An Unsteady Flow Field Prediction Method Based on an Improved Unet Network

Abstract

The present invention relates to the fields of computational fluid dynamics and artificial intelligence, and proposes an unsteady flow field prediction method based on an improved Unet network, specifically including the following steps: Step 1, sample the unsteady flow field; Step 2, construct a hybrid neural network U-ConvLSTM; Step 3, train the hybrid neural network U-ConvLSTM; Step 4, use the hybrid neural network U-ConvLSTM to predict the unsteady flow field.

Description

Technical Field

[0001] The present invention relates to the fields of computational fluid dynamics and artificial intelligence, and specifically to an unsteady flow field prediction method based on an improved Unet network. Background Art

[0002] With the continuous improvement of the performance requirements of aircraft, the shape is becoming more and more complex, and the refined design requirements lead to an increasing demand for the prediction of unsteady flow fields. However, the prediction of unsteady flow fields has always been a difficult problem in the field of computational fluid dynamics (CFD). Currently, the most widely used method is to use traditional CFD methods for prediction. Traditional CFD methods are sensitive to initial conditions, have low calculation accuracy, and the dual-time-step iteration of unsteady calculations requires a large amount of computing resources and computing time. Therefore, there is an urgent need to develop a new tool for quickly and accurately predicting unsteady flow fields. Summary of the Invention

[0003] Aiming at the problem of unsteady flow field prediction in computational fluid dynamics, the present invention proposes an unsteady flow field prediction method based on an improved Unet network based on a hybrid neural network, reducing the computational consumption of the unsteady flow field, so as to achieve a fast and accurate prediction of the unsteady flow field.

[0004] The technical solution of the present invention is as follows:

[0005] The unsteady flow field prediction method based on an improved Unet network is characterized in that it includes the following steps:

[0006] Step 1: Using the object model that needs to perform unsteady flow field prediction, construct an unsteady flow field sample;

[0007] Step 2: Construct a hybrid neural network U-ConvLSTM suitable for unsteady flow field prediction. The hybrid neural network U-ConvLSTM is based on the Unet network model and uses the ConvLSTM layer, convolutional block, and deconvolutional block as constituent modules. The transfer and processing process of the input flow data is as follows:

[0008] Take several flow field data at equal time intervals obtained in Step 1 as the input data of one layer, and perform operations through the convolutional block and the ConvLSTM layer respectively. Among them, the data obtained through the convolutional block operation is passed to the next layer as the input data of the second layer; the data obtained through the ConvLSTM layer operation is used as the transfer data of one layer and is added to the output data of the second layer processed by the deconvolutional block operation through a skip connection using an addition operation to obtain the data as the output data;

[0009] The input data of the second layer are respectively operated through a convolutional block and a ConvLSTM layer. Among them, the data operated through the convolutional block are passed to the next layer as the input data of the third layer; the data operated through the ConvLSTM layer are used as the data passed by the second layer. Through skip connections, an addition operation is performed with the output data of the third layer processed by the deconvolutional block in the third layer to obtain the data as the output data of the second layer, and the output data of the second layer are passed to the upper layer through the deconvolutional block operation;

[0010] The input data of the third layer are respectively operated through a convolutional block and a ConvLSTM layer. Among them, the data operated through the convolutional block are passed to the next layer as the input data of the fourth layer; the data operated through the ConvLSTM layer are used as the data passed by the third layer. Through skip connections, an addition operation is performed with the output data of the fourth layer processed by the deconvolutional block in the fourth layer to obtain the data as the output data of the third layer, and the output data of the third layer are passed to the upper layer through the deconvolutional block operation;

[0011] The input data of the fourth layer are respectively operated through a convolutional block and a ConvLSTM layer. Among them, the data operated through the convolutional block are passed to the next layer as the input data of the fifth layer; the data operated through the ConvLSTM layer are used as the data passed by the fourth layer. Through skip connections, an addition operation is performed with the output data of the fifth layer processed by the deconvolutional block in the fifth layer to obtain the data as the output data of the fourth layer, and the output data of the fourth layer are passed to the upper layer through the deconvolutional block operation;

[0012] The input data of the fifth layer are operated through the ConvLSTM layer to obtain the output data of the fifth layer, and the output data of the fifth layer are passed to the upper layer through the deconvolutional block operation;

[0013] Step 3: Use the unsteady flow field samples constructed in Step 1 to train the hybrid neural network U-ConvLSTM constructed in Step 2;

[0014] Step 4: Fast prediction of the unsteady flow field:

[0015] Input the existing flow field information at several identical time intervals into the hybrid neural network U-ConvLSTM trained in Step 3 to obtain the flow field information at the next moment; input the flow field information at the next moment predicted by the hybrid neural network U-ConvLSTM into the trained hybrid neural network U-ConvLSTM to obtain the new flow field information at the next moment. By repeating this process, continuous prediction of the unsteady flow field is achieved.

[0016] Furthermore, constructing the unsteady flow field samples in Step 1 includes the following steps:

[0017] Step 1.1: Unsteady flow field data preparation: For the object model that requires unsteady flow field prediction, use the unsteady numerical simulation method to obtain the unsteady flow field change process around the object model within a finite number of time steps. Save the flow field information at fixed time step intervals to obtain the flow field information at each computational grid point around the object model at different times.

[0018] Step 1.2: Normalize the flow field information data obtained in Step 1.1. Use the linear normalization method to linearly transform the interval where each physical quantity of the flow field is located to the interval [0, 1]. Then arrange them in chronological order, and select several flow field data in a continuous time series to form a sample. Randomly divide the sample set into two parts according to a certain proportion, one part is the training set, and the other part is the test set.

[0019] Further, in Step 1.1, the unsteady numerical simulation method uses the detached eddy numerical simulation method or the large eddy simulation method.

[0020] Further, in Step 1.1, the flow field information includes pressure, density, flow velocity in the flow direction, and normal velocity.

[0021] Further, in Step 1.2, the proportion of the training set is 70% - 80%, and the proportion of the test set is 20% - 30%.

[0022] Further, in Step 2, the format of the flow field data is a matrix, and the format of the matrix is: number of sequences × number of channels × height × width; where the number of sequences corresponds to the number of flow fields at multiple times, the number of channels corresponds to the number of variables of the multi-variable flow field, and the height and width correspond to the size of the two-dimensional flow field.

[0023] Further, in Step 2, the convolutional block operation includes a convolutional layer, a Batch Normalization layer, and a ReLU layer, and the transposed convolutional block includes a transposed convolutional layer, a Batch Normalization layer, and a ReLU layer.

[0024] Further, in Step 3, the process of training the hybrid neural network U-ConvLSTM is as follows:

[0025] Use several unsteady flow field information at equal time intervals in the unsteady flow field samples constructed in Step 1 as the input of the hybrid neural network, and use the flow field information in the next time interval as the output of the hybrid neural network.

[0026] Use the mean square root (MSE) of the flow field parameters as the loss function, use the Adam steepest descent method as the optimization method, and use the adaptive learning rate adjustment algorithm to train the hybrid neural network U-ConvLSTM, so as to obtain the trained hybrid neural network U-ConvLSTM that can be used for unsteady flow field prediction.

[0027] Further, in step 3, the specific method of the adaptive learning rate adjustment algorithm is as follows: the initial learning rate is set to one-thousandth. If the loss does not decrease after 50 epochs of training, the learning rate is adjusted to half of the previous value.

[0028] Further, in step 3, use the entire flow field data output by the trained hybrid neural network U-ConvLSTM to draw the MSE graph. If the MSE of some regions exceeds the average MSE of the entire flow field region by at least one order of magnitude, separate the samples in these regions, and re-train the hybrid neural network U-ConvLSTM with the Batchsize set according to the size of the regional flow field and the GPU memory, so as to enhance the model accuracy of different regions.

[0029] Beneficial effects

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] The present invention proposes an unsteady flow field prediction method based on an improved Unet network. By adding ConvLSTM layer operations to the convolutional layer of the Unet network model, a hybrid neural network model U-ConvLSTM is constructed, which integrates the advantages of convolutional neural networks and long short-term memory networks, and can achieve fast and accurate prediction of unsteady flow fields. Compared with the existing CFD methods, the U-ConvLSTM model has a huge advantage in computing speed; in addition, the present invention divides the flow field into different regions for training according to the MSE value distribution in the flow field predicted by the first training of the model, which can enhance the prediction accuracy of different regions, ensure the prediction accuracy while realizing the fast prediction of the flow field, and the U-ConvLSTM model has a certain generalization ability and can be extended to the prediction of unsteady flow fields with arbitrary complex shapes.

[0032] The additional aspects and advantages of the present invention will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the present invention. Description of the drawings

[0033] Figure 1 Steps of the unsteady flow field prediction method of the present invention in the embodiment

[0034] Figure 2 Mesh of the unsteady flow field sample of the circular cylinder around which the hybrid neural network model is built in the embodiment

[0035] Figure 3 Partial enlarged view of the mesh of the unsteady flow field sample of the circular cylinder around which the hybrid neural network model is built in the embodiment

[0036] Figure 4Schematic diagram of the hybrid neural network U-ConvLSTM model for unsteady flow field prediction in the embodiment of the present invention

[0037] Figure 5 Comparison diagram of the prediction results of the U-ConvLSTM model and the CFD prediction results of the pressure, density, velocity in the z-axis direction, and velocity in the x-axis direction at each point in the unsteady flow field of circular cylinder flow at a certain moment when Re = 350 in the embodiment Detailed implementation manners

[0038] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present invention and should not be construed as limiting the present invention.

[0039] As Figure 1 shown, in this embodiment, a method for predicting the unsteady flow field around a two-dimensional circular cylinder based on an improved Unet network proposed by the present invention is used for prediction, including the following steps:

[0040] Step 1: Generate unsteady flow field samples for building a hybrid neural network model

[0041] Step 1-1: Prepare the flow field data of the unsteady flow field. The specific steps are as follows:

[0042] The method of detached eddy numerical simulation is used to solve the change process of the unsteady flow field around the circular cylinder within a finite time step, and the flow field of circular cylinder flow is simulated. The calculation conditions are as follows: Ma = 0.2, Re = 150, 200, 250, 300, 350, 400, 450, 500, 550, and 600, T ∞ = 276K. Among them, the calculation results when Re = 150, 200, 250, 300, 400, 450, 500, and 550 are used as the training set, and the calculation results when Re = 350 and 600 are used as the test set.

[0043] Set the dimensionless time step to Δt = 0.01. As Figure 2 and Figure 3 shown, the computational grid uses an O-type structured grid, and the grid quantity is 100 (radial) × 200 (circumferential). To capture the flow conditions in the boundary layer, ensure that the grid Reynolds number calculated based on the height of the first layer of the wall grid is less than 50. The wall boundary condition is an adiabatic and non-slip wall. After the calculation is stable, every 1000 time steps is a cycle, and sampling is performed within two cycles after the calculation is stable under each calculation condition. The specific sampling method is to save the flow field every 10 time steps, including the pressure, density, flow velocity, and normal velocity at each computational grid point.

[0044] In Step 1-2, the processed flow field information data is normalized using the linear normalization method, linearly transforming the data range of each physical quantity in the flow field of the sample set to [0, 1]. Arranged in chronological order, 16 flow field sequences form a sample. In this embodiment, the training set accounts for 80% with a quantity of 1472, and the test set accounts for 20% with a quantity of 368.

[0045] Step 2: Construct a hybrid neural network U-ConvLSTM suitable for unsteady flow field prediction. This network is based on the Unet network model and consists of ConvLSTM layers, convolutional blocks, and deconvolutional blocks as its building blocks.

[0046] As Figure 4 shown, the input data transfer process in Step 2-1 is as follows: The flow field data of the first 15 sequences in a single sample is used as the input data for one layer. It undergoes operations through the convolutional block and the ConvLSTM layer respectively. The data passing through the ConvLSTM layer is used as the transfer data for one layer and is added through skip connections to the output data of the second layer processed by the deconvolutional block to obtain the output data. The data passing through the convolutional block is then transferred to the next layer as the input data for the second layer. The input data for the second layer undergoes operations through the convolutional block and the ConvLSTM layer respectively. The data passing through the ConvLSTM layer is used as the transfer data for the second layer and is added through skip connections to the output data of the third layer processed by the deconvolutional block to obtain the output data for the second layer. The output data for the second layer is then processed by the deconvolutional block and transferred to the upper layer, while the data passing through the convolutional block is transferred to the next layer as the input data for the third layer. The input data for the third layer undergoes operations through the convolutional block and the ConvLSTM layer respectively. The data passing through the ConvLSTM layer is used as the transfer data for the third layer and is added through skip connections to the output data of the fourth layer processed by the deconvolutional block to obtain the output data for the third layer. The output data for the third layer is then processed by the deconvolutional block and transferred to the upper layer, while the data passing through the convolutional block is transferred to the next layer as the input data for the fourth layer. The input data for the fourth layer undergoes operations through the convolutional block and the ConvLSTM layer respectively. The data passing through the ConvLSTM layer is used as the transfer data for the fourth layer and is added through skip connections to the output data of the fifth layer processed by the deconvolutional block to obtain the output data for the fourth layer. The output data for the fourth layer is then processed by the deconvolutional block and transferred to the upper layer, while the data passing through the convolutional block is transferred to the next layer as the input data for the fifth layer. The input data for the fifth layer undergoes operations through the ConvLSTM layer to obtain the output data for the fifth layer, and the output data for the fifth layer is then processed by the deconvolutional block and transferred to the upper layer.

[0047] Step 3: The specific training steps of the hybrid neural network U-ConvLSTM are as follows:

[0048] Step 3-1: The input of the hybrid neural network U-ConvLSTM is 15 unsteady flow field information at equal time intervals in the dataset constructed in Step 1-2, and the output is the flow field information at the next time interval.

[0049] Step 3-2: Using the root mean square of the flow field parameters as the loss function (MSE) and the Adam steepest descent method as the optimization method, the hybrid neural network U-ConvLSTM is trained using an adaptive learning rate adjustment algorithm. The specific method is as follows: The initial learning rate is set to one-thousandth. If the loss does not decrease after 50 epochs of training, the learning rate is adjusted to half of the previous value. Thus, a trained hybrid neural network U-ConvLSTM that can be used for unsteady flow field prediction is obtained.

[0050] Step 3-3: After training is completed, an MSE graph of the entire flow field is plotted. If the MSE in some regions exceeds one to two orders of magnitude of the overall region, the samples in that region are trained with a Batchsize based on the size of the regional flow field and the GPU memory. The samples are separated into different regions and trained separately to enhance the accuracy of different regions. In this embodiment, since the flow field mainly only contains unsteady vortex shedding structures caused by shear layer instability and the flow structure is relatively simple, there is no need for regional training.

[0051] Step 4: Fast prediction of the unsteady flow field: The first 15 flow field information in the training set samples are input into the trained hybrid neural network U-ConvLSTM in Step 3-3, and the flow field information at the next moment is obtained. The flow field information at the next moment predicted by the hybrid neural network U-ConvLSTM is input into the trained hybrid neural network U-ConvLSTM, and the new flow field information at the next moment is obtained. By repeating this process, continuous prediction of the unsteady flow field of cylinder flow around is achieved.

[0052] As Figure 5As shown, the predicted results of the U-ConvLSTM model, the CFD predicted results, and the error comparison of the predicted results of the U-ConvLSTM model compared with the CFD predicted results for the pressure, density, velocity in the z-axis direction, and velocity in the x-axis direction at each location in the unsteady flow field of a circular cylinder at a certain moment when Re = 350 are given. The x-axis is the flow direction, and the z-axis is the direction perpendicular to the flow direction. Compared with the CFD prediction method, the prediction error of the method of the present invention is less than 1%. It can be seen that the method of the present invention can accurately predict the unsteady flow field. Moreover, on the same computer, the calculation time for calculating the flow field at a single moment using the CFD method is 3 minutes, while the calculation time for predicting the flow field at a single moment using the U-ConvLSTM model is only 1.3 seconds. It can be seen that the method of the present invention can greatly improve the calculation efficiency while ensuring the calculation accuracy.

[0053] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. An unsteady flow field prediction method based on an improved Unet network, characterized in that, It includes the following steps: Step 1: Using the object model for which unsteady flow field prediction is required, construct unsteady flow field samples; Step 2: Construct a hybrid neural network U-ConvLSTM suitable for unsteady flow field prediction. The hybrid neural network U-ConvLSTM is based on the Unet network model and uses ConvLSTM layers, convolutional blocks, and deconvolutional blocks as constituent modules. The transfer and processing process for the input flow data is as follows: Take several flow field data at equal time intervals obtained in Step 1 as the input data for one layer, and perform operations through convolutional blocks and ConvLSTM layers respectively. Among them, the data processed through convolutional block operations is passed to the next layer as the input data for the second layer; the data processed through ConvLSTM layer operations is used as the transfer data for one layer and is added through skip connections to the output data of the second layer processed through deconvolutional block operations to obtain the data as the output data; The input data for the second layer is respectively processed through convolutional block operations and ConvLSTM layer operations. Among them, the data processed through convolutional block operations is passed to the next layer as the input data for the third layer; the data processed through ConvLSTM layer operations is used as the transfer data for the second layer and is added through skip connections to the output data of the third layer processed through deconvolutional block operations to obtain the data as the output data for the second layer, and the output data for the second layer is passed to the upper layer through deconvolutional block operations; The input data for the third layer is respectively processed through convolutional block operations and ConvLSTM layer operations. Among them, the data processed through convolutional block operations is passed to the next layer as the input data for the fourth layer; the data processed through ConvLSTM layer operations is used as the transfer data for the third layer and is added through skip connections to the output data of the fourth layer processed through deconvolutional block operations to obtain the data as the output data for the third layer, and the output data for the third layer is passed to the upper layer through deconvolutional block operations; The input data for the fourth layer is respectively processed through convolutional block operations and ConvLSTM layer operations. Among them, the data processed through convolutional block operations is passed to the next layer as the input data for the fifth layer; the data processed through ConvLSTM layer operations is used as the transfer data for the fourth layer and is added through skip connections to the output data of the fifth layer processed through deconvolutional block operations to obtain the data as the output data for the fourth layer, and the output data for the fourth layer is passed to the upper layer through deconvolutional block operations; The input data for the fifth layer is processed through the ConvLSTM layer operation to obtain the output data for the fifth layer, and the output data for the fifth layer is passed to the upper layer through deconvolutional block operations; Step 3: Use the unsteady flow field samples constructed in Step 1 to train the hybrid neural network U-ConvLSTM constructed in Step 2; Step 4: Fast prediction of the unsteady flow field: Input the flow field information of several identical time intervals into the hybrid neural network U-ConvLSTM trained in step 3 to obtain the flow field information at the next moment; input the flow field information at the next moment predicted by the hybrid neural network U-ConvLSTM into the trained hybrid neural network U-ConvLSTM to obtain the new flow field information at the next moment. Repeat this process to achieve continuous prediction of the unsteady flow field.

2. The unsteady flow field prediction method based on an improved Unet network according to claim 1, characterized in that, In step 1, constructing the unsteady flow field samples includes the following steps: Step 1.1: Preparation of unsteady flow field data: For the object model for which unsteady flow field prediction is required, use the unsteady numerical simulation method to obtain the unsteady flow field change process around the object model within a finite number of time steps. Save the flow field information at fixed time step intervals to obtain the flow field information at each computational grid point around the object model at different times. Step 1.2: Normalize the flow field information data obtained in step 1.

1. Use the linear normalization method to linearly transform the interval where each physical quantity of the flow field is located to the interval [0, 1]; then arrange them in chronological order and select several flow field data in a continuous time series to form a sample; randomly divide the sample set into two parts according to a certain proportion, one part is the training set and the other part is the test set.

3. The unsteady flow field prediction method based on an improved Unet network according to claim 2, characterized in that, In step 1.1, the unsteady numerical simulation method uses the detached eddy numerical simulation method or the large eddy simulation method.

4. The unsteady flow field prediction method based on an improved Unet network according to claim 2, characterized in that, In step 1.1, the flow field information includes pressure, density, flow velocity in the flow direction, and normal velocity.

5. The unsteady flow field prediction method based on an improved Unet network according to claim 2, characterized in that, In step 1.2, the proportion of the training set is 70% - 80%, and the proportion of the test set is 20% - 30%.

6. The unsteady flow field prediction method based on an improved Unet network according to claim 1, characterized in that, In step 2, the format of the flow field data is a matrix, and the format of the matrix is: number of sequences × number of channels × height × width; Among them The number of sequences corresponds to the number of flow fields at multiple moments, the number of channels corresponds to the number of variables of the multi-variable flow field, and the height and width correspond to the size of the two-dimensional flow field.

7. The unsteady flow field prediction method based on an improved Unet network according to claim 1, characterized in that, In step 2, the convolution block operation includes a convolutional layer, a Batch Normalization layer, and a ReLU layer, and the deconvolution block includes a deconvolutional layer, a Batch Normalization layer, and a ReLU layer.

8. The unsteady flow field prediction method based on an improved Unet network according to claim 1, characterized in that, In step 3, the process of training the hybrid neural network U-ConvLSTM is as follows: Use several unsteady flow field information at equal time intervals in the unsteady flow field samples constructed in step 1 as the input of the hybrid neural network, and use the flow field information at the next time interval as the output of the hybrid neural network; Use the mean square root (MSE) of the flow field parameters as the loss function, use the Adam steepest descent method as the optimization method, and use the adaptive learning rate adjustment algorithm to train the hybrid neural network U-ConvLSTM, so as to obtain the trained hybrid neural network U-ConvLSTM that can be used for unsteady flow field prediction.

9. The unsteady flow field prediction method based on an improved Unet network according to claim 1, characterized in that, In step 3, the specific method of the adaptive learning rate adjustment algorithm is: set the initial learning rate to one-thousandth. If the loss does not decrease after 50 epochs of training, adjust the learning rate to half of the previous value.

10. The unsteady flow field prediction method based on the improved Unet network according to claim 1, characterized in that, In step 3, draw an MSE graph using the entire flow field data output by the trained hybrid neural network U-ConvLSTM. If the MSE in some regions exceeds the average MSE of the entire flow field region by at least one order of magnitude, separate the samples in these regions, and then retrain the hybrid neural network U-ConvLSTM with the Batchsize determined according to the size of the regional flow field and the GPU memory, so as to enhance the model accuracy of different regions.

Images