Hypersonic flow field density and speed high-precision prediction method, system and equipment based on physical information constraint U-Net, and medium

Abstract

The invention discloses a hypersonic flow field density and speed high-precision prediction method, system and device based on physical information constraint U-Net, and a medium. The method comprises the following steps: obtaining value distribution of instantaneous transverse and longitudinal speeds and densities in target wake flow fields with different Mach numbers; forming a data set by the densities, the transverse speeds and the longitudinal speeds of two continuous moments and the density, the transverse speed and the longitudinal speed of the third moment in the target wake flow field at different moments, and dividing the data set into a training set and a test set; constructing a physical information U-Net flow field prediction model, and defining a continuity equation in an N-S equation as a loss function of the model; training the physical information U-Net flow field prediction model to obtain a physical information U-Net flow field prediction model of which the weight is trained; predicting input data by using the physical information U-Net flow field prediction model of which the weight is trained, and obtaining a prediction result through recursive prediction; the system, the equipment and the medium are used for implementing the method. By means of the method, the output data can meet the physical mechanism, and the accuracy is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of flow field prediction, and in particular relates to a method, system, equipment and medium for high-precision prediction of hypersonic flow field density and velocity based on physical information constraint U-Net. Background Art

[0002] Current numerical simulations of hypersonic flow fields primarily rely on traditional computational fluid dynamics (CFD) methods based on the Navier-Stokes (NS) equations. While these methods have a clear physical foundation, they suffer from low computational efficiency and high hardware resource consumption when dealing with complex three-dimensional flow fields. In recent years, deep learning technology has shown promise in flow field prediction, but purely data-driven models often suffer from issues such as a lack of physical conservation and poor adaptability to boundary conditions. This makes them particularly difficult to meet the accuracy requirements of engineering predictions at high Mach numbers. While existing neural network prediction methods can rapidly generate flow fields, they fail to effectively incorporate the constraints of the fundamental fluid dynamics equations, resulting in a lack of physical consistency in the prediction results. Furthermore, existing technical solutions rely on massive amounts of high-precision numerical simulation data during the training phase, significantly increasing experimental verification cycles and computational costs, limiting the iterative efficiency of hypersonic vehicle aerodynamic design. Furthermore, relevant scientific research requires complete and continuous data sets of flow field parameters such as density and velocity, which experimental measurement methods struggle to meet.

[0003] Patent application publication number CN 118504421 A discloses a flow field prediction method that integrates operator learning and convolutional neural networks. This method predicts flow field data for aircraft with different shapes at different incoming Mach numbers and angles of attack. The flow field prediction scheme that integrates operator learning and convolutional neural networks has good generalization of model shapes and high accuracy. However, because the integration of operator learning and convolutional neural networks lacks time series data processing capabilities, time series prediction cannot be performed. Furthermore, the integration of operator learning and convolutional models is relatively complex, requiring more time and training data. However, the material and time costs required for high-performance flow field experiments are high, making them incapable of meeting large data requirements. Similarly, tasks with small data volumes can encounter data issues such as overfitting and output data that does not conform to physical mechanisms. Summary of the Invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a high-precision prediction method, system, equipment and medium for the density and velocity of hypersonic flow fields based on physical information constraint U-Net, construct a physical information U-Net flow field prediction model, and define the continuity equation in the NS equation as the loss function of the physical information U-Net flow field prediction model. The physical information U-Net flow field prediction model can realize the time series prediction of the hypersonic flow field. By inputting the density, lateral velocity and longitudinal velocity data of the flow field at two consecutive moments, the future flow field evolution is predicted. Benefiting from the continuity equation in the NS equation, the output data can meet the physical mechanism and have high accuracy.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A high-precision prediction method for hypersonic flow field density and velocity based on physical information constrained U-Net includes the following steps:

[0007] Step 1: Obtain the numerical distribution of instantaneous lateral and longitudinal velocities and densities in the wake flow field of the target at different Mach numbers; Compose a data set of the density, lateral velocity, and longitudinal velocity at two consecutive moments and the third moment in the wake flow field of the target at different moments, and divide the data set into a training set and a test set;

[0008] Step 2: construct a physical information U-Net flow field prediction model, which includes four consecutive encoders, namely encoder 1, encoder 2, encoder 3, and encoder 4, and four consecutive decoders, namely decoder 1, decoder 2, decoder 3, and decoder 4, and a reconstruction layer, which outputs data to a desired size; define the continuity equation in the NS equation as the loss function of the physical information U-Net flow field prediction model;

[0009] Step 3: Use the training set divided in step 1 to train the physical information U-Net flow field prediction model constructed in step 2 in the hypersonic target wake flow field at different times. Use the loss function defined in step 2 to supervise the training process. During the training process, predict the test set data at the same time until the loss function of the training set and the root mean square error of the test set are reduced and converged, thus obtaining the physical information U-Net flow field prediction model with trained weights.

[0010] In step 4, the lateral velocity and longitudinal velocity of the target flow field at two consecutive moments and the density data of the target flow field are used as input data, and the physical information U-Net flow field prediction model with trained weights obtained in step 3 is used to predict the input data, and the prediction results are obtained through recursive prediction.

[0011] The specific method of step 1 includes:

[0012] The numerical distribution of instantaneous lateral, longitudinal velocity and density in the wake flow field of targets with different Mach numbers is obtained and stored as matrix data at different times. The matrix data is preprocessed to obtain flow field data with consistent flow field direction and flow field area. The density values ​​and the lateral and longitudinal velocity data distribution of the wake flow field of targets with different Mach numbers at all times are normalized, and the data are named according to the time sequence. The density, lateral velocity and longitudinal velocity of two consecutive moments and the third moment in the wake flow field of targets at different times are combined into a data set, and the data set is divided into a training set and a test set.

[0013] In step 2, the encoder is composed of the first convolutional layer, the maximum pooling layer, and the second convolutional layer in sequence. Encoder 1 has 128 channels and a size of 125×100, encoder 2 has 256 channels and a size of 62×50, encoder 3 has 512 channels and a size of 31×25, and encoder 4 has 512 channels and a size of 15×12;

[0014] The decoder consists of a deconvolution layer, a skip connection, a third convolution layer, and a fourth convolution layer. Decoder 1 has 512 channels and a size of 31×25, decoder 2 has 256 channels and a size of 62×50, decoder 3 has 128 channels and a size of 125×100, and decoder 4 has 64 channels and a size of 250×200.

[0015] The input data is of 6-channel 250*200 size. After passing through the double convolutional layer, the data becomes 64-channel 250*200 size, and then enters four consecutive encoders and four consecutive decoders. The connection relationship between the encoder and decoder is: every time the flow field data passes through an encoder, the data feature size is reduced by half and the number of channels is doubled. Every time the flow field data passes through a decoder, the feature map size is doubled and the number of channels is halved; the flow field data enters the encoder and is compressed layer by layer to the bottleneck size. The decoder starts from the bottleneck size and fuses the encoder features layer by layer through jump connections to gradually restore the original size; finally, the data enters the reconstruction layer and outputs 3-channel 250*200 size data, where the 3 channels represent 3 different physical quantities.

[0016] In step 2, the continuity equation in the NS equation is defined as the loss function of the physical information U-Net flow field prediction model as follows: the continuity equation in the NS equation is used to extract spatial gradient and time gradient features, and discretized using central difference and forward difference methods respectively; the discrete expression is combined with the mean square error to form a new loss function; the loss function is combined with the physical information U-Net flow field prediction model, so that the loss function is reversely propagated in the physical information U-Net flow field prediction model to optimize the weights of the physical information U-Net flow field prediction model.

[0017] In the step 1, pre-processing the matrix data includes locating bad values ​​in the matrix data, and using surrounding values ​​to calculate and replace them according to the continuity equation in the NS equation.

[0018] In step 1, the normalization process is specifically to perform minimum-maximum normalization on all data.

[0019] The present invention also provides a high-precision prediction system for hypersonic flow field density and velocity based on physical information constraint U-Net, comprising:

[0020] The data set acquisition and processing module is used to obtain the numerical distribution of instantaneous lateral and longitudinal velocities and densities in the wake flow field of the target at different Mach numbers; the density, lateral velocity and longitudinal velocity of the target wake flow field at two consecutive moments and the third moment are combined into a data set, and the data set is divided into a training set and a test set;

[0021] A model construction and loss function definition module is used to construct a physical information U-Net flow field prediction model, wherein the physical information U-Net flow field prediction model includes four consecutive encoders, namely encoder 1, encoder 2, encoder 3, and encoder 4, and four consecutive decoders, namely decoder 1, decoder 2, decoder 3, and decoder 4, and a reconstruction layer, which outputs data to a desired size; the continuity equation in the NS equation is defined as the loss function of the physical information U-Net flow field prediction model;

[0022] The model training module is used to train the physical information U-Net flow field prediction model in the hypersonic target wake flow field at different times using the training set. The training process is supervised by the defined loss function. During the training process, predictions are made on the test set data until the loss function of the training set and the root mean square error of the test set are reduced and converged, thereby obtaining the physical information U-Net flow field prediction model with trained weights.

[0023] The data prediction module is used to use the known lateral velocity and longitudinal velocity of the target flow field at two consecutive moments and the density data of the target flow field as input data, use the physical information U-Net flow field prediction model with trained weights to predict the input data, and obtain the prediction results through recursive prediction.

[0024] The present invention also provides a high-precision prediction device for hypersonic flow field density and velocity based on physical information constraint U-Net, comprising:

[0025] Memory: a computer-readable device storing a computer program for the above-mentioned method for high-precision prediction of hypersonic flow field density and velocity based on physical information constraint U-Net;

[0026] Processor: used to implement the high-precision prediction method for hypersonic flow field density and speed based on physical information constraint U-Net when executing the computer program.

[0027] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the high-precision prediction method for hypersonic flow field density and speed based on physical information constraint U-Net.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. The present invention constrains the training process through the continuity equation in the NS equation, so that the fluid structure maintains physical consistency during the prediction process, enhances the prediction accuracy, and improves the adaptability of small data samples.

[0030] 2. The present invention realizes multi-physical field coupling by defining the continuity equation in the NS equation as the loss function of the physical information U-Net flow field prediction model, which can simultaneously predict the density and velocity fields and solve the coupling problem of distributed solution in traditional methods.

[0031] 3. The present invention realizes the function of time-series stable prediction of flow field physical quantities through recursive prediction, solving the problem that traditional convolutional neural networks cannot perform time-series prediction.

[0032] 4. The innovation of the present invention is to introduce physical conservation laws (such as the Navier-Stokes equations) as embedded constraints into the deep learning framework, realizing the coordinated optimization of data-driven and physical mechanism priors. Compared with the traditional purely data-driven time convolution model, the present invention shows significant advantages in the following three aspects: First, by constructing a physical information constraint term based on the continuity equation in the NS equation, the neural network prediction results are forced to follow the basic physical laws of mass conservation and momentum conservation in terms of time and space characteristics, fundamentally ensuring the numerical stability of the physical field evolution process; secondly, based on the physical information fusion mechanism, the continuity equation in the flow field is used as an important component of the training target, which effectively reduces the model's dependence on high-precision labeled data, and can still achieve reliable predictions when there are only sparse observation data or partial boundary conditions; thirdly, by embedding the spatial gradient and time gradient operators of the physical equations into the network architecture, the physical information U-Net flow field prediction model has cross-scale generalization capabilities, and can still maintain robust prediction performance when the input data resolution changes or the flow parameters exceed the training distribution. Experiments show that this modeling paradigm that integrates physical constraints with deep learning exhibits certain generalization properties in tests.

[0033] In summary, the present invention constrains the physical information of the U-Net flow field prediction model by using the continuity equation and combining it with the residual of the NS equation. This allows the network's prediction results to naturally meet the physical characteristics of the flow field, enhancing physical consistency. This approach has the advantage of being able to handle multi-physics problems and simultaneously capture characteristics with large spatial gradients and temporal variations in the flow field. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a structural diagram of the physical information U-Net flow field prediction model provided by an embodiment of the present invention.

[0035] Figure 2 It is a physical information loss function structure for supervising U-Net training provided by an embodiment of the present invention.

[0036] Figure 3 The hypersonic aircraft geometric model provided by the embodiment of the present invention is used for obtaining wake data simulation.

[0037] Figure 4 This is the convergence process of the loss function of the physical information U-Net flow field prediction model training provided by the embodiment of the present invention.

[0038] Figure 5 This is a comparison diagram of the density prediction results and true values ​​of the flow field provided by an embodiment of the present invention, with the left side showing the true values ​​at different consecutive moments and the right side showing the predicted values.

[0039] Figure 6 This is the correlation between the 0.38ms and 0.44ms prediction results provided by the embodiment of the present invention and the true value.

[0040] Figure 7 This is the recursive prediction result of 0.35-0.47ms provided by an embodiment of the present invention using 0.29 and 0.32ms data as input.

[0041] Figure 8 The correlation coefficient of the results from 0.35 to 0.77 ms is provided in the embodiment of the present invention when 0.29 and 0.32 ms data are used as input. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0043] This invention, a high-precision hypersonic flow field density and velocity prediction model based on a physical information-constrained U-Net, combines the laws of physics with the U-Net model to predict wake flow field density and velocity data using the NS equations. This model complements the lack of interpretability in deep learning models by using the conservation equation of the continuity equation. This invention fully utilizes flow field experimental data, reducing both material and time costs.

[0044] like Figure 1 As shown, a high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net includes the following steps:

[0045] like Figure 3 As shown in step 1, the blunt cone has a head radius of 0.25m, a semi-cone angle of 8°, and a bottom diameter of 1m. After model simulation, the instantaneous lateral, longitudinal velocity, and density distributions in the wake flow field of the blunt cone target at different Mach numbers are obtained, and the instantaneous lateral, longitudinal velocity, and density values ​​are stored as data in a format of 250×200. The data acquisition steps are repeated to organize the data at different times into a data set. The data at different Mach numbers and times are divided into a training set and a test set in a ratio of 8:2.

[0046] In step 1, the numerical distribution of instantaneous lateral and longitudinal velocities and densities in the wake flow field of the blunt cone target at different Mach numbers is obtained by wind tunnel, shock tube experimental data or CFD simulation.

[0047] Step 2: Use four consecutive encoders, namely encoder 1, encoder 2, encoder 3, and encoder 4, and four consecutive decoders, namely decoder 1, decoder 2, decoder 3, and decoder 4, to build a physical information U-Net flow field prediction model; define the continuity equation in the NS equation as the loss function of the physical information U-Net flow field prediction model, specifically:

[0048] The encoder consists of the first convolutional layer, the maximum pooling layer, and the second convolutional layer in sequence. Encoder 1 has 128 channels and a size of 125×100, encoder 2 has 256 channels and a size of 62×50, encoder 3 has 512 channels and a size of 31×25, and encoder 4 has 512 channels and a size of 15×12. The purpose is to gradually compress the spatial dimensions of the input flow field (such as velocity field and density field) and extract high-level features.

[0049] The decoder consists of a deconvolution layer, a skip connection, a third convolution layer, and a fourth convolution layer. Decoder 1 has 512 channels and a size of 31×25, decoder 2 has 256 channels and a size of 62×50, decoder 3 has 128 channels and a size of 125×100, and decoder 4 has 64 channels and a size of 250×200. The purpose is to fuse shallow details with deep information and gradually restore spatial resolution.

[0050] The physical information U-Net flow field prediction model is used to generate two-dimensional flow field data. The input layer receives six channels of input, each consisting of three features at two moments, and normalizes the data's numerical structure. Four consecutive encoders perform four sampling passes, reducing the feature size of the flow field data and increasing the number of channels. Next, four decoders restore the reduced data features to the required size, accepting skip connection data from the corresponding encoders during the restoration process. Finally, the data passes through a reconstruction layer, which outputs the data at the desired size.

[0051] In addition, it is necessary to construct the loss function of the physical information U-Net flow field prediction model, such as Figure 2 As shown, the continuity equation in the NS equation is introduced, and the density ρ and the lateral velocity v are calculated by the central difference method. x and longitudinal velocity v y The partial derivative values ​​of are defined as the loss function of the physical information U-Net flow field prediction model by summing these partial derivative values ​​according to the continuity equation.

[0052] In the near-wake flow field of a hypersonic target, the NS equations are usually used to describe the spatial distribution of various parameters in the flow field. The loss function of the physical information U-Net flow field prediction model is constructed by using the continuity equations describing the physical quantities in the NS equations:

[0053]

[0054] Where t represents time, ρ is density, and v y is the component of the physical quantity v in the y direction; the above formula is expressed using the central difference with four surrounding points:

[0055] The expression in the wake domain is:

[0056]

[0057]

[0058] The trail boundary expression is:

[0059]

[0060] The continuity equation in the NS equation is defined as the loss function of the physical information U-Net flow field prediction model, so:

[0061]

[0062] The above loss function can capture spatial features with dramatic changes in physical quantities, such as shock wave contours and parameter gradients at the post-stagnation point position, as well as temporal evolutionary features, such as shock wave expansion and the gradual stabilization of the post-stagnation point. Furthermore, traditional loss calculation methods can guarantee the numerical accuracy of the output. Therefore, the loss function combines the above spatial gradient term with the traditional mean square error loss term:

[0063] The physical information U-Net flow field prediction model calculates the residual module of the entire smooth area and adds the spatial physical quantity differential loss. The specific calculation method is:

[0064] loss=loss p +loss MSE

[0065] Among them, loss MSE The loss value calculated using the above formula is inserted into the physical information U-Net flow field prediction model to supervise the flow field prediction results. The physical information U-Net flow field prediction model has physical theoretical advantages for the non-uniform spatial distribution of wake density and velocity.

[0066] Step 3: Use the training set divided in step 1 to train the physical information U-Net flow field prediction model constructed in step 2 in the hypersonic target wake flow field at different times. At the same time, use the loss function of step 2 to supervise the training process until the loss function converges. The convergence process of the loss value on the training set and the test set is shown in Figure 2. Figure 4 As shown, the physical information U-Net flow field prediction model with trained weights is obtained;

[0067] In step 4, the lateral and longitudinal velocities and density data of the target flow field at two consecutive moments are used as input data. Specifically, the velocity and density data at two consecutive moments are derived from wind tunnel, shock tube experimental data, or CFD simulation data. The input data is predicted using the weighted, physically-informed U-Net flow field prediction model obtained in step 3 to obtain the flow field data distribution for the target flow field at several future moments.

[0068] Use the trained physical information U-Net flow field prediction model to obtain the flow field data of the target flow field at the next moment, such as Figure 5As shown in FIG, it is a comparison diagram of the density prediction result and the true value of the flow field provided by the embodiment of the present invention. The left side is the true value at different consecutive moments, and the right side is the predicted value. The specific prediction process is to use the data of 0.32ms and 0.35ms to predict the flow field data of 0.38ms; use the data of 0.35ms and 0.38ms to get the data of 0.41ms, and so on. Among them, the prediction results of 0.38~0.47ms are highly consistent with the true value. And as Figure 6 The correlation coefficients of the prediction results of 0.38ms and 0.44ms are shown in the figure. The correlation coefficient can reach more than 90%, indicating that the results given by the physical information U-Net flow field prediction model are reliable. In the prediction process, the recursive prediction method is used, and the 0.29 and 0.32ms data can be used as input to obtain the flow field data at 0.35-0.47ms. The results are shown in the figure. Figure 7 As shown in the figure, it can be seen that the lateral velocity, longitudinal velocity, and density data of the target flow field maintain a consistent development process within the time range of 0.35 to 0.47 ms, showing the characteristics of multi-physical quantity coupling.

[0069] To evaluate the effect of recursive training, the prediction cases with 0.29 and 0.32 ms data as input are continuously recursively output until 0.44 ms. The correlation coefficient is statistically calculated as follows: Figure 8 As shown, the correlation of the results after the tenth recursion begins to drop significantly, indicating that the physical information U-Net flow field prediction model of the present invention has the ability to make stable recursive predictions for 8 or 9 times.

[0070] The present invention can predict flow field velocity and density data based on the transverse and longitudinal velocity and density data of the flow field at two moments in time. The physical information U-Net flow field prediction model is capable of predicting the density and velocity data of the target wake flow field over a subsequent period of time using density and velocity data from two consecutive moments. This invention can assist high-performance flow field experimental measurement equipment, addressing the high costs associated with repeated experiments. Therefore, the present invention aligns with theoretical mechanisms and reduces experimental costs.

[0071] The key points and protection points of the present invention are:

[0072] 1. Deep learning model: The deep learning model consists of two main parts: one is the two-dimensional physical information U-Net flow field prediction model, and the other is the loss function defined based on the NS equation.

[0073] 2. Physical information U-Net flow field prediction model: Perform two-dimensional flow field prediction on the input data to obtain the two-dimensional distribution of target density and velocity.

[0074] The present invention mentions constructing a physical mechanism regulation mechanism based on the continuity equation in the NS equation. If other equations in the NS equation are used, similar models can also be constructed and the same functions can be achieved, such as the mass conservation equation.

[0075] The present invention also provides a high-precision prediction system for hypersonic flow field density and velocity based on physical information constraint U-Net, comprising:

[0076] The data set acquisition and processing module is used to obtain the numerical distribution of instantaneous lateral and longitudinal velocities and densities in the wake flow field of the target at different Mach numbers in step 1; the density, lateral velocity and longitudinal velocity of the target wake flow field at two consecutive moments and the third moment are combined into a data set, and the data set is divided into a training set and a test set;

[0077] A model construction and loss function definition module is used to implement the construction of a physical information U-Net flow field prediction model in step 2, wherein the physical information U-Net flow field prediction model includes four consecutive encoders, namely encoder 1, encoder 2, encoder 3, and encoder 4, and four consecutive decoders, namely decoder 1, decoder 2, decoder 3, and decoder 4, and a reconstruction layer, which outputs data to a desired size; and defines the continuity equation in the NS equation as the loss function of the physical information U-Net flow field prediction model;

[0078] The model training module is used to train the physical information U-Net flow field prediction model constructed in step 2 in the hypersonic target wake flow field at different times using the training set divided in step 1 in step 3, supervise the training process using the loss function defined in step 2, and simultaneously predict the test set data during the training process until the loss function of the training set and the root mean square error of the test set are reduced and converged, thereby obtaining the physical information U-Net flow field prediction model with trained weights;

[0079] The data prediction module is used to implement the method in step 4, which uses the known lateral velocity and longitudinal velocity of the target flow field at two consecutive moments and the density data of the target flow field as input data, and uses the physical information U-Net flow field prediction model with trained weights obtained in step 3 to predict the input data, and obtains the prediction result through recursive prediction.

[0080] The present invention also provides a high-precision prediction device for hypersonic flow field density and velocity based on physical information constraint U-Net, comprising:

[0081] Memory: a computer-readable device storing a computer program for the above-mentioned method for high-precision prediction of hypersonic flow field density and velocity based on physical information constraint U-Net;

[0082] Processor: used to implement the high-precision prediction method for hypersonic flow field density and speed based on physical information constraint U-Net when executing the computer program.

[0083] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it can implement the high-precision prediction method for hypersonic flow field density and speed based on physical information constraint U-Net.

Claims

1. A high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net, characterized by: The following steps are involved: Step 1: Obtain the numerical distribution of instantaneous lateral and longitudinal velocities and densities in the wake flow field of the target at different Mach numbers; Compose a data set of the density, lateral velocity, and longitudinal velocity at two consecutive moments and the third moment in the wake flow field of the target at different moments, and divide the data set into a training set and a test set; Step 2: construct a physical information U-Net flow field prediction model, which includes four consecutive encoders, namely encoder 1, encoder 2, encoder 3, and encoder 4, and four consecutive decoders, namely decoder 1, decoder 2, decoder 3, and decoder 4, and a reconstruction layer, which outputs data to a desired size; define the continuity equation in the NS equation as the loss function of the physical information U-Net flow field prediction model; Step 3: Use the training set divided in step 1 to train the physical information U-Net flow field prediction model constructed in step 2 in the hypersonic target wake flow field at different times. Use the loss function defined in step 2 to supervise the training process. During the training process, predict the test set data at the same time until the loss function of the training set and the root mean square error of the test set are reduced and converged, thus obtaining the physical information U-Net flow field prediction model with trained weights. In step 4, the lateral velocity and longitudinal velocity of the target flow field at two consecutive moments and the density data of the target flow field are used as input data, and the physical information U-Net flow field prediction model with trained weights obtained in step 3 is used to predict the input data, and the prediction results are obtained through recursive prediction.

2. The high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net according to claim 1 is characterized in that: The specific method of step 1 includes: The numerical distribution of instantaneous lateral, longitudinal velocity and density in the wake flow field of targets with different Mach numbers is obtained and stored as matrix data at different times. The matrix data is preprocessed to obtain flow field data with consistent flow field direction and flow field area. The density values ​​and the lateral and longitudinal velocity data distribution of the wake flow field of targets with different Mach numbers at all times are normalized, and the data are named according to the time sequence. The density, lateral velocity and longitudinal velocity of two consecutive moments and the third moment in the wake flow field of targets at different times are combined into a data set, and the data set is divided into a training set and a test set.

3. The high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net according to claim 1 is characterized in that: In step 2, the encoder is composed of the first convolutional layer, the maximum pooling layer, and the second convolutional layer in sequence. Encoder 1 has 128 channels and a size of 125×100, encoder 2 has 256 channels and a size of 62×50, encoder 3 has 512 channels and a size of 31×25, and encoder 4 has 512 channels and a size of 15×12; The decoder consists of a deconvolution layer, a skip connection, a third convolution layer, and a fourth convolution layer. Decoder 1 has 512 channels and a size of 31×25, decoder 2 has 256 channels and a size of 62×50, decoder 3 has 128 channels and a size of 125×100, and decoder 4 has 64 channels and a size of 250×200. The input data is of 6-channel 250*200 size. After passing through the double convolutional layer, the data becomes 64-channel 250*200 size, and then enters four consecutive encoders and four consecutive decoders. The connection relationship between the encoder and decoder is: every time the flow field data passes through an encoder, the data feature size is reduced by half and the number of channels is doubled. Every time the flow field data passes through a decoder, the feature map size is doubled and the number of channels is halved; the flow field data enters the encoder and is compressed layer by layer to the bottleneck size. The decoder starts from the bottleneck size and fuses the encoder features layer by layer through jump connections to gradually restore the original size; finally, the data enters the reconstruction layer and outputs 3-channel 250*200 size data, where the 3 channels represent 3 different physical quantities.

4. The high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net according to claim 1 is characterized in that: In step 2, the continuity equation in the NS equation is defined as the loss function of the physical information U-Net flow field prediction model as follows: the continuity equation in the NS equation is used to extract spatial gradient and time gradient features, and discretized using central difference and forward difference methods respectively; the discrete expression is combined with the mean square error to form a new loss function; the loss function is combined with the physical information U-Net flow field prediction model, so that the loss function is reversely propagated in the physical information U-Net flow field prediction model to optimize the weights of the physical information U-Net flow field prediction model.

5. The high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net according to claim 2 is characterized in that: The pre-processing of the matrix data includes locating bad values ​​in the matrix data, and using surrounding values ​​to calculate and replace them according to the continuity equation in the NS equation.

6. The high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net according to claim 2 is characterized in that: The normalization process specifically involves performing minimum-maximum normalization on all data.

7. A high-precision prediction system for hypersonic flow field density and velocity based on physical information constraint U-Net based on the method according to any one of claims 1 to 6, characterized in that: include: The data set acquisition and processing module is used to obtain the numerical distribution of instantaneous lateral and longitudinal velocities and densities in the wake flow field of the target at different Mach numbers; the density, lateral velocity and longitudinal velocity of the target wake flow field at two consecutive moments and the third moment are combined into a data set, and the data set is divided into a training set and a test set; A model construction and loss function definition module is used to construct a physical information U-Net flow field prediction model, wherein the physical information U-Net flow field prediction model includes four consecutive encoders, namely encoder 1, encoder 2, encoder 3, and encoder 4, and four consecutive decoders, namely decoder 1, decoder 2, decoder 3, and decoder 4, and a reconstruction layer, which outputs data to a desired size; the continuity equation in the NS equation is defined as the loss function of the physical information U-Net flow field prediction model; The model training module is used to train the physical information U-Net flow field prediction model in the hypersonic target wake flow field at different times using the training set. The training process is supervised by the defined loss function. During the training process, predictions are made on the test set data until the loss function of the training set and the root mean square error of the test set are reduced and converged, thereby obtaining the physical information U-Net flow field prediction model with trained weights. The data prediction module is used to use the known lateral velocity and longitudinal velocity of the target flow field at two consecutive moments and the density data of the target flow field as input data, use the physical information U-Net flow field prediction model with trained weights to predict the input data, and obtain the prediction results through recursive prediction.

8. A high-precision prediction device for hypersonic flow field density and velocity based on physical information constraint U-Net, characterized by: include: Memory: a computer-readable device storing a computer program for a high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net according to any one of claims 1 to 6; Processor: used to implement the high-precision prediction method for hypersonic flow field density and velocity based on physical information constraint U-Net as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, can implement a high-precision prediction method for hypersonic flow field density and velocity based on physical information constrained U-Net as described in any one of claims 1 to 6.

Images