CN117910140A - Aerodynamic and electromagnetic coupling optimization design method for aircraft based on hybrid conditional generative adversarial network - Google Patents

Abstract

The present invention belongs to the technical field of aircraft design optimization, specifically, an aircraft aerodynamic electromagnetic coupling optimization design method based on a hybrid conditional generative adversarial network, comprising the following steps: step S1: determining the optimization state, optimization target and constraint; step S2: selecting a reference shape, parameterizing the reference aircraft shape, determining the design variables, design space, and extracting samples; step S3: performing aerodynamic stealth performance calculations on sample points to obtain a sample aerodynamic stealth data set; step S4: constructing and verifying an MCGAN model based on the sample aerodynamic stealth data set; step S5: determining a specific optimization mathematical model, using a particle swarm algorithm to perform optimization and solution based on the MCGAN model obtained in step S4, and obtaining an optimized aircraft shape. The MCGAN model of the present invention can have high-precision prediction capabilities under small samples, and the small sample quantity requirement can reduce the calculation cost, thereby being able to directly perform aerodynamic electromagnetic coupling optimization design on the three-dimensional shape of the aircraft.

Description

Aerodynamic and electromagnetic coupling optimization design of aircraft based on hybrid conditional generative adversarial network Method

Technical Field

The present invention relates to the technical field of aircraft design optimization, and specifically to an aircraft aerodynamic electromagnetic coupling optimization design method based on a mixed condition generative adversarial network.

Background technique

With the rapid development of aviation technology, the requirements for aircraft stealth performance are getting higher and higher. The design of aircraft needs to be carried out under the requirements of aerodynamics and stealth. In order to solve the contradiction between high aerodynamic performance and high stealth performance, designers widely use various methods to integrate aerodynamic stealth design of aircraft to achieve a higher performance level.

At present, the aerodynamic stealth optimization design method based on intelligent optimization algorithm (such as genetic algorithm, particle swarm algorithm, etc.) and proxy model is mainly used to perform aerodynamic electromagnetic coupling design on aircraft aerodynamics, and corresponding effects have been achieved. However, there are still some shortcomings, mainly reflected in: 1. It is limited to the optimization of two-dimensional airfoils. In the actual flight process, the performance of the aircraft as a three-dimensional entity will still be different from that of the two-dimensional airfoil, such as the cross-flow effect on the wing surface. Therefore, the result of two-dimensional airfoil optimization cannot represent the final performance of the aircraft; 2. Commonly used proxy models, such as interpolation proxy models (Kriging proxy model, RBF proxy model) and regression proxy models (LS-SVR proxy model), have the problems of low design dimension, poor generalization ability and large training calculation amount. Once the direct three-dimensional shape optimization design is considered, the number of design variables is large, and the prediction accuracy of the above commonly used proxy models will be significantly reduced, resulting in low design accuracy; in addition, due to the large number of design variables, the number of sample points required to construct the proxy model is also increased accordingly. If a high-precision aerodynamic solver and stealth solver are used, the computational cost of constructing the above proxy model is very high, resulting in low design efficiency.

Summary of the invention

In order to solve the above problems existing in the prior art, the present invention proposes an aircraft aerodynamic electromagnetic coupling optimization design method based on a mixed conditional generative adversarial network. By exploring the influence of sample number, learning rate and batch size on the neural network model, a mixed conditional generative adversarial network (MCGAN) model is constructed. The MCGAN model can directly and efficiently perform aerodynamic electromagnetic coupling optimization design for the three-dimensional aircraft shape and obtain good design results.

The present invention is achieved through the following technical solutions:

A method for optimizing the aerodynamic and electromagnetic coupling of an aircraft based on a hybrid conditional generative adversarial network comprises the following steps:

Step S1: Determine the optimization state, optimization target and constraints according to the specific problem of airfoil aerodynamic stealth optimization;

Step S2: selecting a reference aircraft shape, generating a grid around the reference aircraft shape, parameterizing the reference aircraft shape, determining design variables and a design space, and sampling in the design space to obtain a sample;

Step S3: performing aerodynamic stealth performance calculation on the aircraft shape sample obtained in step S2 to obtain an aerodynamic stealth data set of the sample;

Step S4: Based on the sample aerodynamic stealth dataset obtained in step S3, the MCGAN model is constructed by exploring the influence of sample number, learning rate, and batch size on the neural network model, and its reliability is verified;

Step S5: According to the optimization objectives and constraints determined in step S1, a specific optimization mathematical model is determined, and an intelligent optimization algorithm is used to search for an optimal solution in the design space based on the MCGAN model obtained in step S4 to obtain the optimized aircraft shape.

Furthermore, the optimization targets in step S1 include: lift target, drag target, pitch moment target, and radar cross-sectional area target.

Furthermore, step S3 includes the following sub-steps:

Step S31: Based on the RANS method, numerically calculate the aerodynamic characteristics of the aircraft shape in the sample, store the design variables and the calculated aerodynamic data, and obtain the aerodynamic data set of the sample point;

Step S32: Based on the physical optics method, perform stealth calculation on the aircraft shape in the sample, add the calculated stealth data to the aerodynamic data set of the aircraft shape obtained in step S31, and obtain the aerodynamic stealth data set of the aircraft shape.

Furthermore, the aerodynamic data in step S31 includes lift coefficient, drag coefficient, and pitch moment coefficient; and the stealth data in step S32 is radar cross-sectional area.

Furthermore, the MCGAN model constructed in step S4 includes: a generator and a discriminator; the generator is used to generate aerodynamic stealth target parameters, and the discriminator is used to discriminate the authenticity of the aerodynamic stealth target parameters generated by the generator; the specific construction process is as follows:

The training of the model follows the game theory, and the formula is expressed as:

Where D is the discriminator, G is the generator, x is the aerodynamic and stealth target parameters generated by the generator, y and z are the input parameters of the generator, y is the design variable, i.e., the physical constraint, z is the noise, P _data is the original data distribution, P _z is the noise distribution. For a given generator, it is necessary to maximize the discriminator, i.e., the mathematical expectation V(D,G), and integrate V(D,G):

To simplify the expression, let a = P _data (x|y), b = P _G (x|y), D = D (x|y), and maximize the discriminator, that is, maximize the function f (D) = alog (D) + blog (1-D), so we get:

Solutions have to

When and only when P _data (x|y) = P _G (x|y), the discriminator D obtains the optimal solution. At this time, the generator G is the optimal solution of the binary minimax game. At this time, the corresponding x is the target parameter value predicted according to the physical constraint y.

Furthermore, the reliability evaluation indicators for verifying the MCGAN model in step S4 are mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and correlation coefficient (R ^{^2} ); the specific definitions are as follows:

The mean absolute error (MAE) is defined as:

The mean absolute percentage error (MAPE) is defined as:

The root mean square error (RMSE) is defined as:

The correlation coefficient (R ^{^2} ) is defined as:

Where N is the number of data points, _yi and are the true value and the model predicted value respectively; SSR and SST are the regression sum of squares and the total sum of squares respectively; the evaluation criteria are MAE, MAPE, and RMSE. The closer they are to 0 and R ^{^2} is to 1, the better the predictive ability of the model.

Furthermore, the intelligent optimization algorithm used in step S5 is a particle swarm algorithm;

Furthermore, the present invention also provides a computer device, comprising: a processor, a storage medium and a bus, wherein the storage medium stores program instructions executable by the processor, and when the computer device is running, the processor and the storage medium communicate through the bus, and when the processor executes the program instructions, it is used to implement the aircraft aerodynamic electromagnetic coupling optimization design method based on the mixed condition generation adversarial network.

Furthermore, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a computer, it is used to implement the aircraft aerodynamic electromagnetic coupling optimization design method based on the mixed condition generative adversarial network.

Beneficial Effects

The present invention provides an aircraft aerodynamic electromagnetic coupling optimization design method based on a mixed conditional generative adversarial network. The present invention constructs a mixed conditional generative adversarial network (MCGAN) model by exploring the influence of sample quantity, learning rate and batch size on the neural network model. The model is based on binary minimax game theory, and obtains the optimal solution of the generator through dynamic game between a generator and a discriminator. The model can predict high-dimensional and multi-objective aerodynamic stealth performance parameters under small sample conditions, thereby solving high-dimensional constraints and multi-objective optimization problems. Compared with traditional proxy models, the model can have high-precision prediction capabilities under small samples. The reduction in the number of samples required can reduce the computational cost of the optimization design, thereby greatly improving the design efficiency while ensuring the design accuracy, shortening the design cycle, and then being able to directly and efficiently perform aerodynamic electromagnetic coupling optimization design for the three-dimensional aircraft shape.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for optimizing aerodynamic and electromagnetic coupling design of an aircraft shape according to an embodiment of the present invention;

FIG2 is a schematic diagram of an FFD frame and design variables according to an embodiment of the present invention;

FIG3 is a schematic diagram of a calculation flow of aerodynamic stealth performance of an aircraft according to an embodiment of the present invention;

FIG4 is a flow chart of establishing an MCGAN model according to an embodiment of the present invention;

FIG5 is a schematic diagram of the composition of the MCGAN model according to an embodiment of the present invention;

FIG6 is a schematic diagram showing the effect of sample size on model loss provided by an embodiment of the present invention;

FIG7 is a schematic diagram showing the effect of sample size on the root mean square error of a model provided by an embodiment of the present invention;

FIG8 is a schematic diagram showing the effect of the learning rate lr on the loss of the model provided by an embodiment of the present invention;

FIG9 is a schematic diagram showing the effect of the learning rate lr on the root mean square error of the model provided by an embodiment of the present invention;

FIG10 is a schematic diagram showing the effect of batch size bs on the loss of a model provided by an embodiment of the present invention;

FIG11 is a schematic diagram showing the effect of batch size bs on the root mean square error of a model provided by an embodiment of the present invention;

FIG12 is an optimization result of the first optimization case provided by an embodiment of the present invention;

Among them, Figure 12(a) is a cloud diagram of the overall pressure coefficient of the aircraft; Figure 12(b) is a distribution curve of the wing root pressure coefficient; Figure 12(c) is a contour diagram of the pressure coefficient on the upper surface of the aircraft; Figure 12(d) is a distribution curve of the mid-wing pressure coefficient; Figure 12(e) is a schematic diagram of RCS; Figure 12(f) is a distribution curve of the wing tip pressure coefficient;

FIG13 is an optimization result of the second optimization case provided by an embodiment of the present invention;

Among them, Figure 13 (a) is a cloud diagram of the overall pressure coefficient of the aircraft; Figure 13 (b) is a distribution curve of the wing root pressure coefficient; Figure 13 (c) is a contour diagram of the pressure coefficient on the upper surface of the aircraft; Figure 13 (d) is a distribution curve of the mid-wing pressure coefficient; Figure 13 (e) is a schematic diagram of RCS; Figure 13 (f) is a distribution curve of the wing tip pressure coefficient;

Detailed ways

In order to make the technical problems, technical solutions and beneficial effects solved by the present invention clearer and more understandable, and to enable those skilled in the art to better understand the solutions of the present invention, the present invention is further described and fully described in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

As shown in FIG1 , this embodiment uses the aircraft aerodynamic electromagnetic coupling optimization design method based on the hybrid conditional generative adversarial network proposed in the present invention to directly carry out a three-dimensional shape aerodynamic electromagnetic coupling optimization design for a certain aircraft shape, including the following steps:

Step S1: Determine the optimization state, optimization target and constraints according to the specific problem of airfoil aerodynamic stealth optimization;

In this embodiment, the optimization state is Ma=0.8, AOA=2°, and the optimization targets are lift coefficient, drag coefficient, pitch moment coefficient and radar scattering area;

Step S2: selecting a reference aircraft shape, generating a grid around the reference aircraft shape, parameterizing the reference aircraft shape, determining design variables and a design space, and sampling in the design space to obtain a sample;

In this embodiment, a certain flying wing layout aircraft is selected as the reference aircraft shape. Since the selected reference aircraft is bilaterally symmetrical, a half mold is taken, and five sections are selected as characteristic sections along the span direction of the aircraft using the FFD parameterization method to parameterize the three-dimensional shape. The characteristic sections and control points are shown in FIG2 . The Z-coordinates z _i , i=1…n of the nodes of the FFD control frame at the five sections are changed by Δz _i , i＝1…n as design variables, and establish a mapping between the surface grid and the design variables, where the subscript i represents the i-th control node, and n is the number of all FFD box control nodes; the coordinate axis is defined as: the chord line of the wing of the X-axis symmetry plane points to the flow direction, the Y-axis is perpendicular to the symmetry plane and points to the right wing, and the Z-axis is in the symmetry plane, perpendicular to the X-axis and points upward, satisfying the right-hand system; wherein, each section has 5 design variables on the upper and lower surfaces, a total of 50 design variables, that is, the optimization problem of this embodiment is 50-dimensional, which is a typical high-dimensional design variable problem; the design variable range is: ±0.04, forming a design space; then Latin hypercube sampling is used in the design space to obtain 5000 different aircraft shape samples;

By changing the design variables, the surface mesh is subjected to deformation disturbance, so that the surface mesh of the new shape is deformed. Then, the radial basis function (RBF)-infinite interpolation (TFI) mesh deformation method is used to interpolate and deform the space mesh based on the surface mesh of the new shape to obtain the CFD calculation mesh of the new shape.

Step S3: Carry out aerodynamic stealth performance calculation on the aircraft shape sample obtained in step S2 to obtain an aerodynamic stealth data set of the sample; as shown in FIG3 , it specifically includes the following sub-steps:

Step S31: Based on the RANS method, numerically calculate the aerodynamic characteristics of the aircraft shape in the sample, store the design variables and the calculated aerodynamic data, and obtain the aerodynamic data set of the sample point; the aerodynamic data includes the lift coefficient, the drag coefficient, and the pitch moment coefficient;

In this embodiment, the finite volume method is used to solve the three-dimensional NS equation, and the three-dimensional NS equation is:

Among them, ξ,η,ζ are generalized coordinates, is the solution vector, /> is the inviscid flux vector, /> is the viscous flux vector;

In this embodiment, the convection term and pressure term of the RANS equation adopt the third-order upwind bias, the time-marching format adopts the implicit approximate factorization method, the turbulence model adopts SA (Spalart-Allmaras), and the calculation adopts the multi-grid acceleration technology to accelerate convergence;

Step S32: Based on the physical optics method, a stealth calculation is performed on the aircraft shape in the sample, and the calculated stealth data is added to the aerodynamic data set of the aircraft shape obtained in step S31 to obtain an aerodynamic stealth data set of the aircraft shape; the stealth data is a radar cross-sectional area;

In this embodiment, the RCS calculation expression is:

Where k = 2π/λ is the wave number, λ is the wavelength, is the unit outward direction vector at the object surface S,/> The unit vector of the receiving device's electric polarization direction, /> is the magnetic field polarization direction, /> The incident direction unit vector, /> The unit vector of the scattering direction;

RCS has the dimension of area. For convenience, it is usually given in logarithmic form. The expression is:

RCS _dBsm = 10log ₁₀ RCS _m2

In this embodiment, the aircraft yaw angle ±60°, i.e., [0°, 60°] and [300°, 360°], is taken as the main threat area for stealth performance optimization;

Step S4: Based on the sample aerodynamic stealth dataset obtained in step S3, the MCGAN model is constructed by exploring the influence of sample number, learning rate, and batch size on the neural network model, and its reliability is verified;

In this embodiment, as shown in FIG. 4 and FIG. 5 , the MCGAN model includes a generator and a discriminator; the generator is used to generate aerodynamic stealth target parameters, and the discriminator is used to discriminate the authenticity of the aerodynamic stealth target parameters generated by the generator; the specific construction process is as follows:

The training of the model follows the game theory, and the formula is expressed as:

Where D is the discriminator, G is the generator, x is the aerodynamic and stealth target parameters generated by the generator, y and z are the input parameters of the generator, y is the design variable, i.e., the physical constraint, z is the noise, P _data is the original data distribution, P _z is the noise distribution. For a given generator, it is necessary to maximize the discriminator, i.e., the mathematical expectation V(D,G), and integrate V(D,G):

To simplify the expression, let a = P _data (x|y), b = P _G (x|y), D = D (x|y), and maximize the discriminator, that is, maximize the function f (D) = alog (D) + blog (1-D), so we get:

Solutions have to

If and only if P _data (x|y) = P _G (x|y), the discriminator D obtains the optimal solution. At this time, the generator G is the optimal solution of the binary minimax game. At this time, the corresponding x is the target parameter value predicted according to the physical constraint y.

In this embodiment, the detailed configuration of the network structure adopted by the MCGAN model is shown in Table 1 and Table 2. The generator adopts a convolutional pooling layer and a fully connected layer structure, and the discriminator adopts a fully connected layer structure; except for the output layer, the LeakyRelu function is used as the nonlinear activation function of the generator and the discriminator. The LeakyRelu function allows a small non-zero gradient when the input is less than zero, thereby alleviating the gradient vanishing problem in model training; the loss function adopts the mean square error MSE (Mean Squared Error), which, as a convex function, has only one global minimum, which makes the training process more stable and reduces the risk of falling into a local minimum. At the same time, the penalty for large errors is relatively large; the weight calculation of the model is based on the idea of gradient descent, that is:

Where ω is the weight, α is the learning rate, and loss is the loss function.

In terms of optimizer selection, the Adam optimizer, which combines the ideas of the RMSProp and Momentum optimization algorithms, is used, and the initial learning rate is adjusted to 0.0001. Compared with the traditional stochastic gradient descent method, the Adam optimizer can adaptively adjust the learning rate based on historical gradient information, so that a larger learning rate can be used in the early stage of training to converge quickly, and a smaller learning rate can be used in the later stage of training to more accurately find the minimum value of the loss function.

Table 1 Generator hyperparameter selection

Table 2 Discriminator hyperparameter selection

The influence of sample size on the loss and root mean square error (RMSE) of the MCGAN model is shown in Figures 6 and 7. The rule for obtaining the sample size is to randomly select 100, 500, 1500, 2500, 3500 and 4500 sample points from the 5000 sample points obtained in step S2. From the results in the figure, it can be concluded that when the sample size is 100, the loss of the model shows obvious sawtooth oscillation, and the RMSE of the model also has similar performance. Compared with the sample size of 500, the loss of the model converges slightly slower. It can be seen that when the RMSE value drops to a stable value, the iteration cycle of the model with a sample size of 100 is longer. In general, increasing the number of samples is conducive to the model loss reaching a balanced state earlier and reducing the oscillation amplitude of the loss. The RMSE oscillation amplitude is reduced while the number of iterations to reach a stable state is reduced. Increasing the number of samples is conducive to accelerating the convergence speed of the model and stabilizing the training of the model.

The influence of learning rate (lr) on the loss and root mean square error (RMSE) of the model is shown in Figures 8 and 9. From the results in the figures, it can be concluded that when the learning rate is 1e-3, the amplitude of parameter update is too large, causing the model to fail to converge, and the RMSE shows a sudden and substantial increase; when the learning rate is 1e-3, the amplitude of parameter update is too small, the model training is stable, but the convergence is too slow, and the model loss has not converged after 1000 iterations; when the learning rate is 1e-4, the model loss converges stably, the convergence speed is moderate, and the RMSE converges very quickly; Overall, too large or too small learning rates are not conducive to model training. In the actual model training, the learning rate needs to be adjusted for specific data samples so that the model has a faster convergence speed and better stability.

The influence of batch size (bs) on the loss and root mean square error (RMSE) of the model is shown in Figures 10 and 11. From the results in the figures, it can be concluded that when bs is 50, the loss and RMSE of the model both show faster convergence speed and smaller oscillation; when bs is 100, the loss of the model can still converge within 1000 steps, but the convergence speed is slower than when bs is 50, and the RMSE oscillates more strongly in the initial few epochs; when bs is 500, the loss of the model does not converge, the RMSE shows more violent oscillations and the number of epochs to converge increases.

In this embodiment, the reliability evaluation indicators for verifying the MCGAN model are mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and correlation coefficient (R ^{^2} ); the specific definitions are as follows:

The mean absolute error (MAE) is defined as:

The mean absolute percentage error (MAPE) is defined as:

The root mean square error (RMSE) is defined as:

The correlation coefficient (R ^{^2} ) is defined as:

Where N is the number of data points, _yi and are the true value and the model predicted value respectively; SSR and SST are the regression sum of squares and the total sum of squares respectively; the evaluation criteria are MAE, MAPE, and RMSE. The closer they are to 0 and R ^{^2} is to 1, the better the predictive ability of the model.

In this embodiment, the data of 5000 samples are randomly divided into 90% training set and 10% test set, the learning rate is determined to be 1e-4, and the batch size is determined to be 50; the prediction performance of the model is shown in Table 3. It can be seen from the table that the average absolute percentage errors of the four performance indicators are all less than 2.2%, and the correlation coefficients are all greater than 0.999. The average absolute error of the drag coefficient is only 2.95 counts, and the root mean square error is only 0.47× ^10-4 . For the problem of aircraft shape optimization, its proxy accuracy is very high.

Table 3 Model prediction performance

Step S5: According to the optimization objectives and constraints determined in step S1, a specific optimization mathematical model is determined, and an intelligent optimization algorithm is used to search for an optimal solution in the design space based on the MCGAN model obtained in step S4 to obtain the optimized aircraft shape.

In this embodiment, two specific optimization mathematical models are determined to be case 1 and case 2, as shown in Table 4, where case 1 is to minimize the radar scattering area while constraining the aerodynamic performance, and case 2 is to improve the lift coefficient while constraining its stealth performance; the particle swarm optimization algorithm is used to optimize and solve the above two specific optimization mathematical models in the design space based on the aerodynamic stealth proxy model obtained in step S4 to obtain the optimized aircraft shape;

Table 4 Specific optimization mathematical model

The optimization results of case 1 are shown in Figure 12. From the results in the figure, it can be concluded that after the aircraft is optimized by the MCGAN agent, the maximum peak value of RCS in the main threat area, that is, within the yaw angle of ±60°, is reduced, but the cost is that the electromagnetic scattering characteristics of the tail of the aircraft are improved as a whole, except at the valley value of RCS, the valley value is further reduced; the difference in pressure distribution is more obvious in the middle of the wing. The suction peak of the leading edge of the optimized airfoil is reduced, the air flow rate is reduced, it is not easy to generate separation vortex, and the shock wave position is moved backward, making it easier for the aircraft to maintain balance and stability. The area enclosed by the pressure area is increased, and the lift coefficient is increased. The comparison of the aerodynamic stealth performance of case 1 before and after optimization is shown in Table 5. It can be seen that the radar scattering area is reduced by 0.021dBsm, the lift coefficient is increased by 0.0107, the drag coefficient is increased by 7counts, the pitch moment coefficient is reduced by 0.0007, and the moment characteristics are improved.

Table 5 Comparison of optimization results of case 1

The optimization results of case 2 are shown in Figure 13. From the results in the figure, it can be concluded that the airfoil profiles before and after optimization are not much different, and the RCS value in the main threat area is basically unchanged, but the electromagnetic scattering characteristics of the tail of the aircraft are improved overall. Compared with case 1, the shock wave intensity is basically unchanged, and the area surrounded by the three pressures of the wing root, wing center, and wing tip is slightly increased; the aerodynamic stealth performance comparison of case 2 before and after optimization is shown in Table 6. The radar scattering area increases by 0.0009dBsm, the lift coefficient increases by 0.0043, the drag coefficient increases by 2counts, and the pitch moment coefficient decreases by 0.0001.

Table 6 Comparison of optimization results of case 2

Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, substitute and modify the above embodiments within the scope of the present invention without departing from the principles and purpose of the present invention.

Claims (9)

1. An aircraft aerodynamic electromagnetic coupling optimization design method based on a hybrid conditional generative adversarial network is characterized by comprising the following steps: Step S1: Determine the optimization state, optimization target and constraints according to the specific problem of airfoil aerodynamic stealth optimization; Step S2: selecting a reference aircraft shape, generating a grid around the reference aircraft shape, parameterizing the reference aircraft shape, determining design variables and a design space, and sampling in the design space to obtain a sample; Step S3: performing aerodynamic stealth performance calculation on the aircraft shape sample obtained in step S2 to obtain an aerodynamic stealth data set of the sample; Step S4: Based on the sample aerodynamic stealth dataset obtained in step S3, the MCGAN model is constructed by exploring the influence of sample number, learning rate, and batch size on the neural network model, and its reliability is verified; Step S5: According to the optimization objectives and constraints determined in step S1, a specific optimization mathematical model is determined, and an intelligent optimization algorithm is used to search for an optimal solution in the design space based on the MCGAN model obtained in step S4 to obtain the optimized aircraft shape. 2. According to claim 1, the aircraft aerodynamic electromagnetic coupling optimization design method based on a hybrid condition generative adversarial network is characterized in that the optimization objectives in step S1 include: lift target, drag target, pitch moment target, and radar scattering cross-sectional area target. 3. The aircraft aerodynamic electromagnetic coupling optimization design method based on hybrid conditional generative adversarial network according to claim 1, characterized in that step S3 comprises the following sub-steps: Step S31: Based on the RANS method, numerically calculate the aerodynamic characteristics of the aircraft shape in the sample, store the design variables and the calculated aerodynamic data, and obtain the aerodynamic data set of the sample point; Step S32: Based on the physical optics method, perform stealth calculation on the aircraft shape in the sample, add the calculated stealth data to the aerodynamic data set of the aircraft shape obtained in step S31, and obtain the aerodynamic stealth data set of the aircraft shape. 4. According to claim 3, the aircraft aerodynamic electromagnetic coupling optimization design method based on a hybrid condition generative adversarial network is characterized in that the aerodynamic data includes lift coefficient, drag coefficient, and pitch moment coefficient; and the stealth data is radar scattering cross-sectional area. 5. According to the method for optimizing the aerodynamic electromagnetic coupling of an aircraft based on a hybrid conditional generative adversarial network in claim 1, it is characterized in that the MCGAN model constructed in step S4 comprises: a generator and a discriminator; the generator is used to generate aerodynamic stealth target parameters, and the discriminator is used to discriminate the authenticity of the aerodynamic stealth target parameters generated by the generator; the specific construction process is as follows: The training of the model follows the game theory, and the formula is expressed as: Where D is the discriminator, G is the generator, x is the aerodynamic and stealth target parameters generated by the generator, y and z are the input parameters of the generator, y is the design variable, i.e., the physical constraint, z is the noise, P _data is the original data distribution, P _z is the noise distribution. For a given generator, it is necessary to maximize the discriminator, i.e., the mathematical expectation V(D,G), and integrate V(D,G): To simplify the expression, let a = P _data (x|y), b = P _G (x|y), D = D (x|y), and maximize the discriminator, that is, maximize the function f (D) = alog (D) + blog (1-D), so we get: Solutions have to When and only when P _data (x|y) = P _G (x|y), the discriminator D obtains the optimal solution. At this time, the generator G is the optimal solution of the binary minimax game. At this time, the corresponding x is the target parameter value predicted according to the physical constraint y. 6. According to the method for optimizing the aerodynamic and electromagnetic coupling of an aircraft based on a hybrid conditional generative adversarial network in claim 1, it is characterized in that the reliability evaluation indicators for verifying the MCGAN model in step S4 are mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and correlation coefficient (R ^{^2} ); the specific definitions are as follows: The mean absolute error (MAE) is defined as: The mean absolute percentage error (MAPE) is defined as: The root mean square error (RMSE) is defined as: The correlation coefficient (R ^{^2} ) is defined as: Where N is the number of data points, _yi and are the true value and the model predicted value respectively; SSR and SST are the regression sum of squares and the total sum of squares respectively; the evaluation criteria are MAE, MAPE, and RMSE. The closer they are to 0 and R ^{^2} is to 1, the better the predictive ability of the model. 7. The aircraft aerodynamic electromagnetic coupling optimization design method based on a hybrid conditional generative adversarial network according to claim 1 is characterized in that the intelligent optimization algorithm used in step S5 is a particle swarm algorithm. 8. A computer device, comprising: a processor, a storage medium and a bus, wherein the storage medium stores program instructions executable by the processor, characterized in that: when the computer device is running, the processor and the storage medium communicate through the bus, and the processor executes the program instructions to implement any one of the methods described in claims 1 to 7. 9. A computer-readable storage medium having a computer program stored thereon, wherein the computer program is used to implement any one of the methods of claim 1 to claim 7 when the computer is executed.