Rocket aircraft flow field correction method and system based on multi-fidelity data fusion

Abstract

The invention provides a rocket aircraft flow field correction method based on multi-fidelity data fusion, and the method comprises the following steps: 1, collecting low-fidelity data, and carrying out the construction of a low-fidelity data set: simulating the change process of an unsteady flow field around an object in a finite time step through a Reynolds time-average simulation method or a large vortex simulation method; 2, data correlation analysis is carried out, wherein a linear relation model yH = rho (x) yL + delta (x) of a non-viscous flow field yL and a viscous flow field yH of the rocket aircraft is established; step 3, neural network architecture design: constructing a low-fidelity data approximation network NNL; 4, performing hyper-parameter learning and optimization; defining a loss function; 5, performing data acquisition and preprocessing: performing high-fidelity viscous flow field data acquisition, and preprocessing the acquired high-fidelity and low-fidelity flow field data; and step 6, model training and verification.

Description

Technical Field

[0001] This invention relates to the field of rocket vehicle design technology, and in particular to a method and system for correcting the flow field of rocket vehicles based on multi-fidelity data fusion. Background Technology

[0002] Flow field correction methods mainly involve adjusting and optimizing flow field simulation results in computational fluid dynamics (CFD) to improve simulation accuracy. Among these methods, multi-fidelity data fusion technology combines high-fidelity and low-fidelity data and utilizes neural networks to learn the relationships between them, thereby improving the model's prediction accuracy and computational efficiency.

[0003] Data fusion technology involves modeling and integrating data from different sources. By making full use of and rationally allocating data from various measurement methods, and utilizing data estimation, modeling, and data acquisition management, redundant or complementary information from these measurement results in time or space is integrated according to certain criteria to generate more accurate information, more consistent interpretations or descriptions of the measured object, and obtain the optimal result with the highest credibility.

[0004] There are many methods for studying the relationships between high- and low-precision data, such as chaotic polynomial expansion and Gaussian process regression, which have been widely used in engineering. However, these methods have certain limitations, such as only being able to estimate linear relationships between data, dimensionality explosion, and computational difficulties. Neural networks, on the other hand, can effectively solve these problems.

[0005] The accuracy of deep learning modeling relies heavily on large amounts of high-fidelity data. High-fidelity and low-fidelity are two important concepts used to describe data quality, accuracy, and level of detail. They are commonly used to evaluate the reliability and usability of datasets, especially in data preprocessing, model building, and decision support systems. High-fidelity data primarily originates from the viscous flow fields of rocket vehicles, while low-fidelity data mainly comes from the inviscid flow fields of rocket vehicles. In viscous flow fields, viscous forces significantly influence the flow characteristics of the fluid, including boundary layer formation, turbulence, and eddies; high-fidelity data can capture these complex phenomena. In inviscid flow fields, neglecting viscous forces, the Navier-Stokes equations simplify to the Euler equations, making it possible to obtain analytical solutions, which is very useful for theoretical analysis and teaching.

[0006] In practical applications, high-fidelity data is often scarce and expensive, while low-fidelity data is inexpensive and plentiful. Multi-fidelity data fusion technology aims to solve this problem by fusing data of different fidelities, maintaining high modeling accuracy even when high-fidelity data is scarce.

[0007] Multifidelity deep neural networks can adaptively approximate linear or nonlinear relationships between multifidelity data, reducing dependence on regularization coefficients and improving model robustness. This type of model improves the loss function, enabling it to better handle correlations between multifidelity data, whether linear or nonlinear.

[0008] Therefore, in order to try to solve the above problems, it is necessary to invent a method and system for correcting the flow field of rocket spacecraft based on multi-fidelity data fusion. Summary of the Invention

[0009] This invention aims to address the problems of high cost and low efficiency in calculating the viscous flow field of rockets using traditional CFD methods. It provides a method and system for correcting the flow field of rockets based on multi-fidelity data fusion, which improves computational efficiency and accuracy and helps optimize rocket design.

[0010] To address the aforementioned technical problems, this invention provides a method for correcting the flow field of rocket vehicles based on multi-fidelity data fusion, comprising the following steps:

[0011] Step 1: Collect low-fidelity data and construct a low-fidelity dataset:

[0012] The unsteady flow field changes around an object within a finite time step are simulated using Reynolds time-averaged simulation or large eddy simulation, generating the same amount of flow field information as the simulation time step, and constructing the input dataset for a deep neural network.

[0013] Step 2, perform data correlation analysis:

[0014] Establishing the inviscid flow field of the rocket vehicle L With viscous flow field y H linear relationship model y H =ρ(x)y L +δ(x), through statistical analysis of a large amount of flow field data from different parts of the rocket, the scaling factor ρ(x) and error term δ(x) are determined; a generalized autoregressive scheme y is introduced. H =αF l (x H ,y L )+(1-α)F nl (x H ,y L ), α∈[0,1], and based on the characteristics of different stages of rocket flight and flow field regions, a linear mapping F is constructed. l and nonlinear mapping F nl And determine the range of values ​​for the hyperparameter α;

[0015] Step 3, Neural Network Architecture Design:

[0016] Constructing a low-fidelity data approximation network (NN) L Design a deep neural network NN with multiple hidden layers. L (x L ,θ), using LeakyReLU or Swish activation functions, and combining L1 and L2 regularization methods, with weight initialization using He or Xavier initialization methods;

[0017] Design a linear dependence approximation network. It contains 2-4 hidden layers, each with 15-25 neurons. The appropriate activation function is selected based on the linear relationship characteristics, and the weights are initialized using a random normal distribution or a uniform distribution.

[0018] Establish a nonlinear correlation approximation network Set up 4-8 hidden layers, with 20-40 neurons in each layer, and use the Swish or Tanh activation function, L2 regularization and He initialization method;

[0019] Step 4, Hyperparameter learning and optimization:

[0020] Define loss function The Adam or Adagrad adaptive learning rate optimization algorithm is adopted, combined with the momentum term to accelerate convergence. The learning rate is dynamically adjusted according to the gradient changes during the training process. Gradient truncation technique is used to prevent gradient explosion, and early stopping method is used to avoid overfitting.

[0021] Step 5, Data Acquisition and Preprocessing:

[0022] High-fidelity viscous flow field data was acquired, and high-precision CFD methods were used to simulate and calculate the flow field under various flight conditions for different rocket shapes. Adaptive or fine mesh technology was used to obtain detailed flow field parameter distributions as high-fidelity samples. Low-fidelity inviscid flow field data was acquired, and inviscid Euler equation solvers were used to calculate the inviscid flow field under the same rocket shape and flight conditions. A relatively coarse mesh was used to improve the calculation speed, and the data obtained were used as low-fidelity samples.

[0023] The collected high-fidelity and low-fidelity flow field data are preprocessed to remove outliers and erroneous data. Based on the characteristics of the flow field parameters, an appropriate normalization method is selected to map the data to a specific interval.

[0024] Step 6, Model Training and Validation:

[0025] The preprocessed high- and low-fidelity flow field data samples are divided into training, validation, and test sets. The model is trained using mini-batch gradient descent and combined with random deactivation and batch standardization techniques to improve the network's generalization ability and training stability. The number of training iterations, learning rate, and other hyperparameters are dynamically adjusted based on the loss function values ​​on the validation set.

[0026] Multiple error metrics were selected to evaluate the model, and the differences between the model's prediction results on the validation and test sets and the actual viscous flow field data were analyzed. The performance of the model under different flight conditions and rocket shapes was compared.

[0027] In a preferred embodiment, in step 3, the low-fidelity data approximation network NN L It contains 6-10 hidden layers, and the number of neurons in each layer is reasonably set according to the data complexity and computing resources, with the range of 30-50 neurons per layer.

[0028] In a preferred embodiment, in step 3, the value of L1 is 0.001-0.01, and the value of L2 is 0.002-0.02.

[0029] In a preferred embodiment, in the defined loss function,

[0030] The regularization parameter λ should be selected reasonably based on the data size and model complexity, and the range of the regularization parameter λ is 0.005-0.05.

[0031] In a preferred embodiment, in step 6, the preprocessed high-fidelity and low-fidelity flow field data samples are divided into training set, validation set and test set according to the proportions of 70%-85%, 10%-20%, and 5%-10%, respectively.

[0032] In a preferred embodiment, in step 6, the probability value of the random inactivation is 0.1-0.3.

[0033] In a preferred embodiment, the error evaluation model includes root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 ), relative error.

[0034] The present invention also provides a rocket flow field correction system based on multi-fidelity data fusion, which includes a data acquisition module, a data correlation analysis module, a neural network architecture design module, and a hyperparameter learning and optimization module.

[0035] The data acquisition module is used to acquire low-fidelity data, simulate the unsteady flow field change process around the object within a finite time step using the Reynolds time-averaged simulation method or the large eddy simulation method, generate the same amount of flow field information as the simulation time step, and construct the input dataset for the deep neural network.

[0036] The data correlation analysis module is used to establish the inviscid flow field y of the rocket vehicle. L With viscous flow field y H linear relationship model y H=ρ(x)y L +δ(x), through statistical analysis of a large amount of flow field data from different parts of the rocket, the scaling factor ρ(x) and error term δ(x) are determined and incorporated into the generalized autoregressive scheme y H =αF l (x H ,y L )+

[0037] (1-α)F nl (x H ,y L ), α∈[0,1], and based on the characteristics of different stages of rocket flight and flow field regions, a linear mapping F is constructed. l and nonlinear mapping F nl And determine the range of values ​​for the hyperparameter α;

[0038] The neural network architecture design module is used to construct a low-fidelity data approximation network (NN). L Design a deep neural network NN with multiple hidden layers. L (x L The network employs LeakyReLU or Swish activation functions, combined with L1 and L2 regularization, and uses He or Xavier initialization methods for weight initialization. A linearly dependent approximation network is designed. It contains 2-4 hidden layers, each with 15-25 neurons. An appropriate activation function is selected based on the linear relationship characteristics, and the weights are initialized using a random normal distribution or a uniform distribution. A nonlinear correlation approximation network is established. Set up 4-8 hidden layers, with 20-40 neurons in each layer, and use the Swish or Tanh activation function, L2 regularization and He initialization method;

[0039] The hyperparameter learning and optimization module is used to define the loss function.

[0040] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0041] 1. This flow field correction method and system can be used to generate sample datasets needed to build neural networks; build and train deep learning neural network models based on the sample datasets; and use the built deep neural network for viscous flow field correction in rockets. Flow field correction addresses the problems of high cost and low efficiency in calculating the viscous flow field of rockets using traditional CFD methods, improving computational efficiency and accuracy, and assisting in rocket design optimization.

[0042] 2. Significantly improved computational efficiency: By fusing low-fidelity inviscid flow field data, the reliance on high-fidelity viscous flow field calculations is reduced, significantly shortening computation time and improving the iterative efficiency of rocket spacecraft design optimization.

[0043] 3. The calculation accuracy meets the requirements: The model can accurately predict the viscous flow field parameters of the rocket. Compared with the high-fidelity CFD calculation results, the errors of parameters such as velocity and pressure are within an acceptable range, providing a reliable basis for rocket performance analysis and design.

[0044] 4. Wide adaptability: Applicable to different rocket shapes and various flight conditions, it can effectively correct viscous flow fields in different speed ranges such as subsonic, transonic, and supersonic, as well as at different angles of attack and flight altitudes, providing support for the design optimization of the entire flight profile of rocket vehicles.

[0045] 5. High engineering application value: It provides accurate flow field data for multidisciplinary design optimization of rockets, such as aerodynamic performance optimization, thermal protection design, and structural strength analysis, which helps to improve the overall performance and reliability of rockets and reduce research and development costs. Attached Figure Description

[0046] Figure 1 This is a diagram of the subsonic cruise state of a rocket in Embodiment 1 of the present invention. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0048] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "top / bottom," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0049] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed", "equipped", "sleeved / connected", "connected", etc., should be interpreted broadly. For example, "connection" can be a wall-mounted connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0050] This embodiment provides a method for correcting the flow field of a rocket vehicle based on multi-fidelity data fusion, including the following steps:

[0051] Step 1: Collect low-fidelity data and construct a low-fidelity dataset: Simulate the unsteady flow field changes around the object within a finite time step using the Reynolds time-averaged simulation method or the large eddy simulation method, generate the same amount of flow field information as the simulation time step, and construct the input dataset for the deep neural network.

[0052] Step 2, perform data correlation analysis: establish the inviscid flow field y of the rocket spacecraft L With viscous flow field y H linear relationship model y H =ρ(x)y L +δ(x), through statistical analysis of a large amount of flow field data from different parts of the rocket (such as the cylindrical section, conical section, head, nozzle, etc.), the scaling factor ρ(x) and error term δ(x) are determined, providing a basis for subsequent model construction;

[0053] Introducing a generalized autoregressive scheme y H =αF l (x H ,y L )+(1-α)F nl (x H ,y L ), α∈[0,1], where y L For low-fidelity inviscid flow field data, y H This is for high-fidelity viscous flow data. F() represents the correlation function between the high-fidelity and low-fidelity data, which can be linear or nonlinear. Based on the characteristics of different stages of rocket flight (e.g., takeoff, acceleration, cruise, reentry) and flow field regions (e.g., boundary layer, shock wave region, wake region), a linear mapping F is constructed. l and nonlinear mapping F nl And determine the range of values ​​for the hyperparameter α in order to accurately capture the nonlinear characteristics of the flow field.

[0054] Step 3, Neural Network Architecture Design: Constructing a Low-Fidelity Data Approximation Network (NN) LDesign a deep neural network NN with multiple hidden layers. L (x L The network employs activation functions such as LeakyReLU or Swish to enhance nonlinear expressive power, and uses a combination of L1 and L2 regularization to prevent overfitting. The regularization coefficients are adjusted according to the actual situation (e.g., L1 coefficient is 0.001-0.01, and L2 coefficient is 0.002-0.02). The weights are initialized using He or Xavier initialization methods to ensure the stability and efficiency of network training.

[0055] In this embodiment, the low-fidelity data approximation network NN L It contains 6-10 hidden layers, and the number of neurons in each layer is reasonably set according to the data complexity and computing resources, with the range of 30-50 neurons per layer.

[0056] Design a linear dependence approximation network. It contains 2-4 hidden layers, each with 15-25 neurons. An appropriate activation function (such as a linear function or no activation function) is selected based on the characteristics of the linear relationship. The weights are initialized using a random normal distribution or a uniform distribution (the standard deviation or range is determined according to the data characteristics) to accurately learn the linear relationship between high-fidelity and low-fidelity data.

[0057] Establish a nonlinear correlation approximation network Set up 4-8 hidden layers, with 20-40 neurons in each layer. Use activation functions such as Swish or Tanh to handle nonlinear mappings. Apply L2 regularization (L2 coefficient value is 0.001-0.01) and He initialization method to improve the network's ability to capture complex nonlinear flow field characteristics.

[0058] Step 4, Hyperparameter learning and optimization:

[0059] Define loss function in, Based on the data size and model complexity, a reasonable regularization parameter λ (such as 0.005-0.05) should be selected to balance the model's fitting ability and generalization ability.

[0060] The Adam or Adagrad adaptive learning rate optimization algorithm is used, combined with a momentum term (such as momentum coefficient 0.8-0.99) to accelerate convergence. The learning rate is dynamically adjusted according to the gradient changes during training (such as exponential decay or step-size-based decay strategy). Gradient truncation technique is used to prevent gradient explosion. Early stopping method is used (such as stopping training if the loss on the validation set does not decrease after 5-15 consecutive iterations) to avoid overfitting and ensure model performance.

[0061] Step 5, Data Acquisition and Preprocessing:

[0062] High-fidelity viscous flow field data were acquired, and high-precision CFD methods (such as the finite volume method combined with advanced turbulence models like Spalart-Al lmaras and k-ωSST) were used to simulate different rocket shapes (such as different nose cone shapes, body lengths, diameter ratios, and nozzle designs) under various flight conditions (Mach numbers 0.5-10.0, angles of attack -10°-20°, flight altitudes 0-120km, etc.). Adaptive or refined meshing techniques were employed to refine the mesh locally in key areas such as the rocket surface boundary layer, the nose cone shock wave region, and the nozzle interior, obtaining detailed distributions of flow field parameters such as velocity, pressure, temperature, density, and viscous stress as high-fidelity samples.

[0063] Low-fidelity inviscid flow field data is obtained, and the inviscid flow field under the same rocket shape and flight conditions is calculated using the inviscid Euler equation solver. A relatively coarse grid is used to improve the calculation speed, and data such as velocity and pressure are obtained as low-fidelity samples to provide supplementary information for the model and reduce the calculation cost.

[0064] The collected high-fidelity and low-fidelity flow field data are preprocessed to clean the data, removing outliers (such as velocity exceeding the reasonable range, negative pressure values, etc.) and erroneous data to ensure the physical rationality of the data. Based on the characteristics of the flow field parameters, an appropriate normalization method (such as linear normalization, logarithmic normalization, etc.) is selected to map the data to a specific interval (such as [-1,1] or [0,1]), improving the training efficiency and stability of the neural network.

[0065] Step 6, Model Training and Validation:

[0066] Dataset partitioning: The preprocessed high-fidelity and low-fidelity flow field data samples are divided into training set, validation set and test set in proportions of 70%-85%, 10%-20%, and 5%-10%, respectively. This ensures that the training set is large enough to learn the model parameters, the validation set is used to adjust hyperparameters and monitor model performance, and the test set is used to objectively evaluate the accuracy and generalization ability of the model.

[0067] Training process: The model is trained using mini-batch gradient descent (the number of samples in each batch is selected according to computing resources and data characteristics, such as 64-256). This is combined with random deactivation (the probability of random deactivation is 0.1-0.3) and batch standardization techniques to improve the network's generalization ability and training stability. The number of training iterations, learning rate, and other hyperparameters are dynamically adjusted based on the loss function value on the validation set to ensure that the model is continuously optimized during training.

[0068] Choose the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2The model is evaluated using various error metrics, including relative error, to analyze the differences between the model's predictions on validation and test sets and real viscous flow field data. These error metrics quantify and assess the model's accuracy and stability. Furthermore, the model's performance is compared under different flight conditions and rocket shapes to deeply analyze its strengths and limitations, providing a basis for model improvement.

[0069] First Embodiment

[0070] This embodiment provides a correction system for a rocket flow field correction method based on multi-fidelity data fusion. The correction system includes a data acquisition module, a data correlation analysis module, a neural network architecture design module, and a hyperparameter learning and optimization module.

[0071] The data acquisition module is used to acquire low-fidelity data, simulate the unsteady flow field change process around the object within a finite time step using the Reynolds time-averaged simulation method or the large eddy simulation method, generate the same amount of flow field information as the simulation time step, and construct the input dataset for the deep neural network.

[0072] The data correlation analysis module is used to establish the inviscid flow field y of the rocket vehicle. L With viscous flow field y H linear relationship model y H =ρ(x)y L +δ(x), through statistical analysis of a large amount of flow field data from different parts of the rocket, the scaling factor ρ(x) and error term δ(x) are determined and incorporated into the generalized autoregressive scheme y H =αF l (x H ,y L )+

[0073] (1-α)F nl (x H ,y L ), α∈[0,1], and based on the characteristics of different stages of rocket flight and flow field regions, a linear mapping F is constructed. l and nonlinear mapping F nl And determine the range of values ​​for the hyperparameter α;

[0074] The neural network architecture design module is used to construct a low-fidelity data approximation network (NN). L Design a deep neural network NN with multiple hidden layers. L (x L The network employs LeakyReLU or Swish activation functions, combined with L1 and L2 regularization, and uses He or Xavier initialization methods for weight initialization. A linearly dependent approximation network is designed. It contains 2-4 hidden layers, each with 15-25 neurons. An appropriate activation function is selected based on the linear relationship characteristics, and the weights are initialized using a random normal distribution or a uniform distribution. A nonlinear correlation approximation network is established. Set up 4-8 hidden layers, with 20-40 neurons in each layer, and use the Swish or Tanh activation function, L2 regularization and He initialization method;

[0075] The hyperparameter learning and optimization module is used to define the loss function.

[0076] Second Embodiment

[0077] like Figure 1 Provides the subsonic cruise status of a certain type of rocket.

[0078] 1. Data preparation;

[0079] For a certain type of rocket operating under subsonic cruise conditions (Mach number 0.6, angle of attack 1.5°, flight altitude 5km), high-fidelity viscous flow field data was calculated using high-precision CFD software (such as ANSYS Fluent or OpenFOAM) combined with a k-ε turbulence model, while low-fidelity inviscous flow field data was obtained using an inviscid Euler equation solver.

[0080] The collected data is cleaned to remove obvious outliers (such as data points with speeds greater than 1000 m / s or less than -100 m / s), and then linear normalization is used to map parameters such as speed and pressure to the interval [-1, 1].

[0081] 2. Model building and training;

[0082] Based on the aforementioned data correlation analysis, neural network architecture design, and hyperparameter learning and optimization methods, a multi-fidelity neural network model is constructed. Among these, a low-fidelity data approximation network (NN) is used. L The network consists of 8 hidden layers, each with 40 neurons, using the LeakyReLU activation function, an L2 regularization coefficient of 0.002, and He initialization for the weights; it is a linearly correlated approximation network. It contains 3 hidden layers, each with 20 neurons, no activation function, and weights initialized with a random normal distribution (standard deviation 0.01); a nonlinear correlation approximation network. There are 6 hidden layers, each with 30 neurons, using the Swish activation function, L2 regularization coefficient of 0.002, and He initialization weights. The loss function is set to 0.01, and the Adam optimization algorithm is employed with an initial learning rate of 0.0005 and an exponential decay rate of 0.96 per 1000 iterations. Gradient cutoff and early stopping are also used (the algorithm stops when the loss on the validation set stops decreasing after 10 consecutive epochs).

[0083] The processed data is divided into training, validation, and test sets at a ratio of 75%, 15%, and 10%, respectively. The model is trained using the training set, and the hyperparameters are adjusted based on the validation set.

[0084] Model optimization and improvement directions:

[0085] 1. Optimize data fusion strategy;

[0086] This study explores ways to further improve the fusion of high-fidelity and low-fidelity data to enhance the accuracy of models under complex flow field conditions. For example, it investigates data fusion methods based on physical constraints, enabling the fused flow field data to better leverage the advantages of multi-fidelity data while satisfying physical laws such as the conservation of mass, momentum, and energy.

[0087] 2. Evolution of neural network architecture;

[0088] This study considers introducing emerging neural network architectures or modules, such as attention mechanisms and residual connections, to enhance the model's ability to capture key features in the viscous flow field of rockets. The impact of these improvements on the accuracy and computational efficiency of flow field prediction under different flight conditions and rocket shapes is analyzed to determine the most suitable neural network architecture for correcting the viscous flow field of rockets.

[0089] 3. Quantification and handling of uncertainty;

[0090] This research explores how to quantify and handle various uncertainties in models, including data uncertainties (such as measurement errors and CFD calculation errors) and model uncertainties (such as uncertainties in neural network structure and parameters). It proposes corresponding uncertainty quantification methods, such as Bayesian neural networks and Monte Carlo methods, and discusses how to utilize uncertainty information to guide rocket design decisions and improve design reliability.

[0091] The above description is merely a preferred embodiment of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention by those skilled in the art within the scope of the technology disclosed in the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A method for correcting the flow field of a rocket vehicle based on multi-fidelity data fusion, characterized in that: Includes the following steps: Step 1: Collect low-fidelity data and construct a low-fidelity dataset: The unsteady flow field changes around an object within a finite time step are simulated using Reynolds time-averaged simulation or large eddy simulation, generating the same amount of flow field information as the simulation time step, and constructing the input dataset for a deep neural network. Step 2, perform data correlation analysis: Establishing the inviscid flow field of the rocket vehicle L With viscous flow field y H linear relationship model y H =ρ(x)y L +δ(x), through statistical analysis of a large amount of flow field data from different parts of the rocket, the scaling factor ρ(x) and error term δ(x) are determined; a generalized autoregressive scheme y is introduced. H =αF l (x H ,y L )+(1-α)F nl (x H ,y L ), α∈[0,1], and based on the characteristics of different stages of rocket flight and flow field regions, a linear mapping F is constructed. l and nonlinear mapping F nl And determine the range of values ​​for the hyperparameter α; Step 3, Neural Network Architecture Design: Constructing a low-fidelity data approximation network (NN) L Design a deep neural network NN with multiple hidden layers. L (x L ,θ), using LeakyReLU or Swish activation functions, and combining L1 and L2 regularization methods, with weight initialization using He or Xavier initialization methods; Design a linear dependence approximation network. It contains 2-4 hidden layers, each with 15-25 neurons. The appropriate activation function is selected based on the linear relationship characteristics, and the weights are initialized using a random normal distribution or a uniform distribution. Establish a nonlinear correlation approximation network Set up 4-8 hidden layers, with 20-40 neurons in each layer, and use the Swish or Tanh activation function, L2 regularization and He initialization method; Step 4, Hyperparameter learning and optimization: Define loss function The Adam or Adagrad adaptive learning rate optimization algorithm is adopted, combined with the momentum term to accelerate convergence. The learning rate is dynamically adjusted according to the gradient changes during the training process. Gradient truncation technique is used to prevent gradient explosion, and early stopping method is used to avoid overfitting. Step 5, Data Acquisition and Preprocessing: High-fidelity viscous flow field data was acquired, and high-precision CFD methods were used to simulate and calculate the flow field under various flight conditions for different rocket shapes. Adaptive or fine mesh technology was used to obtain detailed flow field parameter distributions as high-fidelity samples. Low-fidelity inviscid flow field data was acquired, and inviscid Euler equation solvers were used to calculate the inviscid flow field under the same rocket shape and flight conditions. A relatively coarse mesh was used to improve the calculation speed, and the data obtained were used as low-fidelity samples. The collected high-fidelity and low-fidelity flow field data are preprocessed to remove outliers and erroneous data. Based on the characteristics of the flow field parameters, an appropriate normalization method is selected to map the data to a specific interval. Step 6, Model Training and Validation: The preprocessed high- and low-fidelity flow field data samples are divided into training, validation, and test sets. The model is trained using mini-batch gradient descent and combined with random deactivation and batch standardization techniques to improve the network's generalization ability and training stability. The number of training iterations, learning rate, and other hyperparameters are dynamically adjusted based on the loss function values ​​on the validation set. Multiple error metrics were selected to evaluate the model, and the differences between the model's prediction results on the validation and test sets and the actual viscous flow field data were analyzed. The performance of the model under different flight conditions and rocket shapes was compared.

2. The rocket spacecraft flow field correction method based on multi-fidelity data fusion according to claim 1, characterized in that: In step 3, the low-fidelity data approximation network NN L It contains 6-10 hidden layers, and the number of neurons in each layer is reasonably set according to the data complexity and computing resources, with the range of 30-50 neurons per layer.

3. The rocket spacecraft flow field correction method based on multi-fidelity data fusion according to claim 2, characterized in that: In step 3, the value of L1 is 0.001-0.01, and the value of L2 is 0.002-0.

02.

4. The rocket vehicle flow field correction method based on multi-fidelity data fusion according to claim 1, characterized in that: In the defined loss function, The regularization parameter λ should be selected reasonably based on the data size and model complexity, and the range of the regularization parameter λ is 0.005-0.

05.

5. The rocket vehicle flow field correction method based on multi-fidelity data fusion according to claim 1, characterized in that: In step 6, the preprocessed high-fidelity and low-fidelity flow field data samples are divided into training set, validation set and test set according to the proportions of 70%-85%, 10%-20%, and 5%-10%.

6. The rocket vehicle flow field correction method based on multi-fidelity data fusion according to claim 5, characterized in that: In step 6, the probability value of random inactivation is 0.1-0.

3.

7. The rocket vehicle flow field correction method based on multi-fidelity data fusion according to claim 6, characterized in that: The error evaluation model includes root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 ), relative error.

8. A rocket vehicle flow field correction system based on multi-fidelity data fusion, characterized in that: The correction system includes a data acquisition module, a data correlation analysis module, a neural network architecture design module, and a hyperparameter learning and optimization module. The data acquisition module is used to acquire low-fidelity data, simulate the unsteady flow field change process around the object within a finite time step using the Reynolds time-averaged simulation method or the large eddy simulation method, generate the same amount of flow field information as the simulation time step, and construct the input dataset for the deep neural network. The data correlation analysis module is used to establish the inviscid flow field y of the rocket vehicle. L With viscous flow field y H linear relationship model y H =ρ(x)y L +δ(x), through statistical analysis of a large amount of flow field data from different parts of the rocket, the scaling factor ρ(x) and error term δ(x) are determined and incorporated into the generalized autoregressive scheme y H =αF l (x H ,y L )+(1-α)F nl (x H ,y L ), α∈[0,1], and based on the characteristics of different stages of rocket flight and flow field regions, a linear mapping F is constructed. l and nonlinear mapping F nl And determine the range of values ​​for the hyperparameter α; The neural network architecture design module is used to construct a low-fidelity data approximation network (NN). L Design a deep neural network NN with multiple hidden layers. L (x L The network employs LeakyReLU or Swish activation functions, combined with L1 and L2 regularization, and uses He or Xavier initialization methods for weight initialization. A linearly dependent approximation network is designed. It contains 2-4 hidden layers, each with 15-25 neurons. An appropriate activation function is selected based on the linear relationship characteristics, and the weights are initialized using a random normal distribution or a uniform distribution. A nonlinear correlation approximation network is established. Set up 4-8 hidden layers, with 20-40 neurons in each layer, and use the Swish or Tanh activation function, L2 regularization and He initialization method; The hyperparameter learning and optimization module is used to define the loss function.

Images