High-altitude arched shock wave flow field estimation method based on convolutional neural network

Abstract

The invention belongs to the technical field of aerodynamic thermal protection of hypersonic aircrafts, and particularly relates to a high-altitude arched shock wave flow field estimation method based on a convolutional neural network, which comprises the following steps: S1, constructing a data set; s2, constructing a high-altitude arch shock wave estimation neural network; s3, data preprocessing; s4, performing model training, optimization and verification; s5, flow field estimation: inputting flight condition parameters to be estimated into the trained and verified high-altitude arch shock wave estimation neural network, and obtaining an estimation result of high-altitude arch shock wave flow field distribution through forward propagation calculation; the problems that an existing high-altitude arch shock wave flow field calculation method is complex in calculation process, long in calculation time consumption, high in test cost and difficult to cover all flight working conditions are solved.

Description

Technical Field

[0001] This invention belongs to the field of aerodynamic thermal protection technology for hypersonic vehicles, specifically relating to a method for predicting the flow field of high-altitude bow-shaped shock waves based on convolutional neural networks. Background Technology

[0002] When hypersonic vehicles fly in near space, they generate strong high-altitude bow shocks at their noses. The interaction between the shock wave and the boundary layer causes a sharp increase in localized heat flux on the vehicle surface, resulting in severe aerodynamic heating and posing a significant challenge to the vehicle's thermal protection design. Accurate and rapid prediction of the flow field distribution behind the high-altitude bow shock is crucial for the design of the vehicle's thermal protection system.

[0003] Traditional flow field prediction methods primarily rely on Direct Carlo Simulation (DSMC). While DSMC offers high accuracy, its computational process is complex and time-consuming, requiring substantial computational resources and specialized technical personnel, making it difficult to meet the demands of rapid iteration and multi-condition analysis in the early stages of aircraft design. Furthermore, wind tunnel testing is costly and cannot cover all flight conditions. Therefore, developing a novel, rapid flow field prediction method that balances computational efficiency and prediction accuracy has become a critical technical challenge in this field. Summary of the Invention

[0004] To address the problems of existing high-altitude bow-shaped shock wave flow field calculation methods being complex, time-consuming, costly to test, and unable to cover all flight conditions, this invention provides a high-altitude bow-shaped shock wave flow field prediction method based on convolutional neural networks.

[0005] This invention is achieved using the following technical solution:

[0006] A method for predicting the flow field of high-altitude bow-shaped shock waves based on convolutional neural networks includes the following steps:

[0007] S1: Construct a dataset, use the Latin hypercube sampling method to sample the input flight parameters, and use the direct simulation Monte Carlo method to calculate the corresponding high-altitude bow shock wave flow field data, forming an input flight parameter set and an output flow field cloud map dataset;

[0008] S2: Construct a high-altitude bow-shaped shock wave prediction neural network. The high-altitude bow-shaped shock wave prediction neural network adopts an encoder-decoder structure, including an asymmetric convolutional encoder for compressing and extracting features and a parallel deconvolutional decoder for reconstructing the flow field cloud map.

[0009] S3: Data preprocessing: The input flight parameter set is used as the input dataset of the high-altitude bow shock wave prediction neural network. The output flow field cloud map dataset is normalized and interpolated and then used as the output dataset of the high-altitude bow shock wave prediction neural network.

[0010] S4: Model training, optimization and validation. The input dataset and output dataset are divided into training set and test set. The high-altitude bow shock wave prediction neural network is trained using the training set and the backpropagation algorithm to optimize the network weights. The loss function is monitored during the training process to prevent overfitting. The performance of the trained model is validated using the test set.

[0011] S5: Flow field prediction. The flight condition parameters to be predicted are input into the high-altitude bow shock prediction neural network that has been trained and verified. The prediction result of the high-altitude bow shock flow field distribution is obtained through forward propagation calculation.

[0012] Furthermore, in step S1, the set of input flight parameters includes the aircraft's nose radius, flight altitude, and flight speed.

[0013] Furthermore, in step S1, the output flow field cloud map dataset includes distribution cloud maps of translational temperature, rotational temperature, vibrational temperature, velocity, pressure, and NO number density.

[0014] Further, in step S2, the asymmetric convolutional encoder includes:

[0015] The three convolutional blocks are set in parallel, with kernel sizes of 3×3, 1×3 and 3×1 respectively. Each convolutional block contains a convolutional layer, a first ReLU activation function layer and a first batch normalization layer.

[0016] The feature fusion layer is used to fuse the features output from the three convolutional blocks;

[0017] The max pooling layer is used to downsample the features fused by the feature fusion layer.

[0018] Furthermore, in step S2, the parallel deconvolution decoder includes two decoding sub-modules with identical structures;

[0019] Each of the decoding submodules includes three cascaded deconvolution blocks, as follows:

[0020] The first deconvolution block uses a 1×1 deconvolution layer for feature dimensionality reduction.

[0021] The second deconvolution block uses a 3×1 deconvolution layer to reduce feature channels;

[0022] The third deconvolution block uses a 1×1 deconvolution layer to recover the feature dimension;

[0023] Each deconvolutional block also includes a second ReLU activation function layer and a second batch normalization layer in sequence;

[0024] The outputs of the two decoding submodules are aggregated through a fully connected layer to serve as the final output of the high-altitude bow shock prediction neural network.

[0025] Further, in step S3, the normalization and interpolation processing of the output flow field cloud map dataset includes:

[0026] Divide the original size of the flow field cloud map by the radius of the aircraft nose to obtain the normalized cloud map size;

[0027] Based on the normalized cloud map size, the inverse distance interpolation method is used to interpolate the data, and the interpolated cloud map data is used as the output dataset of the high-altitude bow shock prediction neural network.

[0028] Further, after step S5, the cloud map size of the estimated result is multiplied by the aircraft nose radius to obtain the actual flow field cloud map size.

[0029] Furthermore, in step S4, the loss function is the mean absolute percentage error.

[0030] A high-altitude bow-shaped shock wave flow field prediction system based on convolutional neural networks is provided to implement the high-altitude bow-shaped shock wave flow field prediction method based on convolutional neural networks as described in this invention. The system includes:

[0031] The data preparation module is used to sample the input flight parameters using the Latin hypercube sampling method and calculate the corresponding high-altitude bow shock wave flow field data using the direct simulation Monte Carlo method, forming an input flight parameter set and an output flow field cloud map dataset.

[0032] The preprocessing module is used to take the input flight parameter set as the input dataset and normalize and interpolate the output flow field cloud map dataset to form the output dataset.

[0033] The neural network module adopts an encoder-decoder structure, including an asymmetric convolutional encoder for compressing and extracting features and a parallel deconvolutional decoder for reconstructing flow field cloud maps, forming a neural network for predicting high-altitude bow shock waves.

[0034] The training module is used to divide the input dataset and the output dataset into a training set and a test set, use the training set and the backpropagation algorithm to train and optimize the high-altitude bow shock prediction neural network, and use the test set to verify the performance of the high-altitude bow shock prediction neural network.

[0035] The prediction module is used to input the flight condition parameters to be predicted into the high-altitude bow shock prediction neural network that has been trained and verified, and to obtain the prediction result of the high-altitude bow shock flow field distribution through forward propagation calculation.

[0036] This invention provides a method for predicting the flow field of high-altitude bow-shaped shock waves based on convolutional neural networks, which has the following advantages compared with existing technologies:

[0037] 1. This invention addresses hypersonic flight conditions in sparse watersheds by constructing a predictive neural network that directly maps flight parameters to the high-altitude bow-shaped shock field, enabling rapid calculation of the high-altitude bow-shaped shock field. This effectively reduces computational steps and shortens computation time.

[0038] 2. By constructing an asymmetric convolutional encoder, this invention can flexibly and accurately extract features of different dimensions and levels within the input data.

[0039] 3. By building a parallel deconvolution decoder, this invention can efficiently recover the feature information output by the asymmetric convolution encoder from multiple channels.

[0040] In summary, this invention enhances the nonlinear expressive power of the network and further improves the prediction accuracy of the model. The final constructed model can calculate the flow field of hypersonic high-altitude bow shock waves in sparse water domains in real time under different flight parameter conditions, thus meeting the computational needs of massive amounts of data. Attached Figure Description

[0041] Figure 1 This is the overall flowchart of the present invention.

[0042] Figure 2 This is a schematic diagram of the loss function of the high-altitude bow-shaped shock wave prediction neural network of the present invention. Figure 1 .

[0043] Figure 3 This is a schematic diagram of the loss function of the high-altitude bow-shaped shock wave prediction neural network of the present invention. Figure 2 .

[0044] Figure 4 This is a schematic diagram comparing the calculation results of the prediction method of the present invention with those of the existing flow field prediction method DSMC. Detailed Implementation

[0045] The present invention will be further explained and described below with reference to the accompanying drawings, embodiments, and comparative examples. Obviously, the described embodiments are only a part of the embodiments, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0046] Example

[0047] A method for predicting the flow field of high-altitude bow-shaped shock waves based on convolutional neural networks is shown in the attached figure. Figure 1 As shown, it includes the following steps:

[0048] S1: Construct a dataset by sampling the input flight parameters using the Latin hypercube sampling method and calculating the corresponding high-altitude bow shock wave flow field data using the Direct Simulation Monte Carlo (DSMC) method, forming an input flight parameter set and an output flow field cloud map dataset; the input flight parameter set includes the aircraft nose radius, flight altitude, and flight speed; the output flow field cloud map dataset includes distribution cloud maps of translational temperature, rotational temperature, vibrational temperature, velocity, pressure, and NO number density.

[0049] S2: Construct a high-altitude bow-shaped shock wave prediction neural network. The high-altitude bow-shaped shock wave prediction neural network adopts an encoder-decoder structure, including an asymmetric convolutional encoder for compressing and extracting features and a parallel deconvolutional decoder for reconstructing flow field cloud maps.

[0050] The neural network for predicting high-altitude bow shock waves is defined as follows:

[0051] ;

[0052] In the formula, T t T represents translational temperature. r Indicates rotational temperature; T v The vibration temperature is represented by u; the velocity is represented by P; and the pressure is represented by N. NO denoted by NO number density; r represents the nose radius of the aircraft; h represents the flight altitude; and v represents the flight speed.

[0053] The asymmetric convolutional encoder includes:

[0054] The three convolutional blocks are set in parallel, with kernel sizes of 3×3, 1×3 and 3×1 respectively. Each convolutional block contains a convolutional layer, a first ReLU activation function layer and a first batch normalization layer.

[0055] The feature fusion layer is used to fuse the features output from the three convolutional blocks;

[0056] Max pooling layers are used to downsample the features fused by the feature fusion layer to obtain more discriminative high-level features.

[0057] The convolutional layer operation is represented as follows:

[0058] ;

[0059] In the formula, O represents the high-altitude bow shock wave flow field matrix; X represents the flight parameter matrix; K represents the convolution kernel; b is the bias constant, applied to all output positions; i and j are the row and column indices of the output matrix O; m and n are the row and column indices of the convolution kernel K; k h k is the kernel length; w is the width of the convolution kernel.

[0060] The first ReLU activation function layer performs pointwise computation on the input elements, outputting all positive numbers and setting negative numbers to zero. This introduces non-linearity, effectively mitigating the vanishing gradient problem.

[0061] The first batch normalization layer normalizes the data and scales and translates it, dynamically adjusting the data distribution to a standard normal space with a mean of 0 and a variance of 1. The normalization transformation is expressed as:

[0062] ;

[0063] ;

[0064] In the formula, The input feature value is the (i, j)th feature of the b-th sample in the current mini-batch; μ B This represents the average of the current small batch. ε is the variance of the current mini-batch; γ is the numerical stability constant; and β are learnable scaling and offset parameters. This is the intermediate normalized value; This is for normalized output.

[0065] The parallel deconvolutional decoder includes two decoding sub-modules with identical structures;

[0066] Each of the decoding submodules includes three cascaded deconvolution blocks, as follows:

[0067] The first deconvolution block uses a 1×1 deconvolution layer for feature dimensionality reduction.

[0068] The second deconvolution block uses a 3×1 deconvolution layer to reduce feature channels;

[0069] The third deconvolution block uses a 1×1 deconvolution layer to recover the feature dimension;

[0070] Each deconvolutional block also includes a second ReLU activation function layer and a second batch normalization layer in sequence;

[0071] The outputs of the two decoding submodules are aggregated through a fully connected layer to serve as the final output of the high-altitude bow shock prediction neural network. The transformation process can be expressed as follows:

[0072] ;

[0073] In the formula, y is the output vector of the fully connected layer; f(·) is the third ReLU activation function; w is the weight matrix of the fully connected layer, where the i-th row corresponds to the i-th output neuron and the j-th column corresponds to the j-th input neuron. The matrix elements w ij That is, the connection weights; x is the input vector, which is formed by concatenating the output features of the two decoding submodules; b is the bias matrix of the fully connected layer.

[0074] S3: Data preprocessing: The input flight parameter set is used as the input dataset of the high-altitude bow shock prediction neural network. The output flow field cloud map dataset is normalized and interpolated, and then used as the output dataset of the high-altitude bow shock prediction neural network.

[0075] The normalization and interpolation processing of the output flow field cloud map dataset includes:

[0076] Divide the original size of the flow field cloud map by the radius of the aircraft nose to obtain the normalized cloud map size;

[0077] Based on the normalized cloud map size, the inverse distance interpolation method is used to interpolate the data, and the interpolated cloud map data is used as the output dataset of the high-altitude bow shock prediction neural network.

[0078] S4: Model Training, Optimization, and Validation. The input and output datasets are divided into training and test sets. The high-altitude bow shock prediction neural network is trained using the training set and a backpropagation algorithm to optimize network weights and biases. The loss function is monitored during training, and early stopping is used to prevent overfitting. The test set is used to validate the performance of the trained model. The loss function is the mean absolute percentage error, and its expression is:

[0079] ;

[0080] In the formula, MAPE is the mean absolute percentage error of the neural network for predicting high-altitude bow shock waves, and n is the number of samples. This represents the estimated value for the i-th sample. Let be the true value of the i-th sample.

[0081] The specific formulas for correcting the weights and biases are as follows:

[0082] ;

[0083] ;

[0084] In the formula, t represents the iteration number; η is the learning rate, i.e., the hyperparameter controlling the update step size; w tThe weight matrix before correction; w t+1 This is the corrected weight matrix; b t b is the bias matrix before correction; t+1 This is the corrected bias matrix.

[0085] S5: Flow field prediction. The flight condition parameters to be predicted are input into the trained and validated high-altitude bow shock prediction neural network. The predicted flow field distribution of the high-altitude bow shock is calculated through forward propagation. The cloud map size of the predicted result is multiplied by the aircraft nose radius to obtain the actual flow field cloud map size.

[0086] Appendix Figure 2 ~Appendix Figure 3 This is a schematic diagram of the loss function of the high-altitude bow-shaped shock wave prediction neural network in this invention. As can be seen from the figure, the high-altitude bow-shaped shock wave prediction neural network has good convergence.

[0087] Appendix Figure 4 This is a schematic diagram comparing the calculation results of the prediction method of the present invention (denoted as CNN in the figure) with the existing flow field prediction method DSMC. As can be seen from the figure, on the high-altitude bow shock flow field dataset under different flight conditions, the present invention shows better robustness and adaptability, and can calculate the flow field of high-altitude bow shock under different flight parameters more accurately.

[0088] A high-altitude bow-shaped shock wave flow field prediction system based on convolutional neural networks is used to implement the prediction method described in this embodiment. The system includes:

[0089] The data preparation module is used to sample the input flight parameters using the Latin hypercube sampling method and calculate the corresponding high-altitude bow shock wave flow field data using the direct simulation Monte Carlo method, forming an input flight parameter set and an output flow field cloud map dataset.

[0090] The preprocessing module is used to take the input flight parameter set as the input dataset and normalize and interpolate the output flow field cloud map dataset to form the output dataset.

[0091] The neural network module adopts an encoder-decoder structure, including an asymmetric convolutional encoder for compressing and extracting features and a parallel deconvolutional decoder for reconstructing flow field cloud maps, forming a neural network for predicting high-altitude bow shock waves.

[0092] The training module is used to divide the input dataset and the output dataset into a training set and a test set, use the training set and the backpropagation algorithm to train and optimize the high-altitude bow shock prediction neural network, and use the test set to verify the performance of the high-altitude bow shock prediction neural network.

[0093] The prediction module is used to input the flight condition parameters to be predicted into the high-altitude bow shock prediction neural network that has been trained and verified, and to obtain the prediction result of the high-altitude bow shock flow field distribution through forward propagation calculation.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the flow field of high-altitude bow-shaped shock waves based on convolutional neural networks, characterized in that: Includes the following steps: S1: Construct a dataset, use the Latin hypercube sampling method to sample the input flight parameters, and use the direct simulation Monte Carlo method to calculate the corresponding high-altitude bow shock wave flow field data, forming an input flight parameter set and an output flow field cloud map dataset; S2: Construct a high-altitude bow-shaped shock wave prediction neural network. The high-altitude bow-shaped shock wave prediction neural network adopts an encoder-decoder structure, including an asymmetric convolutional encoder for compressing and extracting features and a parallel deconvolutional decoder for reconstructing the flow field cloud map. S3: Data preprocessing: The input flight parameter set is used as the input dataset of the high-altitude bow shock wave prediction neural network. The output flow field cloud map dataset is normalized and interpolated and then used as the output dataset of the high-altitude bow shock wave prediction neural network. S4: Model training, optimization and validation. The input dataset and output dataset are divided into training set and test set. The high-altitude bow shock wave prediction neural network is trained using the training set and the backpropagation algorithm to optimize the network weights. The loss function is monitored during the training process to prevent overfitting. The performance of the trained model is validated using the test set. S5: Flow field prediction. The flight condition parameters to be predicted are input into the high-altitude bow shock prediction neural network that has been trained and verified. The prediction result of the high-altitude bow shock flow field distribution is obtained through forward propagation calculation.

2. The method for predicting the flow field of a high-altitude bow-shaped shock wave based on a convolutional neural network according to claim 1, characterized in that: In step S1, the set of input flight parameters includes the aircraft's nose radius, flight altitude, and flight speed.

3. The method for predicting the flow field of a high-altitude bow-shaped shock wave based on a convolutional neural network according to claim 2, characterized in that: In step S1, the output flow field cloud map dataset includes distribution cloud maps of translational temperature, rotational temperature, vibrational temperature, velocity, pressure, and NO number density.

4. The method for predicting the flow field of a high-altitude bow-shaped shock wave based on a convolutional neural network according to claim 1, characterized in that: In step S2, the asymmetric convolutional encoder includes: The three convolutional blocks are set in parallel, with kernel sizes of 3×3, 1×3 and 3×1 respectively. Each convolutional block contains a convolutional layer, a first ReLU activation function layer and a first batch normalization layer. The feature fusion layer is used to fuse the features output from the three convolutional blocks; The max pooling layer is used to downsample the features fused by the feature fusion layer.

5. The method for predicting the flow field of a high-altitude bow-shaped shock wave based on a convolutional neural network according to claim 1, characterized in that: In step S2, the parallel deconvolution decoder includes two decoding sub-modules with identical structures; Each of the decoding submodules includes three cascaded deconvolution blocks, as follows: The first deconvolution block uses a 1×1 deconvolution layer for feature dimensionality reduction. The second deconvolution block uses a 3×1 deconvolution layer to reduce feature channels; The third deconvolution block uses a 1×1 deconvolution layer to recover the feature dimension; Each deconvolutional block also includes a second ReLU activation function layer and a second batch normalization layer in sequence; The outputs of the two decoding submodules are aggregated through a fully connected layer to serve as the final output of the high-altitude bow shock prediction neural network.

6. The method for predicting the flow field of a high-altitude bow-shaped shock wave based on a convolutional neural network according to claim 3, characterized in that: In step S3, the normalization and interpolation processing of the output flow field cloud map dataset includes: Divide the original size of the flow field cloud map by the radius of the aircraft nose to obtain the normalized cloud map size; Based on the normalized cloud map size, the inverse distance interpolation method is used to interpolate the data, and the interpolated cloud map data is used as the output dataset of the high-altitude bow shock prediction neural network.

7. The method for predicting the flow field of a high-altitude bow-shaped shock wave based on a convolutional neural network according to claim 6, characterized in that: After step S5, the cloud map size of the estimated result is multiplied by the aircraft nose radius to obtain the actual flow field cloud map size.

8. The method for predicting the flow field of a high-altitude bow-shaped shock wave based on a convolutional neural network according to claim 1, characterized in that: In step S4, the loss function is the mean absolute percentage error.

9. A high-altitude bow-shaped shock wave flow field prediction system based on convolutional neural networks, characterized in that: The system is used to implement the high-altitude bow-shaped shock wave flow field prediction method based on convolutional neural networks as described in any one of claims 1-8, the system comprising: The data preparation module is used to sample the input flight parameters using the Latin hypercube sampling method and calculate the corresponding high-altitude bow shock wave flow field data using the direct simulation Monte Carlo method, forming an input flight parameter set and an output flow field cloud map dataset. The preprocessing module is used to take the input flight parameter set as the input dataset and normalize and interpolate the output flow field cloud map dataset to form the output dataset. The neural network module adopts an encoder-decoder structure, including an asymmetric convolutional encoder for compressing and extracting features and a parallel deconvolutional decoder for reconstructing flow field cloud maps, forming a neural network for predicting high-altitude bow shock waves. The training module is used to divide the input dataset and the output dataset into a training set and a test set, use the training set and the backpropagation algorithm to train and optimize the high-altitude bow shock prediction neural network, and use the test set to verify the performance of the high-altitude bow shock prediction neural network. The prediction module is used to input the flight condition parameters to be predicted into the high-altitude bow shock prediction neural network that has been trained and verified, and to obtain the prediction result of the high-altitude bow shock flow field distribution through forward propagation calculation.

Images