Multimodal aircraft flow field solution inference method based on multilayer feature enhancement

Abstract

The invention belongs to the field of flow field solution inference, and particularly relates to a multi-modal aircraft flow field solution inference method based on multilayer feature enhancement. The method comprises the steps that flow field numerical solutions of the aircraft under different incoming flow conditions are obtained through simulation; performing modal conversion on the aircraft flow field numerical solution, representing an incoming flow condition as a numerical modal, and representing an aircraft profile as an image modal; a multi-modal feature fusion model for multi-layer feature enhancement of aircraft flow field solution inference is constructed, and a simulated flow field numerical solution is used for model training. According to the method, the incoming flow condition and the aircraft appearance are represented as different modals for feature fusion, and the problem that the model performance is insufficient due to the fact that the capacity of capturing complex flow field features of the model is insufficient and the model fitting difficulty is increased is solved. By integrating the feature enhancement module, the model proposed by the scheme is better in generalization and higher in precision.

Description

Technical Field

[0001] The present invention relates to the field of flow field solution inference, and in particular to a multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement. Background Art

[0002] Flow field data provides detailed flow field characteristics. Through flow field analysis, the performance of aircraft under different flight conditions can be accurately evaluated. This information is of great significance for in-depth understanding of aircraft aerodynamic characteristics and optimization design. The incoming flow conditions determine the initial state and boundary conditions of fluid motion, while the shape information directly affects the interaction between the fluid and the surface of the object. These two types of information have their own unique physical meanings and mechanisms of action in the flow. However, existing methods usually roughly fuse this information into an image and encode it into a single network architecture with a convolution module as the core, and expect to directly output high-precision flow field prediction results. This encoding scheme makes it difficult for the model to distinguish and process the inherent differences between these two types of information. Therefore, the model's ability to capture complex flow field characteristics decreases, the difficulty of model fitting increases, and the effectiveness of the model is reduced. At the same time, it also makes the model overly dependent on a specific combination of information in the training data, lacks sufficient adaptability to new incoming flow conditions and shape combinations, and thus limits the generalization ability of the model.

[0003] Different from the existing flow field prediction methods, the present invention represents the future flow conditions and the aircraft shape information as different modes, and uses the future flow generalization network and the shape feature compensation network to encode the different modal information, making up for the shortcomings of the method of inputting the rough fusion image into a single network structure and directly mapping out a high-precision flow field. At the same time, the feature enhancement module is integrated to solve the problem that the model needs to process a large number of flow field data points at one time, while in fact not all points in the flow field have the same importance, thus dispersing the model performance. Summary of the invention

[0004] In view of the problems of insufficient generalization and low accuracy of existing flow field solution inference methods, the present invention provides a multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement. The future flow conditions and aircraft shape are represented as different modes respectively, and a special network structure is used to encode the different modes, and a feature enhancement module is integrated.

[0005] A multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement, comprising:

[0006] The numerical solution of the flow field of the aircraft under different incoming flow conditions is obtained through simulation;

[0007] Perform modal conversion on the numerical solution of the flow field obtained by simulation, and express the future flow conditions and grid coordinates as numerical modes, and express the aircraft shape as image mode;

[0008] Construct a multi-modal feature fusion model with multi-layer feature enhancement for aircraft flow field solution inference;

[0009] Based on the flow field solution after modal conversion, the numerical mode is first encoded through the incoming flow generalization network to obtain the approximate flow field solution, and then the image mode is encoded through the shape feature compensation network to obtain the final flow field solution, so as to train the flow field solution inference model for aircraft with different shapes under different incoming flow conditions.

[0010] The simulated aircraft flow field numerical solution includes:

[0011] Determine the initial and boundary conditions for each operating condition;

[0012] A body-fitting unstructured grid is used which is encrypted on the aircraft surface, and the grid density is increased in the high gradient area according to the preliminary results of the numerical solution of the flow field;

[0013] Using computational fluid dynamics to simulate the solutions of the Reynolds-averaged Navier-Stokes equations under different operating conditions;

[0014] The known flow field values ​​at the unstructured grid nodes are linearly interpolated to obtain the flow field solution at the Cartesian grid points.

[0015] The modal conversion of the flow field numerical solution includes:

[0016] The incoming flow conditions include the Reynolds number and the angle of attack, which are expressed as the numerical modal form of the incoming flow velocity component by the following formula:

[0017]

[0018] v x =cos(a)×u

[0019] v y =-sin(a)×u

[0020] v=(v x ,v y )

[0021] In the formula, Re is the Reynolds number, ρ is the density, u is the characteristic velocity, L is the characteristic length of the fluid flow, μ is the dynamic viscosity of the fluid, v is the incoming flow velocity, and v x and v y They represent the incoming flow velocity components in the x-direction and the y-direction respectively, and a is the angle of attack;

[0022] For the aircraft shape, the corresponding signed distance field is calculated to represent it in the form of image modality.

[0023] Multi-layer feature enhancement is applied to each layer of the shape feature compensation network, including:

[0024] The multi-layer feature enhancement module extracts the features of the image modality F through the following formula:

[0025] M c =(σ(MLP(MaxPool(F))+MLP(AvgPool(F)))+b)*F

[0026] M s =(σ(CONV([MaxPool(M c ); AvgPool(M c )]))+b)*M c

[0027] Where σ represents the nonlinear activation function, b represents the bias term, and M c represents the result of the channel attention mechanism, M s Represents the result of the spatial attention mechanism, MLP represents a multi-layer perceptron, CONV represents convolution, MaxPool represents maximum pooling, and AvgPool represents average pooling.

[0028] The multi-modal feature fusion model for multi-layer feature enhancement for aircraft flow field solution inference includes:

[0029] The multi-modal feature fusion model with multi-layer feature enhancement consists of a flow generalization network and a shape feature compensation network;

[0030] The incoming velocity component v of the numerical mode is x ,v y ) and the grid coordinate z=(z x ,z y ) input to generalize the network and obtain an approximate flow field solution;

[0031] Reconstructing the approximate flow field solution into a matrix form, and cascading it with the signed distance field corresponding to the aircraft shape in the channel dimension to obtain a feature map containing the approximate flow field solution and the aircraft shape information;

[0032] The feature map containing approximate flow field information and aircraft shape information is input as an image modality into the shape feature compensation network for flow field reconstruction, thereby obtaining the final flow field result.

[0033] Based on the final flow field solution output by the multi-modal feature fusion model with multi-layer feature enhancement and the results in the flow field data set, the mean square error is calculated to optimize the parameters of the above model until a model that meets expectations is obtained.

[0034] The beneficial effects of the present invention are as follows: the present invention proposes a multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement, which solves the problem of insufficient generalization and low precision caused by directly mapping a rough fusion image to a flow field solution using a single network. At the same time, the present invention simplifies the mapping relationship from a specific type of input to the flow field output, so that each part of the network can more easily learn its corresponding mapping function, thereby improving the stability and efficiency of training. The present invention also reduces the difficulty of fitting the overall model, and each sub-network can be independently tuned, reducing the risk of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the implementation manner will be introduced below.

[0036] Figure 1 It is a flow chart of a multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement in an embodiment of the present invention;

[0037] Figure 2 It is a schematic diagram of a framework of a multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement in an embodiment of the present invention;

[0038] Figure 3 It is a schematic diagram of the execution of a feature enhancement module in an embodiment of the present invention;

[0039] Figure 4 The pressure field prediction result provided by the embodiment of the present invention;

[0040] Figure 5 The x-direction velocity field prediction result provided by the embodiment of the present invention;

[0041] Figure 6 This is the y-direction velocity field prediction result provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the technical solution of the present invention clearer, the present invention will be clearly and completely explained below in conjunction with the drawings in the embodiments of the invention. The described embodiments only use a part of the embodiments, rather than all the embodiments.

[0043] like Figure 1 As shown, the present invention provides a multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement, and the present invention is specifically implemented according to the following steps:

[0044] Computational fluid dynamics software is used to simulate the flow field numerical solutions under different aircraft shapes and incoming flow conditions to form a flow field data set; wherein the data set contains the Reynolds number range and attack angle range of aircraft of various shapes, and each data corresponds to a set of Reynolds numbers and attack angles of an aircraft, as well as the flow field numerical solution labels of all grid points of the aircraft under the set of Reynolds numbers and attack angles;

[0045] Preprocess the flow field data set, represent the future flow conditions and grid coordinates as numerical modes, and represent the aircraft shape as image modes;

[0046] Construct a multi-modal feature fusion model with multi-layer feature enhancement for flow field solution inference; wherein the model of the present invention is composed of a flow generalization network and a shape feature compensation network, and the complete model framework is as follows Figure 2 As shown;

[0047] Based on the flow field dataset after modal conversion, the incoming flow conditions and grid coordinates of the numerical mode are used as the input of the incoming flow generalization network. The approximate flow field obtained by the incoming flow generalization network and the aircraft shape of the image mode are used as the input of the shape feature compensation network to train the flow field solution inference model for aircraft with different shapes under different incoming flow conditions.

[0048] The specific implementation method of each step in this embodiment is described below.

[0049] To construct flow field datasets under different aircraft shapes and incoming flow conditions:

[0050] In this example, two-dimensional incompressible, viscous, steady-state flow data and problems are used to train and test the performance of the model in inferring the Reynolds-averaged Navier-Stokes flow field solution. The problem description formula is as follows:

[0051]

[0052]

[0053]

[0054] In the formula represents the time-averaged velocity component of the fluid in the x direction, represents the time-averaged velocity component of the fluid in the y direction, represents the time-averaged pressure of the fluid, ρ represents the fluid density, ν represents the kinematic viscosity of the fluid, and represents the Reynolds stress;

[0055] Based on the above problems, the incompressible flow with Reynolds number range of Re = [5e5, 5e6] and angle of attack range of a = [-22.5°, 22.5°] was considered. The pressure field and velocity field of the two-dimensional flow around the airfoil were calculated using OpenFOAM to obtain the flow field data of each aircraft under different incoming flow conditions;

[0056] In this embodiment, in order to make the obtained flow field data format conform to the input specification of the convolution module, each flow field data is mapped to a (128×128) Cartesian grid.

[0057] Perform mode conversion on the flow field dataset:

[0058] The incoming flow conditions in this embodiment include the Reynolds number and the angle of attack, which are expressed as the numerical modal form of the incoming flow velocity component by the following formula:

[0059]

[0060] v x =cos(a)×u

[0061] v y =-sin(a)×u

[0062] v=(v x ,v y )

[0063] In the formula, Re is the Reynolds number, ρ is the density, u is the characteristic velocity, L is the characteristic length of the fluid flow, μ is the dynamic viscosity of the fluid, v is the incoming flow velocity, and v x and v y They represent the incoming flow velocity components in the x-direction and the y-direction respectively, and a is the angle of attack;

[0064] For the aircraft shape, the corresponding signed distance field is calculated to represent it in the form of image modality.

[0065] Aiming at building a multi-modal feature fusion model with multi-layer feature enhancement for flow field solution inference:

[0066] The multi-modal aircraft flow field solution inference model based on multi-layer feature enhancement in this embodiment is composed of an incoming flow generalization network and an appearance feature compensation network. The incoming flow generalization network is composed of a trunk network and a branch network. The branch network is responsible for processing the numerical modal incoming flow velocity component v = (v x ,v y ) input and output a feature vector [b1,b2,…,b p ] T The backbone network processes spatial variables, and the input is the numerical modal grid coordinate point z = (z x ,zy ), and output another feature vector [t1,t2,…,t p ] T , which is used to process the local dependency of points in the output space. The feature vectors output by these two parts of the network are combined together by dot product, and then a bias b0 is added at the end to obtain the approximate flow field G0, as shown in the following formula:

[0067]

[0068] The shape feature compensation network has a deep encoder-decoder structure and jump connections, which enables the model to capture features at different scales and retain spatial information. In addition, a feature enhancement module is added to each jump connection, such as Figure 3 As shown, the feature enhancement module concentrates the learning focus of the model by sequentially integrating channel attention and spatial attention, thereby improving the performance of the model;

[0069] In this embodiment, the approximate flow field obtained by the future flow generalization network is reconstructed into a (128×128) matrix form, and the signed distance field corresponding to the aircraft shape is cascaded in the channel dimension to obtain a feature map containing approximate flow field information and aircraft shape information, which is input into the shape feature compensation network as an image modality, thereby outputting the final flow field solution. Strided convolution is performed in the encoder part to gradually downsample, so that the network can extract more and more information from the growing number of feature channels. The decoder part increases the spatial resolution while gradually reducing the number of feature channels. The corresponding parts of each layer of the encoder and decoder are connected through jump connections, so that the number of feature channels is doubled, which helps the network to consider more information when reconstructing the decoder part. In addition, a feature enhancement module is added to each jump connection to optimize the model's ability to capture key flow features, thereby improving the accuracy of flow field solution inference;

[0070] Based on the final flow field solution and the results of the flow field data set output by the multi-modal aircraft flow field solution inference model based on multi-layer feature enhancement in this embodiment, the mean square error is calculated to optimize the parameters of the above model until a model that meets expectations is obtained.

[0071] In order to verify the effectiveness of the method of the embodiment of the present invention, the multi-modal aircraft flow field solution inference model based on multi-layer feature enhancement of the present invention is used to predict the flow field for working conditions outside the training range.

[0072] In this embodiment, the extrapolation prediction ability of the model is considered when the Reynolds number and the incoming flow velocity are significantly increased compared to the training range. The airfoil used is MH49 (not involved in training), and the incoming flow velocity components are (173.61, 29.88) m / s. During the model training process, AdamW is selected as the optimizer, the exponential decay rate is (0.5, 0.999), the learning rate is 0.0006, the loss function is MSE, and the number of iterations is 3000. The multi-modal aircraft flow field solution inference model based on multi-layer feature enhancement proposed in the present invention predicts the pressure field under the above working conditions as shown below. Figure 4 As shown, the prediction results of the velocity field in the x direction are as follows Figure 5 As shown, the prediction results of the velocity field in the y direction are as follows Figure 6 This embodiment verifies that the present invention still maintains a high prediction accuracy for working conditions outside the training range. Compared with the prior art, the present invention has better generalization and prediction accuracy.

[0073] The embodiments of the present invention have been disclosed above for the purpose of helping professionals in the relevant field to understand and implement the present invention. The present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement, characterized in that: The steps include: The numerical solution of the flow field of the aircraft under different incoming flow conditions is obtained through simulation; Performing modal conversion on the numerical solution of the aircraft flow field; Construct a multi-modal feature fusion model with multi-layer feature enhancement for aircraft flow field solution inference; Based on the flow field solution after modal conversion, the numerical mode is first encoded through the flow generalization network to obtain the approximate flow field solution, and then the image mode is encoded through the shape feature compensation network to obtain the final flow field solution.

2. The multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement according to claim 1 is characterized in that: The simulated aircraft flow field numerical solution specifically includes: Determine the initial and boundary conditions for each operating condition; A body-fitting unstructured grid is used which is encrypted on the aircraft surface, and the grid density is increased in the high gradient area according to the preliminary results of the numerical solution of the flow field; Using computational fluid dynamics to simulate the solutions of the Reynolds-averaged Navier-Stokes equations under different operating conditions; The known flow field values ​​at the unstructured grid nodes are linearly interpolated to obtain the flow field solution at the Cartesian grid points.

3. The multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement according to claim 1 is characterized in that: Performing modal conversion on the flow field numerical solution specifically includes: The incoming flow condition is expressed as the numerical mode of the incoming flow velocity component by the following formula: V x =cos(a)×u v y =-sin(a)×u v=(v x ,v y ) In the formula, Re is the Reynolds number, ρ is the density, u is the characteristic velocity, L is the characteristic length of the fluid flow, μ is the dynamic viscosity of the fluid, v is the incoming flow velocity, and v x and v y They represent the incoming flow velocity components in the x-direction and the y-direction respectively, and a is the angle of attack; For the aircraft shape, its corresponding signed distance field is calculated to represent it as an image modality.

4. The multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement according to claim 1 is characterized in that: The multi-layer feature enhancement is applied to the shape feature compensation network, specifically including: The multi-layer feature enhancement is applied to each layer in the shape feature compensation network, specifically including: Get the channel weight map and apply it to the corresponding image modality; A spatial weight map is obtained and applied to the channel-weighted image modality; The enhanced image modality is passed to the next layer of the network until the final flow field solution is output.

5. The multi-modal aircraft flow field solution inference method based on multi-layer feature enhancement according to claim 1 is characterized in that: Construct a multi-modal feature fusion model with multi-layer feature enhancement for aircraft flow field solution inference, including: The multi-modal feature fusion model for multi-layer feature enhancement is composed of a flow generalization network and a shape feature compensation network; The incoming velocity component v of the numerical mode is x ,v y ) and the grid coordinate z=(z x ,z y ) input to generalize the network and obtain an approximate flow field solution; Reconstructing the approximate flow field solution into a matrix form, and cascading it with the signed distance field corresponding to the aircraft shape in the channel dimension to obtain an image mode containing the approximate flow field solution and the aircraft shape information; The image mode containing the approximate flow field solution and the aircraft shape information is input into the shape feature compensation network to reconstruct the flow field, thereby obtaining the final flow field solution.

Images