Intelligent adaptive fidelity optimization design method for aviation dispenser, electronic equipment and storage medium

Abstract

The invention discloses an intelligent adaptive fidelity optimization design method for an aviation dispenser, electronic equipment and a storage medium, and belongs to the technical field of intelligent control of aviation dispensers. The invention aims to solve the multidisciplinary overall design problem of the aviation dispenser. The method comprises the steps of data acquisition and preprocessing, construction of a VDNN-TL neural network agent model by using a full-connection neural network, training of the constructed VDNN-TL neural network agent model constructed by the full-connection neural network, and selection and adjustment of training hyper-parameters in the model to realize high-precision fitting of the VDNN-TL neural network agent model constructed by the full-connection neural network. And constructing a VDNN-TL neural network agent model for the trained full-connection neural network, carrying out two rounds of transfer learning, and then correcting a calculation result by adopting fine tuning to obtain an intelligent adaptive fidelity optimization design result of the aviation dispenser. According to the invention, the precision of intelligent adaptive fidelity optimization of the aviation dispenser is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent control of aerial dispensers, and particularly relates to an intelligent adaptive fidelity optimization design method, an electronic device and a storage medium for an aerial dispenser. Background Art

[0002] The overall design of an aerial dispenser is a systematic design process involving multiple disciplines and specialties. During the process of multidisciplinary design, surrogate models are often used to replace the original models for optimization design. Compared with directly obtaining performance data, the application of surrogate models greatly improves the efficiency. However, different surrogate model construction methods may lead to significant differences in performance.

[0003] The Variable-Fidelity Surrogate Model (VFM) uses low-fidelity data to construct a high-precision surrogate model with fewer high-fidelity data points. This method not only reduces the overhead and time of data acquisition, but also improves the modeling accuracy of the surrogate model. Based on different implementation methods, the VFM method can be divided into traditional machine learning methods and deep learning methods. In traditional machine learning methods, the trend characteristics of low-fidelity data are fully retained to assist in constructing a surrogate model with high-fidelity data. However, since it is necessary to calculate the covariance matrix between different fidelity data, it is relatively challenging to process multi-level fidelity data. Summary of the Invention

[0004] The problem to be solved by the present invention is the multidisciplinary overall design problem of an aerial dispenser, and an intelligent adaptive fidelity optimization design method, an electronic device and a storage medium for an aerial dispenser are proposed.

[0005] To achieve the above object, the present invention is realized through the following technical solutions:

[0006] An intelligent adaptive fidelity optimization design method for an aerial dispenser includes the following steps:

[0007] S1. Data collection and preprocessing: Extract the aerodynamic shape data of the aerial dispenser as model samples through experimental design, and obtain sample data with different grid sizes through engineering sampling methods, which are used as low-fidelity sample data, medium-fidelity sample data, and high-fidelity sample data respectively. Then, the data is preprocessed by standardization to obtain a low-fidelity sample data set, a medium-fidelity sample data set, and a high-fidelity sample data set;

[0008] S2. Construct a VDNN-TL neural network surrogate model using a fully connected neural network, including a low-fidelity deep neural network surrogate model LDNN, a medium-fidelity deep neural network surrogate model MDNN, and a high-fidelity deep neural network surrogate model HDNN. The low-fidelity deep neural network surrogate model LDNN, the medium-fidelity deep neural network surrogate model MDNN, and the high-fidelity deep neural network surrogate model HDNN are connected through transfer learning;

[0009] S3. Use the low-fidelity sample dataset, medium-fidelity sample dataset, and high-fidelity sample dataset obtained in step S1 to train the low-fidelity deep neural network surrogate model LDNN, medium-fidelity deep neural network surrogate model MDNN, and high-fidelity deep neural network surrogate model HDNN in the VDNN-TL neural network surrogate model constructed by the fully connected neural network in step S2. Select and adjust the training hyperparameters in the model to achieve high-precision fitting of the VDNN-TL neural network surrogate model constructed by the fully connected neural network;

[0010] S4. Perform two rounds of transfer learning on the trained VDNN-TL neural network surrogate model obtained in step S3. The first round is to transfer from the low-fidelity deep neural network surrogate model LDNN to the medium-fidelity deep neural network surrogate model MDNN, and the second round is to transfer from the medium-fidelity deep neural network surrogate model MDNN to the high-fidelity deep neural network surrogate model HDNN. Then, use fine-tuning to correct the calculation results to obtain the intelligent adaptive fidelity optimization design result of the aerial dispenser.

[0011] Further, in step S1, sample data with grid sizes of 60, 150, and 300 are obtained through engineering sampling methods and used as low-fidelity sample data, medium-fidelity sample data, and high-fidelity sample data, respectively.

[0012] Further, the specific implementation method of step S2 includes the following steps:

[0013] S2.1. The expression for the calculation process of the neurons in the fully connected neural network is:

[0014]

[0015] where y represents the output of the neuron, f(·) is the activation function, ω i is the weight, x i is the input, and b is the bias. By adjusting the weight and bias, the neuron can learn different features;

[0016] S2.2. Construct a low-fidelity deep neural network surrogate model LDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer;

[0017] The first layer of the low-fidelity deep neural network surrogate model is the input layer, which is responsible for inputting feature data to obtain the low-fidelity sample data set of the aerial dispenser. The number of neurons in the input layer is the same as the dimension of the shape structure parameters.

[0018] The second and third layers of the low-fidelity deep neural network surrogate model are feature extraction layers, and the number of neurons in the third layer is twice that of the second layer.

[0019] The fourth to sixth layers of the low-fidelity deep neural network surrogate model are regression adjustment layers, and the number of neurons in each layer is the same.

[0020] The seventh layer of the low-fidelity deep neural network surrogate model is the output layer, and the number of neurons matches the number of output parameters.

[0021] Set the number of neurons in the first layer of the feature extraction layer of the low-fidelity deep neural network surrogate model to n1, the number of neurons in the regression adjustment layer to n2, the learning rate of the low-fidelity deep neural network surrogate model to g1, the batch size of the low-fidelity deep neural network surrogate model to b1, and the activation function of the low-fidelity deep neural network surrogate model to f1.

[0022] S2.3. Construct a medium-fidelity deep neural network surrogate model MDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer. The model structure is that the first layer is the input layer, the second and third layers are feature extraction layers, the fourth to sixth layers are regression adjustment layers, the number of neurons in each layer of the regression adjustment layer is the same, and the seventh layer is the output layer.

[0023] The input layer and feature extraction layer in the medium-fidelity deep neural network surrogate model are obtained through transfer learning of LDNN, and the regression adjustment layer and output layer are obtained through training.

[0024] Set the number of neurons in each neural network layer of the regression adjustment layer in the medium-fidelity deep neural network surrogate model to n3, the activation function of the medium-fidelity deep neural network surrogate model to f2, the learning rate of the medium-fidelity deep neural network surrogate model to g2, and the batch size of the medium-fidelity deep neural network surrogate model is directly transferred from LDNN through learning.

[0025] S2.4. Construct a high-fidelity deep neural network surrogate model HDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer. The model structure is that the first layer is the input layer, the second and third layers are feature extraction layers, the fourth to sixth layers are regression adjustment layers, the number of neurons in each layer of the regression adjustment layer is the same, and the seventh layer is the output layer.

[0026] The input layer, feature extraction layer, and regression adjustment layer in the high-fidelity deep neural network proxy model are obtained through the transfer learning of MDNN. The output layer fine-tunes the neurons in the output layer using the high-fidelity sample dataset. Then, the learning rate of the high-fidelity deep neural network proxy model is set to g3 in HDNN, and the activation function and batch size are the same as those in MDNN.

[0027] Furthermore, the specific implementation method of step S3 includes the following steps:

[0028] S3.1. Set the adjusted hyperparameters during training, including n1, n2, f1, g1, b1, n3, f2, g2, g3;

[0029] S3.2. Set the early stopping strategy during training. When the generalization loss GL(t) is greater than a certain value, stop training. The generalization loss GL(t) is defined as follows:

[0030]

[0031] where E va (t) is the validation set error, and E opt (t) is the lowest validation set error up to time t;

[0032]

[0033] where t' is the time before t;

[0034] S3.3. Set to use the accelerated Adam optimizer during training, combining the advantages of momentum and adaptive learning rate adjustment to accelerate the convergence to the local optimum. The accelerated Adam optimizer is described as follows:

[0035]

[0036] where θ represents the optimization parameter, represents the exponentially weighted average of the past squared gradients, represents the exponentially weighted average of the past gradients, β1 is the exponential decay rate, taken as 0.9; η represents the learning rate, ∈ is a constant, with a value of 10-8; J(θ) is the stochastic objective function; d is the adjustment parameter;

[0037] The learning rate gradually decreases following a cosine function pattern in each training epoch, as shown in the following formula:

[0038]

[0039] where i represents the index number, and represent the maximum and minimum values of the learning rate respectively, and T curRepresents the number of current learning experiences; T i Represents the number of learning times from the i-th warm restart to the (i + 1)-th warm restart;

[0040] S3.4. Set the loss function in the training process to the cross-entropy function Loss, and the expression is:

[0041]

[0042] Among them, Q(j) and P(j) respectively represent the probabilities of the j-th category in the predicted distribution and the true distribution.

[0043] Furthermore, in step S3, Bayesian optimization is combined to optimize the hyperparameters.

[0044] Furthermore, in step S1, with the lift coefficient C L , drag coefficient C D , pitch moment coefficient C M and volume ratio η ν as the four optimization performance indicators of the aerodynamic shape of the aerial dispenser, samples of low-fidelity, medium-fidelity, and high-fidelity models with grid sizes of 300, 150, and 60 are respectively used, and the input parameter data is from the aerodynamic coefficient dataset calculated by CFD.

[0045] Furthermore, in step S1, for the lift coefficient C L , drag coefficient C D , pitch moment coefficient C M and volume ratio η ν as the four optimization performance indicators of the aerodynamic shape of the aerial dispenser, two groups of VDNN-TL models are used for modeling.

[0046] Furthermore, in step S1, the pitch moment coefficient of the aerial dispenser needs to meet certain constraint conditions to ensure good maneuverability and stability characteristics. The optimized mathematical model is expressed as follows:

[0047]

[0048] Among them, x represents the aerodynamic shape parameter variable, X i represents the range of the i-th aerodynamic shape parameter variable, n represents the number of aerodynamic shape parameters, f obj1 (x) is the lift-to-drag ratio optimization index function, f obj2 (x) is the volume ratio optimization index function.

[0049] An electronic device, characterized in that it includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the intelligent adaptive fidelity optimization design method of an aerial dispenser.

[0050] A computer-readable storage medium stores a computer program thereon, characterized in that when the computer program is executed by a processor, it implements the intelligent adaptive fidelity optimization design method of an aerial dispenser.

[0051] Advantages of the present invention:

[0052] The intelligent adaptive fidelity optimization design method of an aerial dispenser according to the present invention is based on an intelligent algorithm of deep learning, which solves the problem that traditional machine learning methods have a strong correlation dependence between low-fidelity data and high-fidelity data. At the same time, it can effectively utilize data of three fidelity levels, filter out large errors in low-fidelity data while retaining useful information, and convert it into medium-fidelity data. Finally, a small amount of high-fidelity data is used for fine-tuning to achieve the accuracy of intelligent adaptive fidelity optimization of the aerial dispenser. Description of the drawings

[0053] Figure 1 It is a flowchart of the intelligent adaptive fidelity optimization design method of an aerial dispenser according to the present invention;

[0054] Figure 2 It is a curve graph of different grid calculation results of the present invention;

[0055] Figure 3 It is a schematic diagram of the neuron calculation principle of the present invention;

[0056] Figure 4 It is a network architecture diagram of the VDNN-TL neural network surrogate model of the present invention;

[0057] Figure 5 It is a network working framework diagram of the VDNN-TL neural network surrogate model of the present invention. Detailed implementation manners

[0058] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners. It should be understood that the specific implementation manners described herein are only used to explain the present invention and are not used to limit the present invention, that is, the specific implementation manners described are only a part of the implementation manners of the present invention, rather than all of the specific implementation manners. Usually, the components of the specific implementation manners of the present invention described and shown in the accompanying drawings herein can be arranged and designed in various different configurations, and the present invention can also have other implementation manners.

[0059] Accordingly, the following detailed description of the specific embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected specific embodiments of the present invention. All other specific embodiments obtained by those skilled in the art based on the specific embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0060] To further understand the content, features and effects of the present invention, the following specific embodiments are exemplified and are accompanied by the attached Figure 1 - Attachment Figure 5 The detailed description is as follows:

[0061] Example 1:

[0062] An intelligent adaptive fidelity optimization design method for an aerial dispenser, comprising the following steps:

[0063] S1. Data acquisition and preprocessing: Extract the aerodynamic shape data of the aerial dispenser as model samples through experimental design, and obtain sample data with different grid sizes by means of engineering sampling, which are used as low-fidelity sample data, medium-fidelity sample data, and high-fidelity sample data respectively. Then, perform standardized preprocessing on the data to obtain a low-fidelity sample data set, a medium-fidelity sample data set, and a high-fidelity sample data set;

[0064] Further, in step S1, sample data with grid sizes of 60, 150, and 300 are obtained by means of engineering sampling, which are used as low-fidelity sample data, medium-fidelity sample data, and high-fidelity sample data respectively; as Figure 2 shown, in order to determine the grid scales for obtaining medium-fidelity and high-fidelity CFD data, a convergence analysis was carried out. Six different grid scales were used to calculate the lift-to-drag ratio under the same shape and conditions. The results show that compared with the smallest grid, the second-level grid produced significantly improved results and also had a shorter calculation time. Therefore, the L2-level grid was selected to obtain medium-fidelity data. For high-fidelity data, since the difference between the results of the L5-level and L6-level grids was not significant, but the calculation time of the latter was significantly longer. Therefore, the L5-level grid was selected to obtain high-fidelity data. The calculation time for the L5-level calculation was 1212 s, while the calculation time for the L2-level calculation was 305 s, meeting our assumption in model validation that the time and cost for obtaining medium and high-fidelity samples differed by at least three times.

[0065] S2. Construct a VDNN-TL neural network surrogate model using a fully connected neural network, including a low-fidelity deep neural network surrogate model LDNN, a medium-fidelity deep neural network surrogate model MDNN, and a high-fidelity deep neural network surrogate model HDNN. The low-fidelity deep neural network surrogate model LDNN, the medium-fidelity deep neural network surrogate model MDNN, and the high-fidelity deep neural network surrogate model HDNN are connected through transfer learning;

[0066] Further, the specific implementation method of step S2 includes the following steps:

[0067] S2.1. The expression for the calculation process of the neurons in the fully connected neural network is:

[0068]

[0069] where y represents the output of the neuron, f(·) is the activation function, ω i is the weight, x i is the input, and b is the bias. By adjusting the weight and bias, the neuron can learn different features;

[0070] Further, the deep learning regression model is established based on the fully connected neural network and transfer learning. A fully connected neural network (FCNN), also known as a multi-layer perceptron, is a commonly used artificial neural network structure composed of multiple layers of neurons. Each neuron is connected to all neurons in the previous layer, forming a fully connected structure. In the FCNN, each neuron is a mathematical model responsible for receiving input signals, performing weighted processing, and activating the output. The input of each neuron is the output of all neurons in the previous layer. After weighted summation, it is then processed by a non-linear activation function to obtain the output of the neuron. This output serves as the input for the neurons in the next layer. The training process of the FCNN is based on the backpropagation algorithm, which adjusts the weights and biases of the network to minimize the loss function. When dealing with large-scale data, the backpropagation algorithm can utilize the data in batches and repeat for multiple epochs. Therefore, a deep learning regression model is adopted to improve the fitting accuracy of the Co-Kriging surrogate model.

[0071] Further, based on the Co-Kriging surrogate model, a triple variable-fidelity surrogate model of the aerial dispenser is constructed using the deep neural network transfer learning technology. The VDNN-TL framework proposed in this embodiment is introduced below.

[0072] S2.2. Construct a low-fidelity deep neural network surrogate model LDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer;

[0073] The first layer of the low-fidelity deep neural network surrogate model is the input layer, which is responsible for inputting feature data and obtaining the low-fidelity sample dataset of the aerial dispenser. The number of neurons in the input layer is the same as the dimension of the shape structure parameters;

[0074] The second and third layers of the low-fidelity deep neural network surrogate model are feature extraction layers, and the number of neurons in the third layer is twice that of the second layer;

[0075] The fourth to sixth layers of the low-fidelity deep neural network surrogate model are regression adjustment layers, and the number of neurons in each layer is the same;

[0076] The seventh layer of the low-fidelity deep neural network surrogate model is the output layer, and the number of neurons matches the number of output parameters;

[0077] Set the number of neurons in the first layer of the feature extraction layer of the low-fidelity deep neural network surrogate model to n1, the number of neurons in the regression adjustment layer to n2, the learning rate of the low-fidelity deep neural network surrogate model to g1, the batch size of the low-fidelity deep neural network surrogate model to b1, and the activation function of the low-fidelity deep neural network surrogate model to f1;

[0078] S2.3. Construct the medium-fidelity deep neural network surrogate model MDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer. The model structure is that the first layer is the input layer, the second and third layers are feature extraction layers, the fourth to sixth layers are regression adjustment layers, the number of neurons in each layer of the regression adjustment layer is the same, and the seventh layer is the output layer;

[0079] The input layer and feature extraction layer in the medium-fidelity deep neural network surrogate model are obtained through transfer learning of LDNN, and the regression adjustment layer and output layer are obtained through training;

[0080] Set the number of neurons in each neural network layer of the regression adjustment layer in the medium-fidelity deep neural network surrogate model to n3, the activation function of the medium-fidelity deep neural network surrogate model to f2, the learning rate of the medium-fidelity deep neural network surrogate model to g2, and the batch size of the medium-fidelity deep neural network surrogate model is directly transferred and learned from LDNN;

[0081] S2.4. Construct the high-fidelity deep neural network surrogate model HDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer. The model structure is that the first layer is the input layer, the second and third layers are feature extraction layers, the fourth to sixth layers are regression adjustment layers, the number of neurons in each layer of the regression adjustment layer is the same, and the seventh layer is the output layer;

[0082] The input layer, feature extraction layer, and regression adjustment layer in the high-fidelity deep neural network proxy model are obtained through the transfer learning of MDNN. The output layer fine-tunes the neurons in the output layer using the high-fidelity sample dataset. Then, the learning rate of the high-fidelity deep neural network proxy model in HDNN is set to g3, and the activation function and batch size are the same as those in MDNN.

[0083] S3. Use the low-fidelity sample dataset, medium-fidelity sample dataset, and high-fidelity sample dataset obtained in step S1 to train the low-fidelity deep neural network proxy model LDNN, medium-fidelity deep neural network proxy model MDNN, and high-fidelity deep neural network proxy model HDNN in the VDNN-TL neural network proxy model constructed in step S2, and select and adjust the training hyperparameters in the model to achieve high-precision fitting of the VDNN-TL neural network proxy model constructed by the fully connected neural network;

[0084] Furthermore, the specific implementation method of step S3 includes the following steps:

[0085] S3.1. Set the adjusted hyperparameters during the training process to include n1, n2, f1, g1, b1, n3, f2, g2, g3;

[0086] Furthermore, on the basis of determining the VDNN-TL network structure, samples with different fidelity levels are used to train the model. The loss function is set to the mean square error between the predicted value of each model and the sample label value. During the training process of each sub-model, the training objective is to make the corresponding fidelity data achieve the optimal regression performance without considering other fidelity levels. The adjustable parameters of each sub-model are introduced below.

[0087] In the VDNN-TL network, there are nine adjustable hyperparameters that need to be optimized. Table 1 lists the adjustable hyperparameters of each fidelity sub-model.

[0088] Table 1 Adjustable Parameters of Sub-Models

[0089]

[0090] S3.2. Set the training process to adopt an early stopping strategy. When the generalization loss GL(t) is greater than a certain value, stop the training. The generalization loss GL(t) is defined as follows:

[0091]

[0092] where, E va (t) is the validation set error, and E opt (t) is the lowest validation set error up to time t;

[0093]

[0094] where t' is the moment before t;

[0095] S3.3. Set to use the Accelerated Adam optimizer during training, combining the advantages of momentum and adaptive learning rate adjustment to accelerate convergence to the local optimum. The Accelerated Adam optimizer is described as follows:

[0096]

[0097] where θ represents the optimization parameter, represents the exponentially weighted average of past squared gradients, represents the exponentially weighted average of past gradients, β1 is the exponential decay rate, taken as 0.9; η represents the learning rate, ∈ is a constant, with a value of 10-8; J(θ) is the stochastic objective function; d is the adjustment parameter;

[0098] The learning rate gradually decreases following a cosine function pattern in each training epoch, as shown in the following equation:

[0099]

[0100] where i represents the index number, and represent the maximum and minimum values of the learning rate respectively, T cur represents the number of times the current learning has experienced; T i represents the number of learning times from the i-th warm restart to the (i + 1)-th warm restart;

[0101] S3.4. Set the loss function to the cross-entropy function Loss during training, and the expression is:

[0102]

[0103] where Q(j) and P(j) represent the probabilities of the j-th category in the predicted distribution and the true distribution respectively.

[0104] Furthermore, in step S3, Bayesian optimization is combined to optimize hyperparameters;

[0105] Combining Bayesian optimization to optimize hyperparameters enables the VDNN-TL variable-fidelity surrogate model to be automatically executed, improving the efficiency of VDNN-TL model construction. The Bayesian hyperparameter optimization steps are as follows:

[0106] 1) Adopt the posterior distribution and update the previous results of the objective function using Gaussian processes.

[0107] 2) Select the best point of the objective function by using the acquisition function

[0108] 3) Identify the proposed sampling points obtained by the acquisition function.

[0109] 4) Use the objective function to obtain the results in the validation set.

[0110] 5) Enhance the best optimized sample points to the selected data.

[0111] 6) Update the statistical Gaussian distribution model.

[0112] Repeat this operation until a certain number of iterations are performed to adjust the validation set to obtain optimized parameters that can achieve better classification.

[0113] S4. Construct a VDNN-TL neural network surrogate model for the trained fully connected neural network obtained in step S3, and perform two rounds of transfer learning. The first round is from the low-fidelity deep neural network surrogate model LDNN to the medium-fidelity deep neural network surrogate model MDNN, and the second round is from the medium-fidelity deep neural network surrogate model MDNN to the high-fidelity deep neural network surrogate model HDNN. Then, use fine-tuning to correct the calculation results to obtain the intelligent adaptive fidelity optimization design results of the aerial dispenser.

[0114] The practical operation using this embodiment is as follows: Use the VDNN-TL surrogate model and high-fidelity CFD simulation to calculate the aerodynamic coefficients and volume ratio. In step S1, the lift coefficient C L , drag coefficient C D , pitching moment coefficient C M and volume ratio η ν are taken as the four optimization performance indicators of the aerodynamic shape of the aerial dispenser. Samples of low-fidelity, medium-fidelity, and high-fidelity models with grid sizes of 300, 150, and 60 are used respectively, and the input parameter data is from the aerodynamic coefficient dataset calculated by CFD.

[0115] Considering that the volume ratio of the aerial dispenser is only related to the shape and has nothing to do with the flight conditions. Therefore, a separate model is used to predict the volume ratio, and finally, two sets of VDNN-TL models are used to complete the modeling of these four performance indicators. The hyperparameter sizes of VDNN-TL are shown in Table 2:

[0116] Table 2 VDNN-TL Hyperparameter Data Table

[0117]

[0118] Further, in step S1, for the lift coefficient C L , drag coefficient C D , pitching moment coefficient C M and volume ratio η νFour optimized performance indicators for the aerodynamic shape of the aerial dispenser are modeled using two groups of VDNN-TL models.

[0119] Furthermore, in step S1, the pitching moment coefficient of the aerial dispenser needs to satisfy certain constraint conditions to ensure good maneuverability and stability characteristics. The optimized mathematical model is expressed as follows:

[0120]

[0121] Among them, x represents the variable of the aerodynamic shape parameter, X i represents the range of the i-th aerodynamic shape parameter variable, n represents the number of aerodynamic shape parameters, f obj1 (x) is the optimization index function of the lift-drag ratio, f obj2 (x) is the optimization index function of the volume ratio.

[0122] The final results are as follows: The prediction errors of the VDNN-TL model for both configurations are below 2%. In addition, after optimization, the lift-drag ratio and volume ratio of the aerial dispenser are increased by 7.86% and 26.2% respectively. The pitching moment coefficient also satisfies the constraint conditions. Compared with directly using CFD for optimization, the optimization efficiency of the VDNN-TL surrogate model is increased by 98.9%.

[0123] Example 2:

[0124] An electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of an intelligent adaptive fidelity optimization design method for an aerial dispenser described in Example 1.

[0125] The computer device of the present invention may be a device including a processor and a memory, such as a single-chip microcomputer including a central processing unit. And, when the processor is used to execute the computer program stored in the memory, it implements the steps of the above-mentioned intelligent adaptive fidelity optimization design method for an aerial dispenser.

[0126] The so-called processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0127] The memory may mainly include a program storage area and a data storage area. Among them, the program storage area may store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area may store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0128] Embodiment 3:

[0129] A computer-readable storage medium stores a computer program thereon, and when the computer program is executed by a processor, it implements the method for optimizing the intelligent adaptive fidelity design of an aerial dispenser described in Embodiment 1.

[0130] The computer-readable storage medium of the present invention can be any form of storage medium readable by the processor of a computer device, including but not limited to non-volatile memory, volatile memory, ferroelectric memory, etc. A computer program is stored on the computer-readable storage medium. When the processor of the computer device reads and executes the computer program stored in the memory, the steps of the above-mentioned method for optimizing the intelligent adaptive fidelity design of an aerial dispenser can be implemented.

[0131] The computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0132] It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising said element.

[0133] Although the present application has been described above with reference to specific embodiments, various improvements can be made thereto and components thereof can be replaced with equivalents without departing from the scope of the present application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in the present application can be combined with each other in any way, and the reason for not exhaustively describing the situations of these combinations in this specification is only to save space and resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An intelligent adaptive fidelity optimization design method for an aerial dispenser, characterized in that: The steps include: S1. Data collection and preprocessing: Through experimental design, the aerodynamic shape data of the aerial dispenser is extracted as a model sample. Through the engineering sampling method, sample data of different grid sizes are obtained, which are used as low-fidelity sample data, medium-fidelity sample data, and high-fidelity sample data, respectively. Then, the data is standardized and preprocessed to obtain low-fidelity sample data sets, medium-fidelity sample data sets, and high-fidelity sample data sets; S2. constructing a VDNN-TL neural network proxy model using a fully connected neural network, including a low-fidelity deep neural network proxy model LDNN, a medium-fidelity deep neural network proxy model MDNN, and a high-fidelity deep neural network proxy model HDNN, wherein the low-fidelity deep neural network proxy model LDNN, the medium-fidelity deep neural network proxy model MDNN, and the high-fidelity deep neural network proxy model HDNN are connected through transfer learning; S3. Use the low-fidelity sample data set, medium-fidelity sample data set, and high-fidelity sample data set obtained in step S1 to train the low-fidelity deep neural network proxy model LDNN, the medium-fidelity deep neural network proxy model MDNN, and the high-fidelity deep neural network proxy model HDNN in the fully connected neural network constructed VDNN-TL neural network proxy model in step S2, and select and adjust the training hyperparameters in the model to achieve high-precision fitting of the fully connected neural network constructed VDNN-TL neural network proxy model; S4. Construct a VDNN-TL neural network proxy model for the trained fully connected neural network obtained in step S3, and perform two rounds of transfer learning. The first round is to migrate from the low-fidelity deep neural network proxy model LDNN to the medium-fidelity deep neural network proxy model MDNN, and the second round is to migrate from the medium-fidelity deep neural network proxy model MDNN to the high-fidelity deep neural network proxy model HDNN. Then, fine-tuning is used to correct the calculation results to obtain the intelligent adaptive fidelity optimization design results of the aerial dispenser.

2. The intelligent adaptive fidelity optimization design method for an aerial dispenser according to claim 1, characterized in that: In step S1, sample data with grid sizes of 60, 150, and 300 are obtained by an engineering sampling method as low-fidelity sample data, medium-fidelity sample data, and high-fidelity sample data, respectively.

3. The intelligent adaptive fidelity optimization design method for an aerial dispenser according to claim 2, characterized in that: The specific implementation method of step S2 includes the following steps: S2.

1. The expression of the computational process of neurons in constructing a fully connected neural network is: Where y represents the output of the neuron, f(·) is the activation function, and ω i is the weight, x i is the input, b is the bias, and neurons learn different features by adjusting weights and biases; S2.

2. Construct a low-fidelity deep neural network proxy model LDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer; The first layer in the low-fidelity deep neural network proxy model is the input layer, which is responsible for inputting feature data and is used to obtain the low-fidelity sample data set of the aerial dispenser. The number of neurons in the input layer is consistent with the dimension of the shape structure parameters. The second and third layers in the low-fidelity deep neural network proxy model are feature extraction layers, and the number of neurons in the third layer is twice that of the second layer; The fourth to sixth layers in the low-fidelity deep neural network proxy model are regression adjustment layers, and the number of neurons in each layer is the same; The seventh layer in the low-fidelity deep neural network proxy model is the output layer, and the number of neurons matches the number of output parameters; The low-fidelity deep neural network proxy model sets the number of neurons in the first layer of the feature extraction layer to n1, the number of neurons in the regression adjustment layer to n2, sets the learning rate of the low-fidelity deep neural network proxy model to g1, the batch size of the low-fidelity deep neural network proxy model to b1, and the activation function of the low-fidelity deep neural network proxy model to f1; S2.

3. Construct a medium-fidelity deep neural network proxy model MDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer. The model structure is that the first layer is the input layer, the second and third layers are feature extraction layers, the fourth to sixth layers are regression adjustment layers, the number of neurons in each layer of the regression adjustment layer is the same, and the seventh layer is the output layer; The input layer and feature extraction layer in the medium-fidelity deep neural network proxy model are obtained through LDNN transfer learning, and the regression adjustment layer and output layer are obtained through training; Set the number of neurons in each neural network layer of the regression adjustment layer in the medium-fidelity deep neural network proxy model to n3, the activation function of the medium-fidelity deep neural network proxy model to f2, the learning rate of the medium-fidelity deep neural network proxy model to g2, and the batch size of the medium-fidelity deep neural network proxy model to transfer learning directly from LDNN; S2.

4. Construct a high-fidelity deep neural network proxy model HDNN, including an input layer, a feature extraction layer, a regression adjustment layer, and an output layer. The model structure is that the first layer is the input layer, the second and third layers are feature extraction layers, the fourth to sixth layers are regression adjustment layers, the number of neurons in each layer of the regression adjustment layer is the same, and the seventh layer is the output layer; The input layer, feature extraction layer, and regression adjustment layer in the high-fidelity deep neural network proxy model are obtained through transfer learning of MDNN, and the neurons in the output layer are fine-tuned through a high-fidelity sample data set. The learning rate of the high-fidelity deep neural network proxy model is set to g3 in HDNN, and the activation function and batch size remain the same as those in MDNN.

4. The intelligent adaptive fidelity optimization design method for an aerial dispenser according to claim 3 is characterized in that: The specific implementation method of step S3 includes the following steps: S3.

1. Set the hyperparameters adjusted during training to include n1, n2, f1, g1, b1, n3, f2, g2, g3; S3.

2. Set the training process to adopt an early stopping strategy. When the generalization loss GL(t) is greater than a certain value, stop training. The generalization loss GL(t) is defined as follows: Among them, E va (t) is the validation set error, E opt (t) is the minimum validation set error at time t; where t' is the time before t; S3.

3. Set to use the accelerated Adam optimizer in the training process, combining the advantages of momentum and adaptive learning rate adjustment to accelerate convergence to the local optimal value. The accelerated Adam optimizer is described as follows: Among them, θ represents the optimization parameter, represents the exponentially decaying average of past squared gradients, represents the exponential decay average of past gradients, β1 is the exponential decay rate, which is taken as 0.9; η represents the learning rate, ∈ is a constant, which is taken as 10-8; J(θ) is the random objective function; d is the adjustment parameter; The learning rate gradually decreases in each training cycle following the cosine function pattern, as shown in the following formula: Among them, i represents the number of indexes, and Respectively represent the maximum and minimum values ​​of the learning rate, T cur Indicates the number of current learning experiences; T i Indicates the number of learning times from the i-th hot restart to the i+1-th hot restart; S3.

4. Set the loss function in the training process to the cross entropy function Loss, expressed as: Among them, Q(j) and P(j) represent the probability of the jth category in the predicted distribution and the true distribution, respectively.

5. The intelligent adaptive fidelity optimization design method for aerial dispensers according to claim 4 is characterized in that: In step S3, the hyperparameters are optimized in combination with Bayesian optimization.

6. The intelligent adaptive fidelity optimization design method for aerial dispensers according to claim 5, characterized in that: In step S1, the lift coefficient C L , drag coefficient C D , pitch moment coefficient C M and volume ratio η ν For the four optimized performance indicators of the aerodynamic shape of the aerial dispenser, samples of low-fidelity, medium-fidelity and high-fidelity models with mesh sizes of 300, 150 and 60 are used respectively, and the input parameter data are derived from the aerodynamic coefficient data set calculated by CFD.

7. The intelligent adaptive fidelity optimization design method for aerial dispensers according to claim 6, characterized in that: In step S1, for the lift coefficient C L , drag coefficient C D , pitch moment coefficient C M and volume ratio η ν Two sets of VDNN-TL models are used to model the four optimized performance indicators of the aerodynamic shape of the aerial dispenser.

8. The intelligent adaptive fidelity optimization design method for aerial dispensers according to claim 7, characterized in that: The pitch moment coefficient of the aerial dispenser in step S1 needs to meet certain constraints to ensure good maneuverability and stability. The optimized mathematical model is expressed as follows: Among them, x represents the aerodynamic shape parameter variable, X i represents the range of the i-th aerodynamic shape parameter variable, n represents the number of aerodynamic shape parameters, f obj1 (x) is the lift-to-drag ratio optimization index function, f obj2 (x) is the volume ratio optimization index function.

9. An electronic device, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the intelligent adaptive fidelity optimization design method for an aerial dispenser as claimed in any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the intelligent adaptive fidelity optimization design method for an aerial dispenser as described in any one of claims 1 to 8 is implemented.

Images