Unsteady flow field neural network prediction method based on data pool

Abstract

The invention discloses an unsteady flow field neural network prediction method based on a data pool, and the method comprises the following specific steps: carrying out the initialization of a data access module based on the data pool, enabling the module to comprise a training data pool, a time memory unit and a reset recording unit, carrying out the preprocessing of a training data set, and taking out the flow field data; initializing a convolutional neural network model based on an attention mechanism, training the model, outputting flow field data of a next time step, and performing back propagation by a neural network; an output return module of the neural network is put back to the data pool, and the module extracts data from the data pool or the training data set according to a reset recording unit for next training; and adopting the trained model to carry out autoregressive flow field prediction. Compared with a traditional algorithm for solving the unsteady algorithm based on the neural network, the method has the advantages that noise hyper-parameters are eliminated, accumulated errors are reduced, and the industrial design iteration speed is increased.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning simulation and relates to a neural network prediction method for fluid simulation, and specifically to an unsteady flow field neural network prediction method based on a data pool. Background Art

[0002] Computational fluid dynamics is one of the important technologies in the field of fluid mechanics in the 21st century. It uses numerical methods to solve the governing equations of fluid mechanics in computers, thereby achieving flow analysis, prediction and control. Traditional finite element method (FEM) and finite difference method (FDM) have complex simulation processes (physical modeling, meshing, numerical discretization, iterative solution, etc.) and high computational costs. For example, the analysis of transonic buffeting problems often requires a large amount of computing resources because it involves high-resolution numerical simulation, nonlinear problems, multi-scale characteristics, etc., and its simulation accuracy will also face challenges.

[0003] The rapid development of neural network methods in recent years has shown that as an alternative to traditional solutions, neural networks solve unsteady flow fields by directly observing and learning from a large number of data sets; however, this end-to-end learning method has many unresolved problems for unsteady flow fields with large state spaces and complex dynamic patterns, such as the problem of cumulative errors. In recent years, many neural network methods based on deep learning frameworks have been used to solve flow field control equations, such as the unsteady incompressible NS equations. These methods rely on the use of noise as a hyperparameter to reduce the error accumulation phenomenon generated by the neural network in the autoregressive process; however, hyperparameters are difficult to set during the training of neural networks. Due to the sensitivity of neural networks to hyperparameters, different hyperparameters will have a very large impact on the results of neural networks, and hyperparameters even determine whether the training of neural networks can be successful to some extent. However, there is no scientific method for the selection of hyperparameters in academia. Most of the time, they are selected based on experience and experiments, so hyperparameters are a direction that still needs to be improved in neural network methods.

[0004] Some methods that do not require hyperparameters to reduce the cumulative error of neural networks have also been proposed, such as multi-time step binding and two-step forward. However, multi-time step binding essentially reduces the cumulative error of neural networks by reducing the number of inferences of the neural network, while the two-step forward method is difficult to work on complex equations such as the control equations of the flow field. In short, eliminating the hyperparameters of unsteady flow field prediction and reducing the cumulative error in the autoregressive model are urgent issues to be solved. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a data pool-based unsteady flow field neural network prediction method, which can be run on different neural network architectures, improve the prediction accuracy of the neural network, eliminate noise hyperparameters, and reduce the cumulative error of the neural network; at the same time, the introduction of the data pool can greatly reduce the training time of the training graph neural network, thereby alleviating the problem of consuming a lot of time in unsteady flow field calculations and realizing rapid prediction of unsteady flow fields.

[0006] In order to solve the above technical problems, the present invention is implemented in the following ways:

[0007] A neural network prediction method for unsteady flow field based on a data pool, the specific steps are as follows:

[0008] S1. Initialize a data access module based on a data pool, wherein the module includes a training data pool, a time memory unit, and a reset recording unit, preprocess the training data set, and retrieve the flow field data;

[0009] S2, initialize the convolutional neural network model based on the attention mechanism and train the model, output the flow field data of the next time step, and the neural network performs back propagation;

[0010] S3, the output of the neural network is returned to the module and put back into the data pool, and the module takes out data from the data pool or the training data set according to the reset record unit for the next round of training;

[0011] S4. Use the trained model to perform autoregressive flow field prediction.

[0012] Furthermore, the specific sub-steps of step S1 are as follows:

[0013] S11, initialize the training data pool, the time memory unit, and reset the recording unit, wherein the training data pool is initialized to a tensor of shape [N, C, H, W], where N represents the number of samples, C represents the number of features of the sample, including the speed in the X and Y directions and the pressure of the flow field, H represents the height of the sample, and W represents the width of the sample;

[0014] S12, the time memory unit is used to store the time step of the data in the training data pool. The time memory unit is initialized to a tensor of shape [N], where the time step is randomly initialized to a random integer from 0 to train_data_set_step (where train_data_set_step is the time series length of the training data set); the reset record unit is used to record the reset sample time step, where a record of True indicates that it has been reset, and a record of False indicates that it has not been reset, where all records are True during initialization;

[0015] S13, preprocess the time step data of the reset record unit. If the corresponding sample reset record is True, take out the data of the corresponding time step from the data set; if the corresponding sample reset record is False, take out the corresponding training data and the true value of the next time step from the training data pool.

[0016] Furthermore, the specific sub-steps of step S2 are as follows:

[0017] S21, training data Attention Unet network receives multi-channel input feature maps through the input layer, downsamples through a series of feature extraction modules to obtain downsampled feature maps, and the downsampled feature maps pass through the attention mechanism module and feature fusion module, and finally obtain the flow field data of the next time step through the output module, and reversely regress with the true value calculation loss of the next time step to optimize the neural network;

[0018] S22, the specific structure of the Attention Unet network used to output the flow field data for the next time step is as follows:

[0019] 1) The input layer receives a multi-channel input feature map with a size of 4×128×128, which contains flow field information with a time step of n;

[0020] 2) Feature extraction module: A 3×3 convolutional layer is used in combination with GroupNorm normalization and ReLU activation function for feature extraction. The feature extraction process uses a convolutional layer with a step size of 2 to perform 4×4, 8×8, 16×16, 32×32, and 64×64 multi-scale feature extraction;

[0021] 3) Attention mechanism module: Attention mechanism is applied to feature maps of different scales to generate attention weights, and the feature maps are multiplied by the corresponding attention weights to highlight key information;

[0022] 4) Feature fusion module: Upsample feature maps of different scales, and concatenate the obtained feature maps with the feature maps in the downsampling process in the channel dimension to obtain a fused feature map;

[0023] 5) Output module: After a 1×1 convolutional layer, the final output feature map is generated, and the flow field data with a time step of n+1 and a feature map size of 3×128×128 is output;

[0024] The specific structure of the attention mechanism module is as follows:

[0025] 1) The input layer receives a multi-channel input feature map with a size of 64×128×128;

[0026] 2) The input feature map passes through a 1×1 convolution layer, and the number of convolution kernels is the same as the number of output channels. The feature map after convolution processing enters the normalization layer for normalization processing;

[0027] 3) Perform feature fusion processing on the normalized feature maps, concatenate the feature maps of different inputs in the channel dimension, and obtain a fused feature map with a size of 128×128×128;

[0028] 4) The fused feature map is processed as follows:

[0029] First branch: The fused feature map passes through the adaptive average pooling layer to change the feature map size to 128×1×1, then passes through the fully connected layer, the ReLU activation function layer, and another fully connected layer, and finally passes through the Sigmoid activation function to output a set of weight coefficients. The weight coefficients are multiplied element by element with the original fused feature map to obtain the first corrected feature map;

[0030] The second branch: The fused feature map passes through a 1×1 convolution layer, and then passes through a Sigmoid activation function to output another set of weight coefficients. The weight coefficients are multiplied element by element with the original fused feature map to obtain the second corrected feature map.

[0031] The first corrected feature map and the second corrected feature map are added in the channel dimension to obtain the final output feature map, and the output feature map size is 128×128×128.

[0032] Furthermore, the specific sub-steps of step S3 are as follows:

[0033] S31, putting the flow field data generated by the model in the S2 process into the training data pool, replacing the flow field data of the previous time step;

[0034] S32, the time memory unit updates the time step after each epoch of the neural network, increases the time step of each sample by one, accumulates the current number of samples into the counter count, calculates the reset number of samples, and finally updates the counter count to the result of taking the modulus of per_train_data_time_step, where per_train_data_time_step is the average length of continuous training for each sample;

[0035] S33. Check the reset sequence list based on the calculated reset sample number. If it is empty, regenerate a sequence list from 0 to the number of samples, randomly shuffle it, take out an index from the reset sequence list, and reset the corresponding file according to the index.

[0036] Furthermore, the specific implementation of step S33 is as follows:

[0037] S331, randomly assigning a time memory unit at the index position to an integer between 0 and train_data_set_steps-per_train_data_time_step, where train_data_set_steps is the time series length of the training data set, and marking the data at the index position as reset;

[0038] S332. Use a mask to filter out data whose time step is greater than the time step of the training sample, reset the time step of these data to 0, and mark these data as reset in the reset record unit.

[0039] Furthermore, the specific sub-steps of step S4 are as follows:

[0040] S41, using the model trained by steps S2 to S3 to predict the unsteady flow field, taking out the data of the initial time step t=0 from the data set, inputting it into the trained neural network, and predicting the flow field data of the next time step t=1;

[0041] S42, inputting the flow field data of the previous time step predicted by the model into the trained neural network, and predicting the flow field data of the next step again;

[0042] S43, repeat step S42 until the stopping time step is set.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The neural network data processing module of the present invention introduces a data pool, which can greatly reduce the training time overhead of training the graph neural network, thereby alleviating the problem of consuming a lot of time in unsteady flow field calculations and realizing rapid prediction of unsteady flow fields; it runs on different neural network architectures to improve the prediction accuracy of the neural network, eliminate noise hyperparameters, and reduce the cumulative error of the neural network; at the same time, this method successfully enhances the robustness of the model by improving the training strategy of the neural network, without relying on the injection of additional noise, and can be integrated into existing graph neural network (GNN) and convolutional neural network models. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a training flow diagram of the present invention;

[0046] Figure 2 It is a schematic diagram of the prediction process of the present invention.

[0047] Figure 3 It is a schematic diagram of the Attention Unet architecture of the present invention;

[0048] Figure 4 is a schematic diagram of the Attention module of the present invention; DETAILED DESCRIPTION

[0049] The specific implementation of the present invention is further described in detail below with reference to the accompanying drawings and specific examples.

[0050] like Figures 1 to 4 As shown in FIG. 1 , a neural network prediction method for unsteady flow field based on a data pool is shown in FIG. 1 , and the specific steps are as follows:

[0051] S1. Initialize the data access module based on the data pool. The module includes a training data pool, a time memory unit and a reset record unit. Preprocess the training data set and extract the flow field data. The specific steps are as follows:

[0052] S11, initialize the training data pool, the time memory unit, and reset the recording unit, wherein the training data pool is initialized to a tensor of shape [N, C, H, W], N represents the number of samples, which is used to support the method of one-step prediction of multiple steps and multiple-step prediction of multiple steps in the neural network, and the default value is 1; C represents the number of features of the sample, including the velocity in the X and Y directions and the pressure of the flow field, H represents the height of the sample, and W represents the width of the sample;

[0053] S12, the time memory unit is used to store the time step of the data in the training data pool. The time memory unit is initialized to a tensor of shape [N], where the time step is randomly initialized to a random integer from 0 to train_data_set_step (where train_data_set_step is the time series length of the training data set); the reset record unit is used to record the reset sample time step, where a record of True indicates that it has been reset, and a record of False indicates that it has not been reset, where all records are True during initialization;

[0054] S13, preprocess the time step data of the reset record unit. If the corresponding sample reset record is True, take out the data of the corresponding time step from the data set; if the corresponding sample reset record is False, take out the corresponding training data and the true value of the next time step from the training data pool.

[0055] S2. Initialize the convolutional neural network model based on the attention mechanism and train the model. The model takes the data taken out from the data pool or training data set in the module as input, downsamples through a series of convolutional layers, normalization layers, and ReLU layers, and then outputs the data of the next time step through a series of upsampling layers and attention modules. The neural network performs back propagation. The specific steps are as follows:

[0056] S21, training data In the AttentionUnet network, the multi-channel input feature map is received through the input layer, and the down-sampled feature map is obtained through a series of feature extraction modules. The down-sampled feature map passes through the attention mechanism module and the feature fusion module, and finally the flow field data of the next time step of the forward process is obtained through the output module, and the true value calculation loss of the next time step is reversely regressed to optimize the neural network; the neural network training steps are as follows:

[0057] fort=1,2,3,...,N do:

[0058] Start timing

[0059] Set the network Net to training mode

[0060] for input,data_target,data_idx in data in a batch

[0061] Calculate learning rate decay The learning rate reduction formula is:

[0062] learning rate = 0.0001*γ current_epoch

[0063] Update the learning rate of optimizer optimizerG

[0064] Zero the gradient of the Net

[0065] Generate output gen as the output of Net for input

[0066] Calculate the mean square error MSE_loss. The mean square error calculation formula is:

[0067] MSE_loss=1 / n∑(data_target-gen) 2

[0068] Back propagation MSE_loss

[0069] Update the number of steps of optimizer optimizerG

[0070] After each batch of data is completed

[0071] Dataset Update Training Data Pool

[0072] Reset time data pool

[0073] Stop the timer

[0074] S22, the specific structure of the AttentionUnet network used to output the flow field data for the next time step is as follows:

[0075] 1) The input layer receives a multi-channel input feature map with a size of 4×128×128, which contains flow field information with a time step of n;

[0076] 2) Feature extraction module: A 3×3 convolutional layer is used in combination with GroupNorm normalization and ReLU activation function for feature extraction. The feature extraction process uses a convolutional layer with a step size of 2 to perform 4×4, 8×8, 16×16, 32×32, and 64×64 multi-scale feature extraction;

[0077] 3) Attention mechanism module: Attention mechanism is applied to feature maps of different scales to generate attention weights, and the feature maps are multiplied by the corresponding attention weights to highlight key information;

[0078] 4) Feature fusion module: Upsample feature maps of different scales, and concatenate the obtained feature maps with the feature maps in the downsampling process in the channel dimension to obtain a fused feature map;

[0079] 5) Output module: After a 1×1 convolutional layer, the final output feature map is generated, and the flow field data with a time step of n+1 and a feature map size of 3×128×128 is output;

[0080] The specific structure of the attention mechanism module is as follows:

[0081] 1) The input layer receives a multi-channel input feature map with a size of 64×128×128;

[0082] 2) The input feature map passes through a 1×1 convolution layer, and the number of convolution kernels is the same as the number of output channels. The feature map after convolution processing enters the normalization layer for normalization processing;

[0083] 3) Perform feature fusion processing on the normalized feature maps, concatenate the feature maps of different inputs in the channel dimension, and obtain a fused feature map with a size of 128×128×128;

[0084] 4) The fused feature map is processed as follows:

[0085] First branch: The fused feature map passes through the adaptive average pooling layer to change the feature map size to 128×1×1, then passes through the fully connected layer, the ReLU activation function layer, and another fully connected layer, and finally passes through the Sigmoid activation function to output a set of weight coefficients. The weight coefficients are multiplied element by element with the original fused feature map to obtain the first corrected feature map;

[0086] The second branch: The fused feature map passes through a 1×1 convolution layer, and then passes through a Sigmoid activation function to output another set of weight coefficients. The weight coefficients are multiplied element by element with the original fused feature map to obtain the second corrected feature map.

[0087] The first corrected feature map and the second corrected feature map are added in the channel dimension to obtain the final output feature map, and the output feature map size is 128×128×128.

[0088] S3, the output of the neural network is returned to the module and put back into the data pool. The module takes out data from the data pool or training data set according to the reset record unit for the next round of training. The specific steps are as follows:

[0089] S31, putting the flow field data generated by the model in the S2 process into the training data pool, replacing the flow field data of the previous time step;

[0090] S32, the time memory unit updates the time step after each epoch of the neural network, increases the time step of each sample by one, accumulates the current number of samples into the counter count, calculates the reset number of samples, and finally updates the counter count to the result of taking the modulus of per_train_data_time_step, where per_train_data_time_step is the average length of continuous training for each sample;

[0091] S33. Check the reset sequence list based on the calculated reset sample number. If it is empty, regenerate a sequence list from 0 to the number of samples, randomly shuffle it, take out an index from the reset sequence list, and reset the corresponding file according to the index.

[0092] The specific implementation of step S33 is as follows:

[0093] S331, randomly assigning a time memory unit at the index position to an integer between 0 and train_data_set_steps-per_train_data_time_step, where train_data_set_steps is the time series length of the training data set, and marking the data at the index position as reset;

[0094] S332. Use a mask to filter out data whose time step is greater than the time step of the training sample, reset the time step of these data to 0, and mark these data as reset in the reset record unit.

[0095] S4. Use the trained model to perform autoregressive flow field prediction. The specific steps are as follows:

[0096] S41, using the model trained by steps S2 to S3 to predict the unsteady flow field, taking out the data of the initial time step t=0 from the data set, inputting it into the trained neural network, and predicting the flow field data of the next time step t=1;

[0097] S42, inputting the flow field data of the previous time step predicted by the model into the trained neural network, and predicting the flow field data of the next step again;

[0098] S43, repeat step S42 until the stopping time step is set.

[0099] The above description is only an implementation mode of the present invention. It is stated again that for ordinary technicians in this technical field, several improvements can be made to the present invention without departing from the principle of the present invention. These improvements are also included in the protection scope of the claims of the present invention.

Claims

1. A neural network prediction method for unsteady flow field based on a data pool, characterized by: The specific steps are as follows: S1. Initialize a data access module based on a data pool, wherein the module includes a training data pool, a time memory unit, and a reset recording unit, preprocess the training data set, and retrieve the flow field data; S2, initialize the convolutional neural network model based on the attention mechanism and train the model, output the data of the next time step, and the neural network performs back propagation; S3, the output of the neural network is returned to the module and put back into the data pool, and the module takes out data from the data pool or the training data set according to the reset record unit for the next round of training; S4. Use the trained model to perform autoregressive flow field prediction.

2. According to claim 1, a method for predicting unsteady flow field based on a neural network of a data pool is characterized by: The specific sub-steps of step S1 are as follows: S11, initialize the training data pool, the time memory unit, and reset the recording unit, wherein the training data pool is initialized to a tensor of shape [N, C, H, W], where N represents the number of samples, C represents the number of features of the sample, including the speed in the X and Y directions and the pressure of the flow field, H represents the height of the sample, and W represents the width of the sample; S12, the time memory unit is used to store the time step of the data in the training data pool. The time memory unit is initialized to a tensor of shape [N], where the time step is randomly initialized to a random integer from 0 to train_data_set_step; the reset record unit is used to record the reset sample time step, and the record is True to indicate that it has been reset, and the record is False to indicate that it has not been reset, where all records are True during initialization; S13, preprocess the time step data of the reset record unit. If the corresponding sample reset record is True, take out the data of the corresponding time step from the data set; if the corresponding sample reset record is False, take out the corresponding training data and the true value of the next time step from the training data pool.

3. According to claim 1, a method for predicting unsteady flow field based on a neural network of a data pool is characterized in that: The specific sub-steps of step S2 are as follows: S21, training data Attention Unet network receives multi-channel input feature maps through the input layer, downsamples through a series of feature extraction modules to obtain downsampled feature maps, and the downsampled feature maps pass through the attention mechanism module and feature fusion module, and finally obtain the flow field data of the next time step through the output module, and reversely regress with the true value calculation loss of the next time step to optimize the neural network; S22, the specific structure of the Attention Unet network used to output the flow field data for the next time step is as follows: 1) The input layer receives a multi-channel input feature map with a size of 4×128×128, which contains flow field information with a time step of n; 2) Feature extraction module: A 3×3 convolutional layer is used in combination with GroupNorm normalization and ReLU activation function for feature extraction. The feature extraction process uses a convolutional layer with a step size of 2 to perform 4×4, 8×8, 16×16, 32×32, and 64×64 multi-scale feature extraction; 3) Attention mechanism module: Perform attention mechanism processing on feature maps of different scales, generate attention weights, and multiply the feature maps with the corresponding attention weights; 4) Feature fusion module: Upsample feature maps of different scales, and concatenate the obtained feature maps with the feature maps in the downsampling process in the channel dimension to obtain a fused feature map; 5) Output module: After a 1×1 convolutional layer, the final output feature map is generated, and the output feature map size is 3×128×128 with flow field data of time step n+1.

4. According to claim 1, a method for predicting unsteady flow field based on a neural network of a data pool is characterized by: The specific sub-steps of step S3 are as follows: S31, putting the flow field data generated by the model in the S2 process into the training data pool, replacing the flow field data of the previous time step; S32, the time memory unit updates the time step after each epoch of the neural network, increases the time step of each sample by one, accumulates the current number of samples into the counter count, calculates the reset number of samples, and finally updates the counter count to the result of taking the modulus of per_train_data_time_step, where per_train_data_time_step is the average length of continuous training for each sample; S33. Check the reset sequence list based on the calculated reset sample number. If it is empty, regenerate a sequence list from 0 to the number of samples, randomly shuffle it, take out an index from the reset sequence list, and reset the corresponding file according to the index.

5. The unsteady flow field neural network prediction method based on a data pool according to claim 4 is characterized by: The specific sub-steps of step S4 are as follows: S41, using the model trained by steps S2 to S3 to predict the unsteady flow field, taking out the data of the initial time step t=0 from the data set, inputting it into the trained neural network, and predicting the flow field data of the next time step t=1; S42, inputting the flow field data of the previous time step predicted by the model into the trained neural network, and predicting the flow field data of the next step again; S43, repeat step S42 until the stopping time step is set.

6. According to claim 3, a method for predicting unsteady flow field using a neural network based on a data pool is characterized by: The specific structure of the attention mechanism module is as follows: 1) The input layer receives a multi-channel input feature map with a size of 64×128×128; 2) The input feature map passes through a 1×1 convolution layer, and the number of convolution kernels is the same as the number of output channels. The feature map after convolution processing enters the normalization layer for normalization processing; 3) Perform feature fusion processing on the normalized feature maps, concatenate the feature maps of different inputs in the channel dimension, and obtain a fused feature map with a size of 128×128×128; 4) The fused feature map is processed as follows: First branch: The fused feature map passes through the adaptive average pooling layer to change the feature map size to 128×1×1, then passes through the fully connected layer, the ReLU activation function layer, and another fully connected layer, and finally passes through the Sigmoid activation function to output a set of weight coefficients. The weight coefficients are multiplied element by element with the original fused feature map to obtain the first corrected feature map; The second branch: The fused feature map passes through a 1×1 convolution layer, and then passes through a Sigmoid activation function to output another set of weight coefficients. The weight coefficients are multiplied element by element with the original fused feature map to obtain the second corrected feature map. The first corrected feature map and the second corrected feature map are added in the channel dimension to obtain the final output feature map, and the output feature map size is 128×128×128.

7. The unsteady flow field neural network prediction method based on a data pool according to claim 4 is characterized by: The specific implementation of step S33 is as follows: S331, randomly assigning a time memory unit at the index position to an integer between 0 and train_data_set_steps-per_train_data_time_step, where train_data_set_steps is the time series length of the training data set, and marking the data at the index position as reset; S332. Use a mask to filter out data whose time step is greater than the time step of the training sample, reset the time step of these data to 0, and mark these data as reset in the reset record unit.

Images