A data augmentation and model training method, system and medium for three-dimensional aircraft aerodynamic coefficient prediction

Abstract

The present invention provides a data augmentation and model training method, system and medium for predicting aerodynamic coefficients of three-dimensional aircraft, relates to the field of rapid prediction of aerodynamic coefficients of aircraft, and solves the scarcity and high cost problems of high-quality CFD data in intelligent CFD; the method comprises: defining a basic aerodynamic data set including shape, working condition and aerodynamic coefficient, and forming an aerodynamic database with sampled point cloud data, working condition and aerodynamic coefficient; randomly sampling point cloud data to obtain augmented point cloud data, and forming an augmented aerodynamic database on the basis of the aerodynamic database; constructing an aerodynamic coefficient prediction neural network, and after pre-training, fine-tuning the neural network using the augmented aerodynamic database to obtain an intelligent model for predicting aerodynamic coefficients of three-dimensional aircraft; the present invention is a low-cost aerodynamic data augmentation solution, which utilizes general three-dimensional data, reduces the model's demand for aerodynamic data, and improves the performance of the intelligent model.

Description

Technical Field

[0001] The present invention relates to the field of rapid prediction of aircraft aerodynamic coefficients, and in particular to a data augmentation and model training method, system and medium for prediction of three-dimensional aircraft aerodynamic coefficients. Background Art

[0002] In recent years, artificial intelligence technology, with its powerful learning and reasoning capabilities, is gradually reshaping various industries and effectively solving many industry problems. Its influence has penetrated into the traditional field of computational fluid dynamics (CFD). Especially driven by large models, intelligent CFD, as an emerging research direction, aims to use artificial intelligence technology to optimize the fluid dynamics simulation process. It is divided into two core strategies: intelligent agent model and intelligent acceleration model. The former attempts to completely replace the traditional CFD process with intelligent models that are data-driven or integrated with physical principles, while the latter is committed to embedding AI models into CFD calculations to accelerate the convergence of the solution process.

[0003] Although intelligent CFD has shown great potential, its development still faces severe challenges, namely the scarcity and high cost of high-quality CFD data. Compared with data-intensive fields such as text and images, the generation of CFD data not only requires high precision, but also requires the simulation of complex physical processes. The calculation of a single case may take hours or even days, especially for the evaluation of aerodynamic performance of aircraft shapes under specific working conditions, the data acquisition cycle is long and the resource consumption is huge. This bottleneck directly restricts the training scale and quality of intelligent agent models and acceleration models, resulting in limited model generalization ability and poor prediction accuracy.

[0004] Most existing solutions rely on fast approximation methods or low-precision solvers to increase the amount of data. Although they can alleviate data hunger to a certain extent, such methods sacrifice the authenticity and reliability of data and cannot fundamentally meet the needs of intelligent CFD for high-quality data sets. Therefore, developing a pneumatic data generation and enhancement technology that is both economical and efficient and guarantees data accuracy has become an urgent need to promote the practicality and precision of intelligent CFD. Summary of the invention

[0005] Based on the current status of the background technology, the purpose of the present invention is to solve the scarcity and high cost of high-quality CFD data in intelligent CFD, and therefore proposes a data augmentation and model training method, system and medium for the prediction of aerodynamic coefficients of three-dimensional aircraft. The present invention is a low-cost aerodynamic data augmentation solution, which expands the existing aerodynamic database and then conducts new model training, utilizes general three-dimensional data, thereby reducing the model's demand for aerodynamic data and effectively improving the performance of the intelligent model.

[0006] The present invention adopts the following technical solutions to achieve the purpose:

[0007] A data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction, the method comprising the following steps:

[0008] S1. Define the basic aerodynamic data set including shape, working conditions, and aerodynamic coefficients;

[0009] S2. Based on the basic aerodynamic data set, the grid points of the three-dimensional aircraft are sampled to obtain point cloud data with a fixed number of points; the point cloud data, working conditions, and aerodynamic coefficients are combined to form an aerodynamic database for intelligent model training;

[0010] S3, randomly sampling the point cloud data with a fixed number of points to obtain augmented point cloud data;

[0011] S4, using the augmented point cloud data, based on the constructed aerodynamic database, together with the working conditions and aerodynamic coefficients, to construct an augmented aerodynamic database;

[0012] S5, constructing aerodynamic coefficient prediction neural network;

[0013] S6. Use a general three-dimensional point cloud data set to pre-train the aerodynamic coefficient prediction neural network, and then use the augmented aerodynamic database to fine-tune the aerodynamic coefficient prediction neural network after pre-training to obtain an intelligent model for the prediction of aerodynamic coefficients of three-dimensional aircraft.

[0014] Further, in step S2, The number of grid points on the wall of the aircraft is recorded as ; Set the number of sampling points , each sampling point corresponds to a sampling area, and the sampling point is the point closest to the center point of the corresponding sampling area; according to the number of sampling points and the number of wall grid points , calculate the The number of grid points in each sampling area for each aircraft , and based on this calculation method, each three-dimensional aircraft shape is obtained from the basic aerodynamic data set Upsampling to obtain sampling points; if the number of grid points in the last sampling area is insufficient , still sample one sampling point in the area to obtain the shape of each three-dimensional aircraft Point cloud data with a fixed number of points ,use Indicates Point cloud data obtained by sampling the wall of the aircraft The points;

[0015] The point cloud data of each 3D aircraft The load cases defined in the basic aerodynamic data set , the first aerodynamic coefficient Together they form a pneumatic database for intelligent model training, For the three-dimensional aircraft shape The total number of is the total number of working conditions; As a sample of the aerodynamic database, The point cloud representation of the aircraft and its Aerodynamic coefficients under various conditions .

[0016] Further, in step S3, according to the set number of sampling points and the number of grid points within each sampling area , in The correspondence of the aircraft shape In the sampling area, random sampling is performed to obtain random points ; Perform one operation in each sampling area to obtain a random point The sampling operation is thus performed After that, we get the point cloud data Different sampling data , complete a random sampling process; repeat the random sampling process Second, get the target The shape of the aircraft corresponds to Different random sampling data , that is, the augmented point cloud data.

[0017] Specifically, in step S5, the aerodynamic coefficient prediction neural network includes a point cloud shape coding network, an operating condition coding network and an aerodynamic coefficient prediction head; the point cloud shape coding network is used for point cloud data processing, the operating condition coding network is used to encode different operating conditions, and the aerodynamic coefficient prediction head is used to predict the aerodynamic coefficient with the parameter support of the point cloud shape coding network and the operating condition coding network.

[0018] Preferably, in step S6, first, an additional training encoding network that is the same as the point cloud shape encoding network is obtained, and a decoding network is connected after the training encoding network to form an autoencoder including an encoding part and a decoding part; the autoencoder is unsupervisedly trained using a general three-dimensional point cloud data set to obtain a pre-trained autoencoder; the parameters of the encoding part in the pre-trained autoencoder are copied to the point cloud shape encoding network of the aerodynamic coefficient prediction neural network to complete the pre-training of the aerodynamic coefficient prediction neural network.

[0019] The present invention also provides a computer device system, including a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the aforementioned data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction.

[0020] The present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the aforementioned data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction.

[0021] In summary, due to the adoption of this technical solution, the beneficial effects of the present invention are as follows:

[0022] The present invention abandons the traditional path of generating new CFD numerical calculation data, shortens the simulation cycle, and reduces the economic and time costs of data augmentation. This not only accelerates the research and development process, but also makes it possible to build large-scale data sets, providing strong support for the deep learning of intelligent CFD models.

[0023] Through the data augmentation method of the present invention, even different point cloud samples of the same shape can show rich changes, effectively avoiding the problem of data duplication in the sample library, and significantly enhancing the diversity and authenticity of the aerodynamic data set. This diversity is a key factor in improving the generalization ability of the model and the prediction accuracy.

[0024] The present invention introduces a pre-training strategy, using widely available 3D shape public data sets to preliminarily train the point cloud shape encoding network, which is not only low-cost but also significantly enhances the encoder's ability to capture shape features. Subsequently, targeted fine-tuning is performed in combination with actual aerodynamic data sets to ensure that the model can accurately learn the subtle differences in aerodynamic characteristics, thereby demonstrating superior performance in the final prediction task. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the overall process of the method of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] Example 1

[0029] A data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction. Figure 1 The method is briefly described in the following figure, which can be viewed simultaneously. The overall steps of the method are summarized as follows:

[0030] S1. Define the basic aerodynamic data set including shape, working conditions, and aerodynamic coefficients;

[0031] S2. Based on the basic aerodynamic data set, the grid points of the three-dimensional aircraft are sampled to obtain point cloud data with a fixed number of points; the point cloud data, working conditions, and aerodynamic coefficients are combined to form an aerodynamic database for intelligent model training;

[0032] S3, randomly sampling the point cloud data with a fixed number of points to obtain augmented point cloud data;

[0033] S4, using the augmented point cloud data, based on the constructed aerodynamic database, together with the working conditions and aerodynamic coefficients, to construct an augmented aerodynamic database;

[0034] S5, constructing aerodynamic coefficient prediction neural network;

[0035] S6. Use a general three-dimensional point cloud data set to pre-train the aerodynamic coefficient prediction neural network, and then use the augmented aerodynamic database to fine-tune the aerodynamic coefficient prediction neural network after pre-training to obtain an intelligent model for the prediction of aerodynamic coefficients of three-dimensional aircraft.

[0036] This embodiment will follow the above steps to introduce in detail how to obtain the specific content of the augmented pneumatic database. This embodiment is applied to the CFD field and can mainly realize the expansion of the existing pneumatic database. In this field, some fast engineering methods or low-precision solvers are usually used to calculate new data and expand the database. There is no low-cost data augmentation method for pneumatic data, so this embodiment fills the gap in this aspect.

[0037] In step S1, The three-dimensional aircraft shape is recorded as ,Will Different working conditions are recorded as ; For each three-dimensional aircraft shape, numerical calculations are performed under each different working condition to obtain The first aerodynamic coefficient is set as ;use Indicates The aircraft is in The aerodynamic coefficient samples under various working conditions are: As a sample in the basic aerodynamic data set, the definition of the basic aerodynamic data set is completed; in the basic aerodynamic data set, there are samples.

[0038] In this field, three-dimensional aircraft are generally represented by grids, usually in a point-surface manner. Since the grid points and grid cells on each aircraft surface are different, but the input length of the intelligent model needs to be fixed, grid point sampling is required for each shape.

[0039] In step S2, The number of grid points on the wall of the aircraft is recorded as ; Set the number of sampling points , each sampling point corresponds to a sampling area, and the sampling point is the point closest to the center point of the corresponding sampling area; according to the number of sampling points and the number of wall grid points , calculate the The number of grid points in each sampling area for each aircraft , and based on this calculation method, each three-dimensional aircraft shape is obtained from the basic aerodynamic data set Upsampling to obtain sampling points; if the number of grid points in the last sampling area is insufficient , still sample one sampling point in the area to obtain the shape of each three-dimensional aircraft Point cloud data with a fixed number of points ,use Indicates Point cloud data obtained by sampling the wall of the aircraft The points;

[0040] The point cloud data of each 3D aircraft The load cases defined in the basic aerodynamic data set , the first aerodynamic coefficient Together they form a pneumatic database for intelligent model training, For the three-dimensional aircraft shape The total number of is the total number of working conditions; As a sample of the aerodynamic database, The point cloud representation of the aircraft and its Aerodynamic coefficients under various conditions .

[0041] Then in step S3, according to the set number of sampling points and the number of grid points within each sampling area , in The correspondence of the aircraft shape In the sampling area, random sampling is performed to obtain random points ; Perform one operation in each sampling area to obtain a random point The sampling operation is thus performed After that, we get the point cloud data Different sampling data , complete a random sampling process; repeat the random sampling process Second, get the target The shape of the aircraft corresponds to Different random sampling data , that is, the augmented point cloud data.

[0042] Since in the specific content of step S3, for the The aircraft shape is obtained Different augmented point clouds, which still represent the The aerodynamic coefficient of the aircraft is theoretically constant. The shape of the aircraft corresponds to Different random sampling data ,and Different working conditions Combined, we get Group 2 aerodynamic coefficient ; According to random sampling data, working conditions and second aerodynamic coefficient, Expressed as The first aircraft The augmented point cloud is obtained by combining the Working conditions and corresponding aerodynamic coefficients The sample completes a combined operation process. It should be noted that in the second aerodynamic coefficient, , , Until The meanings are the same as those defined in step S1. Similarly, it can be used to represent The sample in The aerodynamic performance under various working conditions; the combined operation here also achieves the purpose of augmenting and expanding the data.

[0043] Repeat this combined operation process times, The value ranges from 1 to , thus obtaining The augmented aerodynamic performance dataset of aircraft contains samples of , which constitutes the expanded aerodynamic database. After the operation of step S4, it can be seen that the number of samples has been expanded. times.

[0044] In summary, the above process of this embodiment overcomes the core problems of data scarcity and high cost in aerodynamic data analysis, and can be applied to provide data support in subsequent model training, laying a technical foundation for the practical application of intelligent CFD technology.

[0045] Example 2

[0046] Based on Example 1, this example specifically introduces the contents related to the aerodynamic coefficient prediction neural network in the method.

[0047] In step S5, the aerodynamic coefficient prediction neural network includes a point cloud shape coding network, an operating condition coding network and an aerodynamic coefficient prediction head; the point cloud shape coding network is used for point cloud data processing, the operating condition coding network is used to encode different operating conditions, and the aerodynamic coefficient prediction head is used to predict the aerodynamic coefficient with the parameter support of the point cloud shape coding network and the operating condition coding network.

[0048] Based on the constructed network structure, the model training of step S6 is performed. First, an additional training encoding network identical to the point cloud shape encoding network is obtained, and a decoding network is connected after the training encoding network to form an automatic encoder including an encoding part and a decoding part.

[0049] The autoencoder is unsupervisedly trained using a general 3D point cloud data set (this data set is a public data set, and various types can be easily obtained without difficulty for those skilled in the art; this data set is not an aerodynamic database, and it only includes shape point cloud data, and no aerodynamic coefficient and other related data) to obtain a pre-trained autoencoder. The specific method of unsupervised training is a mature technology in this field, and this embodiment will not be repeated here.

[0050] The parameters of the encoding part in the pre-trained automatic encoder are copied to the point cloud shape encoding network of the aerodynamic coefficient prediction neural network to complete the pre-training of the aerodynamic coefficient prediction neural network.

[0051] After completing the pre-training, the augmented aerodynamic performance data set of the aircraft in the augmented aerodynamic database is input into the aerodynamic coefficient prediction neural network. The aerodynamic coefficient prediction head is fine-tuned in combination with the point cloud shape encoding network and the working condition encoding network to realize the formal training of the aerodynamic coefficient prediction neural network. After the formal training is completed, an intelligent model for the prediction of aerodynamic coefficients of three-dimensional aircraft is obtained and applied to the aerodynamic coefficient prediction process.

[0052] In summary, through intelligent pre-training and fine-tuning mechanisms, combined with the detailed data provided in the augmented aerodynamic database, the accuracy and efficiency of the intelligent model can be greatly improved, which also brings innovative effects to the practical application of intelligent CFD technology.

[0053] Example 3

[0054] Based on the above embodiments, this embodiment provides a computer device system, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to implement the steps of the data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction described in Example 1 or Example 2.

[0055] At the same time, this embodiment also provides a computer-readable storage medium on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction described in Example 1 or Example 2 are implemented.

[0056] In this embodiment, these computer programs / instructions may be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more steps in the method.

Claims

1. A data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction, characterized in that: The method comprises the following steps: S1. Define the basic aerodynamic data set including shape, working conditions, and first aerodynamic coefficient; S2. Based on the basic aerodynamic data set, grid point sampling is performed on the three-dimensional aircraft to obtain point cloud data with a fixed number of points; the point cloud data, the working condition, and the first aerodynamic coefficient are used together to form an aerodynamic database for intelligent model training; S3, randomly sampling the point cloud data with a fixed number of points to obtain augmented point cloud data; S4, using the augmented point cloud data, based on the constructed aerodynamic database, together with the second aerodynamic coefficient and the working conditions in the basic aerodynamic data set, to construct an augmented aerodynamic database; S5, constructing aerodynamic coefficient prediction neural network; S6. Pre-training the aerodynamic coefficient prediction neural network using a general three-dimensional point cloud data set that only includes shape point cloud data, and then fine-tuning the aerodynamic coefficient prediction neural network after pre-training using the augmented aerodynamic database to obtain an intelligent model for aerodynamic coefficient prediction of three-dimensional aircraft; In step S1, The three-dimensional aircraft shape is recorded as ,Will Different working conditions are recorded as ; For each three-dimensional aircraft shape, numerical calculations are performed under each different working condition to obtain The first aerodynamic coefficient is set as ; In step S2, The number of grid points on the wall of the aircraft is recorded as ; Set the number of sampling points , each sampling point corresponds to a sampling area, and the sampling point is the point closest to the center point of the corresponding sampling area; according to the number of sampling points and the number of wall grid points , calculate the The number of grid points in each sampling area for each aircraft , and based on this calculation method, each three-dimensional aircraft shape is obtained from the basic aerodynamic data set Upsampling to obtain sampling points; if the number of grid points in the last sampling area is insufficient , still sample one sampling point in the area to obtain the shape of each three-dimensional aircraft Point cloud data with a fixed number of points ,use Indicates Point cloud data obtained by sampling the wall of the aircraft The points to complete the grid point sampling of the three-dimensional aircraft; In step S4, The shape of the aircraft corresponds to Different random sampling data ,and Different working conditions Combined, we get Group 2 aerodynamic coefficient .

2. The data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction according to claim 1, characterized in that: In step S1, use Indicates The aircraft is in The aerodynamic coefficient samples under various working conditions are: As a sample in the basic aerodynamic data set, the definition of the basic aerodynamic data set is completed; in the basic aerodynamic data set, there are samples.

3. The data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction according to claim 1, characterized in that: The point cloud data of each 3D aircraft The load cases defined in the basic aerodynamic data set , the first aerodynamic coefficient Together they form a pneumatic database for intelligent model training, The three-dimensional aircraft shape The total number of is the total number of working conditions; As a sample of the aerodynamic database, The point cloud representation of the aircraft and its Aerodynamic coefficients under various conditions .

4. The data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction according to claim 3, characterized in that: In step S3, according to the set number of sampling points and the number of grid points within each sampling area , in The correspondence of the aircraft shape In the sampling area, random sampling is performed to obtain random points ; Perform one operation in each sampling area to obtain a random point The sampling operation is thus performed After that, we get the point cloud data Different sampling data , complete a random sampling process; repeat the random sampling process Second, get the target The shape of the aircraft corresponds to Different random sampling data , that is, the augmented point cloud data.

5. The data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction according to claim 4, characterized in that: According to the random sampling data, working conditions and the second aerodynamic coefficient, Expressed as The first aircraft The augmented point cloud is obtained by combining the Working conditions and corresponding aerodynamic coefficients samples, complete a combined operation process; repeat the combined operation process times, The value ranges from 1 to , thus obtaining The augmented aerodynamic performance dataset of aircraft contains samples of , which constitutes the augmented aerodynamic database.

6. The data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction according to claim 1, characterized in that: In step S5, the aerodynamic coefficient prediction neural network includes a point cloud shape coding network, an operating condition coding network and an aerodynamic coefficient prediction head; the point cloud shape coding network is used for point cloud data processing, the operating condition coding network is used to encode different operating conditions, and the aerodynamic coefficient prediction head is used to predict the aerodynamic coefficient with the parameter support of the point cloud shape coding network and the operating condition coding network.

7. The data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction according to claim 6, characterized in that: In step S6, first, an additional training encoding network that is the same as the point cloud shape encoding network is obtained, and a decoding network is connected after the training encoding network to form an autoencoder including an encoding part and a decoding part; the autoencoder is unsupervisedly trained using a general three-dimensional point cloud data set to obtain a pre-trained autoencoder; the parameters of the encoding part in the pre-trained autoencoder are copied to the point cloud shape encoding network of the aerodynamic coefficient prediction neural network to complete the pre-training of the aerodynamic coefficient prediction neural network.

8. The data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction according to claim 6, characterized in that: In step S6, after pre-training is completed, the augmented aerodynamic performance data set of the aircraft in the augmented aerodynamic database is input into the aerodynamic coefficient prediction neural network, and the aerodynamic coefficient prediction head is fine-tuned in combination with the point cloud shape encoding network and the working condition encoding network to realize the formal training of the aerodynamic coefficient prediction neural network; after the formal training is completed, an intelligent model for the prediction of aerodynamic coefficients of three-dimensional aircraft is obtained, which is applied to the aerodynamic coefficient prediction process.

9. A computer system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction as described in claim 1.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the data augmentation and model training method for three-dimensional aircraft aerodynamic coefficient prediction described in claim 1 are implemented.

Images