A wind tunnel aerodynamic data prediction method based on deep attention mechanism

Abstract

The present invention provides a method for predicting wind tunnel aerodynamic data based on a deep attention mechanism. The method includes: extracting aerodynamic state features and aerodynamic characteristic features from historical wind tunnel test aerodynamic data and the corresponding aerodynamic state of a given experimental configuration using an encoder; decoupling the aerodynamic state features and aerodynamic characteristic features using an attention mechanism, and obtaining overall aerodynamic characteristics through multi-level fusion; obtaining a new aerodynamic state for the given experimental configuration, coupling the new aerodynamic state with the overall aerodynamic characteristics, generating wind tunnel aerodynamic data corresponding to the new aerodynamic state using a decoder, and predicting the aerodynamic characteristics of the aircraft under the new aerodynamic state for the given experimental configuration. The present invention can mine the knowledge contained in historical wind tunnel test data and, based on this historical experience, assist in the generation of new aerodynamic state data. The present invention can reduce the frequency of wind tunnel testing or CFD simulations during the design of various experimental configurations, thereby shortening the time and cost.

Description

Technical Field

[0001] The present invention relates to the field of wind tunnel technology, and in particular to a wind tunnel aerodynamic data prediction method based on a deep attention mechanism. Background Art

[0002] Wind tunnel testing uses artificially generated airflow to simulate the effects of airflow on aircraft or other objects moving through the air. Based on the principles of relativity of motion and similarity of flows, wind tunnel testing can simulate the aerodynamic parameters, including drag and lift, experienced by various objects at given speeds, attitudes, and other aerodynamic conditions.

[0003] Wind tunnel testing has a wide range of applications, mainly including the following aspects: First, wind tunnel testing is an indispensable tool in aerodynamic research and aircraft development. Modern aircraft design relies on it to determine the aerodynamic layout and aerodynamic performance of the aircraft. Similarly, wind tunnels are indispensable in the design of automobiles. They are used to test the aerodynamic characteristics of automobiles to help optimize the shape of the car to reduce drag and improve fuel efficiency. They are often used to evaluate the stability of high-speed trains and racing cars. In addition, in the field of construction, wind tunnel testing can simulate wind loads and vibrations on buildings, evaluate the ventilation effect of buildings and their impact on the surrounding wind field.

[0004] The advantages of wind tunnel testing include precise control of experimental conditions, minimal impact from weather conditions and time of day, and highly accurate test results. However, to accurately analyze the aerodynamic performance of a test model, wind tunnel testing is required for each attitude to obtain aerodynamic data such as drag and lift under the corresponding conditions. This typically requires numerous wind tunnel tests, resulting in significant time and labor costs.

[0005] A commonly used method for obtaining aerodynamic data of exterior shapes is CFD (Computational Fluid Dynamics), a numerical method. The CFD method does not require strict experimental environment requirements and can obtain aerodynamic data in a relatively short period of time. However, compared with wind tunnel tests, the aerodynamic data obtained by the CFD method is less accurate and requires a large amount of computing resources. In addition, the key knowledge in the field of aerodynamics that CFD design relies on may have generalization deviations in actual applications, and cannot be corrected using existing wind tunnel test data.

[0006] With the rise of artificial intelligence (AI) technology, a growing number of researchers are applying it to the prediction of aerodynamic data from various experimental models. Some researchers have applied big data analysis and AI techniques to construct an artificial neural network (ANN). By learning from historical wind tunnel test data, they have achieved prediction of aircraft aerodynamic coefficients along the angle of attack dimension. Other researchers have proposed an intelligent prediction method for the airfoil pressure coefficient (Cp) based on a conditional generative adversarial network (cGAN). Given different airfoil geometry information, they can predict the pressure coefficient at various airfoil locations. Other researchers have used adaptive neuro-fuzzy inference systems, artificial neural networks, and radial basis function neural networks to evaluate and replicate wind tunnel data. The adaptive neuro-fuzzy inference system achieved the best replication results. Other researchers have used neural network models to model aerodynamic characteristics at high angles of attack in flight dynamics problems. Results show that recurrent neural networks can better model aircraft aerodynamic characteristics at high angles of attack. Other researchers have proposed a strategy for reconstructing wind pressure on high-rise building cladding based on field measurements, wind tunnel tests, and artificial neural networks. Based on wind tunnel test results, they developed four neural network models to reconstruct wind pressure on a skyscraper model. The reconstruction performance of these models met expectations.

[0007] Currently, existing wind tunnel aerodynamic data prediction methods mainly rely on wind tunnel tests and CFD simulations. These methods have the following problems:

[0008] When conducting wind tunnel tests, not only are operators required to have professional aerodynamic knowledge background, but each test also requires high time and material costs.

[0009] The aerodynamic data obtained by CFD simulation is less accurate than the results of wind tunnel tests, and a large amount of computing resources is required in the process of generating aerodynamic data.

[0010] Due to sparse sampling, the available wind tunnel aerodynamic data are limited. When the data volume is small, it is difficult to obtain effective results using the classic regression prediction model. Summary of the Invention

[0011] An embodiment of the present invention provides a wind tunnel aerodynamic data prediction method based on a deep attention mechanism to effectively improve the efficiency of wind tunnel tests and CFD simulations.

[0012] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0013] According to one aspect of the present invention, a method for predicting wind tunnel aerodynamic data based on a deep attention mechanism is provided, comprising:

[0014] Based on historical wind tunnel test aerodynamic data of a given experimental configuration and the corresponding aerodynamic state, extracting aerodynamic state features and aerodynamic characteristic features of the historical wind tunnel test aerodynamic data using an encoder;

[0015] Using an attention mechanism to decouple the aerodynamic state features and aerodynamic characteristic features, and obtaining the overall aerodynamic characteristics through multi-level fusion;

[0016] obtaining a new aerodynamic state of the given experimental configuration, coupling the new aerodynamic state with the overall aerodynamic characteristics, and generating wind tunnel aerodynamic data corresponding to the new aerodynamic state through a decoder;

[0017] The aerodynamic characteristics of the aircraft in the new aerodynamic state of the given experimental shape are predicted based on the wind tunnel aerodynamic data of the new aerodynamic state.

[0018] Preferably, the extracting of aerodynamic state features and aerodynamic characteristic features of the historical wind tunnel test aerodynamic data and the corresponding aerodynamic state of the given experimental configuration using an encoder includes:

[0019] Assume that a wind tunnel test consists of v-dimensional aerodynamic state information And the corresponding aerodynamic data results with dimension u Composition, obtain m historical wind tunnel test aerodynamic data of a given experimental shape And the aerodynamic state corresponding to the wind tunnel aerodynamic data X data,

[0020] The encoding part includes an encoder and an attention fusion module. The encoder is composed of a three-layer fully connected neural network. The wind tunnel test aerodynamic data X and the corresponding aerodynamic state S are input to the encoder. The encoder performs feature mining on the curve of the wind tunnel test aerodynamic data X to extract the aerodynamic state features and aerodynamic characteristic features corresponding to each set of wind tunnel test aerodynamic data:

[0021] f i s , f i t =F(x i , s i θ en ), (2)

[0022] where f i s , f i t Represent the aerodynamic state characteristics and aerodynamic characteristics, θ en is the encoder parameter, F is the prediction model of wind tunnel aerodynamic data;

[0023] The aerodynamic state characteristic f is Reconstructed into aerodynamic state θ linear is the parameter of the fully connected network, calculated With s i State consistency loss reg , as shown in formula (3);

[0024]

[0025] Through the state consistency loss reg So that from the aerodynamic state characteristics f i s Extracted from With the real aerodynamic state s i As close as possible.

[0026] Preferably, the use of an attention mechanism to decouple the aerodynamic characteristics and obtain overall aerodynamic characteristics through multi-level fusion includes:

[0027] After passing through the encoder, m aerodynamic characteristic features corresponding to m historical wind tunnel test aerodynamic data are obtained The m aerodynamic characteristics As the key and value of the attention fusion module, k learnable tokens are used as the query of the attention fusion module. In the attention fusion module, the aerodynamic characteristics are analyzed through k queries. Perform k different fusions, and splice the extracted k fusion features to obtain the overall aerodynamic characteristics f t :

[0028]

[0029] Among them, t i Represents the i-th fusion feature, θ token and θ attention They are the token parameters and the parameters of the attention layer respectively.

[0030] Preferably, the step of obtaining a new aerodynamic state of the given experimental configuration, coupling the new aerodynamic state with the overall aerodynamic characteristics, and generating wind tunnel aerodynamic data corresponding to the new aerodynamic state through a decoder includes:

[0031] Set the decoder to a three-layer fully connected network to obtain the new aerodynamic state s of the given experimental shape new , the new aerodynamic state s new and the overall aerodynamic characteristics f t Input to the decoder, and the decoder generates the new aerodynamic state snew Corresponding wind tunnel aerodynamic data The decoder process is as follows:

[0032]

[0033] f=cat[f t , g(s new θ pos )], (6)

[0034] Among them, θ pos is the parameter of the fully connected neural network, g(s new θ pos ) is the new aerodynamic state s new Perform feature encoding to make it consistent with the aerodynamic state feature f extracted by the encoder i s Alignment, θ de Represents the parameters of the decoder; cat represents the merged features obtained by concatenating the two input features;

[0035] During the training phase, the aerodynamic data curve generated by the calculation The mean square error with the true value is used as the optimization target:

[0036]

[0037] Through the loss function loss mse The new aerodynamic state s generated new Corresponding wind tunnel aerodynamic prediction data Correction is performed to obtain effective wind tunnel aerodynamic prediction data

[0038] According to another aspect of the present invention, a wind tunnel aerodynamic data prediction system is provided, comprising: a processor for executing the above-mentioned wind tunnel aerodynamic data prediction method based on the deep attention mechanism, a storage device for storing historical wind tunnel aerodynamic data and aerodynamic state characteristics, and a display device for displaying predicted aerodynamic data.

[0039] The processor is further configured to perform preprocessing on the input data, wherein the preprocessing includes cleaning, normalizing and feature extraction of the aerodynamic data.

[0040] As can be seen from the technical solutions provided by the embodiments of the present invention described above, they can mine the knowledge contained in historical wind test data and, based on this historical experience, assist in the generation of new aerodynamic state data. This invention can also reduce the frequency of wind tunnel testing or CFD simulations during various experimental design phases, shortening time and costs while improving efficiency.

[0041] Additional aspects and advantages of the present invention will be set forth in part in the following description, will be obvious from the following description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 A schematic diagram illustrating the implementation principle of a wind tunnel aerodynamic data prediction method based on a deep attention mechanism provided by an embodiment of the present invention;

[0044] Figure 2 This is a processing flow chart of a wind tunnel aerodynamic data prediction method based on a deep attention mechanism proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0046] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the description of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or couplings. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0047] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention pertains. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such herein, will not be interpreted in an idealized or overly formal sense.

[0048] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0049] A deep neural network (DNN) is an artificial neural network composed of multiple hidden layers, each containing multiple neurons. By layer-by-layer processing and nonlinear transformation of input data, DNNs are able to learn and extract complex data features and achieve excellent performance in tasks such as classification and regression. Due to their deep structure, DNNs excel at processing large amounts of data and complex pattern recognition tasks, making them one of the core technologies in the field of deep learning.

[0050] The Attention Mechanism is a deep neural network that mimics cognitive attention. The attention mechanism takes three input components: query, key, and value. By linearly mapping and calculating these components, it focuses attention on the most important parts of the data. Essentially, it shifts attention from global to local.

[0051] Computational Fluid Dynamics (CFD): is the science of using computers to predict the flow of liquids and gases based on the governing equations of conservation of mass, momentum, and energy.

[0052] Example 1

[0053] The implementation principle of the wind tunnel aerodynamic data prediction method based on the deep attention mechanism proposed in the embodiment of the present invention is as follows: Figure 1 As shown in Figure 3, the processing of this method includes two stages: training and inference. In the training stage, the algorithm uses training sample data for supervised learning; in the inference stage, the algorithm predicts the corresponding aerodynamic data based on the given historical aerodynamic data and the new aerodynamic state.

[0054] The processing flow of a wind tunnel aerodynamic data prediction method based on a deep attention mechanism proposed in an embodiment of the present invention is as follows: Figure 2 As shown, the processing steps include the following:

[0055] Step S10: Based on historical wind tunnel test aerodynamic data and corresponding aerodynamic states, an encoder is used to extract aerodynamic state features and aerodynamic characteristic features of the historical wind tunnel test aerodynamic data; wherein the aerodynamic state features are obtained by encoding state information of the experimental appearance (yaw angle, roll angle, Mach number, pitch rudder deflection, yaw rudder deflection, and roll rudder deflection), and the aerodynamic characteristic features are obtained by encoding the six-component force coefficients of the experimental appearance (lift, drag, side force, pitch moment, roll moment, and yaw moment).

[0056] Step S20: decoupling aerodynamic state features and aerodynamic characteristic features using an attention mechanism, and obtaining overall aerodynamic characteristics through multi-level fusion;

[0057] Step S30: obtaining a given new aerodynamic state, re-coupling the new aerodynamic state and the fused overall aerodynamic characteristics, and generating wind tunnel aerodynamic data corresponding to the new aerodynamic state through a decoder.

[0058] Step S40: predicting the aerodynamic characteristics of the aircraft in the new aerodynamic state based on the generated new wind tunnel aerodynamic data.

[0059] Specifically, the above step S10 includes: The problem solved by the present invention is defined as follows: Assume that a wind tunnel test consists of v-dimensional aerodynamic state information And the corresponding aerodynamic data results with dimension u Composition. The m historical wind tunnel test aerodynamic data of a given experimental shape And the aerodynamic state corresponding to the wind tunnel aerodynamic data X Data is used as a reference to generate a given aerodynamic state s new The wind tunnel aerodynamic data is given by . Assume that the prediction model of wind tunnel aerodynamic data is F, and the formal definition of the prediction model is as follows:

[0060]

[0061] in is the wind tunnel aerodynamic data predicted by the prediction model, and Θ is the prediction model parameter.

[0062] The model framework diagram of a wind tunnel aerodynamic data prediction method based on a deep attention mechanism proposed in an embodiment of the present invention is as follows: Figure 2 This invention utilizes an encoder-decoder architecture. The encoding portion decouples the aerodynamic state and aerodynamic characteristics. The decoding portion uses the recoupled aerodynamic state and aerodynamic characteristics to be predicted as input to generate new wind tunnel aerodynamic data. This design effectively mines aerodynamic knowledge from historical aerodynamic data and generates new aerodynamic data by leveraging the relationship between aerodynamic knowledge and aerodynamic state.

[0063] The encoding part consists of two parts: the encoder and the attention fusion module. The following describes these two modules in detail.

[0064] The encoder consists of a three-layer fully connected neural network. The encoder input includes known wind tunnel test aerodynamic data X and the corresponding aerodynamic state S. The encoder's main function is to mine the known historical aerodynamic data curves and extract the aerodynamic state characteristics and aerodynamic characteristics corresponding to each set of wind tunnel test aerodynamic data:

[0065] f i s , f i t =F(x i , s i θ en ), (2)

[0066] where f i s , f i t Respectively represent the aerodynamic state characteristics and aerodynamic characteristic characteristics corresponding to the known curve, θ en are encoder parameters.

[0067] For the aerodynamic state feature f extracted by the encoder i s The present invention further provides i s Reconstructed into aerodynamic state θ linear are the parameters of the fully connected network. Then, calculate With s i State consistency loss reg , as shown in formula (3).

[0068]

[0069] Through the state consistency loss reg So that from the aerodynamic state characteristics f i s Extracted from With the real aerodynamic state s i As close as possible. Thus ensuring f i s For aerodynamic state s i The effectiveness of feature encoding.

[0070] Specifically, the above step S20 includes:

[0071] The attention fusion module learns the overall characteristics of the aircraft by fusing the aerodynamic characteristics of known data, thereby achieving the decoupling of aerodynamic state and aerodynamic characteristics. For a given sample consisting of m known aerodynamic data curves, after passing through the encoder, the corresponding m aerodynamic characteristic features can be obtained. The present invention uses these aerodynamic characteristics As the Key and Value of the attention fusion module, k learnable tokens are used as the Query of the attention fusion module. In the attention fusion module, the aerodynamic characteristics can be analyzed through k queries. Then, the k extracted fusion features are spliced ​​to obtain the overall aerodynamic characteristics f of the final experimental shape. t :

[0072]

[0073] Among them, t i Represents the i-th fusion feature, θ token and θ attention They are the token parameters and the parameters of the attention layer. After attention fusion, the aerodynamic characteristics f representing the overall experimental shape are extracted. t , completing the decoupling of aerodynamic state and aerodynamic characteristics.

[0074] Specifically, the above step S30 includes:

[0075] 2. Decoding part

[0076] 2.1 Decoder

[0077] Like the encoder, the decoder is a three-layer fully connected network that uses the recoupled aerodynamic state to be predicted s new and the overall aerodynamic characteristics f t The features are used as input to generate the corresponding aerodynamic data. The decoder is formally expressed as follows:

[0078]

[0079] f=cat[f t , g(s new ,θ pos )], (6)

[0080] Among them, θ pos is the parameter of the fully connected neural network, g(s new θ pos ) of the new aerodynamic state s new Perform feature encoding to make it consistent with the aerodynamic state feature f extracted by the encoder i s Alignment, θde Describes the decoder parameters. cat is the abbreviation of "concatenate", which means concatenating two input features to obtain the merged features.

[0081] During the training phase, the aerodynamic data curve generated by the calculation The mean square error of the actual wind tunnel test value y is used as the optimization target:

[0082]

[0083] 3. Loss Function

[0084] As mentioned above, the present invention includes two parts of loss, namely the state consistency loss of the encoding part and the mean square error loss of the decoding part. The specific expression is as follows:

[0085] Loss(X, y) = loss mse +λloss reg , (8)

[0086] Among them, λ is a hyperparameter used to control the impact of the two loss functions on model training.

[0087] The inference process is basically the same as the training process. In the inference process, the input of the encoding part is m historical aerodynamic data and its corresponding aerodynamic state These data are first processed by the encoder and attention layers, which are combined with the new aerodynamic state s after the linear encoding layer. new The two inputs are fused and sent to the decoder to generate the final aerodynamic data prediction result.

[0088] This paper proposes a wind tunnel aerodynamic data prediction method based on a deep attention mechanism. This method effectively uses known aerodynamic data to predict wind tunnel aerodynamic data under given aerodynamic conditions. In practical applications, the predicted results achieved a mean absolute error of 2.62e-6, a mean square error of 4.04e-7, a symmetric mean absolute percentage error of 19.96%, and a Pearson correlation coefficient of 0.9972.

[0089] This model extracts high-dimensional representations of aerodynamic state and aerodynamic characteristics through an encoder, and then uses an attention fusion module to capture the aerodynamic characteristics of the entire experimental model, achieving decoupling of the aerodynamic state and aerodynamic characteristics. Finally, a decoder recouples the new aerodynamic state and the aerodynamic characteristics of the entire experimental model to generate new aerodynamic data. This method can efficiently and accurately generate new aerodynamic data, reducing the frequency of wind tunnel testing or CFD simulations during the design of various experimental models, reducing time costs and improving efficiency.

[0090] Example 2

[0091] An embodiment of the present invention also provides a wind tunnel aerodynamic data prediction system, comprising: a processor for executing the above-mentioned wind tunnel aerodynamic data prediction method based on the deep attention mechanism, a storage device for storing historical wind tunnel aerodynamic data and aerodynamic state characteristics, and a display device for displaying predicted aerodynamic data.

[0092] The processor is further used to pre-process the input data, including cleaning, normalization and feature extraction of the aerodynamic data, so as to improve the accuracy of the aerodynamic data prediction model.

[0093] In summary, the present invention proposes a wind tunnel aerodynamic data prediction algorithm based on a deep attention mechanism. This method utilizes artificial intelligence techniques to improve the efficiency and accuracy of aerodynamic data generation. The goal of this invention is to mine the aerodynamic knowledge contained in historical wind tunnel test data to assist in the design of various experimental models, reduce the frequency of wind tunnel testing and CFD simulations, and reduce both time and computational costs.

[0094] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0095] From the above description of the embodiments, it can be seen that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus the necessary general-purpose hardware platform. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0096] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without making any creative efforts.

[0097] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A wind tunnel aerodynamic data prediction method based on deep attention mechanism, characterized in that: include: Based on historical wind tunnel test aerodynamic data of a given experimental configuration and the corresponding aerodynamic state, extracting aerodynamic state features and aerodynamic characteristic features of the historical wind tunnel test aerodynamic data using an encoder; Using an attention mechanism to decouple the aerodynamic state features and aerodynamic characteristic features, and obtaining the overall aerodynamic characteristics through multi-level fusion; obtaining a new aerodynamic state of the given experimental configuration, coupling the new aerodynamic state with the overall aerodynamic characteristics, and generating wind tunnel aerodynamic data corresponding to the new aerodynamic state through a decoder; The aerodynamic characteristics of the aircraft in the new aerodynamic state of the given experimental shape are predicted based on the wind tunnel aerodynamic data of the new aerodynamic state.

2. The method according to claim 1, characterized in that The method of extracting aerodynamic state features and aerodynamic characteristic features of the historical wind tunnel test aerodynamic data and the corresponding aerodynamic state of the given experimental configuration using an encoder includes: Assume that a wind tunnel test consists of v-dimensional aerodynamic state information And the corresponding aerodynamic data results with dimension u Composition, obtain m historical wind tunnel test aerodynamic data of a given experimental shape And the aerodynamic state corresponding to the wind tunnel aerodynamic data X data, The encoding part includes an encoder and an attention fusion module. The encoder is composed of a three-layer fully connected neural network. The wind tunnel test aerodynamic data X and the corresponding aerodynamic state S are input to the encoder. The encoder performs feature mining on the curve of the wind tunnel test aerodynamic data X to extract the aerodynamic state features and aerodynamic characteristic features corresponding to each set of wind tunnel test aerodynamic data: in Represent the aerodynamic state characteristics and aerodynamic characteristics, θ en is the encoder parameter, F is the prediction model of wind tunnel aerodynamic data; The aerodynamic state characteristics Reconstructed into aerodynamic state θ linear is the parameter of the fully connected network, calculated With s i State consistency loss reg , as shown in formula (3); Through the state consistency loss reg From the aerodynamic state characteristics Extracted from With the real aerodynamic state s i As close as possible.

3. The method according to claim 2, characterized in that The use of the attention mechanism to decouple the aerodynamic characteristics and obtain the overall aerodynamic characteristics through multi-level fusion includes: After passing through the encoder, m aerodynamic characteristic features corresponding to m historical wind tunnel test aerodynamic data are obtained The m aerodynamic characteristics As the key and value of the attention fusion module, k learnable tokens are used as the query of the attention fusion module. In the attention fusion module, the aerodynamic characteristics are analyzed through k queries. Perform k different fusions, and splice the extracted k fusion features to obtain the overall aerodynamic characteristics f t : Among them, t i Represents the i-th fusion feature, θ token and θ attention They are the token parameters and the parameters of the attention layer respectively.

4. The method according to claim 3, characterized in that The step of obtaining a new aerodynamic state of the given experimental configuration, coupling the new aerodynamic state with the overall aerodynamic characteristics, and generating wind tunnel aerodynamic data corresponding to the new aerodynamic state through a decoder includes: Set the decoder to a three-layer fully connected network to obtain the new aerodynamic state s of the given experimental shape new , the new aerodynamic state s new and the overall aerodynamic characteristics f t Input to the decoder, and the decoder generates the new aerodynamic state s new Corresponding wind tunnel aerodynamic data The decoder process is as follows: f=cat[f t ,g(s new ;θ pos )],#(6) Among them, θ pos is the parameter of the fully connected neural network, g(s new θ pos ) is the new aerodynamic state s new Perform feature encoding to make it consistent with the pneumatic state features extracted by the encoder Alignment, θ de Represents the parameters of the decoder; cat represents the merged features obtained by concatenating the two input features; During the training phase, the aerodynamic data curve generated by the calculation The mean square error with the true value is used as the optimization target: Through the loss function loss mse The new aerodynamic state s generated new Corresponding wind tunnel aerodynamic prediction data Correction is performed to obtain effective wind tunnel aerodynamic prediction data 5. A wind tunnel aerodynamic data prediction system, characterized in that: include: A processor for executing the wind tunnel aerodynamic data prediction method based on a deep attention mechanism as described in any one of claims 1 to 4, a storage device for storing historical wind tunnel aerodynamic data and aerodynamic state characteristics, and a display device for displaying predicted aerodynamic data.

6. The wind tunnel aerodynamic data prediction system according to claim 5, characterized in that: The processor is further configured to perform preprocessing on the input data, wherein the preprocessing includes cleaning, normalizing and feature extraction of the aerodynamic data.

Images