Aircraft aerodynamic parameter prediction method based on multi-view airfoil concept association learning

Abstract

This invention discloses a method for predicting aerodynamic parameters of aircraft based on multi-view airfoil concept association learning, belonging to the interdisciplinary fields of aerodynamics, deep learning, and computer vision. This method converts a 3D model into a multi-view image sequence through a hybrid projection strategy combining orthogonal and surround view; it utilizes optimal transmission theory to map visual features to a learnable "airfoil concept" space, achieving semanticization and discretization of features; combined with flow field condition encoding, it regresses aerodynamic parameters under the constraints of aerodynamic physical laws; and finally, by demonstrating the matching relationship between input samples and airfoil concepts, it provides instance-based inherent interpretability.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of aerodynamics, deep learning, and computer vision, and in particular to an intelligent method for rapidly predicting the aerodynamic parameters of three-dimensional aircraft and providing inherent interpretability analysis by utilizing multi-view image projection and airfoil concept association. Background Technology

[0002] Accurate acquisition of aircraft aerodynamic parameters is fundamental to aircraft aerodynamic layout design and performance evaluation. Traditionally, aerodynamic parameter acquisition has relied primarily on wind tunnel testing and computational fluid dynamics (CFD) numerical simulations. Wind tunnel testing is costly and time-consuming; while CFD numerical simulations offer high accuracy, each calculation is time-consuming, making it difficult to meet the needs of large-scale scheme selection during the conceptual design phase.

[0003] With the development of artificial intelligence technology, data-driven deep learning methods have been applied to aerodynamic prediction, achieving rapid responses down to the second. However, existing intelligent prediction solutions face serious challenges in practical engineering applications, mainly due to a lack of interpretability.

[0004] Most existing intelligent aerodynamic parameter prediction methods are "black box" models, only outputting numerical results and unable to identify which local features of the aircraft's shape play a decisive role in specific aerodynamic parameters. This makes it difficult for engineers to trust the model and perform targeted optimizations. To alleviate the trust issue, current technologies often employ "posterior interpretability" methods, which, after the model makes a decision, use auxiliary means such as gradient-weighted class activation mapping or occlusion sensitivity analysis to infer the importance of input features. However, these methods have significant drawbacks: posterior interpretation is an approximate fit to the model's decision behavior, rather than the model's true reasoning logic, resulting in a "fidelity gap"; gradient-based interpretation is sensitive to noise and lacks stability; and methods such as occlusion analysis require significant computational overhead for counterfactual inference, leading to inefficiency. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting aerodynamic parameters of aircraft based on multi-view airfoil concept association learning, which addresses the above-mentioned shortcomings. This method is an "intrinsically interpretable" aerodynamic parameter prediction model, which aims to build an "airfoil concept" mechanism that conforms to aerodynamic cognition at the initial design stage. During the reasoning process, it directly uses the learned typical airfoil geometric features to make decisions, thereby providing a transparent, reliable and physically intuitive explanation without relying on external tools for verification.

[0006] This invention is achieved through the following scheme:

[0007] The method for predicting aerodynamic parameters of aircraft based on multi-view airfoil concept association learning includes the following steps:

[0008] Step 1, Construct a hybrid multi-view aircraft geometric projection dataset: Convert the aircraft model in three-dimensional space into a standardized two-dimensional image sequence;

[0009] Step 2, Feature encoding of flow field conditions: Mapping scalar flow field condition parameters to a high-dimensional feature space;

[0010] Step 3, Visual Feature Extraction: Use deep neural networks to extract deep geometric features from each viewpoint image;

[0011] Step 4, Airfoil concept association learning based on optimal transmission: Construct a learnable "airfoil concept pool" and discretize the visual features into an association distribution with airfoil concepts;

[0012] Step 5, Feature Fusion and Aerodynamic Parameter Prediction: Fusion of operating condition information and reconstructed airfoil concept features, regression prediction of aerodynamic parameters;

[0013] Step 6, Loss Function Construction and Model Optimization: Construct a composite loss function that includes prediction error and concept clustering characteristics;

[0014] Step 7, Online Update of Airfoil Concepts: Dynamically update the airfoil concept pool during training to ensure its representativeness;

[0015] Step 8, Generating Intrinsic Interpretability Results: Based on the trained model, output an intuitive interpretation of the prediction results.

[0016] Step 1 specifically includes the following steps:

[0017] Step 1.1: Obtain the 3D mesh model of the aircraft to be predicted, and calculate the 3D geometric bounding box of the 3D mesh model;

[0018] Step 1.2: Based on the 3D geometric bounding box calculated in Step 1.1, the 3D mesh model is scaled and translated as a whole so that its geometric center is located at the origin of the coordinate system and is normalized to a unit sphere space.

[0019] Step 1.3: Using the center of the unit spherical space in Step 1.2 as the origin, set up an orthogonal projection camera group, which includes a predetermined fixed number of orthogonal projection cameras.

[0020] Step 1.4: Configure the line-of-sight directions of the predetermined number of cameras in Step 1.3 to the front view, rear view, left view, top view, and bottom view of the aircraft, respectively, in order to capture the main view geometry of the aircraft;

[0021] Step 1.5: Using the center of the unit spherical space in Step 1.2 as the origin, set up a panoramic projection camera group on the equatorial plane. The camera group contains multiple perspective cameras, with the interval between adjacent cameras being a preset angle and the elevation angle being uniformly set.

[0022] Step 1.6: Using the orthogonal projection camera group set in Step 1.3 and the surround projection camera group set in Step 1.5, perform synchronous off-screen rendering on the normalized 3D mesh model in Step 1.2 to generate multiple 2D projection images.

[0023] Step 1.7: Adjust the resolution of the multiple images generated in Step 1.6 to a uniform value. The pixel values ​​were then normalized to construct a multi-view image set. ;

[0024] Step 1.8: For each aircraft model, set the flow field operating conditions, including the Mach number. and attack angle ;

[0025] Step 1.9: Using a computational fluid dynamics solver, calculate the actual aerodynamic parameter labels under the operating condition in Step 1.8, including the actual drag coefficient. True lift coefficient Actual pitch moment coefficient and the actual center of pressure .

[0026] Step 2 specifically includes the following steps:

[0027] Step 2.1, extract the Mach number set in Step 1.8. and attack angle Construct the original working condition vector ;

[0028] Step 2.2: Construct a working condition encoder, which consists of a multi-layer sensing mechanism;

[0029] Step 2.3, the original working condition vector constructed in Step 2.1 Input the operating condition encoder constructed in step 2.2;

[0030] Step 2.4: After the linear transformation layer and nonlinear activation function mapping of the working condition encoder, a high-dimensional working condition feature vector is output. .

[0031] Step 3 specifically includes the following steps:

[0032] Step 3.1: Construct a shared-weight visual Transformer as the backbone feature extraction network;

[0033] Step 3.2, combine the multi-view image set obtained in Step 1.7 Each image in the process is input into the backbone feature extraction network constructed in step 3.1;

[0034] Step 3.3: Extract the category token vector from the network output in Step 3.2 to obtain a set of visual features containing multiple perspectives. ,in Indicates the first Feature vectors from each perspective.

[0035] Step 4 specifically includes the following steps:

[0036] Step 4.1, initialize an airfoil concept pool. , which includes A learnable concept vector , Indicates the first A vector of airfoil concepts;

[0037] Step 4.2, based on the visual feature set obtained in Step 3.3 and the airfoil concept pool initialized in step 4.1 Calculate the dot product similarity between the feature vector and the concept vector;

[0038] Step 4.3: Based on the similarity calculated in Step 4.2, use the Sinkhorn algorithm to solve for the optimal transmission allocation matrix. The calculation formula is:

[0039]

[0040] in, For the first The assignment vector corresponding to each viewpoint To normalize and scale the vector, To normalize and scale the vector, This is the transpose of the airfoil concept pool matrix. For the first Feature vectors from each perspective The regularization coefficient is used.

[0041] Step 4.4, based on the allocation vector obtained in step 4.3 , the elements Indicates the first The features of the first perspective belong to the first The allocation probability of the airfoil concept is calculated, and a weighted linear combination is performed on the airfoil concept pool in step 4.1 to calculate the probability of the first airfoil concept. Reconstruction features from a single perspective :

[0042]

[0043] in, For the first Reconstructed feature vectors from multiple perspectives For the first A vector of airfoil concepts.

[0044] Step 5 specifically includes the following steps:

[0045] Step 5.1, construct a cross-attention module;

[0046] Step 5.2, the working condition feature vector obtained in step 2.4 As a query vector (Query);

[0047] Step 5.3, reconstruct the feature set obtained in step 4.4. As a key vector (Key) and a value vector (Value);

[0048] Step 5.4: Input the Query from Step 5.2 and the Key and Value from Step 5.3 into the cross-attention module of Step 5.1, calculate the attention weights, perform weighted aggregation, and output the fused feature vector. ;

[0049] Step 5.5: Construct four independent multilayer sensor prediction heads, corresponding to the drag coefficient, lift coefficient, pitch moment coefficient, and pressure center position, respectively.

[0050] Step 5.6, merging the feature vector obtained in step 5.4. Input the four prediction heads from step 5.5 respectively, and the output will be the predicted aerodynamic parameters: predicted drag coefficient. Predicted lift coefficient Predicted pitch moment coefficient and predict the location of the pressure center .

[0051] Step 6 specifically includes the following steps:

[0052] Step 6.1: Based on the predicted values ​​from Step 5.6 and the true labels from Step 1.9, calculate the mean squared error to obtain the data prediction loss. ;

[0053] Step 6.2, based on the allocation probability in step 4.4 and the airfoil concept vector in step 4.1 Calculate the clustering loss to align visual features with their assigned airfoil concepts:

[0054]

[0055] in, For clustering loss, The square of the Euclidean distance. It is the first One visual feature vector;

[0056] Step 6.3, calculate the separation loss to keep the vectors in the airfoil concept pool far apart from each other:

[0057]

[0058] in, For separation loss, For the number of airfoil concepts, and They are the first The and the first A vector of airfoil concepts;

[0059] Step 6.4: The losses calculated in steps 6.1, 6.2, and 6.3 are weighted and summed, and the network parameters are updated using the backpropagation algorithm.

[0060] Step 7 specifically includes the following steps:

[0061] Step 7.1: At the end of each training iteration, the statistics assigned in step 4.4 are... All visual feature vectors of an airfoil concept;

[0062] Step 7.2: Calculate the average vector of all visual feature vectors counted in Step 7.1. ;

[0063] Step 7.3, update the value in step 4.1 using the exponential moving average strategy. Airfoil concept vector :

[0064]

[0065] in, For the updated concept vectors, The concept vector at the current moment. The momentum coefficient, The average vector calculated in step 7.2.

[0066] Step 8 specifically includes the following steps:

[0067] Step 8.1, in the reasoning phase, for the first... Given a given input sample, obtain the allocation probability vector calculated in step 4.3. ;

[0068] Step 8.2: In the vector from step 8.1, identify the vector with the highest probability value. An index of airfoil concepts;

[0069] Step 8.3: Retrieve from the training dataset the real aircraft image fragment that is closest to the airfoil concept vector identified in Step 8.2, and use it as a visual representation of the airfoil concept;

[0070] Step 8.4: Output an explanation report, displaying the visual representative image retrieved in Step 8.3, explaining that the current prediction result is based on the high similarity between the input sample and the airfoil concept, thereby achieving an intrinsic explanation based on case comparison.

[0071] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0072] 1. Built-in interpretable core mechanism, completely eliminating the limitations of posterior backpropagation. From the initial design stage, the model constructs an "airfoil concept pool" consistent with aerodynamic cognition. During inference, an optimal transmission algorithm maps multi-view visual features to an association distribution with airfoil concepts, directly outputting the probability of each airfoil concept assigned to the input sample. This interpretive logic is not an approximate backpropagation after model decision-making, but rather a core component of the decision-making process. It fundamentally eliminates the "fidelity gap" inherent in posterior interpretation and eliminates redundant calculations such as gradient backpropagation or occlusion interference. This ensures both the transparency and traceability of the decision-making logic and maintains interpretive efficiency.

[0073] 2. The interpretation results are intuitive and concrete, providing case-based, actionable decision-making support: Based on the assigned probability vector, the core airfoil concept that plays a crucial role in the prediction results can be accurately located. Subsequently, the real aircraft image fragments closest to the concept vector in the training dataset are retrieved as visual representatives. Compared with the abstract heatmaps of traditional posterior interpretation, this "input sample - airfoil concept - real sample comparison" display method allows engineering designers to intuitively perceive the key shape features affecting aerodynamic parameters, clearly understand the origin of the prediction results, and provide clear case references for targeted optimization of aircraft layout.

[0074] 3. Airfoil concepts align with physical understanding, enhancing the engineering credibility and practicality of explanations: The "airfoil concept pool" is built upon aerodynamic principles and dynamically updated using an online exponential moving average (EMA) strategy. This ensures that each concept closely tracks the actual geometric distribution, exhibiting high intra-class compactness and inter-class separability. The explanation process is not merely a simple data analysis correlation, but rather unfolds around the inherent physical logic of airfoil geometry and aerodynamic performance. This perfectly aligns with engineers' cognitive habits regarding aircraft aerodynamic design, making model decisions not only "explainable" but also "reliable." This helps designers quickly locate key shape details affecting aerodynamic parameters, improving the targeting and efficiency of optimization designs. Attached Figure Description

[0075] Figure 1This is a flowchart of the present invention. Detailed Implementation

[0076] 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.

[0077] 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.

[0078] 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.

[0079] 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.

[0080] Example 1

[0081] like Figure 1 As shown, the present invention provides a technical solution:

[0082] Step 1, Constructing a Hybrid Multi-View Aircraft Geometric Projection Dataset: This step aims to convert the aircraft model in 3D space into a standardized sequence of 2D images to preserve omnidirectional geometric shape information; it specifically includes the following steps:

[0083] Step 1.1: Obtain the 3D mesh model of the aircraft to be predicted and calculate the 3D geometric bounding box of the 3D mesh model.

[0084] Step 1.2: Based on the 3D geometric bounding box calculated in Step 1.1, the 3D mesh model is scaled and translated as a whole so that its geometric center is located at the origin of the coordinate system and is normalized to a unit sphere space.

[0085] Step 1.3: Using the center of the unit spherical space in Step 1.2 as the origin, set up an orthogonal projection camera group, which consists of 5 fixed orthogonal projection cameras.

[0086] Step 1.4: Configure the line-of-sight directions of the five cameras in Step 1.3 to correspond to the front view, rear view, left view, top view, and bottom view of the aircraft, respectively, in order to capture the main view geometry of the aircraft.

[0087] Step 1.5: Using the center of the unit spherical space in Step 1.2 as the origin, set up a panoramic projection camera group on the equatorial plane. The camera group contains 12 perspective cameras, with an interval of 30 degrees between adjacent cameras and a uniform elevation angle of 30 degrees.

[0088] Step 1.6: Using the orthogonal projection camera group set in Step 1.3 and the panoramic projection camera group set in Step 1.5, synchronous off-screen rendering is performed on the normalized 3D mesh model in Step 1.2 to generate a total of 17 2D projection images.

[0089] Step 1.7: Adjust the resolution of the 17 images generated in Step 1.6 to a uniform value. The pixel values ​​were then normalized to construct a multi-view image set. .

[0090] Step 1.8: For each aircraft model, set the flow field operating conditions, including the Mach number. and attack angle .

[0091] Step 1.9: Using a computational fluid dynamics (CFD) solver, calculate the actual aerodynamic parameter labels under the operating condition of Step 1.8, including the actual drag coefficient. True lift coefficient Actual pitch moment coefficient and the actual center of pressure .

[0092] Step 2, Feature Encoding of Flow Field Conditions: This step maps scalar flow field condition parameters to a high-dimensional feature space for subsequent fusion with visual features; it specifically includes the following steps:

[0093] Step 2.1, extract the Mach number set in Step 1.8. and attack angle Construct the original working condition vector .

[0094] Step 2.2: Construct a working condition encoder, which consists of a multilayer perceptron (MLP).

[0095] Step 2.3, the original working condition vector constructed in Step 2.1 Input the operating condition encoder constructed in step 2.2.

[0096] Step 2.4: After the linear transformation layer and nonlinear activation function mapping of the working condition encoder, a high-dimensional working condition feature vector is output. .

[0097] Step 3, Visual Feature Extraction: This step utilizes a deep neural network to extract deep geometric features from each viewpoint image; it specifically includes the following steps:

[0098] Step 3.1: Construct a shared-weights Visual Transformer (ViT) as the backbone feature extraction network.

[0099] Step 3.2, combine the multi-view image set obtained in Step 1.7 Each image in the process is input into the backbone feature extraction network constructed in step 3.1.

[0100] Step 3.3: Extract the class token vector from the network output in Step 3.2 to obtain a visual feature set containing 17 viewpoints. ,in Indicates the first Feature vectors from each perspective.

[0101] Step 4, Airfoil Concept Association Learning Based on Optimal Transmission: This step constructs a learnable "airfoil concept pool" and discretizes visual features into association distributions with airfoil concepts; it specifically includes the following steps:

[0102] Step 4.1, initialize an airfoil concept pool. , which includes A learnable concept vector , Indicates the first A vector of airfoil concepts.

[0103] Step 4.2, based on the visual feature set obtained in Step 3.3 and the airfoil concept pool initialized in step 4.1 Calculate the dot product similarity between the feature vector and the concept vector.

[0104] Step 4.3: Based on the similarity calculated in Step 4.2, use the Sinkhorn algorithm to solve for the optimal transmission allocation matrix. The calculation formula is:

[0105]

[0106] in, For the first The assignment vector corresponding to each viewpoint To normalize and scale the vector, To normalize and scale the vector, This is the transpose of the airfoil concept pool matrix. For the first Feature vectors from each perspective This is the regularization coefficient.

[0107] Step 4.4, based on the allocation vector obtained in step 4.3 , the elements Indicates the first The features of the first perspective belong to the first The allocation probability of the airfoil concept is calculated, and a weighted linear combination is performed on the airfoil concept pool in step 4.1 to calculate the probability of the first airfoil concept. Reconstruction features from a single perspective :

[0108]

[0109] in, For the first Reconstructed feature vectors from multiple perspectives For the first A vector of airfoil concepts.

[0110] Step 5, Feature Fusion and Aerodynamic Parameter Prediction: This step fuses operating condition information with the reconstructed airfoil concept features to predict aerodynamic parameters; it specifically includes the following steps:

[0111] Step 5.1: Construct a cross-attention module.

[0112] Step 5.2, the working condition feature vector obtained in step 2.4 As a query vector.

[0113] Step 5.3, reconstruct the feature set obtained in step 4.4. As key vectors and value vectors.

[0114] Step 5.4: Input the Query from Step 5.2 and the Key and Value from Step 5.3 into the cross-attention module of Step 5.1, calculate the attention weights, perform weighted aggregation, and output the fused feature vector. .

[0115] Step 5.5: Construct four independent multilayer perceptron (MLP) prediction heads, corresponding to the drag coefficient, lift coefficient, pitch moment coefficient, and pressure center position, respectively.

[0116] Step 5.6, merging the feature vector obtained in step 5.4. Input the four prediction heads from step 5.5 respectively, and the output will be the predicted aerodynamic parameters: predicted drag coefficient. Predicted lift coefficient Predicted pitch moment coefficient and predict the location of the pressure center .

[0117] Step 6: Loss Function Construction and Model Optimization: This step constructs a composite loss function that incorporates prediction error and concept clustering characteristics; specifically, it includes the following steps:

[0118] Step 6.1: Based on the predicted values ​​from Step 5.6 and the true labels from Step 1.9, calculate the mean squared error to obtain the data prediction loss. .

[0119] Step 6.2, based on the allocation probability in step 4.4 and the airfoil concept vector in step 4.1 Calculate the clustering loss to align visual features with their assigned airfoil concepts:

[0120]

[0121] in, For clustering loss, The square of the Euclidean distance. It is the first A visual feature vector.

[0122] Step 6.3, calculate the separation loss to keep the vectors in the airfoil concept pool far apart from each other:

[0123]

[0124] in, For separation loss, For the number of airfoil concepts, and They are the first The and the first A vector of airfoil concepts.

[0125] Step 6.4: The losses calculated in steps 6.1, 6.2, and 6.3 are weighted and summed, and the network parameters are updated using the backpropagation algorithm.

[0126] Step 7, Online Update of Airfoil Concepts: This step dynamically updates the airfoil concept pool during training to ensure its representativeness; it specifically includes the following steps:

[0127] Step 7.1: At the end of each training iteration, the statistics assigned in step 4.4 are... All visual feature vectors of an airfoil concept.

[0128] Step 7.2: Calculate the average vector of all visual feature vectors counted in Step 7.1. .

[0129] Step 7.3, update the value in step 4.1 using the Exponential Moving Average (EMA) strategy. Airfoil concept vector :

[0130]

[0131] in, For the updated concept vectors, The concept vector at the current moment. The momentum coefficient, The average vector calculated in step 7.2.

[0132] Step 8, Generating Intrinsic Interpretability Results: This step, based on the trained model, outputs an intuitive interpretation of the prediction results; it specifically includes the following steps:

[0133] Step 8.1, in the reasoning phase, for the first... Given a given input sample, obtain the allocation probability vector calculated in step 4.3. .

[0134] Step 8.2: In the vector from step 8.1, identify the vector with the highest probability value. An index of airfoil concepts.

[0135] Step 8.3: Retrieve from the training dataset the real aircraft image fragment that is closest to the airfoil concept vector identified in Step 8.2, and use it as a visual representation of the airfoil concept.

[0136] Step 8.4: Output an explanation report, displaying the visual representative image retrieved in Step 8.3, explaining that the current prediction result is based on the high similarity between the input sample and the airfoil concept, thereby achieving an intrinsic explanation based on case comparison.

[0137] The key point of this solution is that it proposes a method for predicting aerodynamic parameters of aircraft by integrating multi-view projection, constructs an "airfoil concept pool" containing learnable vectors, and maps continuous visual features into discrete typical airfoil correlation distributions through an optimal transmission algorithm.

[0138] Meanwhile, the key point of this solution is that it proposes an intrinsically interpretable analysis method based on airfoil concept matching. It abandons the posterior heatmap and directly outputs the matching probability between the input sample and the airfoil concept pool based on the allocation matrix during the model inference process. It also displays the corresponding real samples as visual representatives, providing a transparent decision-making basis based on case comparison.

[0139] Meanwhile, the key point of this scheme is that it proposes an online momentum update mechanism for airfoil concept features. During the training process, instead of directly updating the concept pool through gradients, it updates it by exponential moving average based on the mean of the features assigned in the current batch, ensuring that the learned airfoil concepts have high intra-class compactness and inter-class separability.

[0140] The above description is merely 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. A method for predicting aerodynamic parameters of aircraft based on multi-view airfoil concept association learning, characterized in that: Includes the following steps: Step 1, Construct a hybrid multi-view aircraft geometric projection dataset: Convert the aircraft model in three-dimensional space into a standardized two-dimensional image sequence; Step 2, Feature encoding of flow field conditions: Mapping scalar flow field condition parameters to a high-dimensional feature space; Step 3, Visual Feature Extraction: Use deep neural networks to extract deep geometric features from each viewpoint image; Step 4, Airfoil concept association learning based on optimal transmission: Construct a learnable "airfoil concept pool" and discretize the visual features into an association distribution with airfoil concepts; Step 4 specifically includes the following steps: Step 4.1, initialize an airfoil concept pool. , which includes A learnable concept vector , Indicates the first A vector of airfoil concepts; Step 4.2, based on the obtained visual feature set and the airfoil concept pool initialized in step 4.1 Calculate the dot product similarity between the feature vector and the concept vector; Step 4.3: Based on the similarity calculated in Step 4.2, use the Sinkhorn algorithm to solve for the optimal transmission allocation matrix. The calculation formula is: in, For the first The assignment vector corresponding to each viewpoint To normalize and scale the vector, To normalize and scale the vector, This is the transpose of the airfoil concept pool matrix. For the first Feature vectors from each perspective The regularization coefficient is used. Step 4.4, based on the allocation vector obtained in step 4.3 , the elements Indicates the first The features of the first perspective belong to the first The allocation probability of the airfoil concept is calculated, and a weighted linear combination is performed on the airfoil concept pool in step 4.1 to calculate the probability of the first airfoil concept. Reconstruction features from a single perspective : in, For the first Reconstructed feature vectors from multiple perspectives For the first A vector of airfoil concepts; Step 5, Feature Fusion and Aerodynamic Parameter Prediction: Fusion of operating condition information and reconstructed airfoil concept features, regression prediction of aerodynamic parameters; Step 5 specifically includes the following steps: Step 5.1, construct a cross-attention module; Step 5.2, obtain the working condition feature vector As a query vector (Query); Step 5.3, reconstruct the feature set obtained in step 4.

4. As a key vector (Key) and a value vector (Value); Step 5.4: Input the Query from Step 5.2 and the Key and Value from Step 5.3 into the cross-attention module of Step 5.1, calculate the attention weights, perform weighted aggregation, and output the fused feature vector. ; Step 5.5: Construct four independent multilayer sensor prediction heads, corresponding to the drag coefficient, lift coefficient, pitch moment coefficient, and pressure center position, respectively. Step 5.6, merging the feature vector obtained in step 5.

4. Input the four prediction heads from step 5.5 respectively, and the output will be the predicted aerodynamic parameters: predicted drag coefficient. Predicted lift coefficient Predicted pitch moment coefficient and predict the location of the pressure center ; Step 6, Loss Function Construction and Model Optimization: Construct a composite loss function that includes prediction error and concept clustering characteristics; Step 7, Online Update of Airfoil Concepts: Dynamically update the airfoil concept pool during training to ensure its representativeness; Step 8, Generating Intrinsic Interpretability Results: Based on the trained model, output an intuitive interpretation of the prediction results.

2. The method for predicting aerodynamic parameters of an aircraft based on multi-view airfoil concept association learning as described in claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Obtain the 3D mesh model of the aircraft to be predicted, and calculate the 3D geometric bounding box of the 3D mesh model; Step 1.2: Based on the 3D geometric bounding box calculated in Step 1.1, the 3D mesh model is scaled and translated as a whole so that its geometric center is located at the origin of the coordinate system and is normalized to a unit sphere space. Step 1.3: Using the center of the unit spherical space in Step 1.2 as the origin, set up an orthogonal projection camera group, which includes a predetermined fixed number of orthogonal projection cameras. Step 1.4: Configure the line-of-sight directions of the predetermined number of cameras in Step 1.3 to the front view, rear view, left view, top view, and bottom view of the aircraft, respectively, in order to capture the main view geometry of the aircraft; Step 1.5: Using the center of the unit spherical space in Step 1.2 as the origin, set up a panoramic projection camera group on the equatorial plane. The camera group contains multiple perspective cameras, with the interval between adjacent cameras being a preset angle and the elevation angle being uniformly set. Step 1.6: Using the orthogonal projection camera group set in Step 1.3 and the surround projection camera group set in Step 1.5, perform synchronous off-screen rendering on the normalized 3D mesh model in Step 1.2 to generate multiple 2D projection images. Step 1.7: Adjust the resolution of the multiple images generated in Step 1.6 to a uniform value. The pixel values ​​were then normalized to construct a multi-view image set. ; Step 1.8: For each aircraft model, set the flow field operating conditions, including the Mach number. and attack angle ; Step 1.9: Using a computational fluid dynamics solver, calculate the actual aerodynamic parameter labels under the operating condition in Step 1.8, including the actual drag coefficient. True lift coefficient Actual pitch moment coefficient and the actual center of pressure .

3. The method for predicting aerodynamic parameters of an aircraft based on multi-view airfoil concept association learning as described in claim 2, characterized in that: Step 2 specifically includes the following steps: Step 2.1, extract the Mach number set in Step 1.

8. and attack angle Construct the original working condition vector ; Step 2.2: Construct a working condition encoder, which consists of a multi-layer sensing mechanism; Step 2.3, the original working condition vector constructed in Step 2.1 Input the operating condition encoder constructed in step 2.2; Step 2.4: After the linear transformation layer and nonlinear activation function mapping of the working condition encoder, a high-dimensional working condition feature vector is output. .

4. The method for predicting aerodynamic parameters of an aircraft based on multi-view airfoil concept association learning as described in claim 3, characterized in that: Step 3 specifically includes the following steps: Step 3.1: Construct a shared-weight visual Transformer as the backbone feature extraction network; Step 3.2, combine the multi-view image set obtained in Step 1.7 Each image in the process is input into the backbone feature extraction network constructed in step 3.1; Step 3.3: Extract the category token vector from the network output in Step 3.2 to obtain a set of visual features containing multiple perspectives. ,in Indicates the first Feature vectors from each perspective.

5. The method for predicting aerodynamic parameters of an aircraft based on multi-view airfoil concept association learning as described in claim 4, characterized in that: Step 6 specifically includes the following steps: Step 6.1: Based on the predicted values ​​from Step 5.6 and the true labels from Step 1.9, calculate the mean squared error to obtain the data prediction loss. ; Step 6.2, based on the allocation probability in step 4.4 and the airfoil concept vector in step 4.1 Calculate the clustering loss to align visual features with their assigned airfoil concepts: in, For clustering loss, The square of the Euclidean distance. It is the first One visual feature vector; Step 6.3, calculate the separation loss to keep the vectors in the airfoil concept pool far apart from each other: in, For separation loss, For the number of airfoil concepts, and They are the first The and the first A vector of airfoil concepts; Step 6.4: The losses calculated in steps 6.1, 6.2, and 6.3 are weighted and summed, and the network parameters are updated using the backpropagation algorithm.

6. The method for predicting aerodynamic parameters of an aircraft based on multi-view airfoil concept association learning as described in claim 5, characterized in that: Step 7 specifically includes the following steps: Step 7.1: At the end of each training iteration, the statistics assigned in step 4.4 are... All visual feature vectors of an airfoil concept; Step 7.2: Calculate the average vector of all visual feature vectors counted in Step 7.

1. ; Step 7.3, update the value in step 4.1 using the exponential moving average strategy. Airfoil concept vector : in, For the updated concept vectors, The concept vector at the current moment. The momentum coefficient, The average vector calculated in step 7.

2.

7. The method for predicting aerodynamic parameters of an aircraft based on multi-view airfoil concept association learning as described in claim 6, characterized in that: Step 8 specifically includes the following steps: Step 8.1, in the reasoning phase, for the first... Given a given input sample, obtain the allocation probability vector calculated in step 4.

3. ; Step 8.2: In the vector from step 8.1, identify the vector with the highest probability value. An index of airfoil concepts; Step 8.3: Retrieve from the training dataset the real aircraft image fragment that is closest to the airfoil concept vector identified in Step 8.2, and use it as a visual representation of the airfoil concept; Step 8.4: Output an explanation report, displaying the visual representative image retrieved in Step 8.3, explaining that the current prediction result is based on the high similarity between the input sample and the airfoil concept, thereby achieving an intrinsic explanation based on case comparison.

Images