Aviation dispenser aerodynamic configuration optimization method based on deep neural network

Abstract

The invention discloses an aviation dispenser aerodynamic configuration optimization method based on a deep neural network, and relates to the technical field of aerospace. In order to solve the technical defect that in the prior art, an existing deep learning method and a CFD optimization method still have obvious limitation when facing the complex aerodynamic configuration, composed of a plurality of variable parameters, of the aviation dispenser, the deep learning method and the CFD optimization method still have obvious limitation. According to the technical scheme provided by the invention, the method comprises the following steps: collecting the aerodynamic configuration of the aviation dispenser generated by using a deep generative adversarial network; extracting image features of the aerodynamic configuration according to a convolutional neural network and converting the image features into shape parameters; according to the shape parameters, using a multi-task learning model to predict aerodynamic data, including axial force, normal force, pitching moment and pressure center; and optimizing aerodynamic configuration parameters by adopting a differential evolution algorithm according to a result of predicting the aerodynamic data. The method can be applied to design and optimization of the aviation dispenser.

Description

Technical Field

[0001] The invention relates to the field of aerospace technology, and specifically to aerodynamic shape optimization technology of an aviation dispenser. Background Art

[0002] As a key aerial combat platform, aerial dispensers are widely used in modern military operations, especially in standoff strike missions. Their unique combat capability and tactical flexibility make aerial dispensers of great strategic significance in dealing with complex battlefield environments. The aerodynamic characteristics of aerial dispensers directly affect their flight stability, range, maneuverability and mission success rate. Therefore, optimizing their aerodynamic shape and improving their combat performance have become important research directions in the field of aerial dispenser design.

[0003] At present, in the optimization design of the aerodynamic shape of aerial dispensers, traditional methods mainly rely on the experience of designers and computational fluid dynamics (CFD) simulation. The CFD method solves the fluid flow characteristics through numerical simulation, but its calculation process is very complicated, especially in the face of unsteady flow, supersonic flow and other conditions, the time and resources required for CFD calculation are huge. For example, when performing airfoil optimization, traditional CFD methods usually require hundreds to thousands of simulation calculations, and each calculation requires the support of high-performance computing equipment. In addition, the accuracy of the CFD method depends largely on the selection of meshing and physical models, and these factors often require professional experience and long-term adjustments.

[0004] On the other hand, with the development of artificial intelligence technology, especially deep learning technology, more and more studies have begun to try to apply it to aerodynamic shape optimization. As a generative model, deep generative adversarial networks (GANs) perform well in image generation and sample learning, and can generate reasonable aerodynamic shape designs without relying on traditional numerical simulations. Many studies have also proposed aerodynamic shape generation and optimization methods based on deep learning. For example, the aerodynamic shape generation method based on variational autoencoder (VAE) generates new design samples by learning existing aerodynamic design data. However, these methods usually focus on single airfoil design and fail to effectively extend to multi-component aircraft with complex geometries, especially the optimization design of complex shapes such as aerial dispensers.

[0005] In order to further improve the design efficiency, some studies in recent years have proposed hybrid optimization methods that combine deep learning with CFD. By combining deep generative models with CFD simulation results, the process of aerodynamic shape optimization can be accelerated. However, these methods usually focus on airfoil optimization and do not systematically design the aerodynamic shape of aircraft with complex structures and multiple variable parameters. Moreover, their computational efficiency still cannot meet the needs of efficient optimization.

[0006] Although existing deep learning methods and CFD optimization methods have made some progress in single airfoil design and simplified aerodynamic shape optimization, they still have significant limitations when facing complex aerodynamic shapes such as aerial dispensers that consist of multiple variable parameters. First, existing optimization methods are computationally inefficient when dealing with complex geometric shapes, and the optimization process requires a lot of time and resources. Second, traditional deep learning-based optimization methods are mostly only applicable to airfoil design and are difficult to extend to aerial dispensers with multiple degrees of freedom and complex geometric structures. Finally, existing methods lack comprehensive optimization of multiple aerodynamic parameters (such as lift-to-drag ratio, center of pressure, etc.) during the design process, and can usually only optimize a single aerodynamic characteristic. Summary of the invention

[0007] In order to solve the technical defects in the prior art that the existing deep learning method and CFD optimization method still have significant limitations when facing the complex aerodynamic shape of an aerial dispenser composed of multiple variable parameters, the technical solution provided by the present invention is:

[0008] A method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network, comprising:

[0009] Steps for acquiring an aerodynamic shape of an aerial dispenser generated using a deep generative adversarial network;

[0010] A step of extracting image features of the aerodynamic shape and converting them into shape parameters according to a convolutional neural network;

[0011] Based on the shape parameters, a multi-task learning model is used to predict aerodynamic data, including the steps of axial force, normal force, pitching moment and center of pressure;

[0012] The step of optimizing aerodynamic shape parameters using a differential evolution algorithm according to the result of predicting aerodynamic data.

[0013] Further, a preferred embodiment is provided, further comprising

[0014] Steps to verify the optimized aerodynamic shape through DATCOM.

[0015] Furthermore, a preferred embodiment is provided, wherein the generator of the deep generative adversarial network generates an aerodynamic shape by truncating latent variables in a normal distribution and uses label information to display shape features.

[0016] Furthermore, a preferred embodiment is provided, wherein the convolutional neural network includes four convolutional layers and three fully connected layers, which are used to extract aerodynamic shape features and map them into shape parameters.

[0017] Furthermore, a preferred embodiment is provided, wherein the multi-task learning model predicts aerodynamic data through an expert neural network and a gating network.

[0018] Furthermore, a preferred embodiment is provided, wherein the differential evolution algorithm uses 100 populations and 100 iterations to optimize the aerodynamic shape.

[0019] Based on the same inventive concept, the present invention also provides an aerodynamic shape optimization device for an aerial dispenser based on a deep neural network, comprising:

[0020] A module that collects the aerodynamic shape of an aerial dispenser generated using a deep generative adversarial network;

[0021] A module for extracting image features of the aerodynamic shape and converting them into shape parameters based on a convolutional neural network;

[0022] Based on the shape parameters, a multi-task learning model is used to predict aerodynamic data, including modules for axial force, normal force, pitching moment, and center of pressure;

[0023] A module for optimizing aerodynamic shape parameters using a differential evolution algorithm based on the results of the predicted aerodynamic data.

[0024] Based on the same inventive concept, the present invention also provides a computer storage medium for storing a computer program. When the computer program is read by a computer, the computer executes the described method.

[0025] Based on the same inventive concept, the present invention also provides a computer, including a processor and a storage medium. When the processor reads a computer program stored in the storage medium, the computer executes the method described.

[0026] Based on the same inventive concept, the present invention also provides a computer program product, which is a computer program. When the computer program is executed, the method described above is implemented.

[0027] Compared with the prior art, the technical solution provided by the present invention is beneficial in that:

[0028] This solution achieves efficient optimization of the aerodynamic shape of aerial dispensers by combining deep neural networks (especially CWGAN-GP and MMoE-3D models) with traditional aerodynamic optimization methods. First, CWGAN-GP is used to generate a variety of aerodynamic shapes and features are extracted through convolutional neural networks (CNN), which significantly improves the speed and accuracy of aerodynamic shape generation. This approach solves the limitations of traditional methods that rely on experience and CFD simulation, avoids a large number of complex calculations and long simulations, and can provide effective solutions for design more quickly. Compared with traditional experience-based or CFD-based methods, deep learning-based models not only greatly reduce calculation time, but also improve design flexibility and innovation.

[0029] By combining the MMoE-3D model with the differential evolution algorithm (DE), this scheme can efficiently predict and optimize aerodynamic data, further improving the optimization effect of aerodynamic shape. MMoE-3D predicts aerodynamic data through multi-task learning, which can not only handle complex aerodynamic problems under different flight conditions, but also optimize multiple aerodynamic parameters at the same time, such as axial force, normal force, pitch moment and pressure center. Compared with traditional aerodynamic optimization methods, the multi-task learning framework of MMoE-3D can better capture the relationship between various aerodynamic features, reducing the time and resource consumption of optimizing each parameter separately in traditional methods.

[0030] In addition, the differential evolution algorithm in this scheme can adjust the aerodynamic shape parameters in multiple rounds of optimization and obtain the best design solution through continuous iteration of the optimization process. Compared with conventional optimization methods, the DE algorithm has a stronger global search capability during the optimization process, which can avoid falling into the local optimal solution and ensure that the final optimization result has higher performance. Compared with traditional methods (such as CFD-based optimization), DE optimization not only improves computational efficiency, but also obtains more accurate design parameters in a shorter time.

[0031] In general, the optimization scheme based on deep neural networks significantly improves the optimization efficiency and accuracy of aerodynamic shape by combining generative adversarial networks, convolutional neural networks and multi-task learning models. Compared with traditional CFD and empirical design methods, it not only reduces the calculation time, but also improves the design flexibility and innovation, achieving a more efficient optimization effect.

[0032] It can be applied to the design and optimization of aerial dispensers. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the traditional aerodynamic shape design method;

[0034] Figure 2 Design a flow chart for aerodynamic shapes based on deep neural networks;

[0035] Figure 3 It is a three-dimensional schematic diagram of an aerial dispenser;

[0036] Figure 4 A two-dimensional sketch of the aerial dispenser;

[0037] Figure 5 Schematic diagram of the CWGAN-GP structure;

[0038] Figure 6 The following is a sample from the aviation dispenser shape database and a schematic diagram of the aerodynamic shape synthesized using CWGAN-GP;

[0039] Figure 7 This is a schematic diagram of the CNN structure based on the parameterization of the aerial dispenser shape;

[0040] Figure 8 Schematic diagram of the MMoE-3D structure used for aerodynamic data prediction;

[0041] Fig. 9 This is a schematic diagram comparing the aerodynamic shapes of aerial dispensers;

[0042] Fig.10 C L / C D Compare simulation schematics;

[0043] Fig.11 C A Compare simulation schematics;

[0044] Fig.12 C N Compare simulation schematics;

[0045] Fig.13 C m Compare simulation schematics. DETAILED DESCRIPTION

[0046] In order to make the advantages and benefits of the technical solution provided by the present invention more clearly reflected, the technical solution provided by the present invention is now further described in detail in conjunction with the accompanying drawings, specifically:

[0047] Embodiment 1: This embodiment provides an aerodynamic shape optimization method for an aerial dispenser based on a deep neural network, including:

[0048] Steps for acquiring an aerodynamic shape of an aerial dispenser generated using a deep generative adversarial network;

[0049] A step of extracting image features of the aerodynamic shape and converting them into shape parameters according to a convolutional neural network;

[0050] Based on the shape parameters, a multi-task learning model is used to predict aerodynamic data, including the steps of axial force, normal force, pitching moment and center of pressure;

[0051] The step of optimizing aerodynamic shape parameters using a differential evolution algorithm according to the result of predicting aerodynamic data.

[0052] Specifically, the technical solution provided in this embodiment includes:

[0053] A method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network aims to efficiently optimize the aerodynamic shape of an aerial dispenser by combining a deep generative adversarial network (CWGAN-GP) and a multi-task learning model (MMoE-3D) with traditional optimization methods to improve its flight performance and combat capability. The implementation steps of this method are described in detail below in conjunction with specific embodiments.

[0054] 1. Aerodynamic shape generation

[0055] First, a deep generative adversarial network (CWGAN-GP) is used to generate diverse aerodynamic shape designs for aerial dispensers. This process generates different aerodynamic shapes based on latent variables in a truncated normal distribution through a generator, and controls the shape features according to the input label information. Specifically, the label information can determine whether to generate a shape design with or without rectangular wings. The output of the generator is different types of aerodynamic shapes that can meet the aerodynamic requirements under different flight conditions. Through back-propagation training, the generator can gradually optimize the generated aerodynamic shape to ensure that it has reasonable aerodynamic performance.

[0056] 2. Feature extraction and parameterization

[0057] The generated aerodynamic shape is input into a convolutional neural network (CNN), which extracts features from the aerodynamic shape image through multiple convolutional layers and pooling layers. The convolutional layer extracts local features in the aerodynamic shape, and the pooling layer further compresses the image data to retain the most critical information. After feature extraction, the feature map of the aerodynamic shape is input into the fully connected layer, through which the image features are mapped into seven shape parameters, such as the position of the front wing, the position of the rear wing, the shape of the airfoil, etc. These shape parameters serve as input for the next step of predicting aerodynamic data, providing key information for optimization.

[0058] 3. Aerodynamic data prediction

[0059] The extracted shape parameters and flight conditions (such as angle of attack, Mach number, etc.) are input into the multi-task learning model (MMoE-3D), which combines multiple expert neural networks and performs information fusion through a gating network to predict the aerodynamic data corresponding to the aerodynamic shape. The MMoE-3D model can simultaneously predict multiple aerodynamic parameters, including axial force, normal force, pitch moment, and center of pressure. Each task layer is responsible for predicting an aerodynamic parameter, and the prediction results of each task are optimized through the combination of expert neural networks. This step ensures the accuracy of the aerodynamic data and provides an accurate reference for subsequent optimization.

[0060] 4. Aerodynamic shape optimization

[0061] Based on the aerodynamic data obtained in the above steps, the differential evolution algorithm (DE) is used to optimize the aerodynamic shape parameters. The differential evolution algorithm adjusts the seven shape parameters of the aerodynamic shape through multiple rounds of iterative optimization to meet the optimal aerodynamic design goals. In each round of iteration, the DE algorithm calculates the optimization target value based on the aerodynamic data (such as axial force, normal force, etc.) output by the MMoE-3D model, and selects the optimal aerodynamic shape design. During the optimization process, the DE algorithm optimizes the adaptive changes and convergence of the population to ensure that the optimal solution is obtained under 100 iterations and 100 populations.

[0062] 5. Optimization results verification

[0063] Finally, the optimized aerodynamic shape is verified by the traditional aerodynamic optimization software DATCOM. DATCOM verifies whether the optimized shape meets the design requirements by calculating the lift-to-drag ratio, pressure center and other parameters of the aerodynamic shape. If the optimization result fails to meet the target, the aerodynamic shape parameters are further adjusted by returning to the extended training set and fine-tuning the MMoE-3D model until the optimization result meets the requirements. Through this verification process, the reliability and feasibility of the final design are ensured.

[0064] Embodiment 2: This embodiment further limits the aerodynamic shape optimization method of an aerial dispenser based on a deep neural network provided in Embodiment 1, and also includes

[0065] Steps to verify the optimized aerodynamic shape through DATCOM.

[0066] Embodiment 3: This embodiment further limits the method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network provided in embodiment 1. The generator of the deep generative adversarial network generates an aerodynamic shape by truncating latent variables in a normal distribution and uses label information to display shape features.

[0067] Embodiment 4: This embodiment further limits the method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network provided in embodiment 1. The convolutional neural network includes four convolutional layers and three fully connected layers for extracting aerodynamic shape features and mapping them into shape parameters.

[0068] Embodiment 5: This embodiment further limits the method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network provided in embodiment 1. The multi-task learning model predicts aerodynamic data through an expert neural network and a gating network.

[0069] Embodiment 6: This embodiment is a further limitation of the method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network provided in embodiment 1. The differential evolution algorithm uses 100 populations and 100 iterations to optimize the aerodynamic shape.

[0070] Embodiment 7: This embodiment provides an aerodynamic shape optimization device for an aerial dispenser based on a deep neural network, comprising:

[0071] A module that collects the aerodynamic shape of an aerial dispenser generated using a deep generative adversarial network;

[0072] A module for extracting image features of the aerodynamic shape and converting them into shape parameters based on a convolutional neural network;

[0073] Based on the shape parameters, a multi-task learning model is used to predict aerodynamic data, including modules for axial force, normal force, pitching moment, and center of pressure;

[0074] A module for optimizing aerodynamic shape parameters using a differential evolution algorithm based on the results of the predicted aerodynamic data.

[0075] Embodiment 8: This embodiment provides a computer storage medium for storing a computer program. When the computer program is read by a computer, the computer executes the method provided in embodiment 1.

[0076] Embodiment 9: This embodiment provides a computer, including a processor and a storage medium. When the processor reads a computer program stored in the storage medium, the computer executes the method provided in embodiment 1.

[0077] Embodiment 10: This embodiment provides a computer program product, which is a computer program. When the computer program is executed, the method provided in embodiment 1 is implemented.

[0078] Implementation Method XI: Combination Figure 1-13 This embodiment is described in detail by using specific examples to further describe the technical solution provided above. Specifically:

[0079] As an aerial combat platform that can be launched outside the defense zone, the aerial dispenser has attracted widespread attention due to its unique combat capability and tactical flexibility. The current battlefield environment is becoming increasingly complex and changeable, and the performance requirements for military equipment are becoming higher and higher. In this context, the aerodynamic characteristics of the aerial dispenser, as an unpowered aircraft, have a great impact on its combat performance when performing aerial dispensing missions. Therefore, it is necessary to optimize the aerodynamic shape design of the aerial dispenser to improve performance and meet tactical requirements.

[0080] Traditionally, aerodynamic shape optimization relies on the designer's experience and methods such as computational fluid dynamics (CFD) simulation. Although these methods can provide effective design solutions to a certain extent, they are limited by human factors and complex numerical calculations and require a lot of time to calculate. In order to reduce calculation time and cost, methods that combine deep generative models with CFD have begun to appear to speed up the aerodynamic optimization design process. However, this method is usually limited to the optimization of airfoil design and cannot be extended to other engineering design requirements. The shape of an aviation dispenser is more complex and variable than that of a single airfoil.

[0081] This implementation method designs a deep neural network-based framework to optimize the aerodynamic shape of the aerial dispenser. First, CWGAN-GP is applied to learn the aerodynamic design features of existing aerial dispensers and generate the required diversified aerodynamic shapes; CNN is used to extract the features of the aerodynamic shape and convert the aerodynamic shape generated by CWGAN-GP into specific shape parameters; the MMoE-3D model is trained in the subsonic and supersonic ranges to efficiently generate aerodynamic data corresponding to the aerodynamic shape of the aerial dispenser; finally, these three components are integrated and differential evolution (DE) is used to select the optimal aerodynamic shape.

[0082] The goal of optimizing the aerodynamic shape of an aerial dispenser is to find a specific aerodynamic shape with better performance. The traditional aerodynamic shape design method process is as follows: Figure 1 shown.

[0083] The optimization goal can be described as:

[0084]

[0085] Where c represents the flight conditions, and x represents the aerodynamic shape parameters of the aerial dispenser containing n variables, that is, the optimized aerodynamic characteristic parameters can be expressed as:

[0086] x=(x1,x2,x3,L,x n ) (2)

[0087] F(x,c) is the optimization target obtained by CFD calculation. It can be seen that in order to obtain a better aerodynamic shape of the aerial dispenser, a large number of tests are required, so the optimization solution based on deep neural network proposed in this embodiment can effectively solve the above optimization problem. Figure 2 The structural framework of the aerodynamic shape design of the aerial dispenser proposed in this embodiment is demonstrated.

[0088] Step 1: Build an aerodynamic shape design generator based on the existing aerodynamic shape database:

[0089] G=g(z,c,θ) (3)

[0090] z: N(0, 1), z∈(-3, 3), c=0, 1 (4)

[0091] Where G represents the aerodynamic shape design of the aerial dispenser generated by the generator in CWGAN-GP, z is the latent variable sampled from the truncated normal distribution, c represents additional parameter information, and θ represents the network parameters in the generator.

[0092] Step 2: Train CNN to complete the parameterization of the aerodynamic shape of the aerial dispenser and convert G into

[0093]

[0094] Where C represents the relationship function between the aerodynamic shape and shape parameters of the aerial dispenser, and ε represents the network parameters.

[0095] Step 3: Train MMoE-3D to predict the aerodynamic data corresponding to the aerodynamic shape of the aerial dispenser generated by the generator. The output of this process is two sets of models (M) in the Mach number range of Ma = [0.4, 0.6, 0.8] and Ma = [1.2, 1.4, 1.6]. sub and M sup ):

[0096]

[0097] Where M represents the possible nonlinear relationship between aerodynamic data and the aerodynamic shape of the aerial dispenser. M is the aerodynamic shape of the aerial dispenser Coordinate information α and flight conditions f c Prediction results of the aerodynamic data. e represents the network parameters.

[0098] Step 4: Initialize the differential evolution algorithm shape parameters.

[0099] Step 5: Input the generated shape parameters into the CWGAN-GP generator to obtain the aerodynamic shape design of the aerial dispenser.

[0100] Step 6: The convolutional neural network converts the aerodynamic shape design of the aerial dispenser into shape parameters.

[0101] Step 7: Input the aerial dispenser shape parameters, coordinate data and flight conditions into MMoE-3D, obtain the output of the corresponding aerodynamic data, and calculate the optimization target value.

[0102] Step 8: Iterative optimization. Repeat steps 4 to 8 to output the optimized aerodynamic shape design of the aerial dispenser. Use DATCOM to verify whether the optimized aerial dispenser meets the requirements. If not, return to step 3 and expand the data to the training set to fine-tune MMoE-3D, and repeat the optimization process.

[0103] The above process can be described as:

[0104] Optimize F

[0105]

[0106] z: N(0, 1), z∈(-3, 3), c=0, 1 (7)

[0107] Where F is the objective function of the optimization algorithm.

[0108] Common aerial dispensers such as Figure 3 The missile body is 4.1m long, 0.33m in diameter, and has a wingspan of 2.7m. The aerial dispenser is plane-symmetrical, and its aerodynamic shape is simplified as follows: Figure 4 The 2D plot shown in the figure shows different aerodynamic shapes obtained by changing the front wing position, front wing shape, rear wing position and rear wing shape. There are 10 variable shape parameters: LW, LF, CRW, CTW, BKW, BW, CRF, CTF, BKF, BF. The axial force C is calculated using Missile DATCOM software. A , normal force C N , pitch moment C M and pressure center X CP , forming an aerodynamic data set for aerial dispensers. Missile DATCOM is an engineering-level computer program used to estimate the aerodynamic characteristics of aircraft. It combines theoretical and empirical methods, covering the entire speed range from subsonic to hypersonic flight. The experimental data results are consistent with the parameter change trends obtained by ANSYS Fluent.

[0109] A 2D aerodynamic shape image of 80×48 is drawn based on the constructed aerodynamic shape parameter group of the aerial dispenser, and the color of the aerial dispenser is set to black so that CWGAN-GP can better extract image features. The images are divided into two categories, the aerial dispenser with rectangular wings is assigned a label c of 0, and the aerial dispenser without rectangular wings is assigned a label c of 1.

[0110] The optimization algorithm is:

[0111] Deep generative models represent the design space as designs sampled from a formal statistical distribution and use previous designs to estimate the generative model of this distribution, so that the characteristic representation of the complex aerodynamics of aerial dispensers can be learned. GANs are one of the most commonly used deep generative models and are often used to solve engineering design problems. Their key structures are the generator G and the discriminator D. G is responsible for randomly sampling from real data and learning its distribution to generate fake samples that are as real as possible to deceive D, while D is responsible for distinguishing the samples generated by G from the real data as correctly as possible. The generator G and the discriminator D compete with each other. This model setting and training method leads to a GAN that generates real data in the entire parameter space. GAN originally used JS divergence as a metric to be able to measure the difference between the distribution of pseudo sample data and the distribution of real data. The entire generative adversarial network optimization objective function can be written as:

[0112]

[0113] where x belongs to the true distribution, P x , z belongs to the potential distribution P g .

[0114] The WGAN method uses the Wasserstein-1 distance to calculate the distance between two distributions:

[0115]

[0116] Where Lip1 is a 1-Lipschitz constraint that specifies the maximum local change of a continuous function such that ||f||≤1. The WGAN-GP method adds a gradient penalty (GP) to the loss function. The loss function of the discriminator can be written as:

[0117]

[0118] The sample used to calculate GP The linear interpolation between the generated samples and the real samples is performed.

[0119] The CGAN method adds labels as additional parameters in the generator to guide the training process. Once the training is completed, the labels can be changed to make the generator generate corresponding features. Labels are also added to the discriminator to better distinguish the real aerodynamic data. Common aerial dispensers do not contain rectangular wings, but in some special cases rectangular wings can also show better characteristics. Therefore, this embodiment can also design the aerodynamic shape of aerial dispensers with rectangular wings, and divide the aerial dispenser contour data into two categories (aerodispensers with rectangular wings and aerial dispensers without rectangular wings) to meet different optimization requirements. In order to obtain a higher quality aerodynamic shape of an aerial dispenser, CWGAN-GP is used:

[0120]

[0121] Where c represents label information.

[0122] The structure of CWGAN-GP is as follows Figure 5 As shown in the figure, the generator in this model consists of a layer of fully connected neural network, a layer of convolutional neural network and two layers of residual neural network, and the discriminator consists of four layers of fully connected neural network.

[0123] The aerial dispenser shape database and the generated aerial dispenser shape are as follows Figure 6 As shown in the figure, (A) is the aerial dispenser shape database, (B) is the aerial dispenser shape synthesized by CWGAN-GP, the first and second columns are the aerial dispenser shapes with label 0, and the third and fourth columns are the aerial dispenser shapes with label 1. It can be seen that the generated aerial dispenser shapes have reasonable symmetry and diversity. In addition, the generator can create the corresponding aerial dispenser aerodynamic shape very well through the label c. The CWGAN-GP generator can control the latent variables z and the label c, and provide a sufficient number of samples for the optimization of the aerial dispenser aerodynamic shape according to different needs.

[0124] Convolutional neural networks are:

[0125] Ballistic aerodynamic shape parameterization is the basis of optimization design. MMoE-3D can only obtain the mapping relationship between a limited number of variables. Therefore, parameterizing the ballistic aerodynamic shape is a necessary method to accurately describe the geometric shape of the aerial dispenser. Convolutional neural networks (CNNs) have significant advantages in abstracting image structure features. This embodiment uses CNN to parameterize the aerodynamic shape of the aerial dispenser. Figure 7As shown in the figure. In CNN, the input of the convolutional layer is the aerodynamic shape of the aerial dispenser generated by the CWGAN-GP generator, which is 80×48 in size. The convolutional layer uses the convolution kernel to extract the features of the aerodynamic shape of the aerial dispenser. Since the input data is image data representing the outline of the aerial dispenser, there is a certain sparsity in the image, so the maximum pooling is used in the pooling layer to retain the key information in the image. The feature map generated by the pooling layer is then passed to the next convolutional layer to find higher-level features. This process can be described as:

[0126] y c =Maximum pooling (ReLU (k*a c +b c ))(12)

[0127] Where c represents the cth layer, a c is the input feature map of the convolution layer, * represents the convolution operation, k is the convolution kernel, b c is the bias term, y c is the output feature map. Finally, the feature map extracted by the convolutional layer is passed to the fully connected layer. The algorithm used to train CNN is back propagation, and the loss function of CNN can be written as:

[0128]

[0129] Where n is the number of training sets, x represents the shape parameters corresponding to the shape of the aerial dispenser, is the predicted value of CNN.

[0130] The convolutional neural network structure contains four convolutional layers and three fully connected layers, which can well solve the problem of parameterization of the shape of the aerial dispenser. Back propagation establishes the mapping relationship between the shape of the aerial dispenser and seven shape parameters (LW, CTF, BKW, BW, CTW, BF, BKW). The regional features of the aerial dispenser shape design image are extracted through the convolutional layer and pooling layer in the convolutional neural network, and the image features corresponding to different convolution kernels are obtained and combined and expanded into a one-dimensional vector form. Then, a multi-layer neural network is established using a fully connected layer with output nodes.

[0131] The process used for aerodynamic data prediction is:

[0132] In machine learning, in order to achieve optimization indicators, it is usually necessary to first train a single model or a series of models to perform a given task, and then optimize the model by fine-tuning parameters until the performance no longer improves. The MoE method proposes to share some expert knowledge at the bottom layer, combine them through a gateway network, and apply a gate structure attention network for information fusion, which can generalize better performance on the main task. This implementation adopts the MMoE method, which extends the MoE method by using different gates for each task to obtain the sum of different weights of experts.

[0133] For the aerodynamic data prediction model in this article, we need to solve the prediction of four parameters: axial force, normal force, pitch moment and pressure center. We use the absolute value of cosine similarity to calculate the correlation of the four tasks:

[0134]

[0135] Where A and B represent two attribute vectors respectively. S represents the similarity between A and B. Fig.11 As shown in Figure 1, the closer the absolute value of the cosine similarity is to 1, the higher the correlation between the two vectors. This figure shows that the four tasks are interrelated.

[0136] In the aerodynamic prediction part, four prediction parameters are given, including the aerodynamic shape of the aerial dispenser. Coordinate information a and flight conditions c. The coordinate calculation method is:

[0137]

[0138] Where (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.

[0139] The MMoE-3D structure is as follows Figure 8 shown.

[0140] The MMoE-3D structure can be expressed as:

[0141] y k =h k (f k (x)),k=1,2,3,4 (16)

[0142]

[0143] g k (x)=Softmax(wC) (18)

[0144] C=[Input1||Input2] (19)

[0145] where e i is the i-th expert neural network. And g k (x) is a gating network that combines all expert results by linearly transforming the fusion layer and multiplying it by the trainable matrix w. C is used to fuse the aerial dispenser contour parameters and 3D information of the input layer, and the Softmax activation function is used to generate the probability distribution of the expert. k (x) is the sum of all expert weights and is the final output. Each output parameter corresponds to a specific task layer, and the mapping function of the kth task layer is expressed as h k , the output of each output parameter is y k The gating network of each output parameter realizes the selective utilization of n experts through the final output with different weights. The gating networks with different output parameters can learn different combinations of expert modes. The algorithm used to train MMoE-3D is back propagation, and the loss function can be written as:

[0146]

[0147] Where n is the number of training sets, y is the aerodynamic data corresponding to the shape of the aerial dispenser, is the predicted value of MMoE-3D.

[0148] The prediction can be written as:

[0149] M=l sub M sub +l sup M sup (twenty one)

[0150] Among them and M sub , M sup are subsonic and supersonic models, l su b parameter and l sup The calculation is as follows:

[0151]

[0152] Aerodynamic shape optimization simulation steps:

[0153] Combined with the DE extended dataset method, the trained CWGAN-GP and MMoE-3D are used to optimize the latent variable z in combination with the DE method, and DATCOM is used to verify whether the optimized aerial dispenser shape meets the optimization criteria. If it does not meet the criteria, it will be expanded to the original training set to fine-tune MMoE-3D until MMoE-3D can optimize the aerial dispenser shape we need.

[0154] The complex change of the pressure center of an aerial dispenser is one of the important factors that often need to be considered in its shape design. Considering the aerodynamic characteristics, it is usually required that the change of the pressure center with Mach number and angle of attack is as small as possible. The optimization objective can be written as:

[0155]

[0156] Where w1 is the weight of the lift-to-drag ratio, w2 is the weight of the drag term used to maximize the inverse drag value, and w3 is the weight associated with the center of pressure change rate function.

[0157] The aerodynamic shape of the aerial dispenser was optimized under the flight condition of φ=45°, the optimization target value of the baseline profile was 2.31, the number of iterations of differential evolution was 100, the population size was 100, and a comparison was made by using the DATCOM traditional aerodynamic optimization software and the MMoE-3D optimization algorithm proposed in this embodiment. Fig. 9 The aerodynamic shape of the aerial dispenser before and after optimization is shown. After the aerial dispenser is optimized using the DATCOM and MMoE-3D models, the forward movement of the front wing of the aerial dispenser and the increase in its area make it more susceptible to drag. Figure 10-13 As shown in Table 1, the optimization method using MMoE-3D significantly reduces the time, and the comprehensive performance F of the optimized aerial dispenser is almost the same.

[0158] Table 1 Subsonic optimization performance

[0159] Optimization Mode Optimization value Execution times time DATCOM F＝3.449 1 21743s NMoE-3D F＝3.408 13 2325s

[0160] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable modification and improvement of the present invention, combination of implementation modes and equivalent substitution within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network, characterized in that: include: Steps for acquiring an aerodynamic shape of an aerial dispenser generated using a deep generative adversarial network; A step of extracting image features of the aerodynamic shape and converting them into shape parameters according to a convolutional neural network; Based on the shape parameters, a multi-task learning model is used to predict aerodynamic data, including the steps of axial force, normal force, pitching moment and center of pressure; The step of optimizing aerodynamic shape parameters using a differential evolution algorithm according to the result of predicting aerodynamic data.

2. The method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network according to claim 1, characterized in that: Also includes Steps to verify the optimized aerodynamic shape through DATCOM.

3. The method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network according to claim 1, characterized in that: The generator of the deep generative adversarial network generates aerodynamic shapes by truncating latent variables in a normal distribution and uses label information to display shape features.

4. The method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network according to claim 1, characterized in that: The convolutional neural network includes four convolutional layers and three fully connected layers, which are used to extract aerodynamic shape features and map them into shape parameters.

5. The method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network according to claim 1, characterized in that: The multi-task learning model predicts aerodynamic data through an expert neural network and a gating network.

6. The method for optimizing the aerodynamic shape of an aerial dispenser based on a deep neural network according to claim 1, characterized in that: The differential evolution algorithm uses 100 populations and 100 iterations to optimize the aerodynamic shape.

7. An aerodynamic shape optimization device for an aerial dispenser based on a deep neural network, characterized in that: include: A module that collects the aerodynamic shape of an aerial dispenser generated using a deep generative adversarial network; A module for extracting image features of the aerodynamic shape and converting them into shape parameters based on a convolutional neural network; Based on the shape parameters, a multi-task learning model is used to predict aerodynamic data, including modules for axial force, normal force, pitching moment, and center of pressure; A module for optimizing aerodynamic shape parameters using a differential evolution algorithm based on the results of the predicted aerodynamic data.

8. A computer storage medium for storing a computing program, characterized in that: When the computer program is read by a computer, the computer executes the method of claim 1 .

9. A computer, comprising a processor and a storage medium, characterized in that: When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1 .

10. A computer program product, being a computer program, characterized in that When the computer program is executed, the method of claim 1 is implemented.

Images