Intelligent reconstruction method of aircraft aerodynamic characteristics based on multi-source data fusion

Abstract

The present invention discloses a method for intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion, which relates to the field of intelligent fluid mechanics prediction, including: preparing aircraft multi-source aerodynamic characteristics data, establishing an intelligent dimension reduction model of aircraft flow characteristics, preparing a data set for an aircraft aerodynamic characteristics intelligent prediction model for multi-source data fusion, establishing an aircraft aerodynamic characteristics intelligent prediction model for multi-source data fusion, iteratively training the aircraft aerodynamic characteristics intelligent prediction model, using a corresponding verification data set for verification, and selecting a weight model with the best performance to predict subsequent working conditions to be tested. The present invention provides a method for intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion, which can achieve efficient fusion of multi-source data information, thereby constructing a high-precision aircraft aerodynamic characteristics intelligent reconstruction prediction model.

Description

Technical Field

[0001] The present invention relates to the field of intelligent fluid mechanics prediction, and more specifically, to an intelligent reconstruction method of aircraft aerodynamic characteristics based on multi-source data fusion. Background Art

[0002] The prediction of aerodynamic characteristics in the current design and development of aerospace vehicles mainly relies on wind tunnel tests and numerical calculation methods, and some simple corrections are made based on flight test data. Wind tunnel tests are the most reliable method to obtain limited sample data under limited working conditions, but they have the characteristics of high cost, long cycle, and small sample size; high-fidelity CFD numerical simulations usually have large calculation amounts and take a long time. The substantial improvement in the advanced indicators of aircraft has put forward higher requirements for the accuracy of the prediction of aerodynamic characteristics. It is difficult for any single method to accurately predict the aerodynamic characteristics of an aircraft.

[0003] The current method of reconstructing the aerodynamic characteristics of aircraft based on multi-source data fusion mainly achieves the mapping between input parameters and output aerodynamic characteristics by building a machine learning model. In most studies, only information such as flow field parameters and aircraft attitude is used as input parameters, without fully considering the structural characteristics of the aircraft flow field (such as flow field structure and aircraft aerodynamic shape). In addition, when using numerical calculation data fusion, the flow field reduction and feature selection methods used to simplify highly complex nonlinear flow field data still have certain limitations.

[0004] In summary, it is an inevitable requirement to improve the accuracy of prediction of aerodynamic characteristics of new aircraft by conducting data mining on numerically calculated images of aircraft flow field structures with rich flow characteristics, exploring dimensionality reduction methods for complex high-dimensional data, and realizing deep fusion of multi-source data. Summary of the invention

[0005] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.

[0006] In order to achieve these purposes and other advantages of the present invention, a method for intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion is provided, comprising:

[0007] S1. Based on the study of the aerodynamic characteristics of the aircraft, select the corresponding wind tunnel test data and numerical calculation simulation data;

[0008] S2. Based on the image classification algorithm as the network architecture for extracting aircraft flow features, the t-distributed random neighbor embedding algorithm t-SNE is combined as a nonlinear dimensionality reduction method for aircraft flow features to establish an intelligent dimensionality reduction model for aircraft flow features, and obtain the corresponding dimensionality reduction data through iterative training;

[0009] S3. Using a multi-input and multi-output deep artificial neural network as the basic framework, an intelligent prediction model for aircraft aerodynamic characteristics based on multi-source data fusion is established;

[0010] S4. The numerical calculation simulation data in S1 and the dimension reduction data in S2 are used as input parameters of the intelligent prediction model of the aerodynamic characteristics of the aircraft, and iterative training is performed to select the weighted model with the best performance to predict the subsequent working conditions to be tested, thereby completing the intelligent reconstruction of the aerodynamic characteristics of the aircraft.

[0011] Preferably, in S1, the wind tunnel test data includes: axial force I, normal force I, lateral force I, rolling moment I, yaw moment I, pitching moment I;

[0012] The numerical calculation simulation data include: numerical calculation simulation results of aircraft aerodynamic characteristics and pressure distribution cloud diagram of numerical calculation simulation flow field of aircraft;

[0013] The numerical calculation simulation results of the aerodynamic characteristics of the aircraft include: axial force II, normal force II, lateral force II, rolling moment II, yaw moment II, and pitch moment II;

[0014] The numerical calculation simulation flow field pressure distribution cloud map of the aircraft is classified according to the Mach number when used, and data expansion processing is performed.

[0015] Preferably, in S2, the aircraft flow feature extraction network includes: a feature extraction module and a classification module;

[0016] The feature extraction module includes 5 levels of structural layers, each of which includes two consecutive 3×3 convolutional layers and a ReLU activation layer, and the structural layers of adjacent levels are downsampled through a 2×2 maximum pooling layer;

[0017] The classification module includes three fully connected layers, and the output features of the last fully connected layer are used as the input of the t-SNE algorithm.

[0018] Preferably, in S2, before the aircraft flow characteristic intelligent dimension reduction model is iteratively trained, the data consisting of the pressure distribution cloud map of the aircraft numerical simulation flow field in S1 is divided into a training set, a verification set and a test set based on corresponding working conditions;

[0019] When the intelligent dimensionality reduction model of aircraft flow characteristics is iteratively trained, the training set is used as input for iterative training, and the corresponding validation set is used for validation to extract the high-dimensional features corresponding to the last fully connected layer of the optimal weight model, and the t-SNE algorithm is used to perform nonlinear dimensionality reduction of high-dimensional features.

[0020] Preferably, in S3, the aircraft aerodynamic characteristics intelligent prediction model adopts a random search strategy to realize automatic training of the network and hyperparameters, and the hyperparameters of the random search strategy are set as follows:

[0021] The minimum number of network layers is 3, the maximum number of network layers is 10, the minimum number of neurons in the hidden layer is 5, the maximum number of neurons in the hidden layer is 10, the minimum order of magnitude of the regularization parameter is 2, the maximum order of magnitude of the regularization parameter is 5, the learning rate is 0.001, and the total number of iterations is 8000.

[0022] The present invention has at least the following beneficial effects: In view of the limitation of single means and low efficiency of multi-source data fusion in predicting the aerodynamic characteristics of aircraft, the present invention firstly performs data mining on the numerical calculation results of the aircraft flow field structure, that is, performs feature extraction and dimensionality reduction analysis on the flow field image with rich flow characteristics and aircraft geometric configuration information. Then, on this basis, the wind tunnel test data is integrated to achieve efficient fusion of multi-source data information, thereby constructing a high-precision intelligent reconstruction prediction model for aircraft aerodynamic characteristics.

[0023] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of the intelligent reconstruction method of aerodynamic characteristics of an aircraft based on multi-source data fusion in the present invention;

[0025] Figure 2 It is an expanded example diagram of the pressure distribution cloud diagram of the numerical simulation flow field of the aircraft of the present invention;

[0026] Figure 3 This is a schematic diagram of the visualization results of the t-SNE dimensionality reduction algorithm analysis of the present invention. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0028] like Figure 1 As shown, a method for intelligent reconstruction of aerodynamic characteristics of an aircraft based on multi-source data fusion includes:

[0029] Step 1: Prepare multi-source aerodynamic characteristics data of the aircraft, including wind tunnel test data of corresponding working conditions (such as flow field information parameters, aircraft attitude information) and numerical calculation simulation data.

[0030] For example, in one embodiment, for the study of aerodynamic characteristics of an aircraft, the wind tunnel test data includes: axial force, normal force, lateral force, rolling moment, yaw moment, and pitch moment. The numerical calculation simulation data includes: (1) numerical calculation simulation results of aircraft aerodynamic characteristics: axial force, normal force, lateral force, rolling moment, yaw moment, and pitch moment; (2) numerical calculation simulation flow field pressure distribution cloud map of the aircraft, which is classified according to Mach number in subsequent specific use. In addition, in order to enhance the robustness of the deep learning model, the data is expanded. Figure 2 As shown in the figure, under the working conditions of a certain angle of attack of Mach number 0.4, 0.6 and 0.8, 4 kinds of angles are used to output the flow field pressure distribution cloud map to complete the data expansion processing.

[0031] Step 2: Establish an intelligent dimensionality reduction model for aircraft flow features. This model uses the image classification algorithm in the field of computer vision as the aircraft flow feature extraction network architecture, and uses the t-distributed stochastic neighbor embedding (t-SNE) algorithm as the aircraft flow feature nonlinear dimensionality reduction method. The aircraft flow feature extraction network includes a feature extraction module and a classification module. The backbone feature extraction module consists of 5 levels of structural layers, each of which includes two consecutive 3×3 convolutional layers and a ReLU activation layer. The downsampling process between two connected structural layers is implemented using a 2×2 maximum pooling layer. After each level of structural layer processing, the size (width and height) of the feature map of the next level of structural layer is reduced to half of the previous level, while the number of channels of the feature map increases to twice that of the previous level. The classification module consists of three fully connected layers, and finally the image classification is realized through the Softmax activation function. Among them, the output features of the last fully connected layer are used as the input of the subsequent t-SNE algorithm.

[0032] The design details, operation operators and output data flow format of the aircraft flow feature extraction network are shown in Table 1. In this example, the image input size is 512 × 512 and the output category is 3. The data dimension provided to the subsequent t-SNE algorithm is 4096.

[0033] Table 1

[0034]

[0035] During model training, the pre-trained weights of the VGG13 network trained on the large ImageNet dataset are loaded. Since the VGG13 network has the same backbone feature extraction module structure as the aircraft flow feature extraction network in the present invention, loading the pre-trained weight initialization features can effectively improve the training efficiency and effect.

[0036] After multiple optimizations of model training hyperparameters, the optimal hyperparameter settings were determined: the initial learning rate was 1e-3, the minimum learning rate was 1e-4, and the learning rate was adjusted using the step-descent method; the Adam optimizer was used to train the model with a momentum parameter of 0.9; the weight decay coefficient was 5e-4; the total number of training rounds was 50; and the settings for loading pre-trained models were: 5 rounds of model freezing training, 32 batch parameter size for model freezing training, and 16 batch parameter size for the model after thawing.

[0037] After the model is built, the data of the pressure distribution cloud map of the numerical simulation flow field of the aircraft obtained in step 1 is divided into training set, verification set and test set according to the corresponding working conditions and the predetermined ratio. Then the training set is input into the model for iterative training, and the corresponding verification data set is used for verification to select the weight model with the best performance.

[0038] The high-dimensional features of the last fully connected layer of the optimal weight model are extracted, and the t-SNE algorithm is used to perform nonlinear dimensionality reduction of the high-dimensional features. In this example, three Mach numbers are used as features to extract classification targets. In the subsequent study of intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion, 9 embedding space dimensions from 2 to 10 dimensions are set for cross-validation. Finally, it is determined that the embedding dimension is 2 dimensions with the best performance. The visualization results of the t-SNE dimensionality reduction algorithm analysis are shown in the figure below. Figure 3 It can be found that the flow field structure samples with the same Mach number are obviously and intuitively clustered into a cluster, which shows that the aircraft flow feature extraction network proposed in the present invention has excellent aircraft flow feature capability.

[0039] Step 3, prepare a data set for the intelligent prediction model of aircraft aerodynamic characteristics for multi-source data fusion. Specifically, the numerical calculation simulation data of aircraft aerodynamic characteristics obtained in step 1 and the intelligent dimensionality reduction results of aircraft flow characteristics obtained in step 2 are used as input parameters of the model, and the wind tunnel test data obtained in step 1 are used as output parameters of the model. Table 2 is a data example of the intelligent prediction model of aircraft aerodynamic characteristics for multi-source data fusion. Among them, the input parameters are 2 dimensionality reduction feature value results and 6 numerical calculation simulation values. The output parameters are 6 wind tunnel test values.

[0040] Table 2

[0041]

[0042] Step 4: Establish an intelligent prediction model for aircraft aerodynamic characteristics based on multi-source data fusion. This model uses a multi-input and multi-output deep artificial neural network as the basic architecture (such as Figure 1In this example, the hyperparameters of the random search strategy are set as follows: minimum number of network layers: 3; maximum number of network layers: 10; minimum number of neurons in the hidden layer: 5; maximum number of neurons in the hidden layer: 10; minimum order of magnitude of regularization parameter: 2; maximum order of magnitude of regularization parameter: 5; learning rate: 0.001; total number of iterations: 8000. In actual situations, the above parameters can be adjusted based on comprehensive considerations of factors such as the size of the data set, hardware computing power, and time efficiency.

[0043] Step 5. Then, input the training data set constructed in step 3 into the aircraft aerodynamic characteristics intelligent prediction model with multi-source data fusion for iterative training, and use the corresponding verification data set for verification. The weighted model with the best performance can be selected to predict the subsequent working conditions to be tested.

[0044] The prediction results of some working conditions using the aircraft aerodynamic characteristics intelligent prediction model with the selected optimal multi-source data fusion are shown in Table 3. For the prediction of extended and interpolated working conditions, the errors are within 3%, which shows that the aircraft aerodynamic characteristics intelligent reconstruction method based on multi-source data fusion established by the present invention has good prediction accuracy.

[0045] Table 3

[0046]

[0047] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacement and / or modification can be performed according to user needs.

[0048] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims

1. A method for intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion, characterized in that: include: S1. Based on the study of the aerodynamic characteristics of the aircraft, select the corresponding wind tunnel test data and numerical calculation simulation data; S2. Based on the image classification algorithm as the network architecture for extracting aircraft flow features, the t-distributed random neighbor embedding algorithm t-SNE is combined as a nonlinear dimensionality reduction method for aircraft flow features to establish an intelligent dimensionality reduction model for aircraft flow features, and obtain the corresponding dimensionality reduction data through iterative training; S3. Using a multi-input and multi-output deep artificial neural network as the basic framework, an intelligent prediction model for aircraft aerodynamic characteristics based on multi-source data fusion is established; S4. The numerical calculation simulation data in S1 and the dimension reduction data in S2 are used as input parameters of the intelligent prediction model of the aerodynamic characteristics of the aircraft, and iterative training is performed to select the weighted model with the best performance to predict the subsequent working conditions to be tested, thereby completing the intelligent reconstruction of the aerodynamic characteristics of the aircraft.

2. The method for intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion according to claim 1, characterized in that: In S1, the wind tunnel test data include: axial force I, normal force I, lateral force I, rolling moment I, yaw moment I, pitching moment I; The numerical calculation simulation data include: numerical calculation simulation results of aircraft aerodynamic characteristics and pressure distribution cloud diagram of numerical calculation simulation flow field of aircraft; The numerical calculation simulation results of the aerodynamic characteristics of the aircraft include: axial force II, normal force II, lateral force II, rolling moment II, yaw moment II, and pitch moment II; The numerical calculation simulation flow field pressure distribution cloud map of the aircraft is classified according to the Mach number when used, and data expansion processing is performed.

3. The method for intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion according to claim 1, characterized in that: In S2, the aircraft flow feature extraction network includes: a feature extraction module and a classification module; The feature extraction module includes 5 levels of structural layers, each of which includes two consecutive 3×3 convolutional layers and a ReLU activation layer, and the structural layers of adjacent levels are downsampled through a 2×2 maximum pooling layer; The classification module includes three fully connected layers, and the output features of the last fully connected layer are used as the input of the t-SNE algorithm.

4. The method for intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion according to claim 2, characterized in that: In S2, before the aircraft flow characteristic intelligent dimension reduction model is iteratively trained, the data consisting of the pressure distribution cloud map of the aircraft numerical simulation flow field in S1 is divided into a training set, a validation set and a test set based on the corresponding working conditions; When the intelligent dimensionality reduction model of aircraft flow characteristics is iteratively trained, the training set is used as input for iterative training, and the corresponding validation set is used for validation to extract the high-dimensional features corresponding to the last fully connected layer of the optimal weight model, and the t-SNE algorithm is used to perform nonlinear dimensionality reduction of high-dimensional features.

5. The method for intelligent reconstruction of aircraft aerodynamic characteristics based on multi-source data fusion according to claim 1, characterized in that: In S3, the aircraft aerodynamic characteristics intelligent prediction model adopts a random search strategy to realize automatic training of the network and hyperparameters, and the hyperparameters of the random search strategy are set as follows: The minimum number of network layers is 3, the maximum number of network layers is 10, the minimum number of neurons in the hidden layer is 5, the maximum number of neurons in the hidden layer is 10, the minimum order of magnitude of the regularization parameter is 2, the maximum order of magnitude of the regularization parameter is 5, the learning rate is 0.001, and the total number of iterations is 8000.

Images