Aerodynamic data analysis method based on curve cluster dynamic feature alignment

Abstract

The invention provides a curve cluster dynamic feature alignment-based pneumatic data analysis method, which comprises the following steps of: dividing all test train numbers into different test sets according to a Mach number, and dividing the test train numbers in each test set into a plurality of train number groups; constructing a correlation curve cluster of the attack angle and each aerodynamic coefficient under different Mach numbers; converting the correlation curve cluster corresponding to each train number group into a sample graph; obtaining a plurality of effective sample graphs; extracting a pneumatic data feature vector of each effective sample graph; and calculating the similarity between the to-be-analyzed feature vector and the aerodynamic data feature vector of the effective sample graph, and determining a plurality of matched sample graphs related to the new-configuration aircraft according to a descending order of the similarity. According to the method, the dynamic characteristics under the coupling action of the Mach number and the attack angle are reserved through the curve cluster, and the influence of different Mach numbers on the aerodynamic law is captured, so that the aerodynamic law better fits the actual aerodynamic physical characteristics, and a more comprehensive characteristic basis is provided for matching of a new-configuration aircraft.

Description

Technical Field

[0001] This disclosure relates to the field of aerodynamics, and in particular to an aerodynamic data analysis method, apparatus, storage medium, and electronic device based on dynamic feature alignment of curve clusters. Background Technology

[0002] Wind tunnel testing is a crucial step in the selection, finalization, and modification of aircraft. Force measurement testing, as a fundamental and high-percentage type of test, has accumulated a vast amount of historical aerodynamic data. This data records the performance of different aircraft shapes under specific test conditions (angle of attack). Sideslip angle Roll angle , rudder deflection angle ,Mach number The six-component aerodynamic force / moment coefficients under ( ) are the core resources for exploring aerodynamic laws and supporting the analysis of new configurations.

[0003] However, the aerospace field is currently facing a contradiction between the surge in testing volume and the scarcity of wind tunnel resources: on the one hand, the development cycle of aircraft is accelerating, and the demand for wind tunnel testing is increasing dramatically; on the other hand, as a scarce infrastructure, wind tunnels are difficult to meet all testing needs. Therefore, improving the efficiency of wind tunnel test data analysis has become the key to breaking through the bottleneck.

[0004] Traditional aerodynamic data analysis heavily relies on aerodynamic experts with years of experience to manually judge the rationality of aerodynamic data for new configurations and to find reference data with similar shapes from historical data. However, this approach has significant limitations: first, it is inefficient and cannot meet the needs of rapid analysis of massive amounts of experimental data; second, it is highly subjective and prone to misjudgment due to personnel fatigue and differences in experience; and third, it suffers from the small sample problem, as the mapping relationship between aircraft shape and aerodynamic data is a mapping of "high-dimensional geometric features (shape) to low-dimensional coefficients (aerodynamic data)," and even with historical data on thousands of shapes, it is still difficult to find a similar shape that perfectly matches the new configuration, resulting in a low reuse rate of historical data.

[0005] Therefore, the industry is gradually shifting towards an analytical approach based on aerodynamic pattern matching. This approach does not rely on shape similarity but instead uncovers the inherent patterns in the aerodynamic data itself and matches it with historical data exhibiting similar patterns. However, existing technologies still have limitations in feature extraction, which is centered on "individual train trips." They also suffer from insufficient dynamic feature representation, loss of curve cluster correlations, unresolved data heterogeneity, and insufficient feature accuracy, making it difficult to meet the demands of new aircraft configurations for efficient and reliable data analysis. Summary of the Invention

[0006] The purpose of this disclosure is to provide a method, apparatus, storage medium, and electronic device for aerodynamic data analysis based on dynamic feature alignment of curve clusters, in order to solve the problems existing in the prior art.

[0007] The embodiments of this disclosure adopt the following technical solution: an aerodynamic data analysis method based on dynamic feature alignment of curve clusters, comprising: acquiring historical wind tunnel test runs and their test data; dividing all the test runs into subsonic test sets, transonic test sets, and supersonic test sets according to Mach number; further dividing the test runs in each test set into multiple run groups; the test runs in each run group having the same test state parameters other than Mach number; constructing a cluster of correlation curves between angle of attack and each aerodynamic coefficient at different Mach numbers based on the test data of each run group; and converting the cluster of correlation curves corresponding to each run group into... Sample images; perform sliding window translation and cropping on all sample images to obtain multiple valid sample images; train an aerodynamic data feature extraction model based on the valid sample images, and extract the aerodynamic data feature vector of each valid sample image; construct sample images to be analyzed based on the test data of the new configuration aircraft, determine the feature vector to be analyzed, calculate the similarity between the feature vector to be analyzed and the aerodynamic data feature vector of the valid sample images, and determine multiple matching sample images related to the new configuration aircraft in descending order of similarity; based on the matching sample images, determine the historical wind tunnel number corresponding to the matching sample images and the aircraft configuration of the historical wind tunnel number.

[0008] This disclosure also provides an aerodynamic data analysis device based on curve cluster dynamic feature alignment, comprising: a data processing module for acquiring historical wind tunnel test runs and their test data, dividing all the test runs into subsonic test sets, transonic test sets, and supersonic test sets according to Mach number, and dividing the test runs in each test set into multiple run groups, wherein the test runs in each run group have the same test state parameters other than Mach number; a curve cluster construction module for constructing a cluster of correlation curves between angle of attack and each aerodynamic coefficient at different Mach numbers based on the test data of each run group; and an image conversion module for converting the cluster of correlation curves corresponding to each run group into sample images; and a sample image. The image generation module is used to perform sliding window translation and cropping on all sample images to obtain multiple valid sample images; the feature extraction module is used to train an aerodynamic data feature extraction model based on the valid sample images and extract the aerodynamic data feature vector of each valid sample image; the matching analysis module is used to construct the sample image to be analyzed based on the test data of the new configuration aircraft, determine the feature vector to be analyzed, calculate the similarity between the feature vector to be analyzed and the aerodynamic data feature vector of the valid sample images, and determine multiple matching sample images related to the new configuration aircraft in descending order of similarity; the tracing module is used to determine the historical wind tunnel number corresponding to the matching sample image and the aircraft configuration of the historical wind tunnel number based on the matching sample image.

[0009] This disclosure also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described aerodynamic data analysis method based on dynamic feature alignment of curve clusters.

[0010] This disclosure also provides an electronic device, which includes at least a memory and a processor. The memory stores a computer program, and the processor executes the computer program in the memory to implement the steps of the above-described aerodynamic data analysis method based on curve cluster dynamic feature alignment.

[0011] The beneficial effects of this embodiment are as follows: using curve clusters as the core analysis unit, the dynamic characteristics under the coupling effect of Mach number and angle of attack are preserved through curve clusters, the influence of different Mach numbers on aerodynamic laws is captured, and the aerodynamic laws are more in line with actual aerodynamic physical characteristics, providing a more comprehensive feature basis for matching new configuration aircraft; and the combination of curve cluster image representation achieves data format unification and eliminates heterogeneity between data; at the same time, the combination of a trained aerodynamic data feature extraction model achieves efficient feature matching between new configurations and historical samples, which improves matching efficiency and accuracy, and can also avoid experience-based misjudgments caused by reliance on technical experts. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of the aerodynamic data analysis method based on dynamic feature alignment of curve clusters in the first embodiment of this disclosure; Figure 2 This is a schematic diagram illustrating the execution of the aerodynamic data analysis method based on dynamic feature alignment of curve clusters in the first embodiment of this disclosure; Figure 3 This is a schematic diagram of the test samples and their matching recommendation results in the first embodiment of this disclosure; Figure 4 This is a schematic diagram of the aerodynamic data analysis device based on curve cluster dynamic feature alignment in the second embodiment of this disclosure. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0015] To address the problems existing in the prior art, the first embodiment of this disclosure provides an aerodynamic data analysis method based on dynamic feature alignment of curve clusters. Figure 1 A flowchart of the aerodynamic data analysis method is shown. Figure 2 A schematic diagram illustrating the execution of the aerodynamic data analysis method is shown, such as... Figure 1 and Figure 2 As shown, it mainly includes steps S10 to S70: S10: Obtain historical wind tunnel test train numbers and their test data. Based on the Mach number, divide all test train numbers into subsonic test sets, transonic test sets, and supersonic test sets, and further divide the test train numbers in each test set into multiple train number groups.

[0016] The data used in this embodiment covers 329 aircraft shapes (numbered No. 1 to No. 329) and a total of 55,713 wind tunnel test runs. For each run, test state parameters (including angle of attack) were recorded. Roll angle Yaw angle Pitch rudder deflection Roll rudder deflection Yaw rudder deflection and Mach number ) and six-component aerodynamic coefficients (including drag coefficient CA, normal force coefficient CN, yaw force coefficient CZ, roll moment coefficient MXG, yaw moment coefficient MYG, and pitch moment coefficient MZG).

[0017] Based on the varying impacts of airflow states on aerodynamic coefficients across different Mach number ranges, historical wind tunnel test data were divided into three sets according to Mach number, covering the subsonic, transonic, and supersonic Mach ranges, as shown in Table 1. Specifically, the subsonic test set contained test runs with Mach numbers greater than or equal to Mach 0.4 and less than Mach 0.8; the transonic test set contained test runs with Mach numbers greater than or equal to Mach 0.8 and less than Mach 1.2; and the supersonic test set contained test runs with Mach numbers greater than or equal to Mach 1.2 and less than or equal to Mach 4.5. This division into three sets not only reduces data interference across different ranges and improves subsequent matching accuracy but also significantly increases the data sample size.

[0018] Table 1

[0019] Subsequently, the test state parameters within each set are iterated. For test state parameters other than Mach number and angle of attack, the "single-variable control" principle is adopted, changing only one test state parameter at a time. Trains with all test state parameters except Mach number are grouped together, ultimately forming groups of trains within each group that differ only in Mach number. This ensures that the differences in data within each group are solely due to the Mach number. This triggers the subsequent extraction. The coupling characteristics lay the foundation.

[0020] S20, based on the test data of each train group, constructs a cluster of correlation curves between the angle of attack and each aerodynamic coefficient at different Mach numbers.

[0021] For the train number groups obtained above, the train number data corresponding to different Ma in each group form multiple curves. Correlation curve clusters are constructed according to the six-component aerodynamic coefficients. The six-component aerodynamic coefficients include drag coefficient CA, normal force coefficient CN, yaw force coefficient CZ, roll moment coefficient MXG, yaw moment coefficient MYG, and pitch moment coefficient MZG. Together, they form a correlation curve cluster of "same test state, multiple Ma states", which intuitively presents the influence of Ma change on the aerodynamic coefficient-angle of attack relationship and completely preserves the two-dimensional dynamic correlation information.

[0022] For each train group, six curve clusters were constructed, corresponding to six aerodynamic coefficients. Each curve cluster included the same number of curves as the number of test trains in that train group, representing different test Mach numbers. Corresponding to the train group situation shown in Table 1, a total of [number missing] curves were constructed. A cluster of curves.

[0023] S30 converts the cluster of associated curves corresponding to each train group into a sample graph.

[0024] In this embodiment, the constructed curve clusters are processed into images. Due to the different shapes of the aircraft, the curve clusters are constructed as follows: and The value intervals and the number of data points may differ, making it impossible to directly form a unified feature input. Therefore, the curve clusters need to be converted into standard-sized single-channel grayscale sample images through the following steps: First, a numerical matrix is ​​constructed based on bilinear interpolation. Then, based on multiple known data points within the curve family, the angle of attack is... The first dimension is the horizontal plane, with Mach number as the standard. Construct a two-dimensional grid for the second vertical dimension, where the horizontal dimension... Uniformly mapped to 512 pixels, vertical dimension The data is uniformly mapped to 320 pixels, forming a 512×320 two-dimensional grid structure. Subsequently, based on the aerodynamic coefficient-angle of attack data of each train within the curve cluster, a bilinear interpolation algorithm is used to interpolate the data along with all the "..." in the two-dimensional grid. "The aerodynamic coefficient values ​​corresponding to the coordinates are calculated and completed. That is, by interpolating the aerodynamic coefficient values ​​of the unknown coordinates in the grid using the aerodynamic coefficients under adjacent known α and Ma conditions, a 512×320 numerical matrix is ​​finally generated, in which each element takes the corresponding value." "The aerodynamic coefficient under operating conditions enables the transformation of curve clusters from 'discrete data curve sets' to 'continuous two-dimensional numerical matrices,' solving the problems of different train routes." Data discontinuity The problem of heterogeneous curve clusters caused by data mismatch.

[0025] For numerical matrices, the following is adopted: The normalization method maps the values ​​of all elements in the above numerical matrix to... The normalization formula for the interval is:

[0026] in, Represents the original element values ​​of the matrix. , These represent the minimum and maximum values ​​of the aerodynamic coefficients in the numerical matrix, respectively, thus eliminating interference caused by differences in dimensions and numerical ranges among different aerodynamic components. This represents the normalized element values. Then, a one-to-one mapping is established between the normalized values ​​and grayscale values ​​(0~255), i.e., a normalized value of 0 corresponds to grayscale 0, a normalized value of 1 corresponds to grayscale 255, and so on. This transforms the 512×320 numerical matrix into a single-channel grayscale image, where the brightness directly reflects the corresponding element's value. "The magnitude of the aerodynamic coefficient under operating conditions enables the conversion of aerodynamic data into image features, providing a suitable input for subsequent feature extraction."

[0027] S40 performs a sliding window translation and cropping operation on all sample images to obtain multiple valid sample images.

[0028] Considering the input size requirements of the feature extraction model for training samples, this embodiment uses a sliding window translation method to truncate the sample images to form valid sample images. Combining the size of the sample images with the input size requirements of the feature extraction model, each sample image is truncated using a sliding window based on a preset window, according to preset horizontal and vertical step sizes. This results in multiple valid sample images, ensuring a uniform sample size for subsequent feature extraction models while also fully preserving the different characteristics of the curve clusters through the generation of multiple valid sample images. "The aerodynamic characteristics of the region improve sample utilization. For example, with a preset horizontal step size of 72, a preset vertical step size of 32, and a preset window size of 224×224, for a 512×320 single-channel grayscale image, it can be cropped 5 times horizontally, with starting indices of 0, 72, 144, 216, and 288 respectively; and cropped 4 times vertically, with starting indices of 0, 32, 64, and 96 respectively. Each grayscale image can generate 5×4=20 valid sample images."

[0029] Furthermore, for each valid sample image, it can be named using the following rules to facilitate subsequent matching of samples with historical wind tunnel test runs and aircraft architecture: QDFL_No_CLSlabel_phi_beta_d5_d6_d7_d8_startalfa_endalfa_locx_locy_ _..._ _ _..._ The physical meanings of the above-named codes are as follows: QDFL: Six pneumatic components of the balance. Values: CA, CN, CZ, MXG, MYG, MZG.

[0030] No: Represents the number of 329 aircraft, with a value of 1 to 329; CLSlabel: Characterizes different Range, values: A, B, C; phi: Overall roll angle of the aircraft ; beta: Overall yaw angle of the aircraft ; d5: Pitch control deflection of a local component of the aircraft ; d6: Roll deflection angle of local components of the aircraft ; d7: Yaw deflection angle of a local component of the aircraft ; d8: Reserved for future use; startalfa: Minimum angle of attack; endalfa: Maximum angle of attack; locx: Image pixel X index; locy: Image pixel Y index; The first train number; Train number m: The Mach number corresponding to the first train number; : The Mach number corresponding to the m-th train; Where m is the number of trains in the original curve cluster corresponding to the sample image or numerical matrix without bilinear interpolation.

[0031] In some embodiments, the numerical matrix obtained above is associated with the corresponding valid sample images one by one, and metadata information is labeled for each sample. The metadata includes the corresponding historical test vehicle number, aircraft shape, test state parameters, etc., and is not a category label. An unlabeled aerodynamic data sample library is formed based on the labeled numerical matrix and grayscale sample images. The metadata is only used for tracing the source after subsequent matching and recommendation, and does not participate in the training process of self-supervised learning.

[0032] S50 trains an aerodynamic data feature extraction model based on valid sample images and extracts the aerodynamic data feature vector for each valid sample image.

[0033] Based on the aerodynamic data features represented by the valid sample images obtained in the above steps, this embodiment utilizes an aerodynamic data feature extraction model to extract features. The aerodynamic data feature extraction model uses the ViT (Vision Transformer) model as the base model and performs self-supervised learning on the unlabeled sample library constructed above.

[0034] Specifically, the ViT model in this embodiment can be the "google / vit-base-patch16-224-in21k" model pre-trained on the ImageNet-21k dataset. This model is an image classification model based on the ViT architecture, with 12 layers of encoder weights. A decoder for mask reconstruction is added after the original ViT architecture. This decoder is a 4-layer Transformer structure, with each hidden layer having a dimension of 256, and the activation function is GELU. During actual training, the pre-trained model is loaded. To adapt to the model's input requirements, the effective sample images from the single channel are copied and converted to 3 channels. The model divides the 224×224 image into 16×16 patches, for a total of 14×14=196 patches. The global attention of the Transformer Encoder is used to capture… "The association features in two-dimensional space are then used; subsequently, a masked image reconstruction task is employed, randomly masking 75% of the 16×16 patches in a 224×224 aerodynamic grayscale image, with the masked region pixels set to 0. The masked region pixels are reconstructed using a newly added decoder. During training, the weights of the first 10 layers of the encoder representing the location in the base model remain unchanged. The weights of the last 2 layers of the encoder are fine-tuned using a backpropagation strategy to adapt to the pixel distribution of the aerodynamic grayscale image. Simultaneously, the parameters of the patch embedding layer in the base model and the parameters of the 4 Transformer attention layers and fully connected layers in the decoder are updated." The model updates data to capture the two-dimensional spatial correlation between aerodynamic data and angle of attack, and to reconstruct the mask region. The learning rate for the patch embedding layer is set to 5e-6, the learning rate for the attention layer is 1e-5, and the output dimension of the fully connected layer is 256, corresponding to the pixel dimension of a single 16×16 single-channel grayscale patch. The training epochs are determined by ensuring that the MSE loss of the normalized grayscale values ​​of the reconstructed mask region does not decrease for five consecutive epochs. The optimizer uses AdamW with weight decay configured at a weight decay coefficient of 1e-4 to suppress overfitting and ensure model generalization ability. Finally, the encoder part is retained for subsequent feature extraction.

[0035] After training, the decoder used for mask reconstruction is discarded, and the base model with two layers of fine-tuned weights is retained as the trained aerodynamic data feature extraction model to perform feature extraction. Relying on the global attention mechanism of the Transformer Encoder, it accurately captures the "..." from the effective sample images. The spatial correlation features of "" are described. Specifically, the trained aerodynamic data feature extraction model outputs a 768-dimensional aerodynamic data feature vector corresponding to the CLS Token for each 224×224 valid sample image. This feature vector captures different speed ranges of sound through an attention mechanism. "The spatial correlation of aerodynamic flow fields under coupling, such as the gray-scale changes in the shock wave region and the connection characteristics between the separated flow and the mainstream."

[0036] S60: Construct sample maps to be analyzed based on test data of the new configuration aircraft, determine the feature vectors to be analyzed, calculate the similarity between the feature vectors to be analyzed and the aerodynamic data feature vectors of the valid sample maps, and determine multiple matching sample maps related to the new configuration aircraft in descending order of similarity.

[0037] When performing aerodynamic data analysis on a new configuration aircraft, the process of deriving the effective sample image described above is used to construct the sample image to be analyzed corresponding to the experimental data of the new configuration aircraft. Then, a trained aerodynamic data feature extraction model is used to extract features, resulting in the feature vector to be analyzed. Subsequently, the similarity between the feature vector to be analyzed and the aerodynamic data feature vector of the effective sample image is calculated. To eliminate the influence of differences in vector magnitude on the similarity, the feature vector to be analyzed and the aerodynamic data feature vector of the effective sample image are first normalized, based on the following formula:

[0038] in, The feature vector to be normalized, The L2 norm of this eigenvector. The normalized eigenvectors are obtained, resulting in the normalized eigenvectors to be analyzed and the normalized eigenvectors of the aerodynamic data.

[0039] Finally, the cosine similarity between the normalized feature vector to be analyzed and the normalized feature vectors of all aerodynamic data is calculated. Cosine similarity is more sensitive to the differences in high-dimensional vectors and is suitable for aerodynamic feature matching. Based on the similarity calculation results, the most relevant matching sample images to the new configuration aircraft are determined by sorting them in descending order of similarity. In the actual matching results, to balance the number of recommendations and reliability, the number of matching sample images N can be between 3 and 5, and a similarity threshold of 0.8 is set. If there are no samples with a similarity greater than 0.8 in the similarity calculation results, it means that there is no historical data with high similarity in the sample library, and the corresponding results can be directly fed back to the user to avoid low-precision matching misleading subsequent analysis.

[0040] Figure 3 The paper presents three test samples and matching recommendation results generated by the engineering calculation software Datacom. To make the display more intuitive, the single-channel grayscale image was modified into a three-channel color image by copying. Threshold elimination and normalization were ignored. The top-4 results of each test sample are displayed, and the similarity score of each result is given.

[0041] S70, based on the matching sample map, determine the historical wind tunnel number and the aircraft configuration of the historical wind tunnel number corresponding to the matching sample map.

[0042] For the top N samples that meet the similarity threshold, the corresponding historical test vehicle number, aircraft shape and model and complete aerodynamic data can be traced through the sample name. Wind tunnel operators can combine this historical data and aerodynamic laws to infer the range of aerodynamic coefficients for the new configuration aircraft under test conditions, realize auxiliary analysis function, provide auxiliary support for the aerodynamic data analysis of the new configuration under test conditions, and thus improve the efficiency of data analysis.

[0043] This embodiment uses curve clusters as the core analysis unit. By constructing curve clusters, it fully preserves the two-dimensional dynamic features under the coupling effect of Ma and α, capturing the influence of different Ma numbers on aerodynamic laws. This makes the aerodynamic law representation more consistent with actual aerodynamic physical characteristics, providing a more comprehensive feature foundation for matching and recommendation. It overcomes the deficiency of existing technologies that can only represent the one-dimensional static features of a single train. Furthermore, this embodiment achieves data format unification at all levels, from the train set and curve clusters to sample images, through three-level processing: Mach number interval grouping, incomplete bilinear interpolation, and sliding window truncation. Different curve clusters are first transformed into numerical matrices of a specific size through bilinear interpolation, and then transformed into grayscale samples of the same size through sliding window truncation, instead of being directly transformed into size samples that are finally adapted to the input of the base model through bilinear interpolation. This eliminates the differences in the number and distribution of Ma numbers for different aircraft configurations. The heterogeneity caused by the range difference is not only reflected in the sample of atmospheric physical mechanisms, but also in the expansion of the effective scale of data, which greatly improves the accuracy of feature matching. Furthermore, this embodiment only requires the raw data of basic experimental parameters, and achieves "end-to-end" feature extraction through ViT model pre-training and self-supervised learning. The model autonomously captures... The correlation features in two-dimensional space require no expert intervention, significantly reducing reliance on experts. Simultaneously, this embodiment utilizes a sliding window extraction technique to generate multiple valid samples from a single curve cluster, ensuring both uniform sample size input to the ViT model and sufficient preservation of the differences within the curve cluster. The aerodynamic characteristics of the region improve sample utilization and lay a data foundation for the application of ViT models with large-scale parameters. The three-in-one design of the sample library, which integrates "numerical matrix-grayscale image-metadata information", not only supports ViT feature extraction but also facilitates rapid source tracing after matching. The similarity threshold judgment and "no high similarity prompt" function further avoid the misleading effect of low-precision matching and improve the safety and practicality of engineering applications.

[0044] Based on the same inventive concept, the second embodiment of this disclosure provides an aerodynamic data analysis device for dynamic feature alignment of curve clusters. This device can be deployed in any form in an aerodynamic test chamber or aerodynamic data analysis site, for example, as a field processor or a field embedded device, to complete feature vector extraction in conjunction with other equipment in the test chamber. Its structural schematic diagram is shown below. Figure 4As shown, it mainly includes: a data processing module 10, used to acquire historical wind tunnel test runs and their test data, dividing all test runs into subsonic test sets, transonic test sets, and supersonic test sets according to Mach number, and further dividing the test runs in each test set into multiple run groups, with the test runs in each run group having the same test state parameters except for Mach number; a curve cluster construction module 20, used to construct a cluster of correlation curves between angle of attack and each aerodynamic coefficient at different Mach numbers based on the test data of each run group; an image conversion module 30, used to convert the correlation curve clusters corresponding to each run group into sample images; and a sample image generation module 40, used for... All sample images are subjected to sliding window translation to obtain multiple valid sample images; the feature extraction module 50 is used to train an aerodynamic data feature extraction model based on the valid sample images and extract the aerodynamic data feature vector of each valid sample image; the matching analysis module 60 is used to construct the sample images to be analyzed based on the test data of the new configuration aircraft, determine the feature vector to be analyzed, calculate the similarity between the feature vector to be analyzed and the aerodynamic data feature vector of the valid sample images, and determine multiple matching sample images related to the new configuration aircraft in descending order of similarity; the tracing module 70 is used to determine the historical wind tunnel operation number and the aircraft configuration of the historical wind tunnel operation number corresponding to the matching sample images.

[0045] Specifically, in the subsonic test suite, the Mach number of each test vehicle is greater than or equal to Mach 0.4 and less than Mach 0.8; in the transonic test suite, the Mach number of each test vehicle is greater than or equal to Mach 0.8 and less than Mach 1.2; and in the supersonic test suite, the Mach number of each test vehicle is greater than or equal to Mach 1.2 and less than or equal to Mach 4.5. Test state parameters include Mach number, pitch deflection, roll deflection, yaw deflection, roll angle, and yaw angle; aerodynamic coefficients include drag coefficient, normal force coefficient, yaw force coefficient, roll moment coefficient g, yaw moment coefficient, and pitch moment coefficient.

[0046] In some embodiments, the image conversion module 30 is specifically used to construct a numerical matrix with angle of attack as the first dimension, Mach number as the second dimension, and aerodynamic coefficient value as the element value; normalize the numerical matrix; and map the normalized numerical matrix into a single-channel grayscale image as a sample image.

[0047] In some embodiments, the sample image generation module 40 is specifically used to perform sliding window translation and cropping of each sample image based on a preset window according to a preset horizontal step size and a preset vertical step size, so as to obtain multiple valid sample images of the sample image.

[0048] In some embodiments, the feature extraction module 50 is specifically used to: use a ViT model pre-trained based on ImageNet-21k as the base model; freeze the weights of the first 10 layers of the encoder in the ViT model; add a decoder for mask reconstruction after the base model; use mask image reconstruction by a mask autoencoder as the learning task; perform self-supervised learning on effective sample images; and fine-tune the weights of the last two layers of the encoder in the base model through a backpropagation strategy; update the parameters of the Patch embedding layer in the base model and the parameters of the attention layer and fully connected layer of the decoder until the MSE loss function of the normalized grayscale value of the reconstructed mask region meets the preset condition; remove the decoder, and use the base model with the completed weight fine-tuning as the aerodynamic data feature extraction model; and extract features from each effective sample image based on the aerodynamic data feature extraction model to obtain the aerodynamic data feature vector of each effective sample image. It should be noted that the above aerodynamic data feature extraction model is only trained before initial use in actual use; subsequently, the trained aerodynamic data feature extraction model can be directly called for feature extraction.

[0049] In some embodiments, the matching analysis module 60 is specifically used to normalize the feature vector to be analyzed and the aerodynamic data feature vector of the valid sample map respectively to obtain the normalized feature vector to be analyzed and the aerodynamic data normalized feature vector; and to calculate the cosine similarity between the normalized feature vector to be analyzed and all aerodynamic data normalized feature vectors.

[0050] This embodiment uses curve clusters as the core analysis unit. By preserving the dynamic characteristics under the coupling effect of Mach number and angle of attack through curve clusters, it captures the influence of different Mach numbers on aerodynamic laws, making the aerodynamic laws more consistent with actual aerodynamic physical characteristics, and providing a more comprehensive feature basis for matching new configuration aircraft. In addition, by combining the image-based approach of curve clusters, it achieves data format unification and eliminates heterogeneity between data. At the same time, it combines a trained aerodynamic data feature extraction model to achieve efficient feature matching between new configurations and historical samples. While improving matching efficiency and accuracy, it can also avoid experience-based misjudgments caused by reliance on technical experts.

[0051] Based on the same inventive concept, the third embodiment of this disclosure provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the aerodynamic data analysis method based on dynamic feature alignment of curve clusters described in the first embodiment.

[0052] Based on the same inventive concept, the fourth embodiment of this disclosure provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and when the processor executes the computer program in the memory, it implements the steps of the aerodynamic data analysis method based on curve cluster dynamic feature alignment described in the first embodiment.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit them. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A method for aerodynamic data analysis based on dynamic feature alignment of curve clusters, characterized in that, include: Historical wind tunnel test runs and their test data are obtained. Based on the Mach number, all the test runs are divided into subsonic test sets, transonic test sets, and supersonic test sets. The test runs in each test set are further divided into multiple run groups. The test runs in each run group have the same test state parameters except for the Mach number. Based on the test data of each train group, a cluster of correlation curves between the angle of attack and each aerodynamic coefficient at different Mach numbers was constructed. Convert the associated curve clusters corresponding to each train group into sample graphs; A sliding window is used to capture all sample images to obtain multiple valid sample images; A pneumatic data feature extraction model is trained based on the effective sample images, and a pneumatic data feature vector is extracted for each effective sample image. Based on the test data of the new configuration aircraft, a sample map to be analyzed is constructed, and the feature vector to be analyzed is determined. The similarity between the feature vector to be analyzed and the aerodynamic data feature vector of the effective sample map is calculated. Multiple matching sample maps related to the new configuration aircraft are determined in descending order of similarity. Based on the matching sample map, determine the historical wind tunnel number corresponding to the matching sample map and the aircraft configuration of the historical wind tunnel number.

2. The aerodynamic data analysis method according to claim 1, characterized in that, The process of dividing all test runs into subsonic test sets, transonic test sets, and supersonic test sets based on Mach number includes: The Mach number of the test vehicles in the subsonic test set is greater than or equal to Mach 0.4 and less than Mach 0.8; The Mach number of the test vehicles in the transonic test set is greater than or equal to Mach 0.8 and less than Mach 1.2; The Mach number of the test vehicles in the supersonic test set is greater than or equal to Mach 1.2 and less than or equal to Mach 4.

5.

3. The aerodynamic data analysis method according to claim 1, characterized in that, The step of converting the associated curve cluster corresponding to each train group into a sample image includes: Construct a numerical matrix with angle of attack as the first dimension, Mach number as the second dimension, and aerodynamic coefficient values ​​as element values; The numerical matrix is ​​normalized, and the normalized numerical matrix is ​​mapped to a single-channel grayscale image as the sample image.

4. The aerodynamic data analysis method according to claim 3, characterized in that, The process of sliding windowing and cropping all sample images yields multiple valid sample images, including: According to the preset horizontal step size and preset vertical step size, each of the sample images is subjected to sliding window translation based on a preset window to obtain multiple valid sample images of the sample image.

5. The aerodynamic data analysis method according to claim 1, characterized in that, The step of training an aerodynamic data feature extraction model based on the effective sample images and extracting the aerodynamic data feature vector for each effective sample image includes: Using the ViT model pre-trained based on ImageNet-21k as the base model, the weights of the first 10 layers of the encoder in the ViT model are frozen, and a decoder for mask reconstruction is added after the base model. The mask image reconstruction of the mask autoencoder is used as the learning task. Self-supervised learning is performed on the effective sample image, and the weights of the last two layers of the encoder in the base model are fine-tuned through the backpropagation strategy. The parameters of the Patch embedding layer in the base model and the parameters of the attention layer and fully connected layer of the decoder are updated until the MSE loss function of the normalized gray value of the reconstructed mask region meets the preset condition and training stops. The decoder is removed, and the base model for weight fine-tuning is used as the aerodynamic data feature extraction model. Based on the aerodynamic data feature extraction model, feature extraction is performed on each of the valid sample images to obtain the aerodynamic data feature vector of each valid sample image.

6. The aerodynamic data analysis method according to claim 5, characterized in that, The calculation of the similarity between the feature vector to be analyzed and the aerodynamic data feature vector of the effective sample map includes: The aerodynamic data feature vectors of the target feature vector and the effective sample image are normalized respectively to obtain the normalized feature vectors of the target feature vector and the normalized feature vectors of the aerodynamic data. Calculate the cosine similarity between the normalized feature vector to be analyzed and the normalized feature vectors of all the aerodynamic data.

7. The aerodynamic data analysis method according to any one of claims 1 to 6, characterized in that, The test state parameters include Mach number, pitch deflection, roll deflection, yaw deflection, roll angle, and yaw angle. The aerodynamic coefficients include drag coefficient, normal force coefficient, yaw force coefficient, roll moment coefficient, yaw moment coefficient, and pitch moment coefficient.

8. A pneumatic data analysis device based on dynamic feature alignment of curve clusters, characterized in that, include: The data processing module is used to acquire historical wind tunnel test runs and their test data, and to divide all the test runs into subsonic test sets, transonic test sets and supersonic test sets according to the Mach number. The test runs in each test set are further divided into multiple run groups, and the test runs in each run group have the same test state parameters except for the Mach number. The curve cluster construction module is used to construct a cluster of correlation curves between the angle of attack and each aerodynamic coefficient at different Mach numbers based on the test data of each train group. The image conversion module is used to convert the associated curve clusters corresponding to each train group into sample images. The sample image generation module is used to perform sliding window translation and cropping on all sample images to obtain multiple valid sample images; The feature extraction module is used to train an aerodynamic data feature extraction model based on the effective sample images and extract the aerodynamic data feature vector for each effective sample image. The matching analysis module is used to construct a sample map to be analyzed based on the test data of the new configuration aircraft, determine the feature vector to be analyzed, calculate the similarity between the feature vector to be analyzed and the aerodynamic data feature vector of the effective sample map, and determine multiple matching sample maps related to the new configuration aircraft in descending order of similarity. The tracing module is used to determine the historical wind tunnel number and the aircraft configuration of the historical wind tunnel number corresponding to the matching sample image based on the matching sample image.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the aerodynamic data analysis method based on dynamic feature alignment of curve clusters as described in any one of claims 1 to 7.

10. An electronic device, comprising at least a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program on the memory, it implements the steps of the aerodynamic data analysis method based on dynamic feature alignment of curve clusters as described in any one of claims 1 to 7.

Images