GAN-based high-speed aerodynamic heat flow field reconstruction method and system

Abstract

The invention discloses a GAN-based high-speed aerodynamic heat flow field reconstruction method and system, and relates to the technical field of analogue simulation, and the method comprises the steps: collecting historical heat flow field data of a high-speed aircraft under different flight conditions, and constructing a real aerodynamic heat flow field data set; establishing a heat flow field generator and a discriminator through feature extraction and simulation discrimination training; performing grid initialization and adaptive adjustment according to the distribution change of the heat flow field, and dividing into N hierarchical reconstruction regions; alternately training the generator and the discriminator in each region by adopting a balance training strategy, and constructing N heat flow field reconstruction networks; and monitoring flight condition update data, optimizing by using each reconstruction network, and generating a simulated heat flow field. Therefore, the technical effects of simplifying parameter configuration, reducing process cost and improving timeliness are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of simulation, and particularly to a method and system for reconstructing a high-speed aerodynamic heat flow field based on GAN. Background Art

[0002] Due to the influence of shock waves under complex flight conditions, high-speed aircraft are subjected to strong aerodynamic heating. The construction of the heat flow field distribution of high-speed aircraft is crucial for aircraft structure design, thermal protection system optimization, and flight safety optimization.

[0003] Traditional heat flow field reconstruction methods usually rely on numerical simulations or experimental data (such as based on wind tunnel or flight experiments). With the continuous complication of the shapes of high-speed aircraft, the local aerodynamic heating interference of the aircraft gradually intensifies, the phenomenon of turbulent separation and reattachment becomes more significant, and the coupling effect of rarefied gas and real gas effects also becomes more prominent. Under such complex flight conditions, the above traditional methods have technical problems such as complex parameter settings, high computational costs, and poor real-time performance. Summary of the Invention

[0004] The present invention provides a method and system for reconstructing a high-speed aerodynamic heat flow field based on GAN to solve the technical problems of complex parameter settings, high computational costs, and poor real-time performance in the prior art, and to achieve the technical effects of simplifying parameter configuration, reducing process costs, and improving timeliness.

[0005] In a first aspect, the present invention provides a method for reconstructing a high-speed aerodynamic heat flow field based on GAN, wherein the method for reconstructing a high-speed aerodynamic heat flow field based on GAN includes: Collect and obtain a real aerodynamic heat flow field data set, where the real aerodynamic heat flow field data set is historical heat flow field data of a high-speed aircraft under different flight conditions.

[0006] Based on the real aerodynamic heat flow field data set, perform feature extraction training and simulation discrimination training to construct a heat flow field generator and a heat flow field discriminator.

[0007] According to the distribution change data of the target heat flow field, perform grid initialization and adaptive adjustment to obtain a heat flow field distribution grid, and based on the heat flow field distribution grid, divide the reconstruction area of the target heat flow field to obtain N hierarchical reconstruction areas.

[0008] Construct a balanced training strategy, and use the balanced training strategy to alternately balance-train the heat flow field generator and the heat flow field discriminator based on the N hierarchical reconstruction areas to build N heat flow field reconstruction networks.

[0009] Monitor and obtain flight condition update data, and respectively perform heat flow field reconstruction optimization on the flight condition update data based on the N heat flow field reconstruction networks to generate a high-speed aerodynamic simulation heat flow field.

[0010] In a feasible implementation, constructing the heat flux field generator and the heat flux field discriminator includes: Performing Gaussian filtering and data alignment processing on the real aerodynamic heat flux field data set to obtain an available real aerodynamic heat flux field data set.

[0011] Performing associated feature extraction on the available real aerodynamic heat flux field data set according to the associated factors of heat flux field reconstruction to obtain a heat flux field associated feature set.

[0012] Introducing a multi-scale feature fusion module, and using a convolutional neural network and the multi-scale feature fusion module to perform simulation training on the heat flux field associated feature set to construct a heat flux field generator.

[0013] Outputting a simulated heat flux field data set through the heat flux field generator, and performing heat flux field discrimination training based on the real aerodynamic heat flux field data set and the simulated heat flux field data set to generate a heat flux field discriminator.

[0014] In a feasible implementation, constructing the heat flux field generator includes: According to a convolutional neural network and the multi-scale feature fusion module, building a generator network architecture, and the generator network architecture includes a feature input layer, a plurality of feature convolution extraction layers, a multi-scale fusion layer, and a heat flux field output layer.

[0015] Passing the heat flux field associated feature set through the feature input layer and importing it into the plurality of feature convolution extraction layers for multi-scale feature extraction to obtain multi-scale heat flux field feature data.

[0016] Based on the multi-scale fusion layer, splicing and fusing the multi-scale heat flux field feature data to obtain heat flux field simulation data, and outputting it as an output result through the heat flux field output layer.

[0017] Performing heat flux field simulation training based on the heat flux field associated feature set and the heat flux field simulation data to construct the heat flux field generator.

[0018] In a feasible implementation, generating the heat flux field discriminator includes: Building a generator network architecture, and the generator network architecture includes a data input layer, a plurality of feature convolution extraction layers, a fully connected layer, and a discriminant output layer.

[0019] Passing the real aerodynamic heat flux field data set and the simulated heat flux field data set through the data input layer and transmitting them to the plurality of feature convolution extraction layers for associated feature extraction to obtain heat flux field associated feature data.

[0020] Based on the fully connected layer, perform identification evaluation and judgment on the heat flow field correlation feature data, obtain heat flow field simulation judgment information, and output it through the discriminant output layer as the output result.

[0021] Based on the real aerodynamic heat flow field data set, the simulated heat flow field data set, and the heat flow field simulation judgment information, perform evaluation judgment training to generate the heat flow field discriminator.

[0022] In a feasible implementation manner, the obtaining of the heat flow field distribution grid includes: Extract features from the initial distribution characteristics of the target heat flow field to determine the heat flow field change characteristics, where the heat flow field change characteristics include temperature gradient, flow field characteristics, and geometric characteristics.

[0023] Perform grid initialization according to the heat flow field change characteristics to generate an initial uniform grid.

[0024] Calculate the gradient change of each grid cell in the initial uniform grid based on the distribution change data of the target heat flow field according to the heat flow field change characteristics to determine the grid cell change gradient data.

[0025] Based on the grid cell change gradient data, perform adaptive adjustment of the grid density of the initial uniform grid to obtain the heat flow field distribution grid.

[0026] In a feasible implementation manner, the obtaining of N hierarchical reconstruction regions includes: According to the heat flow field distribution grid, determine the grid cell density distribution information.

[0027] Use the grid cell density distribution information to perform regional clustering analysis on the heat flow field distribution grid to obtain the grid region clustering result.

[0028] Set the reconstruction region segmentation threshold according to the heat flow field reconstruction accuracy requirement.

[0029] Based on the reconstruction region segmentation threshold, perform reconstruction region division on the grid region clustering result to obtain the N hierarchical reconstruction regions.

[0030] In a feasible implementation manner, the building of N heat flow field reconstruction networks includes: According to the heat flow field generator and the heat flow field discriminator, design the generator mean square loss function and the discriminator cross entropy loss function.

[0031] Based on the balanced training strategy, perform analysis of the balance coefficients for the N hierarchical reconstruction regions to determine the N initial region training balance coefficients.

[0032] Based on the mean square loss function of the generator and the cross - entropy loss function of the discriminator, alternately balance - train the heat - flow - field generator and the heat - flow - field discriminator based on the N initial regions' training balance coefficients to obtain N heat - flow - field reconstruction networks.

[0033] In a feasible implementation, obtaining the N heat - flow - field reconstruction networks includes: Perform simulated discrimination training on the heat - flow - field generator and the heat - flow - field discriminator respectively based on the N initial regions' training balance coefficients to obtain N initial generators and N initial discriminators.

[0034] Use the mean square loss function of the generator and the cross - entropy loss function of the discriminator to evaluate the losses of the N initial generators and N initial discriminators, obtaining N generator loss data and N discriminator loss data.

[0035] Based on the N generator loss data and N discriminator loss data, perform balanced alternating training on the N initial generators and N initial discriminators to obtain N heat - flow - field reconstruction networks.

[0036] In a feasible implementation, obtaining the N heat - flow - field reconstruction networks includes: Dynamically adjust the N initial regions' training balance coefficients based on the N generator loss data and N discriminator loss data to obtain N region - training balance dynamic coefficients.

[0037] Iteratively balance - update the N initial generators and N initial discriminators according to the N region - training balance dynamic coefficients to obtain N region generators and N region discriminators.

[0038] Connect and combine the N region generators and N region discriminators to obtain the N heat - flow - field reconstruction networks.

[0039] In a second aspect, the present invention also provides a high - speed aerodynamic heat - flow - field reconstruction system based on GAN. Among them, the high - speed aerodynamic heat - flow - field reconstruction system based on GAN includes: A real - data acquisition component, used to collect and obtain a real aerodynamic heat - flow - field data set, and the real aerodynamic heat - flow - field data set is the historical heat - flow - field data of a high - speed aircraft under different flight conditions.

[0040] A tool training component, used to perform feature extraction training and simulated discrimination training based on the real aerodynamic heat - flow - field data set to construct a heat - flow - field generator and a heat - flow - field discriminator.

[0041] A mesh generation component, which is used to initialize and adaptively adjust the mesh according to the distribution change data of the target heat flux field, obtain the heat flux field distribution mesh, and divide the target heat flux field into N hierarchical reconstruction regions based on the heat flux field distribution mesh.

[0042] An equilibrium training component, which is used to construct an equilibrium training strategy, and alternately balance-train the heat flux field generator and the heat flux field discriminator based on the N hierarchical reconstruction regions by using the equilibrium training strategy, and build N heat flux field reconstruction networks.

[0043] An update and optimization component, which is used to monitor and obtain flight condition update data, and respectively perform heat flux field reconstruction and optimization on the flight condition update data based on the N heat flux field reconstruction networks to generate a high-speed aerodynamic simulation heat flux field.

[0044] The present invention discloses a method and system for reconstructing a high-speed aerodynamic heat flux field based on GAN, including: collecting and obtaining historical heat flux field data of a high-speed aircraft under different flight conditions to form a real aerodynamic heat flux field data set; subsequently, performing feature extraction training and simulation discrimination training based on the data set to construct a heat flux field generator and a discriminator; on this basis, initializing and adaptively adjusting the mesh according to the distribution change data of the target heat flux field to generate a heat flux field distribution mesh, and dividing the target heat flux field into N hierarchical reconstruction regions accordingly; further constructing an equilibrium training strategy, and using the strategy to alternately balance-train the heat flux field generator and the discriminator in the N hierarchical reconstruction regions to build N heat flux field reconstruction networks; finally, monitoring and obtaining flight condition update data, and respectively performing heat flux field reconstruction and optimization on the update data based on the N heat flux field reconstruction networks to generate a high-speed aerodynamic simulation heat flux field. The method and system for reconstructing a high-speed aerodynamic heat flux field based on GAN disclosed by the present invention solve the technical problems of complex parameter setting, high computational cost, and poor real-time performance, and achieve the technical effects of simplifying parameter configuration, reducing process cost, and improving timeliness. Description of the Drawings

[0045] Figure 1 It is a schematic flowchart of the method for reconstructing a high-speed aerodynamic heat flux field based on GAN of the present invention.

[0046] Figure 2 It is a schematic structural diagram of the system for reconstructing a high-speed aerodynamic heat flux field based on GAN of the present invention.

[0047] Description of the reference numerals: The real data acquisition component 11, the tool training component 12, the mesh generation component 13, the equilibrium training component 14, the update and optimization component 15. Detailed Embodiments

[0048] The above technical solutions will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments to better understand the above technical solutions. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments for explaining the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention. In addition, it should be noted that for the sake of description, only parts related to the present invention are shown in the drawings rather than all of them.

[0049] Embodiment 1, as Figure 1 is a schematic flow chart of the method for reconstructing a high-speed aerodynamic heat flow field based on GAN of the present invention. Among them, the method for reconstructing a high-speed aerodynamic heat flow field based on GAN includes: S100: Collect and obtain a real aerodynamic heat flow field data set, and the real aerodynamic heat flow field data set is historical heat flow field data of a high-speed aircraft under different flight conditions.

[0050] Specifically, the real aerodynamic heat flow field data set refers to the actual historical heat flow field data obtained by actual measurement of a high-speed aircraft under different flight conditions (such as Mach number, angle of attack, altitude, etc.). Such as the aerodynamic heat flow field data verified to meet the accuracy requirements obtained through wind tunnel experiments, flight experiments or numerical simulations, among which, physical parameters such as temperature, pressure, and heat flow distribution are included.

[0051] By obtaining the real and accurate real aerodynamic heat flow field data set, it provides a reliable data basis for subsequent training of the heat flow field generator and the heat flow field discriminator based on GAN, and further helps to ensure that the generator and the discriminator can learn the real heat flow field distribution characteristics.

[0052] S200: Based on the real aerodynamic heat flow field data set, perform feature extraction training and simulation discrimination training to construct a heat flow field generator and a heat flow field discriminator.

[0053] Specifically, the feature extraction training is a process of extracting feature information related to the heat flow field distribution from the above-obtained real aerodynamic heat flow field data set through a neural network model, so as to obtain a heat flow field generator with the ability to generate a heat flow field; among them, exemplarily, the target features to be extracted include temperature gradient, flow field pattern, geometric features, etc.

[0054] Specifically, the simulation discrimination training is a process of comparing the simulated heat flow field data generated by the above heat flow field generator with the corresponding real data (and the true values of multiple features), so as to improve the ability of the training discriminator to identify the authenticity of the data. In other words, this simulation discrimination training is used to optimize the performance of the generator and the discriminator.

[0055] Through the above steps of feature extraction training, a heat flow field generator based on a neural network is obtained, reducing the manual adjustment of parameters in the study of the heat flow field. The simulation discrimination training improves the performance of the generator and discriminator through an adversarial training mechanism, ensuring that the generated simulated heat flow field data has high authenticity and accuracy, providing high-quality model support for subsequent heat flow field reconstruction.

[0056] In some embodiments, the construction of the heat flow field generator and the heat flow field discriminator includes: Perform Gaussian filtering and data alignment processing on the real aerodynamic heat flow field data set to obtain an available real aerodynamic heat flow field data set; extract associated features of the heat flow field according to the associated factors of heat flow field reconstruction for the available real aerodynamic heat flow field data set to obtain a heat flow field associated feature set; introduce a multi-scale feature fusion module, and use a convolutional neural network and the multi-scale feature fusion module to perform simulation training on the heat flow field associated feature set to construct a heat flow field generator; output a simulated heat flow field data set through the heat flow field generator, and perform heat flow field discrimination training based on the real aerodynamic heat flow field data set and the simulated heat flow field data set to generate a heat flow field discriminator.

[0057] Specifically, Gaussian filtering is used to smooth the data through a Gaussian function to remove noise in the data; the data after Gaussian filtering needs to be data-aligned to adjust data from different sources or in different formats to a unified coordinate system or time series (i.e., generate an available real aerodynamic heat flow field data set) to ensure the consistency and comparability of the data.

[0058] Specifically, according to the associated factors required for heat flow field reconstruction (such as boundary layer thickness, Mach number distribution, surface temperature gradient, etc.), convolutional layers and pooling layers in the convolutional neural network are used to extract representative heat flow field associated features at different levels in the available real aerodynamic heat flow field data set to form a heat flow field associated feature set. Among them, through a convolutional neural network that is good at processing data with a grid structure (such as the heat flow field distribution map on the surface or cross-section of an aircraft), the features of the data can be automatically extracted.

[0059] Specifically, the multi-scale feature fusion module is used to integrate network structures of different scale features, which can capture the changing features of the heat flow field at different spatial scales. By introducing the multi-scale feature fusion module, features of different scales are spliced and fused to enhance the unified expression ability of different scale information (such as local eddy structures and overall flow field trends). Furthermore, combined with the convolutional neural network (CNN) structure and the multi-scale feature fusion module, the heat flow field associated feature set is subjected to simulation training to obtain a heat flow field generator, which can output a simulated heat flow field data set with high fidelity according to the input flight state parameters (including aircraft information, environmental information, flight state information, etc.).

[0060] Furthermore, the simulated heat flow field data set generated by the heat flow field generator is combined with the real aerodynamic heat flow field data set, and the heat flow field discriminator is trained through the cross-entropy loss function so that it can distinguish real data from simulated data (evaluating the authenticity and structural consistency of the generated data).

[0061] In the above process, data preprocessing and feature extraction can effectively improve the quality of the input data and reduce the impact of noise and inconsistency on model training. The introduction of the multi-scale feature fusion module enables the generator to capture the complex changes of the heat flow field at different spatial scales, improving the authenticity and accuracy of the generated data, thus providing high-quality model support for subsequent heat flow field reconstruction.

[0062] In some implementation manners, constructing the heat flow field generator includes: According to the convolutional neural network and the multi-scale feature fusion module, a generator network architecture is built. The generator network architecture includes a feature input layer, multiple feature convolutional extraction layers, a multi-scale fusion layer, and a heat flow field output layer; the heat flow field correlation feature set is imported through the feature input layer into the multiple feature convolutional extraction layers for multi-scale feature extraction to obtain multi-scale heat flow field feature data; based on the multi-scale fusion layer, the multi-scale heat flow field feature data is spliced and fused to obtain heat flow field simulation data, which is output as the output result through the heat flow field output layer; based on the heat flow field correlation feature set and the heat flow field simulation data, heat flow field simulation training is performed to construct the heat flow field generator.

[0063] Specifically, to construct the heat flow field generator, first, according to the convolutional neural network and the multi-scale feature fusion module, the network architecture of the generator is designed. This architecture includes a feature input layer, multiple feature convolutional extraction layers, a multi-scale fusion layer, and a heat flow field output layer. Among them, the feature input layer is used to receive the heat flow field correlation feature set as the input end of the network. The multiple feature convolutional extraction layers use various convolutional layers (such as convolutional kernels of sizes 3×3, 5×5, etc.) to perform multi-scale feature extraction on the input correlation features, refining the local and global features of the heat flow field. The multi-scale fusion layer corresponding to the multi-scale feature fusion module is used to splice and fuse the features extracted by each convolutional layer, enabling the model to take into account both local details and overall structural information. Exemplarily, this layer adopts a structure similar to Inception and residual connection methods to make the gradient propagation more stable and enhance the information expression ability.

[0064] Optionally, batch normalization and ReLU activation functions are introduced after each convolutional channel to enhance the non-linear expression ability.

[0065] Specifically, the network architecture of the above generator is simulated and trained by using the heat flow field correlation feature set and the generated simulation data to continuously optimize its parameters, ensuring that the heat flow field generator can generate a simulated heat flow field closer to the real data.

[0066] In some implementation manners, the heat flow field discriminator includes: Build a generator network architecture, which includes a data input layer, multiple feature convolution extraction layers, a fully connected layer, and a discriminant output layer; transmit the real aerodynamic heat flow field data set and the simulated heat flow field data set through the data input layer to the multiple feature convolution extraction layers for correlation feature extraction to obtain heat flow field correlation feature data; based on the fully connected layer, perform identification evaluation and judgment on the heat flow field correlation feature data to obtain heat flow field simulation judgment information, and output it as an output result through the discriminant output layer; based on the real aerodynamic heat flow field data set, the simulated heat flow field data set, and the heat flow field simulation judgment information, perform evaluation and judgment training to generate the heat flow field discriminator.

[0067] Optionally, a binary classification neural network model is used to construct the heat flow field discriminator. The structure of the discriminator includes a data input layer, multiple feature convolution extraction layers, a fully connected layer, and a discriminant output layer. Among them, the data input layer is the first layer for receiving input data, providing uniformly formatted data input for subsequent feature extraction; the feature convolution extraction layer extracts local features and correlation information from the input data through convolution operations. The above multiple convolution layers are used to capture low-level, intermediate-level, and high-level features in the data respectively, so that the spatial variation, texture, and dynamic characteristics of the heat flow field are fully expressed; the fully connected layer is used to stretch the two-dimensional or three-dimensional feature map extracted by convolution into a one-dimensional vector to realize feature synthesis and information fusion, facilitating subsequent classification or identification judgment; finally, the discriminant output layer, combined with the identification evaluation result output by the fully connected layer, outputs a discriminant information with an appropriate activation function (such as Softmax or Sigmoid), which is used to characterize whether the input heat flow field data is "real" or "simulated".

[0068] Specifically, first, the real aerodynamic heat flux field dataset and the simulated heat flux field dataset are input through the data input layer and transmitted to multiple feature convolution extraction layers to perform convolution operations, automatically extracting the associated features in the heat flux field. For example, the input data includes temperature distribution, flow field vortices, and boundary layer characteristics. A series of feature maps obtained after continuous convolution can reflect the local and global heat flux distribution laws in the input data. Then, the heat flux field associated feature data extracted by the convolution layer is integrated through the fully connected layer, and the fully connected layer is used to evaluate the identity of the integration result (i.e., the heat flux field simulation judgment information), obtaining the heat flux field simulation judgment information, that is, the probability of discriminating whether the input data is "true" or "false" (e.g., [0, 1]). Finally, this result is output through the discriminant output layer as the final classification result of the input heat flux field data.

[0069] Specifically, based on the real aerodynamic heat flux field dataset, the simulated heat flux field dataset, and the obtained heat flux field simulation judgment information, training samples of the discriminator are constructed. Among them, the simulated heat flux field data output by the heat flux field generator is used as a "pseudo sample", and the real aerodynamic heat flux field data is used as a "positive sample"; then, the discriminator is trained through a supervised learning method, and the parameters of the convolution kernel and the fully connected layer are continuously updated through backpropagation, so that the heat flux field discriminator can distinguish real data from simulated data.

[0070] Optionally, the cross-entropy loss function is used to measure the error between the discriminant output and the true class label, and the gradient descent method is combined for parameter optimization until the discriminant accuracy reaches the expected level.

[0071] In addition, a gradient penalty term (such as the gradient penalty in WGAN-GP) can be introduced to improve the smoothness and physical rationality of the heat flux field data output by the generator.

[0072] S300: Perform grid initialization and adaptive adjustment according to the distribution change data of the target heat flux field to obtain the heat flux field distribution grid, and divide the target heat flux field into N hierarchical reconstruction regions based on the heat flux field distribution grid.

[0073] Specifically, the target heat flux field refers to the heat flux field data obtained in actual or simulation, including the distribution and variation characteristics of parameters such as temperature, flow rate, and heat flux in space, as well as the local or global characteristics of the heat flux field formed thereby; according to the target heat flux field distribution change data, a spatial distribution grid can be initially constructed (i.e., grid initialization), and this process involves determining the positions, spacings, and basic structural forms of the grid nodes, providing a basis for subsequent adaptive adjustment.

[0074] Specifically, the grid density is automatically adjusted according to the local change degree of the target heat flow field (i.e., adaptive adjustment). Denser grids are used in areas where the heat flow field changes drastically (with a large gradient) to capture local details; sparser grids are used in areas with gentle changes (with a small gradient), so as to achieve the purpose of balanced data expression and optimized calculation efficiency. The grid of the heat flow field distribution is the spatial distribution structure obtained after grid initialization and adaptive adjustment.

[0075] Furthermore, based on the grid of the heat flow field distribution, the target heat flow field is divided into several regions (i.e., N hierarchical reconstruction regions). The data within each region has high homogeneity and similar change trends, and different reconstruction strategies can be adopted to gradually achieve a fine restoration of the characteristics of the heat flow field.

[0076] In some embodiments, obtaining the grid of the heat flow field distribution includes: Performing feature extraction on the initial distribution characteristics of the target heat flow field to determine the change characteristics of the heat flow field. The change characteristics of the heat flow field include temperature gradient, flow field characteristics, and geometric characteristics; initializing the grid according to the change characteristics of the heat flow field to generate an initial uniform grid; calculating the gradient change of each grid cell in the initial uniform grid based on the distribution change data of the target heat flow field according to the change characteristics of the heat flow field to determine the gradient change data of the grid cell; and performing adaptive adjustment of the grid density on the initial uniform grid based on the gradient change data of the grid cell to obtain the grid of the heat flow field distribution.

[0077] Specifically, the change characteristics of the heat flow field refer to the change characteristics of the heat flow field in spatial distribution, including temperature gradient (the rate of temperature change, reflecting the non-uniformity of local heat flow), flow field characteristics (such as flow velocity, flow direction, and eddy current distribution, etc.), and geometric characteristics (the shape, size, and boundary conditions of the region where the heat flow field is located, etc.).

[0078] Specifically, the gradient change data of the grid cell refers to the gradient value calculated from the distribution change data of the target heat flow field inside each grid cell or between adjacent cells in the initial uniform grid, which is used to measure the drastic degree of the change of the heat flow field in the local area; the grid of the heat flow field distribution is the grid generated after the adaptive adjustment of the grid density of the initial uniform grid. This grid has a high grid density in areas where the heat flow field changes drastically and a low grid density in areas where the heat flow field changes gently, so as to improve the overall data representation accuracy and calculation efficiency.

[0079] Specifically, after obtaining the change characteristics of the target heat flow field, according to the change characteristics of the target heat flow field, an initial uniform grid is generated in the entire target area in combination with a preset grid rule. For example, grid cells are uniformly generated in the target area with a preset grid granularity; then, by analyzing the distribution change data of the target heat flow field, the gradient change data within each grid cell or between adjacent cells in the initial uniform grid is calculated. For example, the difference in temperature between adjacent grid cells is calculated by the numerical difference method, and gradient estimation is performed in combination with the corresponding flow field and geometric information to obtain the change gradient of each grid cell. Furthermore, for grid cells with a change gradient greater than a preset threshold (e.g., greater than 8 °C / m or significant changes in the corresponding flow field / geometric features), the grid is refined and the cell size is reduced (e.g., from 0.5 m to 0.2 m). For grid cells with a lower change gradient, the grid spacing is maintained or appropriately enlarged (e.g., it can be increased to 0.8 m), so as to reduce the calculation density in the low-change area and achieve a balance between local details and global efficiency.

[0080] In summary, through the extraction of the change characteristics of the target heat flow field, the generation of the initial uniform grid, the calculation of the grid cell gradient, and the adaptive adjustment of the grid density based on the gradient data, a heat flow field distribution grid that accurately reflects the distribution change law of the heat flow field can be obtained, providing a solid data basis and geometric partition basis for subsequent hierarchical region reconstruction, data optimization, and simulation analysis.

[0081] In some embodiments, obtaining the N hierarchical reconstruction regions includes: Determining the grid cell density distribution information according to the heat flow field distribution grid; performing regional clustering analysis on the heat flow field distribution grid using the grid cell density distribution information to obtain a grid region clustering result; setting a reconstruction region segmentation threshold according to the heat flow field reconstruction accuracy requirement; and performing reconstruction region division on the grid region clustering result based on the reconstruction region segmentation threshold to obtain the N hierarchical reconstruction regions.

[0082] Specifically, the grid cell density distribution information refers to the distribution density of each grid cell in the heat flow field distribution grid in space, that is, the number and relative spacing of grid cells in different regions. This information can reflect the characteristics of the heat flow field changing violently or gently in a certain area; classifying the heat flow field distribution grid according to the grid cell density distribution information can group grid cells with similar density or local change characteristics into one category, so as to obtain a grid region clustering result for the reconstruction region.

[0083] Specifically, the reconstruction region segmentation threshold is a predetermined quantitative indicator used to distinguish the boundaries of different reconstruction regions, reflecting the requirements for data uniformity and capturing local details within the required region; based on the reconstruction region segmentation threshold, the grid region clustering results can be further analyzed. For example, the preliminary clustering results may obtain 5 major categories of regions. According to the refinement requirements of the preset threshold, some of the categories can be further subdivided into multi-level regions. The final total number is N hierarchical regions. Each region has high data uniformity, which is convenient for subsequent reconstruction.

[0084] S400: constructing a balanced training strategy, using the balanced training strategy to perform alternating balanced training on the thermal flow field generator and the thermal flow field discriminator based on the N hierarchical reconstruction regions, and building N thermal flow field reconstruction networks.

[0085] Specifically, in the generative adversarial network (GAN), the balanced training strategy is a training scheme adopted to balance the training rates of the generator and the discriminator and stabilize the model convergence. This strategy ensures that each reconstruction network achieves the best performance in the process of thermal flow field data reconstruction through alternating training, loss function balance and multi-region adaptation.

[0086] In some embodiments, the step of building N thermal flow field reconstruction networks includes: According to the thermal flow field generator and the thermal flow field discriminator, the generator mean square loss function and the discriminator cross entropy loss function are designed; based on the balanced training strategy, the balance coefficients of the N hierarchical reconstruction regions are analyzed to determine the N initial region training balance coefficients; using the generator mean square loss function and the discriminator cross entropy loss function, based on the N initial region training balance coefficients, the thermal flow field generator and the thermal flow field discriminator are alternately balanced trained to obtain N thermal flow field reconstruction networks.

[0087] Specifically, the generator mean square loss function is used to measure the difference between the simulated thermal flow field data output by the generator and the real thermal flow field data, and the discriminator cross entropy loss function is used to measure the degree of error in the discriminator's classification of real data and generated data.

[0088] Specifically, the balance coefficient refers to the weight coefficient applied to each hierarchical reconstruction region in the balanced training strategy, which is used to adjust the training speed of the generator and the discriminator in different regions to ensure that the two are in a coordinated and optimized state during the alternating training process. Optionally, the balance coefficient required for the initial training of each region can be determined by statistically analyzing the data change characteristics, error distribution and network convergence in each region. For example, in areas with drastic changes, a larger balance coefficient may be required to speed up the learning of the generator and ensure that the generated results have fine details; while in areas with gentle changes, the balance coefficient is appropriately reduced to prevent overtraining.

[0089] Specifically, by using the designed mean square loss function of the generator and the cross-entropy loss function of the discriminator, combined with the initial balance coefficients of each region obtained, the heat flow field generator and the heat flow field discriminator are alternately and balancedly trained; including dynamically adjusting the training step size and weights according to the preset balance coefficients for each hierarchical reconstruction region. First, the generator is trained for several rounds using the mean square loss function of the generator, and then the discriminator is trained using the cross-entropy loss function of the discriminator. The alternating training is repeated until the generator and the discriminator in each region reach the convergence balance state, and the heat flow field reconstruction networks of each region are generated respectively.

[0090] In the above process, the mean square error loss of the generator and the cross-entropy loss of the discriminator are coordinated with the balance coefficient adjustment, which helps the networks in different regions to achieve stable convergence in the alternating training, and avoids training imbalance or mode collapse caused by excessive differences in local data characteristics; integrating the heat flow field reconstruction networks after training in each region can achieve the continuity and consistency of the overall heat flow field while fully restoring local details.

[0091] In some implementation manners, obtaining the N heat flow field reconstruction networks includes: Performing simulated discrimination training on the heat flow field generator and the heat flow field discriminator respectively based on the N initial region training balance coefficients to obtain N initial generators and N initial discriminators; using the mean square loss function of the generator and the cross-entropy loss function of the discriminator to perform loss evaluation on the N initial generators and N initial discriminators to obtain N generator loss data and N discriminator loss data; based on the N generator loss data and N discriminator loss data, performing balanced alternating training on the N initial generators and N initial discriminators to obtain N heat flow field reconstruction networks.

[0092] Specifically, first, based on the N initial region training balance coefficients, simulated discrimination training is respectively performed on the heat flow field generator and the heat flow field discriminator in each region. The obtained initial generator can output a heat flow field pattern similar to the real data, and at the same time, the initial discriminator can preliminarily judge the authenticity of the output data; then, the mean square loss function of the generator and the cross-entropy loss function of the discriminator are respectively used to perform loss evaluation on the obtained N initial generators and N initial discriminators to obtain the corresponding generator loss data (mean square error values) and discriminator loss data (cross-entropy errors), which reflect the accuracy and stability of the initial training state of the networks in each region.

[0093] Further, based on the obtained N generator loss data and N discriminator loss data, an alternating update strategy is adopted to perform balanced alternating training on the initial generators and initial discriminators in each region. That is, first fix the discriminator, and continue to train the generator using the mean square loss function of the generator (i.e., the number of training iterations is determined based on the generator loss data) to make its output closer to the real data; then fix the generator, and train the discriminator using the cross-entropy loss function of the discriminator (i.e., the number of training iterations is determined based on the discriminator loss data) to improve its discrimination accuracy, so that the generator and the discriminator can achieve the balance and stability of the training process while their respective losses drop to the predetermined target values.

[0094] In some embodiments, the obtaining of the N heat flow field reconstruction networks includes: Based on the N generator loss data and N discriminator loss data, dynamically adjust the training balance coefficients of the N initial regions to obtain N region training balance dynamic coefficients; perform iterative balance updates on the N initial generators and N initial discriminators according to the N region training balance dynamic coefficients to obtain N region generators and N region discriminators; concatenate and merge the N region generators and N region discriminators to obtain the N heat flow field reconstruction networks.

[0095] Optionally, to obtain the aforementioned N generator loss data and N discriminator loss data, methods such as weighted average and gradient change rate analysis are used to dynamically adjust the original region training balance coefficients. For example, if the generator loss is significantly higher than the discriminator loss, then increase the training weight of the generator in this region to make it preferentially optimized and adjusted, and vice versa.

[0096] Further, perform iterative balance updates on the N initial generators and N initial discriminators according to the adjusted region training balance dynamic coefficients. Exemplarily, if the dynamic balance coefficients of a certain region are 0.6 (generator) and 0.4 (discriminator), then more training resources or rounds will be allocated for optimizing the generator parameters in each round of alternating training, so as to effectively compensate for its insufficient accuracy.

[0097] Finally, concatenate and merge the updated N region generators and N region discriminators to form N independent heat flow field reconstruction networks. Each network corresponds to a hierarchical reconstruction region and can efficiently reconstruct the heat flow field characteristics of this region. Among them, the merging method can include output-input connection (i.e., the generation result of a certain region is used as the input of the next region), feature fusion (the generated features of multiple regions are spliced at the shared layer), or discriminant collaboration (the results of multiple discriminators are jointly used to guide the generator in the reverse direction).

[0098] S500: Monitor and obtain updated flight condition data, and respectively perform heat flux field reconstruction optimization on the updated flight condition data based on the N heat flux field reconstruction networks to generate a high-speed aerodynamic simulation heat flux field.

[0099] Specifically, the updated flight condition data is a set of real-time state parameters of the aircraft during flight (such as airspeed, angle of attack, Mach number, atmospheric density, etc.), which affects the distribution of the heat flux field; through the updated flight condition data, a new and high-precision heat flux field distribution map can be re-output to compensate for the deficiencies of traditional numerical simulations in terms of real-time performance or resolution.

[0100] In summary, the GAN-based high-speed aerodynamic heat flux field reconstruction method provided by the present invention has the following technical effects: By collecting and obtaining historical heat flux field data of a high-speed aircraft under different flight conditions, a real aerodynamic heat flux field data set is formed; subsequently, based on this data set, feature extraction training and simulation discrimination training are carried out to construct a heat flux field generator and a discriminator; on this basis, according to the distribution change data of the target heat flux field, grid initialization and adaptive adjustment are performed to generate a heat flux field distribution grid, and accordingly the target heat flux field is divided into N hierarchical reconstruction regions; further, a balanced training strategy is constructed, and this strategy is used to alternately balance-train the heat flux field generator and the discriminator in the N hierarchical reconstruction regions to build N heat flux field reconstruction networks; finally, monitor and obtain the updated flight condition data, and respectively perform heat flux field reconstruction and optimization on the updated data based on the N heat flux field reconstruction networks, so as to generate a high-speed aerodynamic simulation heat flux field, thereby achieving the technical effects of simplifying parameter configuration, reducing process costs, and improving timeliness.

[0101] Embodiment 2, as Figure 2 is a schematic structural diagram of the GAN-based high-speed aerodynamic heat flux field reconstruction system of the present invention. For example, Figure 1 In the flow chart of the GAN-based high-speed aerodynamic heat flux field reconstruction method of the present invention can be implemented by a structure such as Figure 2 shown.

[0102] Based on the same concept as the GAN-based high-speed aerodynamic heat flux field reconstruction method in the above embodiment, the GAN-based high-speed aerodynamic heat flux field reconstruction system provided by the present invention further includes: A real data acquisition component 11, configured to collect and obtain a real aerodynamic heat flux field data set, where the real aerodynamic heat flux field data set is historical heat flux field data of a high-speed aircraft under different flight conditions.

[0103] A tool training component 12, configured to perform feature extraction training and simulation discrimination training based on the real aerodynamic heat flux field data set to construct a heat flux field generator and a heat flux field discriminator.

[0104] The mesh generation component 13 is used to perform mesh initialization and adaptive adjustment according to the distribution change data of the target heat flux field, obtain the heat flux field distribution mesh, and perform reconstruction region division on the target heat flux field based on the heat flux field distribution mesh to obtain N hierarchical reconstruction regions.

[0105] The balance training component 14 is used to construct a balance training strategy, and alternately balance train the heat flux field generator and the heat flux field discriminator based on the N hierarchical reconstruction regions by using the balance training strategy, and build N heat flux field reconstruction networks.

[0106] The update and optimization component 15 is used to monitor and obtain flight condition update data, and respectively perform heat flux field reconstruction optimization on the flight condition update data based on the N heat flux field reconstruction networks to generate a high-speed aerodynamic simulation heat flux field.

[0107] In some embodiments, the tool training component 12 includes: The real aerodynamic heat flux field data preprocessing unit is used to perform Gaussian filtering and data alignment processing on the real aerodynamic heat flux field data set to obtain an available real aerodynamic heat flux field data set.

[0108] The heat flux field correlation feature extraction unit is used to extract correlation features from the available real aerodynamic heat flux field data set according to the heat flux field reconstruction correlation factors to obtain a heat flux field correlation feature set.

[0109] The heat flux field generator construction unit is used to introduce a multi-scale feature fusion module, and perform simulation training on the heat flux field correlation feature set by using a convolutional neural network and the multi-scale feature fusion module to construct a heat flux field generator.

[0110] The heat flux field discriminator generation unit is used to output a simulated heat flux field data set through the heat flux field generator, and perform heat flux field discrimination training based on the real aerodynamic heat flux field data set and the simulated heat flux field data set to generate a heat flux field discriminator.

[0111] In some implementation manners, the heat flux field generator construction unit in the tool training component 12 includes: The generator network architecture construction unit is used to construct a generator network architecture according to a convolutional neural network and the multi-scale feature fusion module, and the generator network architecture includes a feature input layer, a plurality of feature convolution extraction layers, a multi-scale fusion layer, and a heat flux field output layer.

[0112] The multi-scale feature extraction unit is used to input the heat flux field correlation feature set through the feature input layer and import it into the plurality of feature convolution extraction layers for multi-scale feature extraction to obtain multi-scale heat flux field feature data.

[0113] A heat flow field simulation data generation unit, configured to splice and fuse the multi-scale heat flow field feature data based on the multi-scale fusion layer, obtain heat flow field simulation data, and output the heat flow field simulation data as an output result through the heat flow field output layer.

[0114] A heat flow field generator construction unit, configured to perform heat flow field simulation training based on the heat flow field correlation feature set and the heat flow field simulation data, and construct the heat flow field generator.

[0115] In some implementation manners, the heat flow field discriminator generation unit in the tool training component 12 includes: A generator network architecture construction unit, configured to construct a generator network architecture, where the generator network architecture includes a data input layer, a plurality of feature convolution extraction layers, a fully connected layer, and a discriminant output layer.

[0116] An associated feature extraction unit, configured to transmit the real aerodynamic heat flow field data set and the simulated heat flow field data set through the data input layer to the plurality of feature convolution extraction layers for associated feature extraction, and obtain heat flow field associated feature data.

[0117] A heat flow field simulation judgment information generation unit, configured to perform identification evaluation and judgment on the heat flow field associated feature data based on the fully connected layer, obtain heat flow field simulation judgment information, and output the heat flow field simulation judgment information as an output result through the discriminant output layer.

[0118] A heat flow field discriminator generation unit, configured to perform evaluation and judgment training based on the real aerodynamic heat flow field data set, the simulated heat flow field data set, and the heat flow field simulation judgment information, and generate the heat flow field discriminator.

[0119] In some embodiments, the mesh division component 13 includes: A heat flow field change feature extraction unit, configured to extract features from the initial distribution characteristics of the target heat flow field to determine heat flow field change features, where the heat flow field change features include temperature gradient, flow field features, and geometric features.

[0120] An initial uniform mesh generation unit, configured to perform mesh initialization according to the heat flow field change features and generate an initial uniform mesh.

[0121] A mesh cell change gradient calculation unit, configured to perform gradient change calculation on each mesh cell in the initial uniform mesh according to the heat flow field change features based on the distribution change data of the target heat flow field, and determine mesh cell change gradient data.

[0122] A heat flow field distribution mesh generation unit, configured to perform adaptive adjustment of the mesh density of the initial uniform mesh based on the mesh cell change gradient data to obtain the heat flow field distribution mesh.

[0123] In some embodiments, the meshing component 13 further includes: A grid cell density distribution information determination unit, configured to determine grid cell density distribution information according to the heat flow field distribution grid.

[0124] A grid region clustering analysis unit, configured to perform region clustering analysis on the heat flow field distribution grid by using the grid cell density distribution information to obtain a grid region clustering result.

[0125] A reconstruction region segmentation threshold setting unit, configured to set a reconstruction region segmentation threshold according to the heat flow field reconstruction accuracy requirement.

[0126] A hierarchical reconstruction region division unit, configured to perform reconstruction region division on the grid region clustering result based on the reconstruction region segmentation threshold to obtain the N hierarchical reconstruction regions.

[0127] In some embodiments, the balance training component 14 includes: A loss function design unit, configured to design a generator mean square loss function and a discriminator cross entropy loss function according to the heat flow field generator and the heat flow field discriminator.

[0128] An initial region training balance coefficient determination unit, configured to perform balance coefficient analysis on the N hierarchical reconstruction regions based on the balance training strategy to determine N initial region training balance coefficients.

[0129] A heat flow field reconstruction network training unit, configured to alternately balance-train the heat flow field generator and the heat flow field discriminator based on the N initial region training balance coefficients by using the generator mean square loss function and the discriminator cross entropy loss function to obtain N heat flow field reconstruction networks.

[0130] In some implementation manners, the heat flow field reconstruction network training unit in the balance training component 14 includes: An initial generator and discriminator training unit, configured to perform simulated discrimination training on the heat flow field generator and the heat flow field discriminator respectively based on the N initial region training balance coefficients to obtain N initial generators and N initial discriminators.

[0131] A generator and discriminator loss evaluation unit, configured to perform loss evaluation on the N initial generators and N initial discriminators by using the generator mean square loss function and the discriminator cross entropy loss function to obtain N generator loss data and N discriminator loss data.

[0132] The heat flow field reconstruction network optimization unit is used to perform balanced alternating training on the N initial generators and N initial discriminators based on the N generator loss data and N discriminator loss data, and obtain N heat flow field reconstruction networks.

[0133] In some implementation manners, the heat flow field reconstruction network optimization unit in the balanced training component 14 further includes: The regional training balance dynamic coefficient adjustment unit is used to dynamically adjust the N initial regional training balance coefficients based on the N generator loss data and N discriminator loss data, and obtain N regional training balance dynamic coefficients.

[0134] The regional generator and discriminator iterative update unit is used to perform iterative balanced update on the N initial generators and N initial discriminators according to the N regional training balance dynamic coefficients, and obtain N regional generators and N regional discriminators.

[0135] The heat flow field reconstruction network obtaining unit is used to concatenate and combine the N regional generators and N regional discriminators to obtain the N heat flow field reconstruction networks.

[0136] It should be understood that the key point of the embodiments mentioned in this specification lies in their differences from other embodiments. The specific embodiments in the foregoing Embodiment 1 are equally applicable to the GAN-based high-speed aerodynamic heat flow field reconstruction system described in Embodiment 2. For the sake of simplicity of the specification, no further elaboration is made here.

[0137] It should be understood that the disclosed embodiments of the present invention and the above descriptions can enable those skilled in the art to implement the present invention using the present invention. At the same time, the present invention is not limited to the part of the embodiments mentioned above. It should be understood that those of ordinary skill in the art can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A high-speed aerodynamic thermal flow field reconstruction method based on GAN, characterized in that: include: Acquiring a real aerodynamic thermal flow field data set, wherein the real aerodynamic thermal flow field data set is historical thermal flow field data of a high-speed aircraft under different flight conditions; Based on the real aerodynamic thermal flow field data set, feature extraction training and simulation discrimination training are performed to construct a thermal flow field generator and a thermal flow field discriminator; Performing grid initialization and adaptive adjustment according to distribution change data of the target thermal flow field to obtain a thermal flow field distribution grid, and dividing the target thermal flow field into reconstruction areas based on the thermal flow field distribution grid to obtain N hierarchical reconstruction areas; Constructing a balanced training strategy, using the balanced training strategy to perform alternating balanced training on the thermal flow field generator and the thermal flow field discriminator based on the N hierarchical reconstruction regions, and building N thermal flow field reconstruction networks; The flight condition update data is monitored and acquired, and the thermal flow field reconstruction optimization is performed on the flight condition update data based on the N thermal flow field reconstruction networks to generate a high-speed aerodynamic simulation thermal flow field.

2. The high-speed aerodynamic thermal flow field reconstruction method based on GAN according to claim 1, characterized in that: The construction of the thermal flow field generator and the thermal flow field discriminator includes: Performing Gaussian filtering and data alignment processing on the real aerodynamic thermal flow field data set to obtain a usable real aerodynamic thermal flow field data set; Extracting correlation features from the available real aerodynamic thermal flow field data set according to thermal flow field reconstruction correlation factors to obtain a thermal flow field correlation feature set; A multi-scale feature fusion module is introduced, and a convolutional neural network and the multi-scale feature fusion module are used to simulate and train the thermal flow field associated feature set to construct a thermal flow field generator; The thermal flow field generator outputs a simulated thermal flow field data set, performs thermal flow field discrimination training based on the real aerodynamic thermal flow field data set and the simulated thermal flow field data set, and generates a thermal flow field discriminator.

3. The high-speed aerodynamic thermal flow field reconstruction method based on GAN as claimed in claim 2, characterized in that: The construction of the thermal flow field generator comprises: According to the convolutional neural network and the multi-scale feature fusion module, a generator network architecture is constructed, wherein the generator network architecture includes a feature input layer, multiple feature convolution extraction layers, a multi-scale fusion layer and a thermal flow field output layer; The thermal flow field associated feature set is introduced into the plurality of feature convolution extraction layers through the feature input layer to perform multi-scale feature extraction to obtain multi-scale thermal flow field feature data; Based on the multi-scale fusion layer, the multi-scale thermal flow field feature data is spliced ​​and fused to obtain thermal flow field simulation data, and output as an output result through the thermal flow field output layer; Thermal flow field simulation training is performed based on the thermal flow field associated feature set and the thermal flow field simulation data to construct the thermal flow field generator.

4. The high-speed aerodynamic thermal flow field reconstruction method based on GAN as claimed in claim 2, characterized in that: The generating of the thermal flow field discriminator comprises: Building a generator network architecture, the generator network architecture includes a data input layer, multiple feature convolution extraction layers, a fully connected layer and a discriminant output layer; The real aerodynamic thermal flow field data set and the simulated thermal flow field data set are transmitted to the multiple feature convolution extraction layers through the data input layer to extract associated features, so as to obtain thermal flow field associated feature data; Based on the fully connected layer, the thermal flow field associated feature data is identified, evaluated and judged to obtain thermal flow field simulation judgment information, and outputted as an output result through the discrimination output layer; Based on the real aerodynamic thermal flow field data set and the simulated thermal flow field data set, as well as the thermal flow field simulation judgment information, evaluation and judgment training is performed to generate the thermal flow field discriminator.

5. The high-speed aerodynamic thermal flow field reconstruction method based on GAN according to claim 1, characterized in that: The step of obtaining a heat flow field distribution grid comprises: Extracting features of the initial distribution characteristics of the target thermal flow field to determine the thermal flow field change features, wherein the thermal flow field change features include temperature gradient, flow field features, and geometric features; Initializing the grid according to the variation characteristics of the thermal flow field to generate an initial uniform grid; According to the change characteristics of the thermal flow field and based on the distribution change data of the target thermal flow field, a gradient change calculation is performed on each grid unit in the initial uniform grid to determine the grid unit change gradient data; The grid density of the initial uniform grid is adaptively adjusted based on the grid unit change gradient data to obtain the thermal flow field distribution grid.

6. The high-speed aerodynamic thermal flow field reconstruction method based on GAN according to claim 1, characterized in that: The N hierarchical reconstruction regions are obtained, including: Determining grid unit density distribution information according to the thermal flow field distribution grid; Using the grid unit density distribution information to perform regional clustering analysis on the thermal flow field distribution grid to obtain a grid regional clustering result; According to the accuracy requirements of thermal flow field reconstruction, set the reconstruction area segmentation threshold; The grid region clustering result is divided into reconstruction regions based on the reconstruction region segmentation threshold to obtain the N hierarchical reconstruction regions.

7. The high-speed aerodynamic thermal flow field reconstruction method based on GAN as claimed in claim 1, characterized in that: The construction of N thermal flow field reconstruction networks includes: According to the thermal flow field generator and the thermal flow field discriminator, a generator mean square loss function and a discriminator cross entropy loss function are designed; Based on the balance training strategy, balance coefficients of the N hierarchical reconstruction regions are analyzed to determine N initial region training balance coefficients; The generator mean square loss function and the discriminator cross entropy loss function are used to perform alternating balanced training on the thermal flow field generator and the thermal flow field discriminator based on the N initial region training balance coefficients to obtain N thermal flow field reconstruction networks.

8. The GAN-based high-speed aerodynamic thermal flow field reconstruction method according to claim 7, characterized in that: The N thermal flow field reconstruction networks are obtained, including: Based on the N initial region training balance coefficients, respectively, the thermal flow field generator and the thermal flow field discriminator are subjected to simulation discrimination training to obtain N initial generators and N initial discriminators; Using the generator mean square loss function and the discriminator cross entropy loss function to perform loss evaluation on the N initial generators and the N initial discriminators, to obtain N generator loss data and N discriminator loss data; Based on the N generator loss data and the N discriminator loss data, the N initial generators and the N initial discriminators are trained in a balanced alternating manner to obtain N thermal flow field reconstruction networks.

9. The high-speed aerodynamic thermal flow field reconstruction method based on GAN as claimed in claim 8, characterized in that: The step of obtaining N thermal flow field reconstruction networks comprises: Based on the N generator loss data and the N discriminator loss data, dynamically adjusting the N initial regional training balance coefficients to obtain N regional training balance dynamic coefficients; Iteratively balancing and updating the N initial generators and the N initial discriminators according to the N regional training balanced dynamic coefficients to obtain N regional generators and N regional discriminators; The N region generators and the N region discriminators are combined in series to obtain the N thermal flow field reconstruction networks.

10. A high-speed aerodynamic thermal flow field reconstruction system based on GAN, characterized in that: The method for reconstructing a high-speed aerodynamic thermal flow field based on GAN according to any one of claims 1 to 9 comprises: A real data acquisition component is used to acquire a real aerodynamic thermal flow field data set, wherein the real aerodynamic thermal flow field data set is historical thermal flow field data of a high-speed aircraft under different flight conditions; A tool training component, used for performing feature extraction training and simulation discrimination training based on the real aerodynamic thermal flow field data set, and constructing a thermal flow field generator and a thermal flow field discriminator; A grid division component, used for performing grid initialization and adaptive adjustment according to the distribution change data of the target thermal flow field to obtain a thermal flow field distribution grid, and performing reconstruction area division on the target thermal flow field based on the thermal flow field distribution grid to obtain N hierarchical reconstruction areas; A balanced training component, used for constructing a balanced training strategy, using the balanced training strategy to perform alternating balanced training on the thermal flow field generator and the thermal flow field discriminator based on the N hierarchical reconstruction regions, and building N thermal flow field reconstruction networks; The updating and optimizing component is used to monitor and obtain the flight condition updating data, and to optimize the thermal flow field reconstruction of the flight condition updating data based on the N thermal flow field reconstruction networks to generate a high-speed aerodynamic simulation thermal flow field.

Images