High-enthalpy rarefied gas hypersonic wind tunnel flow field prediction method based on machine learning

Abstract

The invention discloses a high-enthalpy rarefied gas hypersonic wind tunnel flow field prediction method based on machine learning. In order to solve the problems of too high wind tunnel experiment cost, slow simulation response and the like in the aircraft design process, the invention provides a fast response solution for wind tunnel prediction by using a neural network. Simulation software is adopted to simulate the flow field and the radiation field of the hypersonic aircraft, and flow field and radiation field prediction is performed on the constructed model in combination with a farthest point sampling method, a symmetric function space transformation network technology and a separated attention technology, so that experiment waste caused by improper design of the aircraft model is avoided, and the experiment frequency is reduced; in the wind tunnel test process, the flow field and the radiation field of the wind tunnel are predicted synchronously, the wind tunnel experiment parameters are adjusted according to the prediction result, and disturbance and the like of the experiment are avoided. The method provided by the invention accelerates the scheme iteration process at the initial stage of aircraft design, and can quickly respond to flow field and radiation field prediction.

Description

Technical Field

[0001] The present invention relates to the field of wind tunnel prediction, and in particular to a method for predicting the flow field of a high-enthalpy rarefied gas hypersonic wind tunnel based on machine learning. Background Art

[0002] Aircraft play an important role in today's military. In particular, hypersonic aircraft are crucial for the future development of warfare, and the stealth and detection of aircraft are closely related to aircraft design. The design process of an aircraft involves the constraints and optimizations of multiple disciplines, including the mutual restrictions and balances among multiple disciplines such as load, aerodynamics, structure, and propulsion. Wind tunnel tests play a very important role in the aircraft design process, and small changes in each parameter will affect the final result.

[0003] The cost of wind tunnel experiments is high, especially for the high-enthalpy rarefied gas hypersonic wind tunnel of hypersonic aircraft. Moreover, wind tunnel experiments require many parameters to be set, and different setting conditions are all new working conditions, resulting in multi-working condition characteristics. The flow fields between different working conditions are quite different, and multiple experiments are required. Software simulation can greatly reduce costs and can simulate various complex fluid motion situations. However, software simulation requires a large amount of computing resources and time, and has high requirements for aircraft models, grids, boundary conditions, and turbulence models, and cannot achieve rapid response. At the initial stage of aircraft design, these limitations seriously affect work efficiency. If software simulation can achieve rapid response and interact and cooperate with wind tunnel experiments, it can greatly reduce experimental costs and improve work efficiency.

[0004] With the development of artificial intelligence technology, for the problems of high cost of wind tunnel experiments and slow simulation response in the aircraft design process, a rapidly responsive solution is provided for wind tunnel prediction by using neural networks. The present invention uses the farthest point sampling method, symmetric function space transformation network technology, and separated attention technology to predict the flow field and radiation field of the constructed model, avoiding experimental waste caused by improper aircraft model design and reducing the number of experiments; during the wind tunnel test process, the flow field and radiation field of the wind tunnel are predicted synchronously, and the wind tunnel experiment parameters are adjusted according to the prediction results to avoid disturbances in the experiment. Thus, the process of scheme iteration at the initial stage of design is accelerated, the cost is reduced, and it becomes possible to explore more innovative design schemes under limited resources. Summary of the Invention

[0005] The present invention solves the problems existing in the initial stage of aircraft design by training a complex deep learning model to simulate the flow field and radiation field of hypersonic aircraft, which can significantly reduce the calculation time and resource consumption.

[0006] The present invention mainly solves the above problems through the following steps:

[0007] Step 1: Use existing simulation software to perform mesh generation and simulation on the flow field in the outer space of the aircraft, couple the flow field with the radiation field, and simulate the radiation field results;

[0008] Step 2: Preprocess the point cloud data, control the total number of points in each point cloud, and ensure that the data in the database has consistent dimensions;

[0009] Step 3: Solve the disorder problem of the point cloud data by using a spatial transformation network and function processing, so that the network can learn the symmetry and invariance of the point cloud;

[0010] Step 4: Train the neural network using the training set obtained from the above process, and input the features through the trained neural network to predict the flow field and radiation field;

[0011] Step 5: According to the prediction results of the flow field and radiation field, conduct targeted wind tunnel tests for specific models, specific angles, etc., and perform real-time prediction on the flow field and radiation field under the same experimental conditions during the experiment. Adjust the wind tunnel parameters synchronously according to the prediction results.

[0012] In the simulation of the flow field, the grid adaptation method is used to generate the mesh for the flow field around the aircraft, and the simulation software is used to simulate the flow field. Then, the flow field is coupled with the radiation field, and the three-temperature model and the line-by-line calculation method are used to simulate the radiation field.

[0013] The farthest point sampling method is used to process the original point cloud data of the aircraft to reduce the scale of the point cloud data. The basic idea is to randomly select an initial point in the point cloud as the starting point of the sampling point set, and iteratively select the point with the farthest distance from the current sampling point set among the remaining points, and add it to the set until the number of sampling points reaches the requirement. The input three-dimensional aircraft point cloud is:

[0014] F i =[x m ,y m ,z m T ,m = 1,…,M

[0015] where, F i is the three-dimensional aircraft coordinate; x m ,y m ,z m is the coordinate value of the m-th point in the x, y, z directions; M is the total number of original point cloud points.

[0016]

[0017] where, is the three-dimensional aircraft coordinate; x n ,y n ,z​n is the coordinate value of the nth point in the x, y, and z directions; N is the total number of point cloud points after sampling.

[0018] In order to solve the problem of disordered point cloud data, a symmetric function is used to perform mathematical transformation and keep the characteristics unchanged, ensuring that the storage order of data points in the point cloud model data will not affect the accuracy of the extracted three-dimensional geometric features. The mathematical expression of the symmetric function can be briefly summarized as:

[0019] f(x i )=g(h(x i ))

[0020] Among them, g(x) is a symmetric function, x i is the point cloud, h(x) is the mapping function;

[0021] For point cloud rotation, the spatial transformer network (TN) module is used. The network can learn the symmetry and invariance of the point cloud, so that the network recognition performance is no longer affected by the directionality of the point cloud.

[0022] The spatial transformer network (TN) module uses a three-layer multi-layer perceptron (MLP) to gradually increase the dimension of point cloud features. The first layer increases the features of the input 3D point cloud to 64 dimensions, the second layer increases the 64-dimensional features to 128 dimensions, and the third layer further increases the 128-dimensional features to 1024 dimensions, enabling the TN module to extract high-dimensional feature representations of point clouds.

[0023] The constructed training set is used to train the neural network; a separate attention module is used in the neural network to independently weight the features and parameters, which can achieve more refined information processing.

[0024] The separated attention module allows the model to independently learn the importance of each feature and parameter. The weights are normalized by the softmax function to ensure that the weights are integrated into one. The normalized weights are used for the corresponding features and parameters, and weighted by element-by-element multiplication (Hadamard product). The weighted parameters are reconstructed into a comprehensive matrix representation.

[0025] Weight normalization is done to concatenate the weights of features and parameters:

[0026] W t =concat(W f ,W p )

[0027] Among them, W t is the weight after concatenation, W f is the feature weight, W p is the parameter weight.

[0028] Normalize the concatenated weights using the softmax function:

[0029] W n = softmax(W t )

[0030] where W n is the normalized weight.

[0031] Split the normalized weights:

[0032] (W′ f , W′ p ) = split(W n )

[0033] where W' f is the feature weight after splitting, and W' p is the parameter weight after splitting.

[0034] Apply the split feature weight to the feature:

[0035] F w = F ⊙ W′ f

[0036] where F w is the weighted feature, ⊙ represents element-wise multiplication, and F is the feature value.

[0037] Apply the split parameter weight to the parameter:

[0038] P w = P ⊙ W′ p

[0039] where P w is the weighted parameter, and F is the parameter value.

[0040] Concatenate the weighted feature and parameter:

[0041] C w = concat(F w , P w )

[0042] where C w is the weighted parameter, and F is the parameter value.

[0043] According to the trained neural network, predict the flow field and radiation field of the constructed model to avoid waste of experiments caused by improper design of the aircraft model; at the same time, the wind tunnel test can be carried out targeted according to the prediction results, reducing the number of experiments; during the wind tunnel test, predict the wind tunnel synchronously, and adjust the wind tunnel experiment according to the prediction results to avoid disturbances in the experiment.

[0044] This disclosure provides a new solution to the problems of high cost and slow simulation response in the wind tunnel experiment during the aircraft design process, thereby accelerating the scheme iteration process in the initial stage of design and making it possible to explore more innovative design schemes under limited resources. Description of the Drawings

[0045] Appendix Figure 1 It is a schematic flow chart of a hypersonic wind tunnel flow field prediction method based on machine learning proposed according to the present invention. Detailed Embodiments

[0046] As Figure 1 shown, a hypersonic wind tunnel flow field prediction method based on machine learning includes the following steps:

[0047] S1. Use existing simulation software to perform grid division and simulation on the external space flow field of the aircraft, and couple the flow field with the radiation field to simulate the radiation field results.

[0048] S101. Perform adaptive grid division on the external space flow field of the aircraft geometric model and perform numerical simulation on its flow field.

[0049] S102. Couple the simulation results of the external flow field of the aircraft with the radiation field, and then simulate the radiation field.

[0050] S103. Construct a training set. The input parameters of the training set include altitude, incoming flow Mach number, incident position polarization angle, angle of attack, and initial gas component concentration; the output parameters include the temperature of each microelement, the calculated component number density, and the ultraviolet radiation intensity.

[0051] In the simulation of the S1 flow field, the grid adaptive method is used to perform grid division on the flow field around the aircraft, and the simulation software is used to perform flow field simulation. Then, the flow field is coupled with the radiation field, and the three-temperature model and the line-by-line calculation method are used to simulate the radiation field.

[0052] S2. Preprocess the point cloud data to control the total number of points in each point cloud and ensure that the data in the database has consistent dimensions.

[0053] The farthest point sampling method is used to process the original point cloud data of the aircraft to reduce the scale of the point cloud data. Its basic idea is to randomly select an initial point in the point cloud as the starting point of the sampling point set, and iteratively select the point with the farthest distance from the current sampling point set among the remaining points and add it to the set until the number of sampling points reaches the requirement. The input three-dimensional aircraft point cloud is:

[0054] F i =[x m ,ym ,z m ] T ,m=1,…,M

[0055] Among them, F i is the three-dimensional aircraft coordinate; x m ,y m ,z m is the coordinate value of the mth point in the x, y, and z directions; M is the total number of points in the original point cloud.

[0056]

[0057] in, is the three-dimensional aircraft coordinate; x n ,y n ,z n is the coordinate value of the nth point in the x, y, and z directions; N is the total number of point cloud points after sampling.

[0058] S3, by adopting spatial transformation network and function processing, solves the disorder problem of point cloud data, so that the network can learn the symmetry and invariance of point cloud.

[0059] In order to solve the problem of disordered point cloud data, a symmetric function is used to perform mathematical transformation and keep the characteristics unchanged, ensuring that the storage order of data points in the point cloud model data will not affect the accuracy of the extracted three-dimensional geometric features. The mathematical expression of the symmetric function can be briefly summarized as:

[0060] f(x i )=g(h(x i ))

[0061] Among them, g(x) is a symmetric function, x i is the point cloud, h(x) is the mapping function;

[0062] For point cloud rotation, the spatial transformer network (TN) module is used. The network can learn the symmetry and invariance of the point cloud, so that the network recognition performance is no longer affected by the directionality of the point cloud.

[0063] The spatial transformer network (TN) module uses a three-layer multi-layer perceptron (MLP) to gradually increase the dimension of point cloud features. The first layer increases the features of the input 3D point cloud to 64 dimensions, the second layer increases the 64-dimensional features to 128 dimensions, and the third layer further increases the 128-dimensional features to 1024 dimensions, enabling the TN module to extract high-dimensional feature representations of point clouds.

[0064] S4, training the neural network with the training set obtained according to the above process, and inputting the features into the trained neural network to predict the flow field and the radiation field.

[0065] The constructed training set is used to train the neural network; a separate attention module is adopted in the neural network, aiming to independently perform attention weighting on features and parameters, and more refined information processing can be achieved.

[0066] The separate attention module allows the model to independently learn the respective importance between features and parameters, normalizes the weights through the softmax function to ensure that the sum of the weights is one, and the normalized weights are used for the corresponding features and parameters, and are weighted through element-wise multiplication (Hadamard product), and the weighted parameters are re-concatenated into a comprehensive matrix representation.

[0067] Weight normalization is performed by concatenating the weights of features and parameters:

[0068] W t = concat(W f , W p )

[0069] where W t is the concatenated weight, W f is the feature weight, and W p is the parameter weight.

[0070] The softmax function is used to normalize the concatenated weights:

[0071] W n = softmax(W t )

[0072] where W n is the normalized weight.

[0073] The normalized weights are split:

[0074] (W′ f , W′ p ) = split(W n )

[0075] where W' f is the split feature weight, and W' p is the split parameter weight.

[0076] The split feature weight is applied to the feature:

[0077] F w = F ⊙ W′ f

[0078] where F w is the weighted feature, ⊙ represents element-wise multiplication, and F is the feature value.

[0079] Apply the segmented parameter weights to the parameters:

[0080] P w = P ⊙ W′ p

[0081] where P w is the weighted parameter and F is the parameter value.

[0082] Concatenate the weighted feature and the parameter:

[0083] C w = concat(F w , P w )

[0084] where C w is the weighted parameter and F is the parameter value.

[0085] S5. According to the prediction results of the flow field and the radiation field, conduct targeted wind tunnel tests for specific models, specific angles, etc., and perform real-time prediction of the flow field and the radiation field during the experiment, and synchronously adjust the wind tunnel parameters according to the prediction results.

[0086] Based on the trained neural network, predict the flow field and the radiation field of the constructed model to avoid experimental waste caused by improper design of the aircraft model; at the same time, the wind tunnel parameters can be adjusted targeted according to the prediction results for wind tunnel tests, reducing the number of experiments; during the wind tunnel test, synchronously predict the flow field and the radiation field of the wind tunnel, and adjust the wind tunnel experiment according to the prediction results to avoid disturbances in the experiment.

[0087] Through the above method, problems such as excessive cost of wind tunnel experiments and slow simulation response in the process of aircraft design are solved, a new solution is provided, neural network prediction is used to reduce the number of wind tunnel tests, accelerate the scheme iteration process in the initial stage of design, and make it possible to explore more innovative design schemes under limited resources.

Claims

1. A method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning, characterized in that: The following steps are involved: Step 1: Use existing simulation software to mesh and simulate the flow field outside the aircraft, and couple the flow field with the radiation field to simulate the radiation field results; Step 2: preprocess the point cloud data to control the total number of points in each point cloud and ensure that the data in the database has consistent dimensions; Step 3: By using a spatial transformation network and function processing, the disorder problem of point cloud data is solved so that the network can learn the symmetry and invariance of the point cloud; Step 4, training the neural network with the training set obtained according to the above process, and inputting the features into the trained neural network to predict the flow field and radiation field; Step 5: Based on the prediction results of the flow field and radiation field, targeted wind tunnel tests are carried out for specific models, specific angles, etc., and real-time predictions are performed during the experiment, and wind tunnel parameters are adjusted synchronously according to the prediction results.

2. The method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning according to claim 1 is characterized by: In the step 1, in the simulation of the flow field, the grid adaptive square technology is used to grid the flow field around the aircraft, the flow field is simulated by using simulation software, and then the flow field is coupled with the radiation field, and the radiation field is simulated by using a three-temperature model and a line-by-line calculation method.

3. The method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning according to claim 1 is characterized by: In step 2, the farthest point sampling method is used to process the original point cloud data of the aircraft. The basic idea is to randomly select an initial point in the point cloud as the starting point of the sampling point set, and iteratively select the point farthest from the current sampling point set from the remaining points to add it to the set until the number of sampling points reaches the requirement, so as to reduce the scale of the point cloud data.

4. The method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning according to claim 1 is characterized in that: In step 3, a symmetric function is used to perform mathematical transformation while keeping the characteristics unchanged. The mathematical expression of the symmetric function can be briefly summarized as follows: f(x i )=g(h(x i )) Among them, g(x) is a symmetric function, x i is the point cloud, h(x) is the mapping function; For point cloud rotation, the spatial transformer network (TN) module is used so that the network recognition performance is no longer affected by the directionality of the point cloud.

5. The method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning according to claim 1, characterized in that: In step 4, the constructed training set is used to train the neural network; a separate attention module is used in the neural network to independently weight the features and parameters, so as to achieve more refined information processing.

6. The method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning according to claim 1, characterized in that: In step 5, the flow field and radiation field of the constructed model are predicted based on the trained neural network to avoid experimental waste due to improper design of the aircraft model; at the same time, wind tunnel tests can be carried out in a targeted manner based on the prediction results to reduce the number of experiments; during the wind tunnel test, the wind tunnel is predicted synchronously, and the wind tunnel experiment is adjusted based on the prediction results to avoid disturbances in the experiment.

7. The method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning according to claim 1, characterized in that: The spatial transformer network (TN) module uses a three-layer multi-layer perceptron (MLP) to gradually increase the dimension of point cloud features. The first layer increases the features of the input 3D point cloud to 64 dimensions, the second layer increases the 64-dimensional features to 128 dimensions, and the third layer further increases the 128-dimensional features to 1024 dimensions, so that the TN module can extract high-dimensional feature representations of point clouds.

8. The method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning according to claim 1, characterized in that: In the neural network, a separate attention module is used to independently weight the features and parameters, which can achieve more refined information processing.

9. The method for predicting high enthalpy rarefied gas high-speed wind tunnel flow field based on machine learning according to claim 1, characterized in that: The separated attention module allows the model to independently learn the importance of features and parameters. The weights are normalized by the softmax function to ensure that the weights are integrated into one. The normalized weights are used for the corresponding features and parameters, and weighted by element-by-element multiplication (Hadamard product). The weighted parameters are reassembled into a comprehensive matrix representation. This module enhances the understanding and processing capabilities of features and parameters, and at the same time, can independently adjust the weight between the two to provide a more accurate information basis for prediction.

Images