Aerodynamic property similarity search method, device, equipment and storage medium

Abstract

The application provides a method and device for aerodynamic characteristic similarity retrieval, equipment and a storage medium, wherein the method comprises: converting aerodynamic data into an aerodynamic feature image, intercepting the aerodynamic feature image to obtain an aerodynamic feature subgraph of an input size of an adaptive mask autoencoder; constructing a mask autoencoder, setting an aerodynamic specific mask strategy for distinguishing the importance of different positions in the mask autoencoder, processing the aerodynamic feature subgraph by using the aerodynamic specific mask strategy to obtain a mask subgraph; training the mask autoencoder by using the mask subgraph to obtain an aerodynamic characteristic feature extraction model; extracting feature vectors of each component of the aerodynamic data of the aircraft by using the aerodynamic characteristic feature extraction model, independently retrieving each feature vector to obtain a similarity ranking result, and fusing the similarity ranking result to obtain a final retrieval result. The application constructs a complete and detailed aerodynamic data similarity retrieval including aerodynamic data preprocessing, feature learning, similarity retrieval and result fusion.

Description

Technical Field

[0001] This invention relates to the field of pneumatic data processing technology, and in particular to a method, apparatus, device and storage medium for pneumatic characteristic similarity retrieval. Background Technology

[0002] In the process of aircraft design and optimization, accurately retrieving historical aircraft with similar aerodynamic characteristics is of great reference value for the design of new aircraft.

[0003] In related technologies, the reliance on designers' expertise and experience is crucial, leading to high subjectivity, low efficiency, and difficulty in handling large-scale data. Traditional machine learning-based methods, such as principal component analysis or linear discriminant analysis, require extensive labeled data for supervised learning, and the labeling of aerodynamic data necessitates the participation of domain experts, resulting in high costs and time consumption. While deep learning-based supervised methods, such as convolutional neural networks, can automatically learn features, they still require substantial labeled data. The features learned by these methods often lack physical interpretability and fail to reflect the physical laws of aerodynamic data, resulting in limited reliability of retrieval results in practical engineering applications. Existing unsupervised learning methods, such as autoencoders, typically employ random masking strategies, failing to fully utilize the physical characteristics of aerodynamic data. Aerodynamic data exhibits clear physical laws, such as Mach number range characteristics, shock wave phenomena, and stall characteristics. Random masking may lead to the loss of important physical information, resulting in low retrieval accuracy and failing to meet practical engineering needs. Based on the above summary, existing solutions in related technologies involve dynamic feature alignment methods for curve clusters and hybrid feature mining methods. These methods have achieved certain results in specific scenarios, but there is still room for improvement in unsupervised feature learning and the integration of physical knowledge. They cannot solve the core pain points of "loss of key aerodynamic information due to random masking and lack of physical interpretability of features".

[0004] Based on the above analysis of the development status of this technology field, the existing technologies lack an aerodynamic characteristic similarity retrieval scheme that can be directly implemented, has high retrieval accuracy, and is highly interpretable. This scheme would organically integrate aerodynamic physical laws with deep learning models by setting an aerodynamic specific mask strategy that can distinguish the importance of different positions within the mask autoencoder. Summary of the Invention

[0005] The purpose of this invention is to provide a method, apparatus, device, and storage medium for aerodynamic characteristic similarity retrieval, in order to solve the aforementioned problems in the prior art.

[0006] According to a first aspect of the present invention, an aerodynamic characteristic similarity retrieval method is provided, comprising: The aerodynamic data is converted into an aerodynamic feature image, and the aerodynamic feature image is cropped to obtain an aerodynamic feature sub-image that is adapted to the input size of the mask autoencoder. Construct a mask autoencoder, set an aerodynamic specific masking strategy to distinguish the importance of different positions in the mask autoencoder, and use the aerodynamic specific masking strategy to process the aerodynamic feature sub-map to obtain the mask sub-map; A masked autoencoder is trained using a masked subgraph to obtain an aerodynamic feature extraction model. The aerodynamic characteristic feature extraction model is used to extract the feature vectors of each component of the aircraft's aerodynamic data. Each feature vector is retrieved independently to obtain the similarity ranking results. The similarity ranking results are then merged to obtain the final retrieval results.

[0007] According to a second aspect of the present invention, an aerodynamic characteristic similarity retrieval device is provided, comprising: The image processing module is used to convert aerodynamic data into aerodynamic feature images, and to crop the aerodynamic feature images to obtain aerodynamic feature sub-images that are adapted to the input size of the mask autoencoder. The masking strategy module is used to construct a masking autoencoder. In the masking autoencoder, aerodynamic specific masking strategies are set to distinguish the importance of different positions. The aerodynamic specific masking strategies are used to process the aerodynamic feature sub-maps to obtain the masked sub-maps. The model training module is used to train a mask autoencoder using a mask subgraph to obtain an aerodynamic feature extraction model. The retrieval application module is used to extract feature vectors of each component of the aircraft's aerodynamic data through an aerodynamic characteristic feature extraction model, perform independent retrieval on each feature vector to obtain similarity ranking results, and fuse the similarity ranking results to obtain the final retrieval result.

[0008] According to a third aspect of the present invention, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the aerodynamic characteristic similarity retrieval method provided in the first aspect of the present disclosure.

[0009] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which an information transmission implementation program is stored, which, when executed by a processor, implements the steps of the aerodynamic characteristic similarity retrieval method provided in the first aspect of the present disclosure.

[0010] The technical solution provided by this invention includes the following beneficial effects: The designed aerodynamic-specific masking strategy, combined with the characteristics of the Mach number interval, implements differentiated masking to accurately protect key aerodynamic regions such as shock waves, stalls, and extreme points. It deeply integrates aerodynamic physical knowledge with deep learning models, guiding the model to learn aerodynamic features with clear physical meaning. This solves the core defects of existing random or uniform masking methods, such as easy loss of key aerodynamic information and lack of physical interpretability of features, significantly improving feature reliability and engineering practicality, thereby increasing the credibility of search results. Independent searches are conducted for the aerodynamic meaning and engineering importance of each balance component, comprehensively capturing the feature differences of each component and avoiding feature omissions. Differentiated weights are assigned to each component based on engineering practice, and the search results are integrated through weighted fusion. This solves the problems of weak generalization ability and insufficient search accuracy of existing direct overall search or single component search methods, making the final results more in line with the actual needs of aircraft design and improving search accuracy.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. 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 drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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 characteristic similarity retrieval method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall process of an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the conversion of pneumatic data into image format according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the aerodynamic specific masking strategy for identifying key shock regions according to an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the identification of stall critical regions using an aerodynamic specific masking strategy according to an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the identification of critical regions at extreme points using a pneumatically specific masking strategy according to an embodiment of the present invention. Figure 7 This is a schematic diagram of the six-component independent retrieval matching recommendation results according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the weighted fusion matching recommendation result 1 based on the number of trains after an embodiment of the present invention; Figure 9 This is a schematic diagram of the weighted fusion matching recommendation result 2 based on the number of trains in an embodiment of the present invention; Figure 10 This is a schematic diagram of the aerodynamic characteristic similarity retrieval device according to an embodiment of the present invention; Figure 11 This is a schematic diagram of an electronic device according to an embodiment of the present invention. 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] Method Example 1 According to embodiments of the present invention, a method for retrieving aerodynamic characteristics similarity is provided. Figure 1 This is a flowchart of the aerodynamic characteristic similarity retrieval method according to an embodiment of the present invention, such as... Figure 1 As shown, the aerodynamic characteristic similarity retrieval method according to an embodiment of the present invention specifically includes: In step S110, the aerodynamic data is converted into an aerodynamic feature image, and the aerodynamic feature image is cropped to obtain an aerodynamic feature sub-image adapted to the input size of the mask autoencoder, specifically including: The original aerodynamic data is obtained from the historical wind tunnel test database. In this embodiment of the invention, the original aerodynamic data includes the balance six-component coefficients and other flight condition parameters of different aircraft under different flight conditions. The balance six-component 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. The flight condition parameters include Mach number Ma, angle of attack α, pitch deflection dp, roll deflection dr, and yaw deflection dy of local components of the model, and roll angle φ and yaw angle β of the model as a whole. The range of Mach number Ma is 0.3-4.5, and the range of angle of attack α is -10° to +30°. Preferably, the aerodynamic data is divided into Mach number ranges. Based on the aerodynamic characteristics, the Mach number is divided into three ranges: subsonic range, transonic range, and supersonic range. Within each Mach number range, the wind tunnel test data is grouped according to the grouping rule of "the test conditions are fixed except for the Mach number Ma". This ensures that each group of data corresponds to only one Mach number range, laying the foundation for the subsequent generation of basic grayscale maps of single-range aerodynamic characteristics. The specific grouping method is as follows: traverse all test state parameters of all trains within the same Mach number range for the same aircraft shape, including roll angle φ, yaw angle β, pitch deflection dp, roll deflection dr, and yaw deflection dy. Based on the principle of single-variable control, trains with identical roll angle φ, yaw angle β, pitch deflection dp, roll deflection dr, and yaw deflection dy, but different Mach numbers, are grouped together to form a set of trains in each group that differ only in Ma, and the Ma of each group of trains belongs to the same Mach number range.

[0016] The aerodynamic data were processed sequentially, including curve cluster construction, bilinear interpolation, normalization, and grayscale mapping, to obtain aerodynamic feature images. The aerodynamic feature images are basic grayscale images of aerodynamic characteristics with a size of 512×320. The aerodynamic feature images correspond to one of the three Mach number ranges: subsonic, supersonic, and transonic. The Mach number range for the subsonic range is 0.3-0.8 Ma, for the transonic range it is 0.8-1.2 Ma, and for the supersonic range it is 1.2-4.5 Ma. For each set of trains, curve clusters are constructed based on the six components of aerodynamic data: with the angle of attack α as the abscissa and the corresponding aerodynamic coefficient as the ordinate, each train corresponds to an "α-aerodynamic coefficient" curve, and multiple trains together form a curve cluster for this aerodynamic component; the curve cluster is used to present the influence of Mach number changes on the aerodynamic coefficient-angle of attack relationship within the same Mach number range, fully preserving the two-dimensional dynamic correlation information, and each curve cluster corresponds to only one Mach number range.

[0017] Because the angle of attack α and Mach number Ma may differ in the intervals and number of data points among the curve clusters constructed from different aircraft shapes, a unified feature input cannot be formed. Therefore, it is necessary to convert the curve clusters corresponding to each Mach number interval into standard-sized single-channel aerodynamic characteristic grayscale images. Each aerodynamic characteristic grayscale image corresponds to only one Mach number interval. Then, the aerodynamic characteristic input sub-image adapted to the model input is obtained through sliding window cropping, thus completing the conversion of aerodynamic data to image format. The specific complete steps are as follows: The first step is to construct the numerical matrix using bilinear interpolation.

[0018] A two-dimensional mesh is constructed with the angle of attack α as the horizontal dimension and the Mach number Ma as the vertical dimension: the horizontal dimension uniformly maps α to 512 pixels, and the horizontal coordinate increases with the value of α from small to large; the vertical dimension uniformly maps Ma to 320 pixels, and the vertical coordinate increases with the value of Ma from small to large, forming a 512×320 two-dimensional mesh structure. Based on the α-aerodynamic coefficient data of each train within the curve cluster, a bilinear interpolation algorithm is used to interpolate and complete the aerodynamic coefficient values ​​corresponding to all α-Ma coordinates in the two-dimensional grid. By interpolating the aerodynamic coefficients under adjacent known α and Ma conditions, the aerodynamic coefficient values ​​of unknown coordinates in the grid are obtained, ultimately generating a 512×320 numerical matrix. This numerical matrix and the subsequent aerodynamic characteristic grayscale map correspond to only one Mach number interval. The value of each element in the numerical matrix is ​​the aerodynamic coefficient under the corresponding α-Ma condition, realizing the transformation of the curve cluster from a discrete set of data curves to a continuous two-dimensional numerical matrix, solving the problem of curve cluster heterogeneity caused by discontinuous α data and mismatched Ma data for different trains. The second step is normalization and grayscale mapping.

[0019] The Min-Max normalization method is used to map the values ​​of all elements in the 512×320 numerical matrix to the interval between 0 and 1. The normalization formula is x_norm = (x-x_min) / (x_max-x_min), where x is the original element value of the matrix, and x_min and x_max are the minimum and maximum values ​​of the aerodynamic coefficients in the numerical matrix, respectively. This eliminates the interference caused by differences in the dimensions and numerical ranges of different aerodynamic components. Subsequently, the normalized values ​​are mapped one-to-one with the gray values. The gray values ​​range from 0 to 255, that is, a normalized value of 0 corresponds to gray value 0, and a normalized value of 1 corresponds to gray value 255. The 512×320 numerical matrix is ​​transformed into a single-channel aerodynamic characteristic grayscale image. Each aerodynamic characteristic grayscale image corresponds to a single Mach number interval. The brightness of the grayscale image directly reflects the magnitude of the aerodynamic coefficient under the corresponding α-Ma condition, realizing the transformation of aerodynamic data into image features.

[0020] The third step is to use a sliding window to extract aerodynamic feature images. Based on the architectural requirements of the mask autoencoder, its input size is set to 224×224 pixels. A sliding window translation method is used to extract samples from the 512×320 single-channel aerodynamic characteristic grayscale image. The horizontal truncation step size is set to 72 pixels and the vertical truncation step size is set to 32 pixels. Each time, a sub-image with the same input size as the mask autoencoder is truncated and defined as an aerodynamic characteristic input sub-image. Truncation stops when the remaining horizontal or vertical pixels are less than 224 pixels. All the truncated aerodynamic characteristic input sub-images are used as valid sample images of the curve cluster, resulting in a 224×224 aerodynamic feature sub-image. This ensures that the sample size of the input model is uniform and that the aerodynamic features of different "α-Ma" regions in the curve cluster are fully preserved through the generation of multiple sub-images, improving the sample utilization rate and completely preserving the coupled aerodynamic features of Mach number and angle of attack.

[0021] In step S120, a mask autoencoder is constructed. An aerodynamic-specific masking strategy that distinguishes the importance of different positions is set within the mask autoencoder. This strategy is then used to process the aerodynamic feature sub-map to obtain the mask sub-map. Specifically, this includes: Construct a mask autoencoder consisting of an encoder and a decoder; This section describes the architectural design of the mask autoencoder, which works in conjunction with an aerodynamic-specific masking strategy. The aerodynamic-specific masking strategy can be considered as part of the mask autoencoder, together achieving effective learning of key aerodynamic features and improving retrieval accuracy.

[0022] The encoder adopts a 6-layer Transformer structure, with each layer including 8 attention heads, a hidden layer dimension of 512, a feedforward network dimension of 2048, and the activation function is GELU. Preferably, in this embodiment of the invention, the position encoding adopts a learnable two-dimensional position encoding, which simultaneously encodes the angle of attack and Mach number position information; the decoder adopts a 4-layer Transformer structure, with a hidden layer dimension of 256.

[0023] The encoder receives the masked sub-image and extracts feature vectors. During training, the decoder performs an inverse transformation on the features extracted by the encoder to restore the pixel information of the masked area. This is then stitched together with the unmasked area to form a complete reconstructed image. In other words, the image is reconstructed based on the feature vectors. Self-supervised training provides support for minimizing reconstruction error. The encoder and decoder collaborate through learnable two-dimensional position encoding to ensure the accuracy of feature transfer and the preservation of aerodynamic position information.

[0024] The aerodynamic-specific masking strategy is designed in conjunction with aerodynamic physical characteristics and works in collaboration with the mask autoencoder to guide the mask autoencoder to learn aerodynamic features that conform to aerodynamic physical laws, thereby avoiding the loss of key aerodynamic information. Specifically, it includes four core links: interval adaptation, accurate identification of key regions, probability matrix construction, and directional continuous masking. Strategy for obtaining aerodynamic specific masks: The first step is to adapt the mask block size and determine the global mask ratio: The Mach number range corresponding to a single aerodynamic characteristic sub-image is clearly defined. The aerodynamic characteristic sub-image is obtained by cropping a 512×320 aerodynamic characteristic base grayscale image through a sliding window. After cropping, the Mach number range of the image is consistent with the original aerodynamic characteristic image. The mask block size parameter and global mask ratio are determined based on the Mach number range to which the aerodynamic feature submap belongs. If it belongs to the subsonic or supersonic range, the mask block size parameter k ranges from 14×14 to 28×28. For the aerodynamic feature submap corresponding to the subsonic or supersonic range, a large continuous area mask is used, and the global mask ratio ranges from 0.50 to 0.70. If it belongs to the transonic range, the mask block size parameter k ranges from 7×7 to 14×14. For the aerodynamic feature submap corresponding to the transonic range, a small continuous area mask is used, and the global mask ratio ranges from 0.40 to 0.60. Step 2: Identification of critical aerodynamic areas: Key region identification based on automatic feature sub-map; Shock regions are identified through aerodynamic characteristic gradient detection. The gradient values ​​of adjacent elements in the aerodynamic feature sub-map are calculated, and the gradient threshold is determined based on aerodynamic physical characteristics. If the gradient value is higher than the gradient threshold, it is identified as a shock region. Stall regions are identified through the variation law of aerodynamic coefficients. If the aerodynamic coefficient decreases within a preset time interval and falls below the stall determination threshold, it is identified as a stall region. A decrease within the preset time interval means that the aerodynamic coefficient needs to decrease abruptly. The coordinate regions corresponding to the peak and valley values ​​of the aerodynamic coefficients in the aerodynamic characteristic sub-map are identified as extreme point regions. The region size can be set to a 15×15 pixel block centered on the extreme point. The stall regions and extreme point regions of subsonic and supersonic speeds are set as key regions, as are the shock regions, stall regions, and extreme point regions of transonic speeds. The third step is to calculate the number of mask blocks and initialize the mask matrix: Combining the aerodynamic characteristics of the input submap (224×224 size), the global mask ratio, and the mask block size k, calculate the number of target mask blocks N. Set the number of target mask blocks N to (224×224×global mask ratio) / (k×k) rounded up, and construct a mask matrix of size (224-k)×(224-k) with all 1s. Randomly select 2N points in the mask matrix and generate random numbers from 0 to 1 evenly distributed to replace the corresponding point values. Keep the values ​​of the remaining positions unchanged. Step 4: Generation of the basic probability matrix and the final probability matrix: Based on the identification results of key regions, a 224×224 basic probability matrix with the same size as the input aerodynamic characteristic submap is generated. The values ​​of non-key regions are set to the global mask ratio, that is, the value range of each element of the non-key regions corresponding to the subsonic and supersonic ranges is 0.50-0.70, and the value range of each element of the non-key regions corresponding to the transonic range is 0.40-0.60. The mask probability of key regions needs to be lower than the global mask ratio of the corresponding range, and the value range of key regions is set to 0.10 to 0.20. Generate a final probability matrix of size (224-k)×(224-k) that is exactly the same as the size of the mask matrix. Set the value of each point in the final probability matrix to the average mask ratio of the corresponding mask block in the basic probability matrix, and the position of each element corresponds one-to-one with the starting position of the lower left corner of the corresponding mask block in the basic probability matrix. This is equivalent to the k×k mask block sliding on a 224×224 size. Step 5: Execution of aerodynamic specificity mask determination: After the final probability matrix is ​​generated, the final probability matrix is ​​compared with the mask matrix point by point. If the value at the corresponding position in the mask matrix is ​​less than the value at the corresponding position in the final probability matrix, the mask block at that position is determined to be masked. Otherwise, no masking operation is performed, and a mask sub-image is obtained. In other words, if the value at the corresponding position in the mask matrix is ​​greater than or equal to the value at the corresponding position in the final probability matrix, the mask block at that position retains the original aerodynamic characteristic input sub-image features and is not masked.

[0025] The implementation process of the aerodynamic specific masking strategy can be summarized as follows: First, determine the global masking ratio of the corresponding interval of a single aerodynamic characteristic sub-map. Then, identify the aerodynamic key regions of the aerodynamic characteristic sub-map. Based on the identification results, construct a probability matrix and complete the adaptive masking by combining the directional continuous masking rules along the angle of attack and Mach number directions. If there are aerodynamic key regions, reduce their masking probability; if there are no aerodynamic key regions, no additional processing is performed. Aerodynamic key regions include shock wave regions, stall regions, and extreme point regions of each component. Reducing the masking probability specifically means setting a value lower than the global masking ratio of the interval for pixels in key regions, while pixels in non-key regions use the global masking ratio of the corresponding interval as their masking probability.

[0026] In step S130, a masked autoencoder is trained using a masked subgraph to obtain an aerodynamic characteristic feature extraction model, specifically including: The masked autoencoder is trained in a self-supervised manner using a masked sub-image. In the self-supervised training, the encoder is responsible for extracting and encoding the key aerodynamic features of the aerodynamic feature sub-image, and the decoder is responsible for image reconstruction based on the features output by the encoder. The training objective is to minimize the pixel reconstruction error between the decoder-reconstructed image and the original aerodynamic feature sub-image, and no manual data annotation is required throughout the process. In this embodiment of the invention, the pixel-level difference between the reconstructed image and the original aerodynamic characteristic sub-image is calculated using the mean square error loss function, and the AdamW optimizer is used in conjunction with a cosine annealing learning rate scheduler. After training, the encoder of the masked autoencoder is set as the aerodynamic feature extraction model, which can convert the input aerodynamic feature subgraph into a 512-dimensional feature vector.

[0027] In step S140, feature vectors of each component of the aircraft's aerodynamic data are extracted using an aerodynamic characteristic feature extraction model. Each feature vector is independently retrieved to obtain a similarity ranking result. The similarity ranking results are then fused to obtain the final retrieval result, specifically including: The specific components of the aircraft aerodynamic data 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, which are the six components of the balance data.

[0028] The aerodynamic data of all aircraft in the historical database are grouped according to Mach number intervals, and then converted into 512-dimensional feature vectors through the aerodynamic characteristic feature extraction model. These feature vectors are then stored separately according to the six components of the balance, thus constructing multiple historical data feature vector libraries.

[0029] For a new aircraft to be retrieved, its aerodynamic data is converted into aerodynamic characteristic sub-maps. First, the aerodynamic data is grouped according to Mach number intervals to ensure that each group of data belongs to only one of the three intervals: subsonic, supersonic, and transonic. For a group of data within the same Mach number interval, six aerodynamic characteristic sub-maps are generated according to the six components of the balance according to the processing methods such as aerodynamic data grouping, curve cluster construction, bilinear interpolation, normalization and grayscale mapping, and sliding window truncation. These six aerodynamic characteristic sub-maps with the same Mach number interval but different components are input into the trained feature extraction model to obtain six 512-dimensional feature vectors. Each feature vector corresponds to a curve cluster of a certain aerodynamic component within the corresponding Mach number interval. Feature vectors are obtained through an aerodynamic characteristic feature extraction model. The similarity between the feature vectors and the corresponding feature vectors of aircraft in the historical database is independently retrieved. Cosine similarity is used as a metric to obtain the similarity ranking results. For each aerodynamic component feature vector of the new aircraft, the cosine similarity between it and the corresponding component feature vector in the same Mach number interval in the historical data feature vector library is calculated. The six aerodynamic components are independently retrieved, and a preset number of most similar curve clusters are output for each component. The similarity ranking results of each component are collected, and the similarity ranking results that exist in each component are selected as the intersection result set. The comprehensive similarity score of each sample in the intersection result set is calculated by weighted fusion, and the final search result is determined based on the comprehensive similarity score. In this embodiment of the invention, a six-component balance is used for independent retrieval to obtain the similarity ranking results of each of the six components. Candidates that appear multiple times under different components indicate that they are similar to the sample to be detected in all dimensions. Therefore, only the samples that appear repeatedly in the initial ranking results of the six components are selected for further judgment, similar to an intersection operation. Different weights are assigned according to the importance of different aerodynamic components in aircraft design, and the data of historical aircraft with similar aerodynamic laws are finally obtained by ranking them according to the comprehensive similarity score. Preferably, the independent similarity ranking results under the components can be used as an explanatory explanation to facilitate the similarity research under independent components.

[0030] Method Example 2 This invention is based on the Python 3.12 development environment, uses the PyTorch deep learning framework, and runs on a workstation equipped with an NVIDIA RTX 4090 GPU. Figure 2 This is a schematic diagram of the overall process of an embodiment of the present invention, as shown below. Figure 2 The figure shows the overall process of aerodynamic characteristic similarity retrieval.

[0031] The first step is data preparation and preprocessing, curve cluster construction and image processing.

[0032] 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. One round of testing is called a run. Each run records the test state parameters (angle of attack α, roll angle φ, yaw angle β, pitch deflection dp, roll deflection dr, yaw deflection dy, Mach number Ma) and six-component aerodynamic coefficients (drag coefficient CA, normal force coefficient CN, yaw force coefficient CZ, roll moment coefficient MXG, yaw moment coefficient MYG, pitch moment coefficient MZG). Before grouping, the historical wind tunnel data needs to be divided into three sets according to Mach number: subsonic range [0.4~0.8], transonic range [0.8~1.2], and supersonic range [1.2~4.5]. Within each set, the 55,713 test data points are further grouped according to the grouping rule of "keeping all test conditions except Mach number Ma". First, the test condition parameters of all test data points are traversed. The test condition parameters include roll angle φ, yaw angle β, pitch deflection dp, roll deflection dr, and yaw deflection dy. Following the principle of "single-variable control", only one test condition parameter is changed each time. Test data points with "φ, β, dp, dr, and dy are completely the same, only Mach number is different" are grouped together to form a set of test data points in each group where only Mach number differs. The number of test data points in each range is shown in Table 1. Table 1 Number of train services combined

[0033] For each set of train services mentioned above, curve clusters are constructed based on the six components of aerodynamic data: with the angle of attack α as the abscissa and the corresponding aerodynamic coefficient as the ordinate. The 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. Each train service corresponds to one "α-aerodynamic coefficient" curve. Multiple train services together form the curve cluster of this aerodynamic component, which contains a total of 6 × (4139 + 3422 + 5614) = 79050 curve clusters.

[0034] Figure 3 This is a schematic diagram illustrating the conversion of pneumatic data into an image format according to an embodiment of the present invention, such as... Figure 3 As shown, for each curve cluster, it is converted into a fixed-size single-channel grayscale sample image through bilinear interpolation, normalization, and sliding window cropping. The specific operations are as follows: Based on multiple known data within the curve cluster, with the angle of attack α as the horizontal dimension and the Mach number Ma as the vertical dimension, the bilinear interpolation method is used to complete the aerodynamic coefficient values ​​corresponding to all (α, Ma) coordinates in the 512×320 grid, ultimately generating a 512×320 numerical matrix.

[0035] The numerical matrix is ​​normalized using the Min-Max method according to Formula 1: Formula 1; in, Represents the original element values ​​of the matrix. , The minimum and maximum values ​​of the aerodynamic coefficients in the numerical matrix are represented respectively. The normalized numerical matrix is ​​then mapped to a single-channel grayscale image, with each grayscale value corresponding to a normalized aerodynamic coefficient. A normalized value of 0 corresponds to a grayscale value of 0, and a normalized value of 1 corresponds to a grayscale value of 255.

[0036] Each 512×320 grayscale image is cropped into a 224×224 sub-image using a sliding window rule with a horizontal step size of 72 and a vertical step size of 32. The horizontal cropping can be done 5 times, with starting indices of 0, 72, 144, 216, and 288 respectively; the vertical cropping can be done 4 times, with starting indices of 0, 32, 64, and 96 respectively. Each grayscale image can generate 5×4=20 valid sample images.

[0037] The second step is the implementation and model building of the aerodynamic specific masking strategy.

[0038] The model is built based on the Transformer architecture to construct a masked autoencoder. The encoder adopts a 6-layer Transformer structure with 8 attention heads in each layer, a hidden layer dimension of 512, a feedforward network dimension of 2048, and the activation function is GELU. The decoder adopts a 4-layer Transformer structure with a hidden layer dimension of 256, which is used for the reconstruction of the masked aerodynamic characteristic input sub-pattern. The position encoding adopts a learnable two-dimensional position encoding, which simultaneously encodes the angle of attack and Mach number position information.

[0039] The implementation process of the aerodynamic specificity masking strategy is as follows: Step 2.1, Mask block size adaptation and global mask ratio determination: The Mach number range corresponding to a single aerodynamic characteristic input sub-plot is clearly defined. The aerodynamic characteristic input sub-plot is obtained by sliding window cropping from a 512×320 aerodynamic characteristic base grayscale image. The Mach number range is one of the three Mach number ranges: subsonic, supersonic, and transonic. The Mach number range of the subsonic range is 0.3-0.8 Ma, the Mach number range of the transonic range is 0.8-1.2 Ma, and the Mach number range of the supersonic range is 1.2-4.5 Ma. Determine the k-value for a single 224×224 aerodynamic characteristic input sub-image, where k is the mask block size parameter. For aerodynamic characteristic input sub-images corresponding to the subsonic or supersonic speed range, use large continuous area masks, with the k-value selected to adapt to the aerodynamic characteristics of the subsonic and supersonic speed ranges, ranging from 14×14 to 28×28 pixels. For aerodynamic characteristic input sub-images corresponding to the transonic speed range, use small continuous area masks, with the k-value selected to adapt to the aerodynamic characteristics of the transonic speed range, ranging from 7×7 to 14×14 pixels. The global mask ratio is determined based on the aerodynamic characteristics of the corresponding Mach number range. The global mask ratio ranges from 0.50 to 0.70 for the subsonic and supersonic ranges, and from 0.40 to 0.60 for the transonic range.

[0040] Step 2.2, Identification of critical aerodynamic areas: Figure 4 This is a schematic diagram of the aerodynamic specific masking strategy for identifying key shock regions according to an embodiment of the present invention, as shown below. Figure 4 As shown, shock wave regions are identified through aerodynamic characteristic gradient detection. The gradient values ​​of adjacent elements in the numerical matrix corresponding to the aerodynamic characteristic input sub-map are calculated. A gradient threshold is determined based on the aerodynamic physical characteristics. Regions with gradient values ​​exceeding this threshold are considered shock wave regions. Figure 4 The area circled in bold red in the fourth sub-image is identified as the critical region of the shock wave; for example... Figure 4As shown, the stall region is identified by the variation law of the aerodynamic coefficient. The stall judgment threshold is determined according to the aerodynamic physical characteristics. The region where the aerodynamic coefficient suddenly drops and is below the stall judgment threshold is judged as the stall region. The threshold is a normalized value. Figure 5 This is a schematic diagram illustrating the stall critical region identification using an aerodynamic specific masking strategy according to an embodiment of the present invention. Figure 5 The area preserved in the right-hand diagram is the stall area; Figure 6 This is a schematic diagram of the critical region identification of extreme points using the aerodynamic specific masking strategy according to an embodiment of the present invention, as shown below. Figure 6 As shown, the extreme point region is obtained by traversing the numerical matrix corresponding to the aerodynamic characteristic input sub-map and identifying the coordinate regions corresponding to the peak and valley values ​​of the aerodynamic coefficient. The region size is a 15×15 pixel block centered on the extreme point. Figure 6 The 15×15 block reserved on the right is the extreme point region.

[0041] Step 2.3, Calculation of the number of mask blocks and initialization of the mask matrix: Combining the 224×224 size of the aerodynamic characteristic submap, the global mask ratio, and the mask block size k, calculate the number of target mask blocks N. The calculation formula is: N = (224×224×global mask ratio) / (k×k). After the calculation, round N up. Construct a (224-k)×(224-k) all-1 mask matrix. Randomly and uniformly select 2×N points in the entire matrix and generate random numbers from 0 to 1 to replace the corresponding point values. Keep the values ​​of the remaining positions unchanged.

[0042] Step 2.4, Generation of the basic probability matrix and the final probability matrix: Based on the aerodynamic critical region identification results, a 224×224 basic probability matrix with the same input size as the aerodynamic characteristic input sub-map is generated. The value rules for each element of the basic probability matrix are as follows: the mask probability of non-critical regions adopts the global mask ratio of the corresponding Mach number interval, where the value range of each element for non-critical regions corresponding to subsonic and supersonic intervals is 0.50-0.70, and the value range of each element for non-critical regions corresponding to transonic intervals is 0.40-0.60; the mask probability of critical regions is lower than the global mask ratio of the corresponding interval, and the specific value range of each element in this region is 0.10-0.20.

[0043] Generate a final probability matrix with the exact same size as the mask matrix, which has a size of (224-k)×(224-k). The value of each element in this matrix is ​​the average mask probability of the corresponding mask block in the basic probability matrix, and the position of each element corresponds one-to-one with the starting position of the lower left corner of the corresponding mask block in the basic probability matrix. Step 2.5, Aerodynamic Specificity Mask Determination: After the final probability matrix is ​​generated, the values ​​of the matrix and the mask matrix are compared point by point. If the value at the corresponding position in the mask matrix is ​​less than the value at the corresponding position in the final probability matrix, the mask block at that position is determined to be masked. This position corresponds to the first point at the lower left corner of the mask block, and the size of the mask block is k×k. If the value at the corresponding position in the mask matrix is ​​greater than or equal to the value at the corresponding position in the final probability matrix, the mask block at that position retains the original aerodynamic characteristic input sub-image features and is not masked.

[0044] The third step is self-supervised training of the model and feature extraction.

[0045] The mean squared error loss function and AdamW optimizer were used, with a base learning rate of 3e-4, weight decay of 0.05, batch size of 32, and training for 200 epochs. A cosine annealing learning rate scheduler was used, and the training objective was to minimize the decoder reconstruction error.

[0046] After training, the encoder, acting as a feature extraction model, can convert the 224×224 aerodynamic characteristic input subgraph into a 512-dimensional feature vector. The aerodynamic characteristic input subgraphs corresponding to the dataset shown in Table 1 are input into the feature extraction model to extract 512-dimensional feature vectors, and a historical data feature vector library is constructed according to aircraft shape, aerodynamic components, and Mach number range.

[0047] The fourth step is similarity retrieval and result output.

[0048] For a new aircraft to be retrieved, its aerodynamic data is grouped according to Mach number intervals, processed into images, and then 512-dimensional feature vectors of the six aerodynamic characteristic input sub-images are extracted. Using cosine similarity as a metric, the similarity between the feature vectors of the six aerodynamic components and the feature vectors of the same component and interval in the historical data feature vector database is calculated. For each component, the top 20 most similar candidate results are output. Figure 7 This is a schematic diagram of the six-component independent retrieval matching recommendation results according to an embodiment of the present invention, as shown below. Figure 7 As shown, the six aerodynamic components independently retrieve the aerodynamic characteristic subgraphs of their respective top three most similar candidate results.

[0049] In this embodiment of the invention, the actual processing uses channel copying to convert the grayscale image into a similar format. Figures 3 to 7 The color image shown.

[0050] Finally, multi-component fusion is performed to generate the final retrieval results.

[0051] The similarity ranking results of the six components are fused. Different weights are assigned to different aerodynamic components based on their importance in aircraft design. A weighted intersection fusion method is used to obtain historical train data with similar aerodynamic patterns. In this example, a weight vector w = [0.12, 0.22, 0.18, 0.08, 0.18, 0.22] is used, corresponding to the six components: drag coefficient Ca, normal force coefficient Cn, yaw force coefficient Cz, roll moment coefficient Mxg, yaw moment coefficient Myg, and pitch moment coefficient Mzg. A weighted comprehensive similarity is calculated, and the final search results are output. Figure 8 This is a schematic diagram of the weighted fusion matching recommendation result 1 based on the number of trains in an embodiment of the present invention. Figure 9 This is a schematic diagram of the weighted fusion matching recommendation result 2 based on the number of trains after an embodiment of the present invention. Figure 8 and Figure 9The display shows the search results by train number after weighted intersection fusion. Dashed lines represent train numbers to be matched, solid lines represent recommended matching train numbers, and the same color indicates the similarity of the same matching result under different weights. In summary, addressing the existing problems, this invention proposes an aerodynamic characteristic similarity retrieval method. It employs a self-supervised training mode, relying on the image reconstruction task of a masked autoencoder to autonomously learn aerodynamic features without manual annotation, significantly reducing the manpower and time costs of aerodynamic data processing, avoiding the subjective errors of manual annotation, providing a feasible path for efficient processing of large-scale aerodynamic data, and improving the efficiency of technology implementation. The designed aerodynamic-specific masking strategy, combined with Mach number range characteristics, implements differentiated masking to accurately protect key aerodynamic regions such as shock waves, stalls, and extreme points. It deeply integrates aerodynamic physical knowledge with deep learning models, guiding the model to learn aerodynamic features with clear physical meaning. This solves the core defects of existing random or uniform masking methods, such as easy loss of key aerodynamic information and lack of physical interpretability of features, significantly improving feature reliability. To enhance engineering practicality and improve the reliability of search results, this method utilizes the Transformer architecture to clearly define key parameters such as the number of encoder and decoder layers, the number of attention heads, and the dimension of hidden layers. Combined with learnable two-dimensional position encoding, it balances the ability to learn complex aerodynamic features with training stability. Detailed descriptions of the architecture and key parameters ensure that those skilled in the art can accurately reproduce the model building and training process. Independent searches are conducted for the aerodynamic significance and engineering importance of each of the six components of the balance, comprehensively capturing the differences in features of each component and avoiding feature omissions. Differentiated weights are assigned to each component based on engineering realities, and the search results are integrated through weighted fusion. This addresses the problems of weak generalization and insufficient search accuracy in existing direct overall search or single-component search methods, making the final results more aligned with the actual needs of aircraft design and improving search accuracy. The overall solution constructs a complete and detailed aerodynamic data similarity retrieval system, encompassing "aerodynamic data preprocessing - feature learning - similarity retrieval - result fusion." This addresses the problems of fragmented processes, strong subjectivity, and low efficiency in existing methods. The method of this invention forms a fully automated closed loop from raw aerodynamic data to the final retrieval result. Each process conforms to the physical laws of aerodynamics and engineering requirements, effectively improving the efficiency, consistency, and engineering applicability of aerodynamic characteristic similarity retrieval.

[0052] Device Examples According to an embodiment of the present invention, an aerodynamic characteristic similarity retrieval device is provided. Figure 10 This is a schematic diagram of the aerodynamic characteristic similarity retrieval device according to an embodiment of the present invention, as shown below. Figure 10 As shown, the aerodynamic characteristic similarity retrieval device according to an embodiment of the present invention specifically includes: Image processing module 100 is used to convert aerodynamic data into aerodynamic feature images, and to crop the aerodynamic feature images to obtain aerodynamic feature sub-images adapted to the input size of the mask autoencoder. Specifically, it is used for: The aerodynamic data are sequentially processed including curve cluster construction, bilinear interpolation, normalization, and grayscale mapping to obtain aerodynamic feature images; where the aerodynamic feature images correspond to one of the three Mach number ranges: subsonic, supersonic, and transonic. A sliding window was used to crop the aerodynamic feature image, resulting in a 224×224 aerodynamic feature sub-image.

[0053] Masking strategy module 102 is used to construct a masking autoencoder. It sets aerodynamic-specific masking strategies to distinguish the importance of different positions within the autoencoder. These strategies are then used to process the aerodynamic feature sub-images to obtain the masked sub-images. Specifically, this is used for: Construct a mask autoencoder consisting of an encoder and a decoder; The encoder uses a 6-layer Transformer structure, with each layer containing 8 attention heads, a hidden layer dimension of 512, a feedforward network dimension of 2048, and GELU activation function; the decoder uses a 4-layer Transformer structure, with a hidden layer dimension of 256. The encoder receives the masked sub-image and extracts the feature vectors, and the decoder reconstructs the image based on the feature vectors during training.

[0054] Strategy for obtaining aerodynamic specific masks: The mask block size parameter and global mask ratio are determined based on the Mach number range to which the aerodynamic feature submap belongs. If it belongs to the subsonic or supersonic range, the value of the mask block size parameter k is in the range of 14×14 to 28×28, and the value of the global mask ratio is in the range of 0.50-0.70. If it belongs to the transonic range, the value of the mask block size parameter k is in the range of 7×7 to 14×14, and the value of the global mask ratio is in the range of 0.40-0.60. Calculate the gradient values ​​of adjacent elements in the aerodynamic feature submap. If the gradient value is higher than the gradient threshold, it is identified as a shock wave region. If the aerodynamic coefficient decreases within a preset time interval and is lower than the stall determination threshold, it is identified as a stall region. Identify the coordinate regions corresponding to the peak and valley values ​​of the aerodynamic coefficient as extreme point regions. Set the stall regions and extreme point regions of subsonic and supersonic speeds as key regions. Set the shock wave regions, stall regions, and extreme point regions of transonic speeds as key regions. Set the number of target mask blocks N to (224×224×global mask ratio) / (k×k) rounded up, construct a mask matrix of size (224-k)×(224-k) with all 1s, and randomly select 2N points in the mask matrix to generate random numbers of 0 to 1 evenly distributed to replace the corresponding point values; Based on the recognition results of the key regions, a basic probability matrix of size 224×224 is generated. The values ​​of non-key regions are set as the global mask ratio, and the values ​​of key regions are set to the range of 0.10 to 0.20. A final probability matrix of size (224-k)×(224-k) is generated. The values ​​of each point in the final probability matrix are set as the average mask ratio of the corresponding mask block in the basic probability matrix. The position of each point in the final probability matrix corresponds to the starting position of the lower left corner of the corresponding mask block in the basic probability matrix. The final probability matrix is ​​compared with the mask matrix point by point. If the value at the corresponding position in the mask matrix is ​​less than the value at the corresponding position in the final probability matrix, the mask block at that position is determined to be masked. Otherwise, no masking operation is performed, and a mask sub-image is obtained.

[0055] Model training module 104 is used to train a masked autoencoder using a masked subgraph to obtain an aerodynamic feature extraction model, specifically for: The masked autoencoder is trained using a masked submap, with the training objective set as minimizing the pixel reconstruction error between the decoder-reconstructed image and the original aerodynamic feature submap. After training, set the encoder of the masked autoencoder to the aerodynamic feature extraction model.

[0056] The retrieval application module 106 is used to extract feature vectors of each component of the aircraft's aerodynamic data through an aerodynamic characteristic feature extraction model, independently retrieve similarity ranking results for each feature vector, and fuse the similarity ranking results to obtain the final retrieval result. Specifically, it is used for: The similarity between the independently retrieved feature vectors and the corresponding feature vectors of aircraft in the historical database is used, and the cosine similarity is used as the metric to obtain the similarity ranking results. The similarity ranking results of each component are collected, and the similarity ranking results that exist in each component are selected as the intersection result set. The comprehensive similarity score of each sample in the intersection result set is calculated by weighted fusion, and the final search result is determined based on the comprehensive similarity score.

[0057] The specific components of the aircraft aerodynamic data 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.

[0058] In summary, to address the existing problems, this invention presents an aerodynamic characteristic similarity retrieval device. Employing a self-supervised training model, it autonomously learns aerodynamic features through image reconstruction tasks using a masked autoencoder, eliminating the need for manual annotation. This significantly reduces the manpower and time costs of aerodynamic data processing, avoids subjective errors in manual annotation, and provides a feasible path for efficient processing of large-scale aerodynamic data, improving the efficiency of technology implementation. The designed aerodynamic-specific masking strategy, combined with Mach number range characteristics, implements differentiated masking to accurately protect key aerodynamic regions such as shock waves, stalls, and extreme points. It deeply integrates aerodynamic physical knowledge with a deep learning model, guiding the model to learn aerodynamic features with clear physical meaning. This solves the core defects of existing random or uniform masking methods, such as the easy loss of key aerodynamic information and the lack of physical interpretability of features, significantly improving feature reliability. To enhance engineering practicality and improve the reliability of search results, this method utilizes the Transformer architecture to clearly define key parameters such as the number of encoder and decoder layers, the number of attention heads, and the dimension of hidden layers. Combined with learnable two-dimensional position encoding, it balances the ability to learn complex aerodynamic features with training stability. Detailed descriptions of the architecture and key parameters ensure that those skilled in the art can accurately reproduce the model building and training process. Independent searches are conducted for the aerodynamic significance and engineering importance of each of the six components of the balance, comprehensively capturing the differences in features of each component and avoiding feature omissions. Differentiated weights are assigned to each component based on engineering realities, and the search results are integrated through weighted fusion. This addresses the problems of weak generalization and insufficient search accuracy in existing direct overall search or single-component search methods, making the final results more aligned with the actual needs of aircraft design and improving search accuracy. The overall solution constructs a complete and detailed aerodynamic data similarity retrieval system, encompassing "aerodynamic data preprocessing - feature learning - similarity retrieval - result fusion." This addresses the problems of fragmented processes, strong subjectivity, and low efficiency in existing methods. The method of this invention forms a fully automated closed loop from raw aerodynamic data to the final retrieval result. Each process conforms to the physical laws of aerodynamics and engineering requirements, effectively improving the efficiency, consistency, and engineering applicability of aerodynamic characteristic similarity retrieval.

[0059] Electronic device examples Figure 11 This is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device 1100 may include at least one processor 1110 and a memory 1120. The processor 1110 can execute instructions stored in the memory 1120. The processor 1110 is communicatively connected to the memory 1120 via a data bus. In addition to the memory 1120, the processor 1110 can also be communicatively connected to an input device 1130, an output device 1140, and a communication device 1150 via the data bus.

[0060] Processor 1110 can be any conventional processor, such as a commercially available CPU. Processors may also include graphics processing units (GPUs), field-programmable gate arrays (FPGAs), systems on chips (SOCs), application-specific integrated circuits (ASICs), or combinations thereof.

[0061] The memory 1120 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0062] In this embodiment of the present disclosure, the memory 1120 stores executable instructions, and the processor 1110 can read the executable instructions from the memory 1120 and execute the instructions to implement all or part of the steps of any of the aerodynamic characteristic similarity retrieval methods in the above exemplary embodiments.

[0063] Computer-readable storage medium embodiments In addition to the methods and apparatus described above, exemplary embodiments of this disclosure may also be a computer program product or a computer-readable storage medium storing the computer program product, the computer product including computer program instructions that can be executed by a processor to implement all or part of the steps described in any of the aerodynamic characteristic similarity retrieval methods in the exemplary embodiments described above.

[0064] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. Programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages, and scripting languages ​​(e.g., Python). The program code can be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0065] Computer-readable storage media may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: static random access memory (SRAM) having one or more electrically connected wires; electrically erasable programmable read-only memory (EEPROM); erasable programmable read-only memory (EPROM); programmable read-only memory (PROM); read-only memory (ROM); magnetic storage; flash memory; magnetic disk or optical disk; or any suitable combination thereof.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A method for retrieving aerodynamic characteristic similarity, characterized in that, include: The aerodynamic data is converted into an aerodynamic feature image, and the aerodynamic feature image is cropped to obtain an aerodynamic feature sub-image that is adapted to the input size of the mask autoencoder. A mask autoencoder is constructed, and an aerodynamic specific masking strategy that distinguishes the importance of different positions is set in the mask autoencoder. The aerodynamic specific masking strategy is used to process the aerodynamic feature sub-image to obtain the mask sub-image. The mask autoencoder is trained using the mask subgraph to obtain an aerodynamic feature extraction model. The aerodynamic characteristic feature extraction model is used to extract feature vectors of each component of the aircraft's aerodynamic data. Each feature vector is independently retrieved to obtain a similarity ranking result. The similarity ranking results are then fused to obtain the final retrieval result.

2. The method according to claim 1, characterized in that, The step of converting aerodynamic data into an aerodynamic feature image and cropping the aerodynamic feature image to obtain an aerodynamic feature sub-image adapted to the input size of the mask autoencoder specifically includes: The aerodynamic data is subjected to a series of processes including curve cluster construction, bilinear interpolation, normalization, and grayscale mapping to obtain the aerodynamic feature image; wherein the aerodynamic feature image corresponds to one of the three Mach number ranges: subsonic, supersonic, and transonic. A sliding window is used to crop the aerodynamic feature image to obtain a 224×224 aerodynamic feature sub-image.

3. The method according to claim 1, characterized in that, The construction of the mask autoencoder specifically includes: constructing the mask autoencoder, which consists of an encoder and a decoder; The encoder employs a 6-layer Transformer structure, with each layer including 8 attention heads, a hidden layer dimension of 512, a feedforward network dimension of 2048, and GELU activation function; the decoder employs a 4-layer Transformer structure, with a hidden layer dimension of 256. The encoder receives a masked sub-image and extracts feature vectors, and the decoder reconstructs the image based on the feature vectors during training.

4. The method according to claim 1, characterized in that, The step of setting an aerodynamic-specific masking strategy in the mask autoencoder to distinguish the importance of different positions, and processing the aerodynamic feature sub-map using the aerodynamic-specific masking strategy to obtain the mask sub-map specifically includes: Obtain the aerodynamic specific mask strategy: The mask block size parameter and global mask ratio are determined based on the Mach number range to which the aerodynamic feature submap belongs. If it belongs to the subsonic or supersonic range, the value range of the mask block size parameter k is 14×14 to 28×28, and the value range of the global mask ratio is 0.50-0.

70. If it belongs to the transonic range, the value range of the mask block size parameter k is 7×7 to 14×14, and the value range of the global mask ratio is 0.40-0.

60. Calculate the gradient values ​​of adjacent elements in the aerodynamic feature submap. If the gradient value is higher than the gradient threshold, it is identified as a shock wave region. If the aerodynamic coefficient decreases within a preset time interval and is lower than the stall determination threshold, it is identified as a stall region. Identify the coordinate regions corresponding to the peak and valley values ​​of the aerodynamic coefficient as extreme point regions. Set the stall regions and extreme point regions of subsonic and supersonic speeds as key regions, and set the shock wave regions, stall regions, and extreme point regions of transonic speeds as key regions. Set the number of target mask blocks N to (224×224×global mask ratio) / (k×k) rounded up, construct a mask matrix of size (224-k)×(224-k) with all 1s, and randomly select 2N points in the mask matrix to generate random numbers of 0 to 1 evenly distributed to replace the corresponding point values; Based on the recognition results of the key regions, a basic probability matrix of size 224×224 is generated. The values ​​of non-key regions are set to the global mask ratio, and the values ​​of the key regions are set to the range of 0.10 to 0.

20. A final probability matrix of size (224-k)×(224-k) is generated. The values ​​of each point in the final probability matrix are set to the average mask ratio of the corresponding mask block in the basic probability matrix. The position of each point in the final probability matrix corresponds to the starting position of the lower left corner of the corresponding mask block in the basic probability matrix. The final probability matrix and the mask matrix are compared point by point. If the value at the corresponding position in the mask matrix is ​​less than the value at the corresponding position in the final probability matrix, the mask block at that position is determined to perform a masking operation. Otherwise, no masking operation is performed, and the mask sub-image is obtained.

5. The method according to claim 1, characterized in that, The step of training the masked autoencoder using the masked subgraph to obtain the aerodynamic feature extraction model specifically includes: The masked autoencoder is trained using the masked sub-image, with the training objective set as minimizing the pixel reconstruction error between the decoder-reconstructed image and the original aerodynamic feature sub-image. After training, the encoder of the masked autoencoder is set as the aerodynamic feature extraction model.

6. The method according to claim 1, characterized in that, The process of independently searching each feature vector to obtain a similarity ranking result, and then fusing the similarity ranking results to obtain the final search result, specifically includes: The similarity between the independently retrieved feature vector and the corresponding feature vector of the aircraft in the historical database is obtained by using cosine similarity as a metric. The similarity ranking results of each component are collected, and the similarity ranking results that exist in each component are selected as the intersection result set. The comprehensive similarity score of each sample in the intersection result set is calculated by weighted fusion, and the final retrieval result is determined based on the comprehensive similarity score.

7. The method according to claim 1, characterized in that, The specific components of the aircraft aerodynamic data 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.

8. A pneumatic characteristic similarity retrieval device, characterized in that, include: The image processing module is used to convert aerodynamic data into aerodynamic feature images, and to crop the aerodynamic feature images to obtain aerodynamic feature sub-images adapted to the input size of the mask autoencoder. A masking strategy module is used to construct a masking autoencoder. An aerodynamic-specific masking strategy is set in the masking autoencoder to distinguish the importance of different positions. The aerodynamic-specific masking strategy is used to process the aerodynamic feature sub-image to obtain a mask sub-image. The model training module is used to train the mask autoencoder using the mask subgraph to obtain an aerodynamic feature extraction model. The retrieval application module is used to extract feature vectors of each component of the aircraft aerodynamic data through the aerodynamic characteristic feature extraction model, perform independent retrieval on each feature vector to obtain similarity ranking results, and fuse the similarity ranking results to obtain the final retrieval result.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the aerodynamic characteristic similarity retrieval method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the aerodynamic characteristic similarity retrieval method as described in any one of claims 1 to 7.

Images