Aircraft aerodynamic characteristics prediction method, device and medium
The MMoE-3D model addresses the limitations of single-task learning by incorporating three-dimensional data and a gating mechanism, enhancing aircraft aerodynamic prediction accuracy and speed through personalized expert contributions.
Patent Information
- Application Number
- CN202210344410.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-31
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2042-03-31
AI Technical Summary
Existing methods for predicting aircraft aerodynamic characteristics, such as BP neural networks, are limited by single-task learning, ignoring three-dimensional information, and are not suitable for handling multiple tasks simultaneously, leading to reduced learning effectiveness and accuracy, especially with insufficient data.
A multi-expert hybrid method (MMoE-3D model) is introduced, which divides the shared layer into multiple experts with a gating mechanism, incorporating three-dimensional data to predict aerodynamic characteristics, allowing personalized combinations of expert contributions and reducing the need for additional model parameters.
The MMoE-3D model enhances prediction accuracy and speed by effectively handling multiple tasks independently, leveraging three-dimensional data to improve convergence and reduce training time, outperforming traditional methods in accuracy and efficiency.
Smart Images

Figure CN114818477B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of aircraft aerodynamic characteristics prediction, and more specifically, to an aircraft aerodynamic characteristics prediction method, device and medium. Background Art
[0002] Structural optimization of a missile refers to designing its aerodynamic layout based on the missile's aerodynamic performance and the expected goals set, that is, adjusting the missile's geometric shape parameters and determining the shape, size and installation position of each aerodynamic component (warhead, body, wings and rudder, etc.) in the missile's mechanical structure.
[0003] The various aerodynamic parameters of a missile during flight can directly reflect the aerodynamic performance of the missile, and the excellent aerodynamic performance can reflect whether the aerodynamic layout of the missile is reasonable and whether it needs further optimization. There are three main traditional methods for directly obtaining the aerodynamic parameters of a missile: theoretical calculation method, wind tunnel test method, and flight test method. The calculation process of the theoretical calculation method is extremely complicated, and the cost of the wind tunnel test method and the flight test method is extremely expensive and the test cycle is very long.
[0004] The traditional engineering fitting method of aerodynamic modeling is to use mathematical methods to obtain the mathematical relationship between aerodynamic force and flight state, aerodynamic shape and other parameters based on the physical characteristics and statistical laws of aerodynamic force. This method is more suitable for local aerodynamic modeling with a small range of state changes. For aircraft with a large range of state parameters and complex aerodynamic shapes, the relationship between aerodynamic force and state parameters and other parameters will be highly nonlinear, and it is difficult to give an accurate description using traditional mathematical modeling methods; with the development of artificial intelligence neural networks, aerodynamic modeling based on neural networks has attracted the research of many scholars and achieved a lot of results. The aerodynamic modeling process of neural networks does not need to consider the nonlinearity between input and output. Through machine learning of a large amount of aerodynamic data, an aerodynamic model is established. It has good learning ability and nonlinear fitting ability, and is an effective method for dealing with nonlinear function approximation.
[0005] Therefore, a prediction method based on artificial intelligence has gradually come into people's view. This method starts from the parameterized geometric shape of the aircraft, uses numerical methods to obtain the aerodynamic characteristics of the aircraft under the typical aerodynamic shape, and establishes neural network models, Kriging models, SVM and other models between the aerodynamic characteristics and the shape parameters. Then, the prediction model is used for optimization design, thus avoiding a large number of numerical calculations. Compared with traditional machine learning, the BP neural network fitting function has the advantage of high accuracy.
[0006] BP neural network is a feedforward neural network, which consists of three layers: input layer, hidden layer, and output layer. Figure 1As shown in the figure, the BP neural network reduces the function value of the error function through two processes: forward transmission of the signal and reverse transmission of the error, and continuously adjusts and optimizes the weights and thresholds. The neural network is composed of a large number of interconnected nodes. Each node represents a specific output function, and the connection between each two nodes represents the weighted value of the signal passing through the connection, that is, the weight. Each layer of nodes weights the input information and performs a nonlinear transformation before outputting it. Its output value is used as the input value of the next layer, and so on until the final output layer.
[0007] The essence of the back propagation method is to minimize the mean square distance between the network class prediction and the actual class label of the data sample. Hornik et al. proved that for hidden layers containing enough neurons, the BP neural network can fit a continuous function of arbitrary complexity with arbitrary precision, and due to the role of weights, the neural network has better robustness in a noisy environment and has good predictive ability for untrained data.
[0008] However, the defects of the common BP neural network method are also prominent. It is only suitable for single-task learning, that is, it can only predict a single task at a time and cannot learn multiple tasks together. Figure 2a As shown in the figure, this method can solve the problem, but this single-task learning method is redundant, ignoring the connection between multiple tasks and losing some information between data or model parameters. Especially when the amount of data for a single task is insufficient, it is difficult for single-task learning to obtain sufficient data distribution information through training, so the learning effect will decline to varying degrees.
[0009] At the same time, the traditional BP neural network prediction method ignores the three-dimensional information of the aircraft. Zhang Ruimin et al. established an aerodynamic model of the aircraft at a high angle of attack based on the BP neural network. Yuan Zhijie et al. used the thinking evolution algorithm to optimize the BP neural network, which has a strong ability to fit the aerodynamic parameters of the missile. The above studies only considered the plane parameters of the aircraft and ignored the importance of three-dimensional data for predicting aerodynamic data.
[0010] Therefore, the present invention considers a prediction method of three-dimensional multi-task learning to solve the above problems. Summary of the invention
[0011] The present invention is provided to solve the above problems existing in the prior art. Therefore, a method, device and medium for predicting aerodynamic characteristics of an aircraft are needed to ensure the prediction speed while improving the prediction accuracy of the aerodynamic data of the aircraft.
[0012] According to a first embodiment of the present invention, there is provided a method for predicting aerodynamic characteristics of an aircraft, the method comprising:
[0013] The three-dimensional coordinate information is expressed by calculating the vertex coordinates of the eight airfoils using formula (1):
[0014]
[0015] Among them, (x α,Φ ,y α,Φ , z α,Φ ) represents the coordinate point with an attack angle of α and a roll angle of φ, and (x0, y0, z0) represents the coordinate point with an attack angle of 0 and a roll angle of 0;
[0016] The MMoE-3D model is constructed according to formula (2)-formula (5):
[0017] y k =h k (f k (x)), k = 1, 2, 3, 4 Formula (2)
[0018]
[0019] g k (x) = Softmax(wC) Formula (4)
[0020] C=Concat(Input1, Input2) Formula (5)
[0021] Among them, e i represents the i-th expert neural network, g(x) represents a gating network that combines the results of all experts, and the k-th output g of g(x) k (x) represents expert e k The probability of the value, and C represents the merging layer, which is used to fuse the aircraft plane parameters, flight conditions and 3D coordinate information. The Softmax activation function is used to generate multiple probability distributions of experts and finally output f k (x) is the sum of the weights of all experts. Each subtask corresponds to a specific task layer. The mapping function of the kth task layer is expressed as h k , the output of each subtask is y k ,The gating network of each task realizes the selective utilization of k experts through the final output of different weights, and the gating networks of different tasks learn different combinations of expert modes;
[0022] The aerodynamic characteristics of the aircraft are predicted by the MMoE-3D model.
[0023] According to a second embodiment of the present invention, there is provided an aircraft aerodynamic characteristics prediction device, the device comprising a processor, the processor being configured to:
[0024] The three-dimensional coordinate information is expressed by calculating the vertex coordinates of the eight airfoils using formula (1):
[0025]
[0026] Among them, (x α,Φ ,y α,Φ , z α,Φ ) represents the coordinate point with an attack angle of α and a roll angle of φ, and (x0, y0, z0) represents the coordinate point with an attack angle of 0 and a roll angle of 0;
[0027] The MMoE-3D model is constructed according to formula (2)-formula (5):
[0028] y k =h k (f k (x)), k = 1, 2, 3, 4 Formula (2)
[0029]
[0030] g k (x) = Softmax(wC) Formula (4)
[0031] C=Concat(Input1, Input2) Formula (5)
[0032] Among them, e i represents the i-th expert neural network, g(x) represents a gating network that combines the results of all experts, and the k-th output g of g(x) k (x) represents expert e k The probability of the value, and C represents the merging layer, which is used to fuse the aircraft plane parameters, flight conditions and 3D coordinate information. The Softmax activation function is used to generate multiple probability distributions of experts and finally output f k (x) is the sum of the weights of all experts. Each subtask corresponds to a specific task layer. The mapping function of the kth task layer is expressed as h k , the output of each subtask is y k ,The gating network of each task realizes the selective utilization of k experts through the final output of different weights, and the gating networks of different tasks learn different combinations of expert modes;
[0033] The aerodynamic characteristics of the aircraft are predicted by the MMoE-3D model.
[0034] According to a third aspect of the present invention, a non-transitory computer-readable storage medium storing instructions is provided. When the instructions are executed by a processor, the method according to various embodiments of the present invention is performed.
[0035] According to the aircraft aerodynamic characteristics prediction method, device and medium of each scheme of the present invention, modeling is carried out through a multi-expert hybrid method, breaking away from the mindset of completely sharing the entire hidden layer, consciously dividing the shared layer into multiple experts, and introducing a gating mechanism. Each task target uses an independent gating weight, so that each expert contributes differently to each target, and the shared layer can be used in a personalized combination. Compared with the shared bottom layer multi-task model, it better handles the problem of differentiation of different targets and does not require more model parameters. Compared with the traditional neural network aerodynamic characteristics prediction method, the present invention adds three-dimensional data to the input data to make the prediction more accurate. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] In the drawings, which are not necessarily drawn to scale, the same reference numerals may describe similar parts in different views. The same reference numerals with letter suffixes or different letter suffixes may represent different instances of similar parts. The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the specification and claims, are used to illustrate the embodiments of the invention. When appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the present apparatus or method.
[0037] Figure 1 Shown is the BP neural network structure diagram.
[0038] Figure 2a A flowchart showing the traditional single-task learning method for predicting aerodynamic data.
[0039] Figure 2b A flowchart showing a conventional multi-task learning method for predicting aerodynamic data.
[0040] Figure 3 A flow chart for predicting aerodynamic data based on the MMoE-3D model is shown.
[0041] Figure 4 The subsonic model convergence process is shown.
[0042] Figure 5 The supersonic model convergence process is shown.
[0043] Figure 6a A randomly selected shape from the test set is shown.
[0044] Figure 6b A comparison of the predicted values and true values of MMoE and MMoE-3D at subsonic speeds is shown.
[0045] Figure 6cThe comparison between the predicted values and the true values of MMoE and MMoE-3D at supersonic speed is shown. DETAILED DESCRIPTION
[0046] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and specific embodiments, but are not intended to limit the present invention. For the various steps described herein, if there is no necessity for a causal relationship between each other, the order in which they are described as examples herein should not be regarded as a limitation, and those skilled in the art should know that they can be adjusted in order, as long as the logic between them is not destroyed, resulting in the inability to implement the entire process.
[0047] Figure 2b FIG. 4 shows a flowchart of a conventional multi-task learning method for predicting aerodynamic data. Figure 2b As shown, the most common multi-task learning idea is to share the most abstract representation layer at the bottom, and then differentiate different tasks at the upper layer. This method of sharing the bottom layer works better for tasks with high correlation, but this method easily causes a certain task goal to be greatly affected by another task goal. The present invention believes that learning to predict aerodynamic data should try to prevent multiple tasks from affecting each other. In other words, even if the learning effect of a certain coordinate is poor due to interference, it will not cause the learning effect of other coordinates to decline.
[0048] Therefore, the embodiment of the present invention adopts a multi-expert hybrid method for modeling, breaking away from the mindset of completely sharing the entire hidden layer, consciously dividing the shared layer into multiple experts, and introducing a gating mechanism. Each task target uses an independent gating weight, so that each expert contributes differently to each target, and the shared layer can be used in personalized combination. Compared with the shared underlying multi-task model, it better handles the problem of differentiation of different targets and does not require more model parameters.
[0049] Considering that the original MMoE model only considers the plane parameters and flight data of the aircraft, the embodiment of the present invention proposes a method for predicting the aerodynamic characteristics of an aircraft based on the original MMoE model. In the aerodynamic prediction part, four tasks can be given. The input is the aerodynamic shape x of the missile, the coordinate information a and the flight condition c. The embodiment of the present invention calculates the vertex coordinates of eight wing surfaces to represent the three-dimensional information, and the coordinates can be calculated as:
[0050]
[0051] Among them, (x α,Φ ,y α,Φ , z α,Φ) represents the coordinate point with an attack angle of α and a roll angle of φ, and (x0, y0, z0) represents the coordinate point with an attack angle of 0 and a roll angle of 0;
[0052] Figure 3 The schematic diagram of the structure of the MMoE-3D model is shown in FIG. Figure 3 As shown, the MMoE-3D model is constructed according to formula (2)-formula (5):
[0053] y k =h k (f k (x)), k = 1, 2, 3, 4 Formula (2)
[0054]
[0055] g k (x) = Softmax(wC) Formula (4)
[0056] C=Concat(Input1, Input2) Formula (5)
[0057] Among them, e i represents the i-th expert neural network. There are 16 expert neural networks in this task. The k-th output g of g(x) k (x) represents expert e k The probability of the value, and g(x) represents a gating network that aggregates the results of all experts. Specifically, the gating network g(x) generates weight distribution on the i-th expert according to the input, and the final output is the weighted sum of all expert outputs. The merging layer c is linearly transformed by dot multiplication with a trainable matrix w. The merging layer c is used to fuse the aircraft plane parameters, flight conditions, and 3D coordinate information. The Softmax activation function is used to generate 16 probability distributions of the experts, and the final output is f k (x) is the sum of the weights of all experts. Each subtask corresponds to a specific task layer, and the mapping function of the kth task layer is expressed as h k , the output of each subtask is y k The gating network of each task realizes the selective utilization of k experts through the final output with different weights, and the gating networks of different tasks can learn different combinations of expert modes.
[0058] Finally, the MMoE-3D model is used to predict the aerodynamic characteristics of the aircraft.
[0059] In some embodiments, the method used to train the MMoE-3D model is back-propagation, and the loss function can be written as:
[0060]
[0061] Where n is the number of training sets, y is the aerodynamic data corresponding to the missile shape, is the predicted value of the MMoE-3D model.
[0062] Exemplarily, the specific parameter settings of the MMoE-3D model are shown in Table 1.
[0063] Table 1. Specific parameter settings of MMoE-3D model
[0064]
[0065]
[0066] In some implementations, separate models are constructed for subsonic and supersonic speeds, respectively. The prediction results of the MMoE-3D model are shown in formula (7):
[0067] M=l sub M sub +l sup M sup Formula (7)
[0068] Among them, M sub and M sup are the subsonic model and the supersonic model respectively. The parameter l is calculated by formula (8): sub and l sup :
[0069]
[0070] The following embodiment of the present invention conducts a comparative experiment between the method of the present invention and multi-task learning in the prior art to verify the effectiveness and advancement of the method provided by the embodiment of the present invention.
[0071] MMoE is compared with multi-task learning (hard parameter sharing) to predict 4 tasks. In order to ensure the rationality of the experiment, the number of neurons in each layer of the two models is set to the same, and other basic hyperparameters are also set to the same. The comparison results are shown in Figure 2. Figure 4 , 5As shown. MMoE has basically converged at epoch = 300, while the loss of MTL at epoch = 500 is still fluctuating greatly, with no convergence trend at all. Whether it is MMoE or MTL, the loss of the test set is decreasing after adding 3D data. The loss value of the MMoE-3D test is the smallest among the four comparison models, which proves the effectiveness of 3D data for training models. An early stopping mechanism was added in the actual training, and the training duration required for each model was also different. As shown in Table 2, after adding 3D data, the training time for a single epoch is longer, but because of its fast convergence speed, the total training time will not be too long. From Figure 6a-6c From the results shown in the figure, at subsonic speed, the calculation results of MMoE-3D and DATCOM are different for the axial force C A , normal force C N , pitch moment C m and pressure center X cp The fitting effect is better. At supersonic speed, the pressure center X cp The fitting ability is slightly worse, but it is basically consistent.
[0072] Table 2. Model comparison
[0073]
[0074] An embodiment of the present invention further provides an aircraft aerodynamic characteristics prediction device, the device comprising a processor, the processor being configured to:
[0075] The three-dimensional coordinate information is expressed by calculating the vertex coordinates of the eight airfoils using formula (1):
[0076]
[0077] Among them, (x α,Φ ,y α,Φ , z α,Φ ) represents the coordinate point with an attack angle of α and a roll angle of φ, and (x0, y0, z0) represents the coordinate point with an attack angle of 0 and a roll angle of 0;
[0078] The MMoE-3D model is constructed according to formula (2)-formula (5):
[0079] y k =h k (f k (x)), k = 1, 2, 3, 4 Formula (2)
[0080]
[0081] g k (x) = Softmax(wC) Formula (4)
[0082] C=Concat(Input1, Input2) Formula (5)
[0083] Among them, e i represents the i-th expert neural network, g(x) represents a gating network that combines the results of all experts, and the k-th output g of g(x) k (x) represents expert e k The probability of the value, and C represents the merging layer, which is used to fuse the aircraft plane parameters, flight conditions and 3D coordinate information. The Softmax activation function is used to generate multiple probability distributions of experts and finally output f k (x) is the sum of the weights of all experts. Each subtask corresponds to a specific task layer. The mapping function of the kth task layer is expressed as h k , the output of each subtask is y k ,The gating network of each task realizes the selective utilization of k experts through the final output of different weights, and the gating networks of different tasks learn different combinations of expert modes;
[0084] The aerodynamic characteristics of the aircraft are predicted by the MMoE-3D model.
[0085] It should be noted that the processor may be a processing device including one or more general-purpose processing devices, such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), etc. More specifically, the processor may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor running other instruction sets, or a processor running a combination of instruction sets. The processor may also be one or more special-purpose processing devices, such as an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a system on a chip (SoC), etc.
[0086] The processor may be communicatively coupled to the memory and configured to execute computer executable instructions stored thereon to perform the aircraft aerodynamic characteristic prediction method according to various embodiments of the present invention.
[0087] In some embodiments, the processor is further configured to train the MMoE-3D model by a back-propagation method.
[0088] In some embodiments, the processor is further configured to train the MMoE-3D model using the following formula (6):
[0089]
[0090] Where n is the number of training sets, y is the aerodynamic data corresponding to the missile shape, is the predicted value of the MMoE-3D model.
[0091] In some embodiments, the processor is further configured to output the prediction result of the MMoE-3D model through the following formula (7):
[0092] M=l sub M sub +l sup M sup Formula (7)
[0093] Among them, M sub and M sup are the subsonic model and the supersonic model respectively. The parameter l is calculated by formula (8): sub and l sup :
[0094]
[0095] The aircraft aerodynamic characteristics prediction devices described in various embodiments of the present invention are consistent in effect and method, and will not be repeated here.
[0096] An embodiment of the present invention further provides a non-transitory computer-readable medium storing instructions, and when the instructions are executed by a processor, the method according to any embodiment of the present invention is executed.
[0097] In addition, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present invention with equivalent elements, modifications, omissions, combinations (e.g., various embodiments intersecting schemes), adaptations or changes. The elements in the claims will be interpreted broadly based on the language adopted in the claims, and are not limited to the examples described in this specification or during the implementation of this application, and the examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered as examples only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.
[0098] The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. For example, those of ordinary skill in the art can use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features can be grouped together to simplify the present invention. This should not be interpreted as a feature of an invention that is not claimed for protection being necessary for any claim. On the contrary, the subject matter of the present invention may be less than all the features of the embodiments of a specific invention. Thus, the following claims are incorporated into the specific embodiments as examples or embodiments, wherein each claim is independently used as a separate embodiment, and it is considered that these embodiments can be combined with each other in various combinations or arrangements. The scope of the present invention should be determined with reference to the full scope of the equivalent forms of the attached claims and these claims.
Claims
1. A method for predicting aerodynamic characteristics of an aircraft, characterized in that: The method comprises: The three-dimensional coordinate information is expressed by calculating the vertex coordinates of the eight airfoils using formula (1): Among them, (x α,Φ ,y a,Φ , z α,Φ ) represents the coordinate point with an attack angle of α and a roll angle of φ, and (x0, y0, z0) represents the coordinate point with an attack angle of 0 and a roll angle of 0; The MMoE-3D model is constructed according to formula (2)-formula (5): y k =h k (f k (x)), k = 1, 2, 3, 4 Formula (2) g k (x) = Softmax(wC) Formula (4) C=Concat(Input1, Input2) Formula (5) Among them, e i represents the i-th expert neural network, g(x) represents a gating network that combines the results of all experts, and the k-th output g of g(x) k (x) represents expert e k The probability of the value, and C represents the merging layer, which is used to fuse the aircraft plane parameters, flight conditions and 3D coordinate information. The Softmax activation function is used to generate multiple probability distributions of experts and finally output f k (x) is the sum of the weights of all experts. Each subtask corresponds to a specific task layer. The mapping function of the kth task layer is expressed as h k , the output of each subtask is y k ,The gating network of each task realizes the selective utilization of k experts through the final output of different weights, and the gating networks of different tasks learn different combinations of expert modes; The aerodynamic characteristics of the aircraft are predicted by the MMoE-3D model.
2. The method according to claim 1, characterized in that After constructing the MMoE-3D model, the method further comprises: The MMoE-3D model is trained by the back-propagation method.
3. The method according to claim 1 or 2, characterized in that: The MMoE-3D model is trained by the following formula (6): Where n is the number of training sets, y is the aerodynamic data corresponding to the missile shape, is the predicted value of MMoE-3D.
4. The method according to claim 1, characterized in that: The method further comprises: The MMoE-3D model outputs the prediction result through the following formula (7): M=l sub M sub +l sup M sup Formula (7) Among them, M sub and M sup are the subsonic model and the supersonic model respectively. The parameter l is calculated by formula (8): sub and l sup :
5. An aircraft aerodynamic characteristics prediction device, characterized in that: The apparatus comprises a processor configured to: The three-dimensional coordinate information is expressed by calculating the vertex coordinates of the eight airfoils using formula (1): Among them, (x α,Φ ,y α,Φ , z α,Φ ) represents the coordinate point with an attack angle of α and a roll angle of φ, and (x0, y0, z0) represents the coordinate point with an attack angle of 0 and a roll angle of 0; The MMoE-3D model is constructed according to formula (2)-formula (5): y k =h k (f k (x)), k = 1, 2, 3, 4 Formula (2) g k (x) = Softmax(wC) Formula (4) C=Concat(Input1, Input2) Formula (5) Among them, e i represents the i-th expert neural network, g(x) represents a gating network that combines the results of all experts, and the k-th output g of g(x) k (x) represents expert e k The probability of the value, and C represents the merging layer, which is used to fuse the aircraft plane parameters, flight conditions and 3D coordinate information. The Softmax activation function is used to generate multiple probability distributions of experts and finally output f k (x) is the sum of the weights of all experts. Each subtask corresponds to a specific task layer. The mapping function of the kth task layer is expressed as h k , the output of each subtask is y k ,The gating network of each task realizes the selective utilization of k experts through the final output of different weights, and the gating networks of different tasks learn different combinations of expert modes; The aerodynamic characteristics of the aircraft are predicted by the MMoE-3D model.
6. The device according to claim 5, characterized in that The processor is further configured to train the MMoE-3D model by a back-propagation method.
7. The device according to claim 5 or 6, characterized in that The processor is further configured to train the MMoE-3D model by the following formula (6): Where n is the number of training sets, y is the aerodynamic data corresponding to the missile shape, is the predicted value of the MMoE-3D model.
8. The device according to claim 5, characterized in that The processor is further configured to output a prediction result of the MMoE-3D model through the following formula (7): M=l sub M sub +l sup M sup Formula (7) Among them, M sub and M sup are the subsonic model and the supersonic model respectively. The parameter l is calculated by formula (8): sub and l sup :
9. A non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, executes the method according to any one of claims 1 to 4.
