mainstream approach for predicting aerodynamic parameters of two-dimensional airfoils is to construct end-to-end black-box models based on deep neural networks. However, these models cannot quantify the specific contribution of different airfoil geometric features to the final prediction results. This invention proposes a technical solution that integrates interpretable shape feature learning with quantitative representation. Summary of the Invention

[0006] The purpose of this invention is to provide an interpretable shape feature learning and quantification method for aerodynamic parameter prediction, addressing the above-mentioned shortcomings. This method solves the problem that black-box models in mainstream methods cannot quantify the specific contribution of different airfoil geometric features to the final prediction results.

[0007] This invention is achieved through the following scheme: An interpretable shape feature learning and quantization representation method for aerodynamic parameter prediction includes the following steps: Step 1: Define the training dataset for airfoil aerodynamic parameters; Step 2: Standardized image representation of airfoil geometry data; preprocess the airfoil coordinates defined in Step 1 to generate standardized airfoil image data; Step 3: Construct an aerodynamic prediction module to achieve end-to-end aerodynamic parameter prediction. Design an aerodynamic prediction module that extracts airfoil shape features and incoming flow condition features based on the standardized image generated in Step 2 and the aerodynamic parameters defined in Step 1, respectively, and inputs the two features into the prediction head after feature fusion, and outputs the prediction results of airfoil aerodynamic parameters. Step 4: Construct an airfoil concept learning module to learn interpretable shape concept representations. Introduce the airfoil concept learning module into the deep convolutional network of the aerodynamic prediction module in Step 3. By grouping and constraining the convolutional kernels, learn interpretable airfoil concept features with clear semantic information. Step 5: Construct a quantitative distillation proxy module to realize the quantitative interpretation of the prediction results. Utilize the knowledge distillation strategy to construct a quantitative distillation proxy module, taking the interpretable airfoil concept features learned in Step 4 as input, and by imitating the output of the aerodynamic prediction module, transform the single aerodynamic prediction output result into a quantitative weighted sum of the contributions of different airfoil concepts.

[0008] Step 1 specifically involves, for a given set of N two-dimensional airfoil coordinates... Under M different flow field conditions The following CFD numerical calculations are performed to obtain the corresponding aerodynamic parameter label values. N×M sets of sample data were obtained, forming a pre-training dataset for the flow field. .

[0009] Step 2 includes the following steps: Step 2.1: Standardize coordinate points; Step 2.2: Airfoil curve fitting; Step 2.3: Standardized image generation.

[0010] Step 2 specifically includes the following steps: Step 2.1: For the N two-dimensional airfoil coordinates in Step 1 The number of coordinate points for each airfoil profile is uniformly interpolated to a preset value through interpolation; Step 2.2: Use the non-uniform rational B-spline method to fit the interpolated coordinate points in Step 2.1 to generate a smooth and continuous airfoil curve; Step 2.3: Based on the fitted airfoil curves from Step 2.2, generate single-channel grayscale airfoil images with a preset resolution to obtain the airfoil image dataset. ,in This is the standardized airfoil image corresponding to the i-th sample.

[0011] Step 3 includes the following steps: Step 3.1: Extraction of upper-level features of the airfoil; Step 3.2: Generating shape feature vectors; Step 3.3: Generation of coordinate feature vectors; Step 3.4: Design of geometric constraint loss function; Step 3.5: Extraction of incoming flow condition features; Step 3.6: Feature fusion and prediction output; Step 3.7: Design of parameter prediction loss function.

[0012] Step 3 is as follows: Step 3.1: Standardize the airfoil image from Step 2.3 Extracting high-level feature maps of the airfoil from the input deep convolutional neural network. ; Step 3.2: Analyze the upper-level feature map of the airfoil obtained in Step 3.1. Perform a global average pooling operation to obtain a compact airfoil shape feature vector. ; Step 3.3: Convert the airfoil shape feature vector obtained in Step 3.2 into... The input is fed into a two-layer multilayer perceptron to obtain coordinate feature vectors used to reconstruct the airfoil coordinates. ; Step 3.4: Design the geometric coordinate point loss function , used to measure the coordinate feature vector of the prediction Compared with the actual airfoil coordinates The differences between them are used to constrain the network to learn airfoil features containing complete geometric information; Step 3.5: Change the flow field condition parameters from Step 1. The input is processed by a two-layer multilayer perceptron, and the resulting flow conditional feature vector is obtained. ; Step 3.6: Convert the airfoil shape feature vector With the incoming flow conditional eigenvector After splicing and fusion, the fused feature vector is input into the prediction head, and the aerodynamic parameter prediction results are output. ; Step 3.7: Use mean squared error as a parameter to predict the loss function. Used to measure prediction results With tag value The differences between them were analyzed, and the network parameters of the aerodynamic prediction module were optimized.

[0013] Step 4 includes the following steps: Step 4.1: Grouping convolutional kernels; Step 4.2: Feature map similarity calculation; Step 4.3: Design of convolution kernel decoupling loss function.

[0014] Step 4 specifically includes the following steps: Step 4.1: For the deep convolutional network used to extract airfoil features in Step 3.1, extract all convolutional kernels from its last convolutional layer. Divide into K groups, denoted as ,in , ; Step 4.2: Convert the output feature map of the previous convolutional layer. The input is fed into the last convolutional layer to generate the airfoil high-level feature map. Calculate the similarity between the feature maps output by any two convolutional kernels in the last convolutional layer. ; Step 4.3: Design the convolution kernel loss function The aim is to increase the similarity of the output feature maps of convolutional kernels within the same group, while reducing the similarity of the output feature maps of convolutional kernels between different groups. By minimizing this loss function, convolutional kernels within the same group express similar semantic features, while convolutional kernels in different groups focus on different airfoil regions. Therefore, the loss function is defined as:

[0015] in, and The calculation method is as follows:

[0016] .

[0017] Step 5 includes the following steps: Step 5.1: Strength calculation is required for the airfoil concept; Step 5.2: Training the quantitative distillation agent module; Step 5.3: Design of the distillation loss function; using the weights obtained after training. This allows for the direct quantification of the contribution of the i-th airfoil concept to the final aerodynamic parameter prediction results, thus achieving an interpretable quantitative representation of the prediction results.

[0018] Step 5 specifically includes the following steps: Step 5.1: Calculate the existence strength of the airfoil concept, and calculate the existence strength of each airfoil high-level feature map learned in Step 4. This intensity indicates the degree of activation of a specific airfoil concept in the input image:

[0019] in, This represents the conceptual feature diagram of the d-th airfoil. In spatial coordinates The value on; Step 5.2: Training the Quantitative Distillation Proxy Module, constructing the Quantitative Distillation Proxy Module based on a deep convolutional network. The airfoil upper-level feature map from step 4.2 As input to the network, learn the airfoil concept weights. ;

[0020] Step 5.3: Design the distillation loss function, design the distillation loss... By minimizing the prediction results of the parameter prediction module The difference between the output of the quantitative distillation agent module and the output of the agent module is used to train the agent module to learn concept weights.

[0021] Where b represents the bias term, This represents the weight of the i airfoil concepts. This represents the strength of the existence of i airfoil concepts; K is the number of airfoil concepts; Weights obtained after training It can directly quantify the contribution of the i-th airfoil concept to the final aerodynamic parameter prediction result, thus realizing an interpretable quantitative representation of the prediction result.

[0022] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This solution can quickly and accurately predict aerodynamic parameters while providing interpretable evidence of the prediction results in a quantitative form. It has significant advantages such as high prediction accuracy, fast speed and strong model interpretability, and has great application potential in the field of aircraft design.

[0023] 2. High-precision and high-efficiency prediction performance: This solution inherits the advantages of data-driven deep learning models, and can quickly and accurately predict the aerodynamic parameters of two-dimensional airfoils under complex flow field conditions. Its computational efficiency far exceeds that of traditional CFD methods, meeting the needs of rapid engineering iteration. 3. Strong inherent interpretability: This solution is the first to introduce concept learning and knowledge distillation mechanisms into the field of airfoil aerodynamic parameter prediction. Through the synergistic effect of the airfoil concept learning module and the quantification distillation proxy module, the influence of different regions (concepts) of the airfoil on the final aerodynamic parameter prediction results can be clearly explained in the form of quantified contribution values, fundamentally solving the technical problem of the "black box" inability to interpret existing intelligent prediction models; 4. Significantly enhance user trust: By providing intuitive and quantifiable explanatory evidence (i.e., the contribution value of each airfoil concept), this solution enables expert users to deeply understand the model's prediction logic and effectively judge the rationality of the prediction process, thereby greatly improving the credibility and reliability of the model's prediction results. 5. Optimized feature learning and representation capabilities: This scheme ensures that the features learned by the network have stronger semantic information and discriminability through the design of the convolution kernel loss function in the airfoil concept learning module, avoiding the redundancy and hybridity of traditional convolution features, and helping to improve the model's generalization ability and prediction accuracy. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall network architecture of the present invention. Detailed Implementation

[0025] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0026] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0027] In the description of this invention, it should be understood that the terms "upper," "lower," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a predetermined orientation, or be constructed and operated in a predetermined orientation. Therefore, they should not be construed as limitations on this invention.

[0028] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0029] Example 1 like Figure 1As shown, the present invention provides a technical solution: An interpretable shape feature learning and quantization representation method for aerodynamic parameter prediction includes the following steps: Step 1: Define a training dataset for the aerodynamic parameters of a two-dimensional airfoil; Given N two-dimensional airfoil coordinates Under M different flow field conditions (Each set of flow field conditions includes angle of attack (AOA), Reynolds number (RE), wall distance, and flow field abscissa and ordinate points.) CFD numerical calculations are performed to obtain the corresponding aerodynamic parameter label values. N×M sets of sample data were obtained, forming a pre-training dataset for the flow field. .

[0030] Step 2: Standardized image representation of airfoil geometry data. The airfoil coordinates defined in Step 1 are preprocessed to generate standardized airfoil image data to meet the input requirements of the deep convolutional neural network. This specifically includes the following steps: Step 2.1: Coordinate point unification, for the N two-dimensional airfoil coordinates in Step 1 The number of coordinate points for each airfoil profile is uniformly interpolated to a preset value to ensure the consistency of geometric data dimensions for different airfoils. Step 2.2: Airfoil curve fitting. The non-uniform rational B-spline (NURBS) method is used to fit the coordinate points interpolated in Step 2.1 to eliminate the non-smooth profile problem caused by coordinate discreteness and generate a smooth and continuous airfoil curve. Step 2.3: Standardized Image Generation. Based on the airfoil curves fitted in Step 2.2, uniformly generate single-channel grayscale airfoil images with a preset resolution, obtaining the airfoil image dataset. ,in This is the standardized airfoil image corresponding to the i-th sample.

[0031] Step 3: Construct an aerodynamic prediction module to achieve end-to-end aerodynamic parameter prediction. The aerodynamic prediction module is designed based on the standardized airfoil image data generated in Step 2 and the flow field condition parameters defined in Step 1. It extracts airfoil shape features and incoming flow condition features respectively, fuses these features, and inputs them into the prediction head to output the predicted airfoil aerodynamic parameters. Specifically, it includes the following steps: Step 3.1: High-level feature extraction of the airfoil, based on the standardized airfoil image from Step 2.3. Input into a deep convolutional neural network (CNN) to extract the upper-level feature map of the airfoil. ; Step 3.2: Generating the shape feature vector, based on the airfoil high-level feature map obtained in Step 3.1. Perform a global average pooling (GAP) operation to obtain a compact airfoil shape feature vector. ; Step 3.3: Generate coordinate feature vectors. This involves generating the airfoil shape feature vectors obtained in Step 3.2. The input is fed into a two-layer MLP to obtain coordinate feature vectors used to reconstruct the airfoil coordinates. ; Step 3.4: Design the geometric constraint loss function, design the geometric coordinate point loss function. , used to measure the coordinate feature vector of the prediction Compared with the actual airfoil coordinates The differences between them are used to constrain the network to learn airfoil features containing complete geometric information. .

[0032] Step 3.5: Incoming flow condition feature extraction, extracting the flow field condition parameters from Step 1. The input is processed by a two-layer MLP, and the incoming flow condition eigenvector is obtained after nonlinear transformation. ; Step 3.6: Feature Fusion and Prediction Output, converting the airfoil shape feature vector... With the incoming flow conditional eigenvector After splicing and fusion, the fused feature vector is input into the prediction head, and the aerodynamic parameter prediction results are output. ,

[0033] in, This indicates a feature concatenation operation. This represents a prediction head consisting of five MLP layers; Step 3.7: Design of the parameter prediction loss function. The mean squared error (MSE) is used as the parameter prediction loss function. Used to measure prediction results With tag value The differences between them were analyzed, and the network parameters of the aerodynamic prediction module were optimized. .

[0034] Step 4: Construct an airfoil concept learning module to learn interpretable shape concept representations. Introduce the airfoil concept learning module into the deep convolutional network of the aerodynamic prediction module in Step 3. By grouping and constraining the convolutional kernels, learn interpretable airfoil concept features with clear semantic information.

[0035] Step 4.1: Grouping Convolutional Kernels. For the deep convolutional network used to extract airfoil features in Step 3.1, group all convolutional kernels in its last convolutional layer. Divide into K groups, denoted as ,in , ; Step 4.2: Feature map similarity calculation, using the output feature map of the previous convolutional layer. The input is fed into the last convolutional layer to generate the airfoil high-level feature map. Calculate the similarity between the feature maps output by any two convolutional kernels in the last convolutional layer. ,

[0036] in, and These represent the feature maps output by the i-th and j-th convolutional kernels, respectively. Represents the similarity measure function. This represents the Pearson correlation coefficient (PCC). and This represents the standard deviation of the corresponding feature map. Represents covariance; Step 4.3: Design the convolution kernel decoupling loss function. The aim is to improve the similarity of the output feature maps of convolutional kernels within the same group (intra-group similarity). At the same time, it reduces the similarity of the output feature maps of convolutional kernels between different groups (inter-group similarity). By minimizing this loss function, convolutional kernels within the same group express similar semantic features (i.e., learn the same airfoil concept), while convolutional kernels in different groups focus on different airfoil regions (i.e., learn different airfoil concepts). Therefore, the loss function is defined as:

[0037] in, and The calculation method is as follows:

[0038]

[0039] Step 5: Construct a quantitative distillation proxy module to achieve quantitative interpretation of the prediction results. Utilizing a knowledge distillation strategy, a quantitative distillation proxy module is constructed. Taking the interpretable airfoil concept features learned in Step 4 as input, it transforms the single aerodynamic prediction output into a quantitatively weighted sum of contributions from different airfoil concepts by mimicking the output of the aerodynamic prediction module. The specific steps are as follows: Step 5.1: Calculate the existence strength of the airfoil concept, and calculate the existence strength of each airfoil high-level feature map learned in Step 4. This intensity indicates the degree of activation of a specific airfoil concept in the input image:

[0040] in, This represents the conceptual feature diagram of the d-th airfoil. In spatial coordinates The value on; Step 5.2: Training the Quantitative Distillation Proxy Module, constructing the Quantitative Distillation Proxy Module based on a deep convolutional network. The airfoil upper-level feature map from step 4.2 As input to the network, learn the airfoil concept weights. ;

[0041] Step 5.3: Design the distillation loss function, design the distillation loss... By minimizing the prediction results of the parameter prediction module The difference between the output of the quantitative distillation agent module and the output of the agent module is used to train the agent module to learn concept weights.

[0042] Where b represents the bias term, This represents the weight of the i airfoil concepts. Let K represent the strength of the existence of i airfoil concepts, and K be the number of airfoil concepts.

[0043] Weights obtained after training This allows for the direct quantification of the contribution of the i-th airfoil concept to the final aerodynamic parameter prediction results, thus achieving an interpretable quantitative representation of the prediction results.

[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An explainable shape feature learning and quantized representation method for aerodynamic parameter prediction, characterized in that: Comprising the following steps: Step 1: defining the airfoil aerodynamic parameter training dataset; Step 2: standardized image representation of airfoil geometry data; preprocessing the airfoil coordinates defined in step 1 to generate standardized airfoil image data; Step 3: constructing an aerodynamic prediction module to realize end-to-end aerodynamic parameter prediction, designing an aerodynamic prediction module based on the standardized images generated in step 2 and the aerodynamic parameters defined in step 1, extracting airfoil shape features and incoming flow condition features respectively, and inputting the two features into the prediction head after feature fusion, and outputting the prediction result of the airfoil aerodynamic parameter; Step 4: constructing an airfoil concept learning module to learn interpretable shape concept representation, introducing an airfoil concept learning module to the deep convolutional network of the aerodynamic prediction module in step 3, learning interpretable airfoil concept features with clear semantic information by grouping and constraining the convolutional kernels; Step 5: constructing a quantitative distillation proxy module to realize quantitative interpretation of the prediction result, using a knowledge distillation strategy to construct a quantitative distillation proxy module, taking the interpretable airfoil concept features learned in step 4 as input, and converting the single aerodynamic prediction output into a quantitative weighted sum of different airfoil concept contributions by imitating the output of the aerodynamic prediction module.

2. The method of claim 1, wherein: Step 1 is specifically, for a given N two-dimensional airfoil coordinates , under M different flow field conditions , CFD numerical calculation is carried out to obtain the corresponding aerodynamic parameter label value , N x M groups of sample data are obtained to form a flow field pre-training data set .

3. The method of claim 2, wherein: Step 2 comprises the following steps: Step 2.1: coordinate point unification; Step 2.2: airfoil curve fitting; Step 2.3: standardized image generation.

4. The method of claim 3, wherein: Step 2 specifically comprises the following steps: Step 2.1: For N two-dimensional airfoil coordinates in step 1 , uniform interpolation of each airfoil profile coordinate point number to a preset value by interpolation; Step 2.2: using the non-uniform rational B-spline method to fit the interpolated coordinate points in step 2.1 to generate smooth and continuous airfoil curves; Step 2.3: Based on the fitted airfoil curve in step 2.2, uniformly generate single-channel gray airfoil images with a preset resolution to obtain an airfoil image dataset wherein is the standardized airfoil image corresponding to the i-th sample.

5. The method of claim 1-4, wherein the method is directed to aerodynamic parameter prediction. Step 3 comprises the following steps: Step 3.1: airfoil high-level feature extraction; Step 3.2: shape feature vector generation; Step 3.3: coordinate feature vector generation; Step 3.4: geometric constraint loss function design; Step 3.5: incoming flow condition feature extraction; Step 3.6: feature fusion and prediction output; Step 3.7: parameter prediction loss function design.

6. The method for interpretable shape feature learning and quantized representation oriented to aerodynamic parameter prediction as claimed in claim 5, characterized in that: The steps of step 3 are specifically: Step 3.1: Standardizing the wing profile images in Step 2.3 In the input deep convolutional neural network, extracting high-level feature maps of wing profiles ; Step 3.2: Obtaining the wing profile feature map for the high-level feature of the wing profile of the step 3.1 performing a global average pooling operation to obtain a compact wing profile feature vector ; Step 3.3: The airfoil shape feature vector from step 3.2 is input into a two layer multi-layer perceptron to obtain a coordinate feature vector for reconstructing the airfoil coordinates input into a two layer multi-layer perceptron to obtain a coordinate feature vector for reconstructing the airfoil coordinates ; Step 3.4: Designing the geometric coordinate point loss function , to measure the difference between the predicted coordinate feature vector and the real airfoil coordinates to constrain the network to learn airfoil features containing complete geometric information; Step 3.5: input the flow field condition parameters in step 1 into a two-layer multi-layer perceptron to obtain a flow condition feature vector after nonlinear transformation Input a two-layer multi-layer perceptron, after nonlinear transformation processing, get the flow condition feature vector ; Step 3.6: the wing profile characteristic vector with the incoming flow condition characteristic vector After splicing and fusion, the fused characteristic vector is input into the prediction head, and the aerodynamic parameter prediction result is output ; Step 3.7: Use mean square error as parameter prediction loss function , to measure the difference between the prediction result and the label value , and optimize the network parameters of the aerodynamic prediction module.

7. The method of claim 1-4, wherein the method is directed to aerodynamic parameter prediction, and the method is characterized in that: Step 4 comprises the following steps: Step 4.1: convolution kernel grouping; Step 4.2: feature map similarity calculation; Step 4.3: convolution kernel decoupling loss function design.

8. The method of claim 7, wherein: Step 4 specifically comprises the following steps: Step 4.1 : For the deep convolutional network used to extract airfoil features in step 3.1, all the convolutional kernels in the last convolutional layer are replaced by a set of K convolutional kernels are divided into K groups, denoted as where , ; Step 4.2: input the output feature maps of the last convolutional layer to the last convolutional layer to generate airfoil high-level feature maps Step 4.3: calculate the similarity between the feature maps output by any two convolutional kernels in the last convolutional layer Step 4.4: select the airfoil high-level feature maps with the highest similarity as the final airfoil high-level feature maps Step 4.5: input the final airfoil high-level feature maps into the fully connected layer to generate the airfoil shape Step 4.3: Designing the convolution kernel loss function The loss function is designed to increase the similarity of the output feature maps of the convolution kernels in the same group, while reducing the similarity of the output feature maps of the convolution kernels in different groups. By minimizing this loss function, the convolution kernels in the same group express similar semantic features, while the convolution kernels in different groups focus on different airfoil regions. Therefore, the loss function is defined as: wherein and are calculated as follows, 。 9. The method of claim 1-4, wherein the method is directed to aerodynamic parameter prediction, and the method is characterized in that: Step 5 comprises the following steps: Step 5.1: airfoil concept existence intensity calculation; Step 5.2: quantitative distillation proxy module training; Step 5.3: Distillation loss function design; weight obtained after training The contribution of the ith airfoil concept to the final aerodynamic parameter prediction result can be directly quantified, thereby realizing an interpretable and quantitative representation of the prediction result.

10. The method of claim 9, wherein: Step 5 specifically comprises the following steps: Step 5.1: Presence strength computation for airfoil concepts, computing the presence strength of each airfoil high-level feature map learned in step 4 The strength indicates how much a particular airfoil concept is activated in the input image: wherein represents the dth airfoil concept feature map values in spatial coordinates on the spatial coordinates Step 5.2: Quantization distillation agent module training, building a quantization distillation agent module based on a deep convolutional network The airfoil high-level feature map in step 4.2 is taken as the input of the network The airfoil concept weight is learned as the input of the network ; Step 5.3: Distillation loss function design, design distillation loss by minimizing the difference between the prediction results of the parameter prediction module and the quantized distillation agent module output results, the agent module learns the concept weight, where b represents a bias term, denotes the weight of the i-th airfoil concept, denotes the presence strength of the i-th airfoil concept; K is the number of airfoil concepts; the weights obtained after training The contribution of the ith airfoil concept to the final aerodynamic parameter prediction result can be directly quantified, thereby achieving an interpretable and quantitative representation of the prediction result.

Images