A digital feature extraction method for aircraft aerodynamic shape based on feature matching

By using feature matching, two convolutional networks are constructed to extract features of the aerodynamic shape of the aircraft. This solves the problem that existing technologies cannot effectively combine different representation methods, achieves efficient extraction of common features, simplifies the network structure, and supports the study of the aerodynamic characteristics of hypersonic aircraft.

CN118821327BActive Publication Date: 2024-11-15CHINA AERODYNAMICS RES AND DEV CENT ULTRA-HIGH SPEED AERODYNAMICS RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411295732.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-18
Publication Date
2024-11-15
Estimated Expiration
2044-09-18

AI Technical Summary

Technical Problem

Existing deep learning-based methods for extracting aerodynamic features of aircraft fail to effectively combine the advantages of different representation methods, resulting in complex network structures, low modeling efficiency, and difficulty in determining effective features.

Method used

The feature matching method is adopted. Two convolutional networks NetA and NetB with the same structure are constructed to extract features from the symbolic distance field SDF and the contour boundary Mask, respectively. The matching feature Fmatched is obtained through feature crossing and feature matching. The common features are extracted by combining the advantages of various shape representation methods.

Benefits of technology

This method enables the extraction of common features from different shape representation methods, simplifies the network structure, improves modeling efficiency, and provides technical support for the study of aerodynamic characteristics of hypersonic vehicles.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118821327B_ABST
    Figure CN118821327B_ABST
Patent Text Reader

Abstract

The present invention belongs to the research field of intelligent fluid mechanics of hypersonic aircraft, and discloses a method for extracting digital features of aircraft aerodynamic shape based on feature matching, including performing digital feature representation of aircraft aerodynamic shape; constructing a convolutional network; performing feature intersection; performing feature matching; and performing reconstruction. The method for extracting digital features of aircraft aerodynamic shape based on feature matching of the present invention can combine the advantages of a variety of different shape representation methods, and extract the most original and unified potential shape features through feature intersection and feature matching, ensuring that no matter which shape representation method is used, the common features in different shape representation methods, that is, the invariant shape features, can be extracted, providing technical support for the study of aerodynamic characteristics of hypersonic aircraft.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the research field of intelligent fluid mechanics of hypersonic aircraft, and specifically relates to a method for extracting digital features of an aircraft's aerodynamic shape based on feature matching. Background Art

[0002] Many applications in the field of aerodynamics, such as flow field prediction, aircraft aerodynamic performance prediction, flow field structure identification, etc., all involve the description of aircraft aerodynamic shape. How to reasonably describe the aerodynamic shape of an aircraft and establish the relationship between aerodynamic data and aerodynamic shape has become one of the key issues in the field of aerodynamics. At present, in the problem of extracting the aerodynamic shape features of aircraft, the solutions are mainly divided into three categories: parameterization, polynomial description, and deep learning-based methods. Among them, the deep learning-based method is superior to the previous two in terms of performance and efficiency. Therefore, it has become a research hotspot in the field of intelligent fluid mechanics.

[0003] Among the deep learning-based methods, the most typical solution is the feature extraction method based on convolutional neural networks. However, most of the deep learning-based methods only use a single method to describe the aerodynamic shape, such as the shape contour mask, signed distance field (SDF), etc., without combining the advantages of different representation methods; secondly, some deep learning-based methods directly mix the original tensor representing the shape with the flow field data, resulting in a complex network structure, low modeling efficiency, and difficulty in determining whether effective aerodynamic shape features are extracted.

[0004] Currently, there is an urgent need to develop a digital feature extraction method for aircraft aerodynamic shape based on feature matching. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for extracting digital features of an aircraft's aerodynamic shape based on feature matching.

[0006] The principle of the digital feature extraction method of the aircraft aerodynamic shape based on feature matching is shown in Figure 2It is necessary to construct two convolutional networks NetA and NetB with exactly the same structure. Convolutional networks NetA and NetB respectively contain an encoder and a decoder. First, the encoder of convolutional network NetA is used to extract features of the signed distance field (SDF) representing the shape, and the encoder of convolutional network NetB is used to extract features of the contour boundary Mask. Then, in the process of feature extraction, feature intersection is performed. Next, feature matching is performed on FA and FB output by the two encoders, so that the outputs of the two encoders are infinitely close, and the matching feature Fmatched is obtained, and the matching feature Fmatched is used as the common output of the two encoders. Finally, the matching feature Fmatched is simultaneously input to the decoders of the two networks for decoding, and the signed distance field SDF and the contour boundary Mask are reconstructed.

[0007] See Feature Crossover and Feature Matching Figure 3 Feature crossing refers to exchanging the output feature map Feature Map A of the encoder of the convolutional network NetA and the output feature map Feature Map B of the encoder of the convolutional network NetB layer by layer, Layer1, Layer2, etc., exchanging half the number of output feature maps in each layer. Feature matching refers to constraining the output FA and output FB of the output layer Output Layer through the loss function Loss MSE, so that the output FA and output FB are infinitely close, achieving feature matching, and obtaining the matching feature Fmatched.

[0008] The method for extracting digital features of an aircraft aerodynamic shape based on feature matching of the present invention comprises the following steps:

[0009] S10. Perform digital representation of the aerodynamic shape of the aircraft;

[0010] According to the shape of the aircraft, the corresponding signed distance field SDF and contour boundary Mask are generated. The SDF and Mask maintain the same dimension and size. The contour boundary Mask defines the boundary of the aircraft as 1, and all other places are 0.

[0011] S20. Construct a convolutional network;

[0012] Construct two convolutional networks Net with exactly the same structure A and convolutional network Net B , Convolutional Network Net A and convolutional network Net BEach contains an encoder Encoder and a decoder Decoder;

[0013] Among them, the convolutional network Net A With the signed distance field SDF as input and output, the convolutional network Net A The encoder Encoder performs dimension reduction encoding on the input signed distance field SDF, and then the decoder Decoder decodes and reconstructs the result after dimension reduction and outputs the original signed distance field SDF;

[0014] Similarly, the convolutional network Net B With the contour boundary Mask as input and output, the convolutional network Net B The encoder Encoder performs dimension reduction encoding on the input contour boundary Mask, and then the decoder Decoder decodes and reconstructs the dimension reduction result and outputs the original contour boundary Mask;

[0015] S30. Perform feature crossover;

[0016] In the convolutional network Net A and convolutional network Net B During the encoding process, the output feature map Feature Map A of the encoder of the convolutional network NetA and the output feature map Feature MapB of the encoder of the convolutional network NetB are exchanged layer by layer in each hidden layer Layer1, Layer2..., and half of the output feature maps are exchanged in each layer; the exchanged output feature maps are used as the input of the next layer of the network, until the output FA and output FB of each encoder are obtained respectively;

[0017] S40. Perform feature matching;

[0018] The convolutional network Net A The output of FA and convolutional network Net B The output FB is subjected to feature matching in the form of MSE constraint through the loss function Loss to obtain the matching feature Fmatched; the matching feature Fmatched is a common feature in different shape representation methods, that is, an invariant shape feature;

[0019] S50. Perform reconstruction;

[0020] The matching features Fmatched obtained by S40 are input into the convolutional network Net A and convolutional network Net A In the decoder Decoder, the signed distance field SDF and the contour boundary Mask are reconstructed respectively.

[0021] The digital feature extraction method of the aircraft aerodynamic shape based on feature matching of the present invention can combine the advantages of a variety of different shape representation methods, and extract the most original and unified potential shape features through feature intersection and feature matching, thereby ensuring that no matter which shape representation method is used, the common features in different shape representation methods, that is, the invariant shape features, can be extracted, thereby providing technical support for the study of the aerodynamic characteristics of hypersonic aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The present invention is a flow chart of a method for extracting digital features of an aircraft aerodynamic shape based on feature matching;

[0023] Figure 2 It is a schematic diagram of the principle of the method for extracting digital features of aerodynamic shape of an aircraft based on feature matching of the present invention;

[0024] Figure 3 It is a schematic diagram of the feature intersection and feature matching process in the digital feature extraction method of the aircraft aerodynamic shape based on feature matching of the present invention. DETAILED DESCRIPTION

[0025] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0026] Example: Figure 1 As shown, the method for extracting digital features of an aircraft aerodynamic shape based on feature matching of the present invention comprises the following steps:

[0027] S10. Perform digital representation of the aerodynamic shape of the aircraft;

[0028] According to the shape of the aircraft, the corresponding signed distance field SDF and contour boundary Mask are generated. The SDF and Mask maintain the same dimension and size. The contour boundary Mask defines the boundary of the aircraft as 1, and all other places are 0.

[0029] S20. Construct a convolutional network;

[0030] Construct two convolutional networks Net with exactly the same structure A and convolutional network Net B , Convolutional Network Net A and convolutional network Net B Each contains an encoder Encoder and a decoder Decoder;

[0031] Among them, the convolutional network Net A With the signed distance field SDF as input and output, the convolutional network Net AThe encoder Encoder performs dimension reduction encoding on the input signed distance field SDF, and then the decoder Decoder decodes and reconstructs the result after dimension reduction and outputs the original signed distance field SDF;

[0032] Similarly, the convolutional network Net B With the contour boundary Mask as input and output, the convolutional network Net B The encoder Encoder performs dimension reduction encoding on the input contour boundary Mask, and then the decoder Decoder decodes and reconstructs the dimension reduction result and outputs the original contour boundary Mask;

[0033] S30. Perform feature crossover;

[0034] In the convolutional network Net A and convolutional network Net B During the encoding process, the output feature map Feature Map A of the encoder of the convolutional network NetA and the output feature map Feature MapB of the encoder of the convolutional network NetB are exchanged layer by layer in each hidden layer Layer1, Layer2..., and half of the output feature maps are exchanged in each layer; the exchanged output feature maps are used as the input of the next layer of the network, until the output FA and output FB of each encoder are obtained respectively;

[0035] S40. Perform feature matching;

[0036] The convolutional network Net A The output of FA and convolutional network Net B The output FB is subjected to feature matching in the form of MSE constraint through the loss function Loss to obtain the matching feature Fmatched; the matching feature Fmatched is a common feature in different shape representation methods, that is, an invariant shape feature;

[0037] S50. Perform reconstruction;

[0038] The matching features Fmatched obtained by S40 are input into the convolutional network Net A and convolutional network Net A In the decoder Decoder, the signed distance field SDF and the contour boundary Mask are reconstructed respectively.

[0039] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the description and the embodiments. For those familiar with the art, all the features disclosed in the present invention, or the steps in all the methods or processes disclosed, except for mutually exclusive features and / or steps, can be combined in any way without departing from the principles of the present invention. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A method for extracting digital features of an aircraft's aerodynamic shape based on feature matching, characterized in that: The following steps are involved: S10. Perform digital representation of the aerodynamic shape of the aircraft; According to the shape of the aircraft, the corresponding signed distance field SDF and contour boundary Mask are generated. The SDF and Mask maintain the same dimension and size. The contour boundary Mask defines the boundary of the aircraft as 1, and all other places are 0. S20. Construct a convolutional network; Construct two convolutional networks Net with exactly the same structure A and convolutional network Net B , Convolutional Network Net A and convolutional network Net B Each contains an encoder Encoder and a decoder Decoder; Among them, the convolutional network Net A With the signed distance field SDF as input and output, the convolutional network Net A The encoder Encoder performs dimension reduction encoding on the input signed distance field SDF, and then the decoder Decoder decodes and reconstructs the result after dimension reduction and outputs the original signed distance field SDF; Similarly, the convolutional network Net B With the contour boundary Mask as input and output, the convolutional network Net B The encoder Encoder performs dimension reduction encoding on the input contour boundary Mask, and then the decoder Decoder decodes and reconstructs the dimension reduction result and outputs the original contour boundary Mask; S30. Perform feature crossover; In the convolutional network Net A and convolutional network Net B During the encoding process, the output feature map Feature Map A of the encoder of the convolutional network NetA and the output feature map Feature Map B of the encoder of the convolutional network NetB are exchanged layer by layer in each hidden layer Layer1, Layer2..., and half of the output feature maps are exchanged in each layer; the exchanged output feature maps are used as the input of the next layer of the network, until the output FA and output FB of each encoder are obtained respectively; S40. Perform feature matching; The convolutional network Net A The output of FA and convolutional network Net B The output FB is subjected to feature matching in the form of MSE constraint through the loss function Loss to obtain the matching feature Fmatched; the matching feature Fmatched is a common feature in different shape representation methods, that is, an invariant shape feature; S50. Perform reconstruction; The matching features Fmatched obtained by S40 are input into the convolutional network Net A and convolutional network Net A In the decoder Decoder, the signed distance field SDF and the contour boundary Mask are reconstructed respectively.