An interpretable intelligent analysis method for three-dimensional aircraft aerodynamic parameter prediction

Abstract

The application discloses an interpretable intelligent analysis method for three-dimensional aircraft aerodynamic parameter prediction, and belongs to the technical field of aerodynamics, deep learning and computer vision, which first converts a three-dimensional model into a two-dimensional image sequence through standardized orthogonal projection, constructs a deep neural network containing five parallel branches to extract geometric features, combines with flow field working condition coding, carries out training by introducing a fine-grained physical constraint step, generates an aerodynamic sensitivity heat map by using an attention mechanism, and verifies the interpretation result by using occlusion sensitivity analysis; the scheme reduces the dimension of a complex three-dimensional problem through standardized orthogonal projection, combines with the global modeling capability of ViT, maintains the second-level prediction speed, and realizes higher prediction accuracy than traditional point cloud networks.

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 and providing visual interpretation and analysis of aerodynamic parameters of three-dimensional aircraft based on multi-view images and physical information-constrained deep neural networks. Background Technology

[0002] Accurate acquisition of aircraft aerodynamic parameters (such as lift coefficient, drag coefficient, pitch moment coefficient, etc.) is fundamental to aircraft aerodynamic layout design and performance evaluation. Traditionally, the acquisition of aerodynamic parameters has mainly relied on wind tunnel testing and computational fluid dynamics (CFD) numerical simulation.

[0003] While wind tunnel testing yields the most reliable results, the construction and operation costs of these facilities are extremely high, model fabrication and testing cycles are lengthy, and limitations in test section size and Reynolds number simulation capabilities make large-scale scheme selection difficult in the early stages of aircraft conceptual design. CFD numerical simulation methods simulate flow fields by solving the Navier-Stokes equations. While offering high accuracy, these methods require high-quality mesh generation techniques and involve complex iterative solutions. For complex 3D aircraft shapes, a single high-precision calculation can often take hours or even days, consuming enormous computational resources. In the optimization design phase, where thousands of shape options need to be evaluated, the "speed bottleneck" of CFD methods severely restricts the efficiency of design iterations. Existing deep learning models are mostly "black box" models, only outputting numerical results and unable to identify which local features of the aircraft shape (such as wing leading edge shape and nose curvature) play a decisive role in specific aerodynamic parameters, making it difficult for engineers to trust the models and perform targeted optimizations.

[0004] Currently, for the task of predicting the aerodynamic parameters of three-dimensional aircraft, data-driven deep learning-based aerodynamic prediction methods aim to construct a mapping relationship between the aircraft's shape and aerodynamic parameters through neural networks to achieve rapid predictions within seconds. However, most existing intelligent prediction schemes only focus on the numerical mapping accuracy from input to output and cannot demonstrate the decision-making mechanism inside the model. Engineering designers cannot know which local features of the aircraft the model is based on for its predictions. This lack of interpretability makes it difficult for designers to trust the model's prediction results and to perform targeted shape optimization based on the model's feedback, thus limiting its application in practical engineering.

[0005] Therefore, there is an urgent need to develop a three-dimensional aerodynamic parameter prediction framework for aircraft that is both highly accurate and efficient, and has interpretability. Summary of the Invention

[0006] The purpose of this invention is to provide an interpretable intelligent analysis method for predicting aerodynamic parameters of three-dimensional aircraft, addressing the aforementioned shortcomings and solving the challenges of interpretability in existing intelligent prediction models.

[0007] This invention is achieved through the following scheme:

[0008] An interpretable intelligent analysis method for predicting aerodynamic parameters of three-dimensional aircraft includes the following steps:

[0009] Step 1, Construct a standardized multi-view aircraft geometric projection dataset: Using computer graphics technology, convert 3D aircraft geometric models of different sizes into standardized multi-view 2D image sequences;

[0010] Step 2, Vectorized encoding of flight conditions: Mapping scalar flow field condition parameters to a high-dimensional feature space to obtain condition feature vectors;

[0011] Step 3, Construct a multi-branch twin visual feature extraction network: Construct a neural network containing a predetermined number of parallel branches. Each branch adopts a visual Transformer architecture to process the projected images from a predetermined number of viewpoints to extract the geometric features of the latent spatial association.

[0012] Step 4, Multimodal Feature Fusion and Aerodynamic Parameter Prediction: The geometric features extracted in Step 3 are fused with the working condition feature vector extracted in Step 2, and the aerodynamic parameters are regressed through the prediction head.

[0013] Step 5, Calculation of loss function based on physical information constraints: Derive intermediate physical quantities according to aerodynamic formulas and calculate composite loss function;

[0014] Step 6, interpretability analysis based on built-in attention mechanism: Using the self-attention mechanism built into the trained ViT network, an aerodynamic sensitivity heatmap is generated and verified by combining occlusion analysis to achieve a visual interpretation of the prediction results.

[0015] Step 7, interpretability verification based on occlusion sensitivity: Introduce occlusion sensitivity analysis and quantify the impact of image regions on prediction results through systematic occlusion experiments.

[0016] Step 1 specifically includes the following steps:

[0017] Step 1.1, 3D model normalization processing: Obtain the 3D mesh model of the aircraft and calculate the 3D bounding box of the model. Scale and translate the model as a whole to normalize it into a unit cube space to ensure the scale consistency of all samples.

[0018] Step 1.2, Rendering environment configuration: The rendering environment is built using VTK-based off-screen rendering technology. The background is set to a solid color, and the surface material of the model is set to standard diffuse or no lighting mode to highlight the geometric contours.

[0019] Step 1.3, Orthographic Projection Camera Setup: Set the camera relative to the center of the unit cube bounding box. A fixed orthogonal projection camera, The camera captures the front, rear, left, top, and bottom views of the corresponding aircraft.

[0020] Step 1.4, Multi-view image rendering: Using the camera configured in Step 1.3, render the aircraft model normalized in Step 1.1 according to the configuration in Step 1.2 to generate a two-dimensional projection image in RGB format;

[0021] Step 1.5, Image Unification and Storage: Adjust the resolution of the image rendered in Step 1.4 to a uniform value. ,in Image height, Let be the image width, and save it as a PNG format; let be the set of multi-view images of the k-th sample. ,in Indicates the first Image matrix from multiple perspectives;

[0022] Step 1.6, Tag and Operating Condition Acquisition: Set different Mach numbers for each aircraft model. and attack angle As the flow field operating conditions, the actual aerodynamic parameters under these conditions are calculated using a high-precision CFD solver. C x Represents the axial force coefficient, C z Represents the normal force coefficient, C D Represents the drag coefficient, C L Represents the lift coefficient, C m Represents the pitching moment coefficient, C cp This indicates the location of the pressure center.

[0023] Step 2 specifically includes the following steps:

[0024] Step 2.1, Construction of working condition vector: Extract the Mach number from step 1.6 and attack angle Construct the original working condition vector ;

[0025] Step 2.2, construct a working condition encoder, which consists of a multilayer sensing mechanism: The encoder constructed in Step 2.1... The input is an MLP, which passes through a linear transformation layer and a nonlinear activation function to map it to a dimension of... Working condition feature vector .

[0026] Step 3 specifically includes the following steps:

[0027] Step 3.1, Image segmentation: For the first segment in step 1.5... Image matrix from multiple perspectives Divide it spatially into The size is Non-overlapping image patches;

[0028] Step 3.2, Image Patch Flattening and Projection: Flatten each image patch from Step 3.1 into a one-dimensional vector, and map it to a feature vector through a linear projection layer with shared weights: Let the first... The feature vectors of each image patch are ,in ;

[0029] Step 3.3, Category Token Insertion: In step 3.2, feature vector... Insert a learnable category token vector into the head. It is used to aggregate global information;

[0030] Step 3.4, Positional Encoding Overlay: Generating and Adding to Feature Vectors Corresponding positional encoding vector , and Adding each element one by one yields the input sequence. ;

[0031] Step 3.5, Multi-layer Transformer Encoding: Encode the input sequence... Input to contain In the Transformer encoder of the layer, feature interaction is achieved through a multi-head self-attention mechanism and a feedforward network to obtain the output sequence. ;

[0032] Step 3.6, Global Feature Extraction: From the output sequence Extract the first vector as the geometric feature vector of that viewpoint. ;

[0033] Step 3.7: Perform steps 3.1 to 3.6 for each of the five perspectives.

[0034] Step 4 specifically includes the following steps:

[0035] Step 4.1, Feature Concatenation: Combine the geometric feature vectors output in Step 3.6. Compared with the working condition feature vector output in step 2.2 By concatenating the channels, a fused feature vector is obtained. ;

[0036] Step 4.2, Prediction Mapping: Construct a multi-head prediction network, where each head corresponds to a gaseous parameter to be predicted; the prediction head is composed of an MLP: the parameters obtained in Step 4.1 are... Inputting the data into a multi-head prediction network, the predicted aerodynamic parameters are regressed through a fully connected layer. , , and ,in For drag coefficient, For lift coefficient, For pitch moment coefficient, This is the location of the pressure center.

[0037] Step 5 includes the following steps:

[0038] Step 5.1, calculate and derive the axial force coefficient;

[0039] Step 5.2, calculate and derive the normal force coefficient;

[0040] Step 5.3, calculate and derive the pitching moment coefficient;

[0041] Step 5.4, calculate the actual physical constraint labels;

[0042] Step 5.5, calculate the data-driven loss term;

[0043] Step 5.6, calculate the physical constraint loss term;

[0044] Step 5.7, calculate the total loss function.

[0045] Specifically, step 5 includes the following steps:

[0046] Step 5.1, using the drag coefficient predicted in step 4.2 and lift coefficient Combined with the angle of attack in step 2.1 Calculate and derive the axial force coefficient :

[0047]

[0048] Step 5.2, using the drag coefficient predicted in step 4.2 and lift coefficient Combined with angle of attack Calculate and derive the normal force coefficient :

[0049]

[0050] Step 5.3, using the normal force coefficient derived in Step 5.2. and the predicted pressure center location in step 4.2 Combined with the preset reference length and reference center location Calculate and derive the pitching moment coefficient :

[0051] ;

[0052] Step 5.4: Utilize the actual pneumatic tag obtained from the CFD calculations in Step 1.6. In and and use the formula Calculate the true axial force coefficient Using formula Calculate the true normal force coefficient ;

[0053] Step 5.5: Calculate the predicted aerodynamic parameters obtained in step 4.2. , , and Labels of the actual aerodynamic parameters obtained in step 1.6 C in D C L C m C cp The mean squared error is used to obtain the data-driven loss. :

[0054] ;

[0055] Step 5.6: Calculate the derived axial force coefficients from steps 5.1-5.3. Derivation of normal force coefficient And the derivation of the pitching moment coefficient Labels of the actual aerodynamic parameters obtained in step 1.6 C x C z C m The mean square error is used to obtain the physical constraint loss. :

[0056] ;

[0057] Step 5.7: Weight the data-driven loss and the physical constraint loss to obtain the total loss used for backpropagation. :

[0058]

[0059] in These are the physical constraint weighting coefficients.

[0060] Step 6 specifically includes the following steps:

[0061] Step 6.1, Attention Weight Extraction: During the model inference stage, from the perspective of... The last layer of the ViT branch is extracted from the attention matrix. ;

[0062] Step 6.2, Extract the category token attention row: the self-attention matrix extracted from Step 6.1 Extract the first row to obtain the attention weight vector. ;

[0063] Step 6.3, Spatial Reorganization: Reorganize the vector obtained in Step 6.2 Reorganized into a two-dimensional matrix according to the spatial distribution of image patches. ;

[0064] Step 6.4, Interpolation Upsampling: Use the bilinear interpolation algorithm to upsample the matrix. Upsampled to the same resolution as the original input image A heat map was obtained. ;

[0065] Step 6.5, Visual Overlay: Overlay the heatmap obtained in Step 6.4. Normalize and map to a pseudo-color image, then overlay it onto the first... Image matrix from multiple perspectives The upper part displays the aerodynamically sensitive area.

[0066] Step 7 includes the following steps:

[0067] Step 7.1, define the occlusion mask;

[0068] Step 7.2, perform sliding occlusion;

[0069] Step 7.3, generating the occluded image;

[0070] Step 7.4, each occluded image The data is input into the trained model, while keeping other viewpoint images and operating conditions unchanged. Forward propagation is then performed to obtain new aerodynamic parameter prediction vectors. ;

[0071] Step 7.5, Sensitivity index calculation;

[0072] Step 7.6, Sensitivity Map Construction;

[0073] Step 7.7, double comparison verification.

[0074] Specifically, step 7 includes the following steps:

[0075] Step 7.1, define a size as Square occlusion mask block Its pixel value is uniformly set to 0 to cover the image content;

[0076] Step 7.2, in the section to be verified Image matrix from multiple perspectives Above, set the sliding step size. Moving occlusion mask block Iterate through all regions of the image;

[0077] Step 7.3, for each sliding position Generate a corresponding occlusion image ;

[0078] Step 7.4, each occluded image The data is input into the trained model, while keeping other viewpoint images and operating conditions unchanged. Forward propagation is then performed to obtain new aerodynamic parameter prediction vectors. ;

[0079] Step 7.5, calculate the new aerodynamic parameter prediction vector. Compared with the original unoccluded prediction value The absolute error change between the two values ​​is used as the sensitivity score for that location. :

[0080]

[0081] in The L1 norm of a vector is used to comprehensively measure the magnitude of change of all aerodynamic parameters.

[0082] Step 7.6: Calculate the sensitivity scores for all locations. Mapping back to the corresponding image coordinates, an occlusion sensitivity map is constructed. ;

[0083] Step 7.7, use the heatmap generated in step 6.4. Compared with the occlusion sensitivity map generated in step 7.6 A comparative analysis was conducted; if the overlap of the highlighted areas of the two is greater than 95%, the reliability of the model interpretation results was cross-validated, confirming that the occlusion sensitivity map is a key geometric feature affecting aerodynamic performance.

[0084] In step 1, a visual camera and a lidar are specifically used.

[0085] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0086] 1. Fast prediction speed and high accuracy: By reducing the dimensionality of complex 3D problems through standardized orthogonal projection and combining ViT's global modeling capabilities, it achieves higher prediction accuracy than traditional point cloud networks while maintaining a prediction speed of seconds.

[0087] 2. Strong physical consistency: Through fine-grained physical derivation steps and constraint loss, the model is forced to learn the rigorous mathematical relationships between lift, drag, axial force, normal force and torque, which significantly improves the model's generalization ability and physical credibility beyond the training data.

[0088] 3. High reliability and interpretability: Not only does it generate intuitive heatmaps through an attention mechanism, but it also specifically introduces occlusion sensitivity analysis as a verification method. This dual verification mechanism effectively avoids the randomness of a single interpretation method, accurately identifies key geometric components affecting aerodynamic performance, and significantly improves engineers' confidence in the model's decisions. Attached Figure Description

[0089] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0090] 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.

[0091] 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.

[0092] 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.

[0093] 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.

[0094] Example 1

[0095] like Figure 1 As shown, the present invention provides a technical solution:

[0096] An interpretable intelligent analysis method for predicting aerodynamic parameters of three-dimensional aircraft includes the following steps:

[0097] Step 1: Construct a standardized multi-view aircraft geometric projection dataset. This step aims to utilize computer graphics techniques to convert 3D aircraft geometric models of different sizes into a standardized multi-view 2D image sequence; it specifically includes the following steps:

[0098] Step 1.1, 3D model normalization. Obtain the 3D mesh model of the aircraft and calculate the 3D bounding box of the model. Scale and translate the entire model to normalize it to a unit cube space, ensuring scale consistency across all samples.

[0099] Step 1.2, Rendering Environment Configuration. The rendering environment is set up using off-screen rendering technology based on VTK (Visualization Toolkit). The background is set to a solid color, and the model surface material is set to standard diffuse or no-light mode to highlight the geometric contours.

[0100] Step 1.3, Orthographic Projection Camera Setup. Set the camera relative to the center of the unit cube bounding box. A fixed orthogonal projection camera, The camera captures the front, rear, left, top, and bottom views of the corresponding aircraft.

[0101] Step 1.4, Multi-view image rendering. Using the camera configured in Step 1.3, the normalized aircraft model from Step 1.1 is rendered according to the configuration in Step 1.2, generating 5 two-dimensional projection images in RGB format.

[0102] Step 1.5, Image Unification and Storage. The resolution of the five images rendered in Step 1.4 is uniformly adjusted to... ,in Image height, Let be the image width, and save it as a PNG file. Let the set of multi-view images for the k-th sample be denoted as . ,in Indicates the first An image matrix from multiple perspectives.

[0103] Step 1.6, Tag and Operating Condition Acquisition: Set different Mach numbers for each aircraft model. and attack angle As the flow field operating condition, the actual aerodynamic parameters under this condition are calculated using a high-precision CFD solver. C x Represents the axial force coefficient, C z Represents the normal force coefficient, C D Represents the drag coefficient, C L Represents the lift coefficient, C m Represents the pitching moment coefficient, C cp This indicates the location of the pressure center.

[0104] Step 2, Vectorization Encoding of Flight Conditions. This step maps scalar flow field parameters to a high-dimensional feature space to obtain condition feature vectors. Specifically, it includes the following steps:

[0105] Step 2.1, Construction of the working condition vector. Extract the Mach number from Step 1.6. and attack angle Construct the original working condition vector .

[0106] Step 2.2: Construct a working condition encoder, which consists of a multilayer perceptron (MLP). The encoder constructed in Step 2.1... The input is an MLP, which passes through a linear transformation layer and a nonlinear activation function to map it to a dimension of... Working condition feature vector .

[0107] Step 3: Construct a five-branch twin visual feature extraction network. This step constructs a neural network with five parallel branches. Each branch adopts a Visual Transformer (VIT) architecture to process the projected images from the five viewpoints, extracting geometric features with latent spatial relationships. Specifically, it includes the following steps:

[0108] Step 3.1, image segmentation. For step 1.5... Image matrix from multiple perspectives Divide it spatially into The size is Non-overlapping image patches.

[0109] Step 3.2, Image Patch Flattening and Projection. Each image patch from Step 3.1 is flattened into a one-dimensional vector and mapped to a feature vector through a linear projection layer with shared weights. Let the... The feature vectors of each image patch are ,in .

[0110] Step 3.3, Category Token Insertion. (This is related to step 3.2, Feature Vector.) Insert a learnable class token vector at the head. It is used to aggregate global information.

[0111] Step 3.4, positional encoding overlay. This generates the feature vector. Corresponding positional encoding vector , and Adding each element one by one yields the input sequence. .

[0112] Step 3.5, Multi-layer Transformer Encoding. The input sequence... Input to contain In the Transformer encoder of the layer, feature interaction is achieved through a multi-head self-attention mechanism and a feedforward network to obtain the output sequence. .

[0113] Step 3.6, Global Feature Extraction. From the output sequence Extract the first vector (i.e., the vector corresponding to the category token) as the geometric feature vector of that viewpoint. .

[0114] Step 3.7: Perform steps 3.1 to 3.6 on each of the five viewpoints to obtain the results. .

[0115] Step 4, Multimodal Feature Fusion and Aerodynamic Parameter Prediction. This step fuses the geometric features extracted in Step 3 with the operating condition feature vector extracted in Step 2, and regresses the aerodynamic parameters using a prediction head. Specifically, it includes the following steps:

[0116] Step 4.1, Feature Concatenation. The geometric feature vector output from Step 3.6 is then concatenated. Compared with the working condition feature vector output in step 2.2 By concatenating the channels, a fused feature vector is obtained. .

[0117] Step 4.2, Prediction Mapping. Construct a multi-head prediction network, with each head corresponding to aerodynamic parameter to be predicted. The prediction heads are composed of MLPs. The parameters obtained in Step 4.1 are then mapped... Inputting the data into a multi-head prediction network, the predicted aerodynamic parameters are regressed through a fully connected layer. (Drag coefficient) (lift coefficient) (Pitching moment coefficient) and (Position of the pressure center).

[0118] Step 5: Calculation of the loss function based on physical information constraints. This step derives intermediate physical quantities based on aerodynamic formulas and calculates the composite loss function. Specifically, it includes the following steps:

[0119] Step 5.1, calculate and derive the axial force coefficient. Utilize the drag coefficient predicted in Step 4.2. and lift coefficient Combined with the angle of attack in step 2.1 Calculate and derive the axial force coefficient :

[0120]

[0121] Step 5.2, calculate and derive the normal force coefficient. Utilize the drag coefficient predicted in Step 4.2. and lift coefficient Combined with angle of attack Calculate and derive the normal force coefficient :

[0122]

[0123] Step 5.3, calculate and derive the pitching moment coefficient. Utilize the normal force coefficient derived in Step 5.2. and the predicted pressure center location in step 4.2 Combined with the preset reference length and reference center location Calculate and derive the pitching moment coefficient :

[0124]

[0125] Step 5.4: Utilize the actual pneumatic tag obtained from the CFD calculations in Step 1.6. In and and use the formula Calculate the true axial force coefficient Using formula Calculate the true normal force coefficient .

[0126] Step 5.5, calculate the data-driven loss term. Calculate the predicted aerodynamic parameters obtained in Step 4.2. , , and Labels of the actual aerodynamic parameters obtained in step 1.6 C in D C L C m C cpThe mean squared error is used to obtain the data-driven loss. :

[0127]

[0128] Step 5.6, calculate the physical constraint loss term. Calculate the axial force coefficients derived in steps 5.1-5.3. Derivation of normal force coefficient And the derivation of the pitching moment coefficient Labels of the actual aerodynamic parameters obtained in step 1.6 C x C z C m The mean square error is used to obtain the physical constraint loss. :

[0129]

[0130] Step 5.7, calculate the total loss function. The data-driven loss and the physical constraint loss are weighted and summed to obtain the total loss used for backpropagation. :

[0131]

[0132] in These are the physical constraint weighting coefficients.

[0133] Step 6, Interpretability Analysis Based on Built-in Attention Mechanism. This step utilizes the built-in self-attention mechanism of the trained ViT network to generate an aerodynamic sensitivity heatmap, which is then validated using occlusion analysis to achieve a visual interpretation of the prediction results. Specifically, it includes the following steps:

[0134] Step 6.1, Attention Weight Extraction. During the model inference stage, from the perspective of... The last layer of the ViT branch is extracted from the attention matrix. .

[0135] Step 6.2, extract the category token attention row. This is derived from the self-attention matrix extracted in Step 6.1. Extract the first row (corresponding to the attention of the corresponding category token to other tokens) to obtain the attention weight vector. .

[0136] Step 6.3, Spatial Reorganization. The vector obtained in Step 6.2... Reorganized into a two-dimensional matrix according to the spatial distribution of image patches. .

[0137] Step 6.4, Interpolation Upsampling. The matrix is ​​upsampled using a bilinear interpolation algorithm. Upsampled to the same resolution as the original input image A heat map was obtained. .

[0138] Step 6.5, Visualization Overlay. The heatmap obtained in Step 6.4 is then overlaid. Normalize and map to a pseudo-color image, then overlay it onto the first... Image matrix from multiple perspectives The upper part displays the aerodynamically sensitive area.

[0139] Step 7, Interpretability Verification Based on Occlusion Sensitivity. To verify the accuracy of the heatmap generated in Step 6, this step introduces occlusion sensitivity analysis, quantifying the impact of image regions on the prediction results through systematic occlusion experiments. Specifically, it includes the following steps:

[0140] Step 7.1, define the occlusion mask. Define a mask with a size of... Square occlusion mask block Its pixel value is uniformly set to 0 to cover the image content.

[0141] Step 7.2, perform sliding occlusion. In the section to be verified... Image matrix from multiple perspectives Above, set the sliding step size. Moving occlusion mask block Iterate through all regions of the image.

[0142] Step 7.3, Occlusion Image Generation. For each sliding position... Generate a corresponding occlusion image .

[0143] Step 7.4, each occluded image The data is input into the trained model, while keeping other viewpoint images and operating conditions unchanged. Forward propagation is then performed to obtain new aerodynamic parameter prediction vectors. .

[0144] Step 7.5, Sensitivity Index Calculation. Calculate the new aerodynamic parameter prediction vector. Compared with the original unoccluded prediction value The absolute error change between the two values ​​is used as the sensitivity score for that location. :

[0145]

[0146] in The L1 norm of a vector is used to comprehensively measure the magnitude of change of all aerodynamic parameters.

[0147] Step 7.6, Sensitivity Map Construction. Calculate the sensitivity scores for all locations. Mapping back to the corresponding image coordinates, an occlusion sensitivity map is constructed. .

[0148] Step 7.7, Double Comparison Verification. The heatmap generated in Step 6.4... Compared with the occlusion sensitivity map generated in step 7.6 A comparative analysis was conducted. If the overlap between the highlighted areas (i.e., the high attention weight area and the high sensitivity area) of the two is greater than 95%, then the reliability of the model interpretation results was cross-validated, confirming that the occlusion sensitivity map is a key geometric feature affecting aerodynamic performance.

[0149] This solution provides a framework for predicting the aerodynamic parameters of a 3D aircraft by integrating multi-view and physical information constraints. It includes five parallel and structurally independent Visual Transformer (ViT) branch networks, which independently extract global geometric feature vectors from five fixed orthogonal viewpoints of the aircraft. These feature vectors are then fused with operational features to form a unified multimodal feature vector. The aerodynamic parameters of the 3D aircraft are directly predicted by regression through a multi-head prediction network. At the same time, a physical derivation link is constructed based on aerodynamic formulas to constrain the model output to meet the physical conservation relationship of aerodynamic vector transformation.

[0150] This solution is based on an interpretability analysis method for aircraft aerodynamic prediction using the built-in attention mechanism of ViT. It utilizes the self-attention matrix output by the last layer Transformer encoder in the ViT network architecture to extract the attention weight vectors of the category tokens for all image patches. These weight vectors are then reconstructed into a two-dimensional matrix according to the original spatial positions of the image patches. Bilinear interpolation upsampling is then used to generate an attention heatmap reflecting the contribution of aircraft geometry to aerodynamic parameters, which is then highlighted on the original projected image. This method requires no additional training or external models; it directly utilizes the internal parameters of the prediction model for interpretation.

[0151] This scheme is based on an interpretability analysis method for aircraft aerodynamic prediction using occlusion sensitivity verification. It establishes an occlusion sensitivity analysis process, defines a fixed-size mask block, and performs sliding occlusion on the input image by setting a step size, generating a series of locally occluded image samples. The occluded images are then input into a trained model for re-prediction. The L1 norm distance between the predicted aerodynamic parameter vector and the original predicted vector is calculated as the sensitivity score for that region. The sensitivity scores of all regions are mapped back to the image space to construct an occlusion sensitivity map, which is used to accurately locate key regions affecting aerodynamic performance and can also serve as a cross-validation method for the built-in attention heatmap.

[0152] 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 interpretable intelligent analysis method for predicting aerodynamic parameters of three-dimensional aircraft, characterized in that: Includes the following steps: Step 1, Construct a standardized multi-view aircraft geometric projection dataset: Using computer graphics technology, convert 3D aircraft geometric models of different sizes into standardized multi-view 2D image sequences; Step 2, Vectorized encoding of flight conditions: Mapping scalar flow field condition parameters to a high-dimensional feature space to obtain condition feature vectors; Step 3, Construct a multi-branch twin visual feature extraction network: Construct a neural network containing a predetermined number of parallel branches. Each branch adopts a visual Transformer architecture to process the projected images from a predetermined number of viewpoints to extract the geometric features of the latent spatial association. Step 4, Multimodal Feature Fusion and Aerodynamic Parameter Prediction: The geometric features extracted in Step 3 are fused with the working condition feature vector extracted in Step 2, and the aerodynamic parameters are regressed through the prediction head. Step 5, Calculation of loss function based on physical information constraints: Derive intermediate physical quantities according to aerodynamic formulas and calculate composite loss function; Specifically, step 5 includes the following steps: Step 5.1, using the drag coefficient predicted in Step 4 and lift coefficient Combined with the angle of attack in step 2 Calculate and derive the axial force coefficient : Step 5.2, using the drag coefficient predicted in Step 4 and lift coefficient Combined with angle of attack Calculate and derive the normal force coefficient : Step 5.3, using the normal force coefficient derived in Step 5.

2. and the predicted pressure center location in step 4 Combined with the preset reference length and reference center location Calculate and derive the pitching moment coefficient : ； Step 5.4, using the labels of the actual aerodynamic parameters obtained from the CFD calculation in Step 1. In and and use the formula Calculate the true axial force coefficient Using formula Calculate the true normal force coefficient ; Step 5.5: Calculate the predicted aerodynamic parameters obtained in Step 4. , , and Labels of the actual aerodynamic parameters obtained in step 1 C in D C L C m X cp The mean squared error is used to obtain the data-driven loss. : ； Step 5.6: Calculate the derived axial force coefficients from steps 5.1-5.

3. Derivation of normal force coefficient And the derivation of the pitching moment coefficient Labels of the actual aerodynamic parameters obtained in step 1 C x C z C m The mean square error is used to obtain the physical constraint loss. : ； Step 5.7: Weight the data-driven loss and the physical constraint loss to obtain the total loss used for backpropagation. : in These are the physical constraint weighting coefficients; Step 6, interpretability analysis based on built-in attention mechanism: Using the self-attention mechanism built into the trained ViT network, an aerodynamic sensitivity heatmap is generated and verified by combining occlusion analysis to achieve a visual interpretation of the prediction results. Step 7, interpretability verification based on occlusion sensitivity: Introduce occlusion sensitivity analysis and quantify the impact of image regions on prediction results through systematic occlusion experiments.

2. The interpretable intelligent analysis method for predicting aerodynamic parameters of three-dimensional aircraft as described in claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1, 3D model normalization processing: Obtain the 3D mesh model of the aircraft and calculate the 3D bounding box of the model. Scale and translate the model as a whole to normalize it into a unit cube space to ensure the scale consistency of all samples. Step 1.2, Rendering environment configuration: The rendering environment is built using VTK-based off-screen rendering technology. The background is set to a solid color, and the surface material of the model is set to standard diffuse or no lighting mode to highlight the geometric contours. Step 1.3, Orthographic Projection Camera Setup: Set the camera relative to the center of the unit cube bounding box. A fixed orthogonal projection camera, The camera captures the front, rear, left, top, and bottom views of the corresponding aircraft. Step 1.4, Multi-view image rendering: Using the camera configured in Step 1.3, render the aircraft model normalized in Step 1.1 according to the configuration in Step 1.2 to generate a two-dimensional projection image in RGB format; Step 1.5, Image Unification and Storage: Adjust the resolution of the image rendered in Step 1.4 to a uniform value. ,in Image height, Let be the image width, and save it as a PNG format; let be the set of multi-view images of the k-th sample. ,in Indicates the first Image matrix from multiple perspectives; Step 1.6, Tag and Operating Condition Acquisition: Set different Mach numbers for each aircraft model. and attack angle As the flow field operating conditions, the actual aerodynamic parameters under these conditions are calculated using a high-precision CFD solver. C x Represents the axial force coefficient, C z Represents the normal force coefficient, C D Represents the drag coefficient, C L Represents the lift coefficient, C m Represents pitch moment coefficient, X cp This indicates the location of the pressure center.

3. The interpretable intelligent analysis method for predicting aerodynamic parameters of three-dimensional aircraft as described in claim 2, characterized in that: Step 2 specifically includes the following steps: Step 2.1, Construction of working condition vector: Extract the Mach number from step 1.6 and attack angle Construct the original working condition vector ; Step 2.2, construct a working condition encoder, which consists of a multilayer sensing mechanism: The encoder constructed in Step 2.1... The input is an MLP, which passes through a linear transformation layer and a nonlinear activation function to map it to a dimension of... Working condition feature vector .

4. The interpretable intelligent analysis method for predicting aerodynamic parameters of three-dimensional aircraft as described in claim 3, characterized in that: Step 3 specifically includes the following steps: Step 3.1, Image segmentation: For the first segment in step 1.5... Image matrix from multiple perspectives Divide it spatially into The size is Non-overlapping image patches; Step 3.2, Image Patch Flattening and Projection: Flatten each image patch from Step 3.1 into a one-dimensional vector, and map it to a feature vector through a linear projection layer with shared weights: Let the first... The feature vectors of each image patch are ,in ; Step 3.3, Category Token Insertion: In step 3.2, feature vector... Insert a learnable category token vector into the head. It is used to aggregate global information; Step 3.4, Positional Encoding Overlay: Generating and Adding to Feature Vectors Corresponding position encoding vector to Adding each element one by one yields the input sequence. ; Step 3.5, Multi-layer Transformer Encoding: Encode the input sequence... Input to contain In the Transformer encoder of the layer, feature interaction is achieved through a multi-head self-attention mechanism and a feedforward network to obtain the output sequence. ; Step 3.6, Global Feature Extraction: From the output sequence Extract the first vector as the geometric feature vector of that viewpoint. ; Step 3.7: Perform steps 3.1 to 3.6 on each of the five viewpoints to obtain the results. .

5. The interpretable intelligent analysis method for predicting aerodynamic parameters of a three-dimensional aircraft as described in claim 4, characterized in that: Step 4 specifically includes the following steps: Step 4.1, Feature Concatenation: Combine the geometric feature vectors output in Step 3.

6. Compared with the working condition feature vector output in step 2.2 By concatenating the channels, a fused feature vector is obtained. ; Step 4.2, Prediction Mapping: Construct a multi-head prediction network, where each head corresponds to a gaseous parameter to be predicted; the prediction head is composed of an MLP: the parameters obtained in Step 4.1 are... Inputting the data into a multi-head prediction network, the predicted aerodynamic parameters are regressed through a fully connected layer. , , and ,in For drag coefficient, For lift coefficient, For pitch moment coefficient, This is the location of the pressure center.

6. The interpretable intelligent analysis method for predicting aerodynamic parameters of a three-dimensional aircraft as described in claim 5, characterized in that: Step 6 specifically includes the following steps: Step 6.1, Attention Weight Extraction: During the model inference stage, from the perspective of... The last layer of the ViT branch is extracted from the attention matrix. ; Step 6.2, Extract the category token attention row: the self-attention matrix extracted from Step 6.1 Extract the first row to obtain the attention weight vector. ; Step 6.3, Spatial Reorganization: Reorganize the vector obtained in Step 6.2 Reorganized into a two-dimensional matrix according to the spatial distribution of image patches. ; Step 6.4, Interpolation Upsampling: Use the bilinear interpolation algorithm to upsample the matrix. Upsampled to the same resolution as the original input image A heat map was obtained. ; Step 6.5, Visual Overlay: Overlay the heatmap obtained in Step 6.

4. Normalize and map to a pseudo-color image, then overlay it onto the first color image. Image matrix from multiple perspectives The upper part displays the aerodynamically sensitive area.

7. The interpretable intelligent analysis method for predicting aerodynamic parameters of a three-dimensional aircraft as described in claim 6, characterized in that: Step 7 includes the following steps: Step 7.1, define the occlusion mask; Step 7.2, perform sliding occlusion; Step 7.3, generating the occluded image; Step 7.4, each occluded image The data is input into the trained model, while keeping other viewpoint images and operating conditions unchanged. Forward propagation is then performed to obtain new aerodynamic parameter prediction vectors. ; Step 7.5, Sensitivity index calculation; Step 7.6, Sensitivity Map Construction; Step 7.7, double comparison verification.

8. The interpretable intelligent analysis method for predicting aerodynamic parameters of a three-dimensional aircraft as described in claim 7, characterized in that: Specifically, step 7 includes the following steps: Step 7.1, define a size as Square occlusion mask block Its pixel value is uniformly set to 0 to cover the image content; Step 7.2, in the section to be verified Image matrix from multiple perspectives Above, set the sliding step size. Moving occlusion mask block Iterate through all regions of the image; Step 7.3, for each sliding position Generate a corresponding occlusion image ; Step 7.4, each occluded image The data is input into the trained model, while keeping other viewpoint images and operating conditions unchanged. Forward propagation is then performed to obtain new aerodynamic parameter prediction vectors. ; Step 7.5, calculate the new aerodynamic parameter prediction vector. Compared with the original unoccluded prediction value The absolute error change between the two values ​​is used as the sensitivity score for that location. : in The L1 norm of a vector is used to comprehensively measure the magnitude of change of all aerodynamic parameters. Step 7.6: Calculate the sensitivity scores for all locations. Mapping back to the corresponding image coordinates, an occlusion sensitivity map is constructed. ; Step 7.7, use the heatmap generated in step 6.

4. Compared with the occlusion sensitivity map generated in step 7.6 A comparative analysis was conducted; if the overlap of the highlighted areas of the two is greater than 95%, the reliability of the model interpretation results was cross-validated, confirming that the occlusion sensitivity map is a key geometric feature affecting aerodynamic performance.

Images