Autonomous generation method for aerodynamic shape skin grid of aircraft

Abstract

The invention discloses an aircraft aerodynamic shape body-fitted grid autonomous generation method, which comprises the following steps of: 1, acquiring body-fitted grid data for forming an aircraft three-dimensional aerodynamic shape body-fitted grid, wherein the body-fitted grid data comprises an aircraft aerodynamic shape input control parameter set and an output three-dimensional coordinate data set of each node of the body-fitted grid; 2, preprocessing the body-fitted grid data set to obtain body-fitted grid features; and 3, inputting the body-fitted grid characteristics into a grid generation model to obtain an aircraft body-fitted grid. According to the method, the deep learning model is introduced to autonomously generate the grids, so that the time cost of manually dividing the grids by researchers is greatly reduced, and the improvement of the iterative optimization speed of aircraft design is facilitated; according to the method, the three-dimensional body-fitted grid of the aircraft is reconstructed in a parameter driving mode, the complexity of grid division in the CFD simulation field of the aircraft is greatly reduced, a large amount of repeated work of researchers is saved, and the simulation iteration period of aerodynamic design of the aircraft is shortened.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer aided design (CAD) and computational fluid dynamics (CFD), and particularly relates to a method for autonomously generating a body-fitted grid of an aerodynamic shape of an aircraft. BACKGROUND

[0002] In CFD simulation, the quality of a body-fitted grid directly affects the calculation accuracy and efficiency. Traditional body-fitted grid generation relies on parameterized modeling, solving of elliptic partial differential equations or algebraic interpolation methods, and has problems such as complex calculation, long time consumption, and the need for manual adjustment. In recent years, intelligent grid generation technology has gradually emerged, such as adaptive grid methods based on gradient optimization or genetic algorithms, but it still faces challenges such as difficulty in optimizing high-dimensional parameter space and difficulty in handling complex geometric topology.

[0003] Deep learning technology has shown great potential in image generation, point cloud processing and other fields. Some studies have attempted to use convolutional neural networks (CNN) or graph neural networks (GNN) for two-dimensional grid generation, but there are still problems in end-to-end prediction of three-dimensional body-fitted grids.

[0004] (1) The three-dimensional geometric topology is complex, and the model needs to consider global features and local details;

[0005] (2) The continuity and smoothness of the grid point coordinates are difficult to directly guarantee by a conventional network;

[0006] (3) The generated grid needs to satisfy the geometric constraint conditions of Delaunay triangulation. SUMMARY

[0007] The application provides a method for autonomously generating a body-fitted grid of an aerodynamic shape of an aircraft, which is a method for autonomously generating an aircraft body-fitted grid based on parameter-driven deep learning. The method uses a parameter-driven approach to reconstruct a three-dimensional body-fitted grid of an aircraft, significantly reducing the complexity of grid division in the field of aircraft CFD simulation, saving a large amount of repetitive work for researchers, shortening the simulation iteration cycle of aircraft aerodynamic design, and having high precision and orthogonality of traditional grid division methods.

[0008] To achieve the above purpose, the application provides the following scheme:

[0009] A method for autonomously generating a body-fitted grid of an aerodynamic shape of an aircraft, comprising the following steps:

[0010] Step S1, obtaining body-fitted grid data for forming a three-dimensional body-fitted grid of an aircraft: including an input control parameter set of an aerodynamic shape of the aircraft and an output three-dimensional coordinate data set of each node of the body-fitted grid;

[0011] Step S2, preprocessing the skin mesh dataset to obtain skin mesh features;

[0012] Step S3, inputting the skin mesh features into a mesh generation model to obtain the aircraft skin mesh;

[0013] The mesh generation model comprises a mesh initialization model, a post-processing module and a mesh reconstructor connected in sequence;

[0014] The mesh initialization model is used to generate a mesh tensor;

[0015] The post-processing module is used to reverse normalize the mesh tensor to obtain a mesh coordinate point set, and to adopt Delaunay triangulation on the mesh coordinate point set to obtain face index information;

[0016] The mesh reconstructor receives the mesh coordinate point set and the face index information, and outputs the aircraft skin mesh according to the mesh coordinate point set and the face index information;

[0017] Wherein:

[0018] The mesh initialization model is a mesh initialization model trained according to skin mesh features of a known three-dimensional aerodynamic shape skin mesh of an aircraft.

[0019] The scheme further comprises: the preprocessing of the skin mesh data is: using a maximum and minimum normalizer to respectively fit an input control parameter set and an output three-dimensional coordinate data set to obtain input normalized data and output normalized data as skin mesh feature data sets, and the normalization calculation formula is:

[0020]

[0021] Wherein: and are the maximum and minimum values of a single feature, and for the input, the maximum and minimum values of each control parameter; for the output, the maximum and minimum values of the coordinates in the three-dimensional coordinate space, and the single feature is the skin mesh feature.

[0022] The scheme further comprises: the reverse normalization formula is:

[0023] Wherein: and are the maximum and minimum values of a single feature;

[0024] In the process of restoring the output coordinate points, abnormal point filtering is performed: the local density of all mesh points based on the near neighbor distance is calculated, the outlier points with a density lower than a threshold value are removed, and the missing points are supplemented by interpolation.

[0025] The scheme is further: the Delaunay triangulation is based on the coordinates after the inverse normalization, and the triangulation mesh satisfying the empty circumscribed sphere criterion is generated: the Bowyer-Watson algorithm is used to insert the mesh point by point, and then the face index information is obtained, and after the boundary constraint in the mesh reconstructor, the aircraft skin mesh is obtained.

[0026] The scheme is further: the method further comprises: after obtaining the aircraft skin mesh, using Laplace smoothing for mesh updating, and the formula is:

[0027]

[0028] Wherein:

[0029] The new value of the embedding vector of node i after updating is represented;

[0030] The original value of the embedding vector of node i is represented;

[0031] The learning rate is represented, which is used to control the amplitude of each update, =0.2;

[0032] The set of neighbor nodes of node i is represented, and j is a node in the set;

[0033] The embedding vector of node j is represented;

[0034] The size of the neighbor node set of node i is represented, and the number of neighbor nodes.

[0035] The scheme is further: the training process of the initial mesh model is: the skin mesh feature data set of the known aircraft three-dimensional aerodynamic shape skin mesh is divided into training set and test set, the total loss of mesh generation is set, the mesh feature data of the training set and the test set are repeatedly input, and after the total loss of mesh generation is less than the set value, the initial mesh model is formed.

[0036] The scheme is further: the total loss of mesh generation is composed of coordinate regression error and geometric smoothness constraint, and the calculation formula is:

[0037]

[0038]

[0039]

[0040] Wherein:

[0041] L is the total loss;

[0042] l mse is the mean square error;

[0043] l curvature is the curvature regularization term, which is based on the average curvature of the grid edge calculated based on the predicted coordinates, and inhibits surface distortion;

[0044] and is the normal vector of the two endpoints;

[0045] and is the endpoint coordinate;

[0046] is set to 0.1;

[0047] N represents the total number of generated grid nodes;

[0048] x i is the grid node coordinate value;

[0049] represents the model predicted grid node coordinate value.

[0050] The scheme further comprises: the number of samples of the known aircraft three-dimensional aerodynamic body grid is at least 2000; the total loss of grid generation is 1%, and if the total loss of grid generation cannot meet the requirements, the training parameters and model parameters are optimized, and the model is trained multiple times until the model accuracy meets the requirements.

[0051] The beneficial effects of the present application are: the present application introduces a deep learning model to autonomously generate a grid, which greatly reduces the time cost of manually dividing the grid by researchers, and helps to improve the iteration optimization speed of aircraft design; the parameter-driven method is used to realize the reconstruction of the three-dimensional body grid of the aircraft, which greatly reduces the complexity of the grid division in the field of aircraft CFD simulation, saves a lot of repetitive work of researchers, shortens the simulation iteration cycle of aircraft aerodynamic design, and has high precision and orthogonality of the traditional grid division method.

[0052] The application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flowchart of the method of the present application;

[0054] Figure 2 is a schematic diagram of model training;

[0055] Figure 3 is a comparison chart of grid generation effect.​ DETAILED DESCRIPTION

[0056] An aircraft aerodynamic shape body-fitted mesh autonomous generation method is a parameter-driven deep learning-based aircraft body-fitted mesh autonomous generation method, as shown in FIG. 1, the method comprises the following steps: Figure 1

[0057] Step S1, obtaining body-fitted mesh data for constructing an aircraft three-dimensional aerodynamic shape body-fitted mesh : including an input control parameter set of an aircraft aerodynamic shape and an output three-dimensional coordinate data set of each node of the body-fitted mesh;

[0058] Step S2, pre-processing the body-fitted mesh data set to obtain body-fitted mesh features;

[0059] Step S3, inputting the body-fitted mesh features into a mesh generation model to obtain an aircraft body-fitted mesh;

[0060] The mesh generation model comprises a mesh initial model, a post-processing module and a mesh reconstructor connected in sequence;

[0061] The mesh initial model is used to generate a mesh tensor (the mesh tensor is a mesh data matrix);

[0062] The post-processing module is used to reverse the mesh tensor to obtain a mesh coordinate point set by de-normalization, and to obtain face index information by Delaunay triangulation on the mesh coordinate point set;

[0063] The mesh reconstructor receives the mesh coordinate point set and the face index information, and outputs the aircraft body-fitted mesh according to the mesh coordinate point set and the face index information; Figure 3 A comparison diagram showing the effect of the traditional method mesh generation left and the method generation right.

[0064] The mesh initial model is a mesh initial model trained according to the body-fitted mesh features of a known aircraft three-dimensional aerodynamic shape body-fitted mesh.

[0065] In the embodiment, the pre-processing of the body-fitted mesh data is: using a maximum and minimum normalizer to respectively fit the input control parameter set and the output three-dimensional coordinate data set, to obtain input normalized data and output normalized data as the body-fitted mesh feature data set, and the normalization calculation formula is:

[0066]

[0067] Wherein: And ​These are the maximum and minimum values ​​of a single feature, respectively. For the input, they are the maximum and minimum values ​​of each control parameter; for the output, they are the maximum and minimum values ​​of the coordinates in the three-dimensional coordinate space. The single feature is the body-fitted mesh feature. Wherein: for the input... There are several control parameters, denoted as a vector. By mapping each feature value to the interval [0, 1] using the above maximum-minimum normalization method, the problem of scale inconsistency between different features is solved; the output coordinate matrix is ​​then used. ( The total number of grid points is also normalized according to the above formula.

[0068] The inverse normalization formula is as follows:

[0069]

[0070] In the set of output coordinate points obtained by restoration, outlier filtering is performed during the coordinate restoration process: the local density of all grid points is calculated based on the nearest neighbor distance, outliers with a density below the threshold are removed, and missing points are supplemented by interpolation.

[0071] The Drone triangulation is based on inversely normalized coordinates, on which a triangular mesh satisfying the space circumsphere criterion is generated: the Bowyer-Watson algorithm is used to insert mesh points point by point to obtain surface index information, and the spacecraft body mesh is obtained after boundary constraints in the mesh reconstructor.

[0072] In this embodiment, the method further includes: mesh quality optimization to further improve the geometric quality of the mesh cells and meet the requirements of CFD simulation. After obtaining the body-fitted mesh of the aircraft, Laplace smoothing is first used to iteratively smooth the internal nodes and update the mesh. The formula is as follows:

[0073]

[0074] in:

[0075] This represents the new value of the embedding vector (or feature vector) of node i after the update;

[0076] Represents the original value of the embedding vector (or feature vector) of node i;

[0077] This represents the learning rate, used to control the magnitude of each update. =0.2;

[0078] Let j represent the set of neighboring nodes of node i, where j is a node in this set;

[0079] Represents the embedding vector of node j;

[0080] This represents the size of the set of neighboring nodes of node i, and the number of neighboring nodes;

[0081] After 10 iterations, using this formula, the node... Based on adjacent nodes ( belong neighborhood The system updates coordinates based on the positional relationships of the nodes, optimizes node distribution, then performs edge folding optimization, merging shorter sides with lengths less than a threshold (5% of the average side length) to reduce the number of narrow triangles; finally, it performs quality assessment and calculates the orthogonality of the mesh cells. Aspect Ratio ,make sure , Mesh that meets the requirements is used as the output mesh. If the requirements are not met, the total mesh generation loss needs to be readjusted, the initial mesh model needs to be retrained, and the mesh tensor needs to be generated using the retrained initial mesh model. The post-processing module is then executed: the mesh tensor is denormalized to restore the mesh coordinate point set, and the mesh coordinate point set is triangulated using De Röne triangulation to obtain the face index information. The mesh reconstructor receives the mesh coordinate point set and the face index information, and outputs the aircraft body-fitted mesh based on the mesh coordinate point set and the face index information.

[0082] In this embodiment, the training process for generating the initial grid model is as follows: Figure 2 As shown, a body-fitted mesh dataset of a standard aircraft is obtained. This dataset is then preprocessed using normalization to obtain a body-fitted mesh feature dataset. The body-fitted mesh feature dataset of a known three-dimensional aerodynamic shape of an aircraft is divided into a training set and a test set. Specifically, the input control parameters of the aerodynamic shape of a standard aircraft obtained through experimental testing, and the body-fitted mesh of the aircraft obtained using traditional methods are acquired. The output three-dimensional coordinate data of each node of the obtained aircraft body-fitted mesh are obtained. The input control parameters and output three-dimensional coordinate data are then preprocessed using normalization to form the three-dimensional body-fitted mesh dataset of the aircraft. ; is the obtained known 3D body mesh dataset of the aircraft. The input normalizer is obtained by fitting the input control parameters and the output three-dimensional coordinate data using a max-min normalizer, respectively. and output normalizer Then, the 3D body-fitted mesh dataset of the aircraft was processed. The body-fitted mesh feature dataset is obtained by normalization. Finally, the body-fitted mesh feature dataset is divided. For the training set and the test set , set the grid generation total loss , input the grid feature data of the training set and the test set repeatedly, and perform loss calculation, precision evaluation, and grid quality inspection, including training parameter optimization and model parameter optimization, to form the initial network model after the grid generation total loss is less than the set value ; the loss is the error, which can evaluate the model by calculating the error between the model prediction tensor and the corresponding data in the data set.

[0083] The model architecture of the grid initial model of the method can select transformer architecture or CNN architecture, and a deep learning model based on transformer architecture is constructed to train the non-linear mapping of control parameters to grid coordinates. The transformer architecture needs to first map the parameter group to a 512-dimensional hidden feature vector, generate a serialized "parameter token" through linear projection, input into a ViT module containing 3 Transformers Encoders, and extract global geometric features. Each layer includes multi-head self-attention mechanism (16 heads) and feedforward neural network (hidden layer dimension 2048), which together determine the parameter amount and expression ability of the model. Finally, it is restored to the grid dimension through the full connection layer. The CNN architecture needs to first add dimensions to the input parameter group to form pseudo-time sequence data, then extract features through one-dimensional convolution and expand the dimension to 256, and finally restore the grid dimension through the full connection layer.

[0084] The total loss function used in neural network training is, in an embodiment: the grid generation total loss is composed of coordinate regression error and geometric smoothness constraint, and the calculation formula is:

[0085]

[0086] The coordinate regression error uses mean square error (MSE), which describes the error size by calculating the square of the difference between the predicted grid point three-dimensional coordinates and the true coordinates:

[0087]

[0088] The curvature regular term calculates the average curvature of the grid based on the predicted coordinates to suppress surface distortion:

[0089]

[0090] Where:

[0091] L is the total loss;

[0092] lmse is the mean square error;

[0093] l curvature is the curvature regularization term, which is calculated based on the average curvature of the edges of the mesh, and is used to suppress surface distortion;

[0094] and is the normal vector of the edge between two endpoints;

[0095] and is the coordinate of the endpoint;

[0096] is set to 0.1;

[0097] N represents the total number of generated mesh nodes;

[0098] x i is the coordinate value of the mesh node;

[0099] represents the predicted coordinate value of the mesh node by the model.

[0100] Training strategy: the model training uses the Adam optimizer, the initial learning rate is 0.001, and is reduced to 0.5 times of the original value every 50 rounds; the batch size is set to 64, and the training rounds are 500 rounds.

[0101] wherein: the number of samples of the known aircraft three-dimensional aerodynamic shape body mesh feature data set is at least 2000; the total loss of mesh generation is 1%, if the total loss of mesh generation cannot meet the requirements, the training parameters and model parameters are optimized, and the model is trained multiple times until the model accuracy meets the requirements.

[0102] The method constructs a deep learning model based on the transformer architecture to train a nonlinear mapping of control parameters to mesh coordinates.

[0103] First, the model hyperparameters need to be set, including the model dimension of 512, the number of multi-head self-attention mechanism heads of 16, the number of encoder layers of 6, the number of decoder layers of 6, and the dimension of the feedforward network of 2048. These parameters jointly determine the parameter amount and expression ability of the model.

[0104] Then the input encoding unit part of the model is constructed, first using a linear layer to map the input features from 5 dimensions to 512 dimensions. The formula of this layer is wherein is the input feature vector, is the weight matrix, is the bias vector, is the output tensor of this layer. Then, the position encoding is added to the encoded input by a positional encoder using a sinusoidal encoding to capture the position information in the sequence elements. Positional encoding can help the model understand the relative position relationship of the elements in the sequence.

[0105] The part of the decoding unit has a learnable query vector with shape (46, 512), which is expanded according to the batch size of the input at each forward propagation. Similarly, the decoder input is also subjected to position encoding.

[0106] The transformer architecture described in the method is composed of an encoder and a decoder, which are stacked with 6 encoding units and 6 decoding units respectively. Each encoder layer contains a multi-head self-attention mechanism and a feedforward network. The multi-head self-attention mechanism allows the model to focus on different parts of the input sequence in different representation subspaces in parallel, and its formula is

[0107]

[0108] 、 and are the query, key and value matrices respectively; is a learnable weight matrix. In addition to containing a multi-head self-attention mechanism and a feedforward network, each decoding unit in the decoder also contains an encoder-decoder attention mechanism to focus on the output of the encoder.

[0109] Finally, the output layer of the model uses a linear layer to expand the output of the model from 512 dimensions to three times the number of grid points as the final prediction result.

[0110] In the method, the Delaunay triangulation process based on inverse normalized coordinates. First, construct a super-triangle containing all discrete points as the initial grid. When inserting a new point , traverse all triangles. If there exists a circumscribed circle containing , delete the triangle and collect its boundary to form a hole. Then complete the local triangulation by connecting the hole boundary points with . Then, for the geometric features of the aircraft, mark the contour edges as constraint edges, and use the constraint edge detection algorithm to ensure that these edges always exist as triangle edges during the triangulation process. Finally, in the incremental optimization stage, based on the Lawson discrimination method, all edges are evaluated. When there exist adjacent triangles and satisfy , perform edge flip operation to reduce the number of narrow triangles and improve the quality of the grid.

[0111] The aircraft aerodynamic shape body grid autonomous generation method is an aircraft body grid autonomous generation method based on parameter-driven deep learning. In training, first, the target aircraft shape control parameters are sampled, high-quality grids corresponding to the target aircraft shape control parameters are obtained respectively, and a data set is formed. Then, the data set is preprocessed, such as normalization, and a deep learning model is used for fitting to obtain a model that can autonomously predict three-dimensional body grid coordinates through the shape control parameters. In application, the output of the model is reversed and normalized, and then Delaunay triangulation is used to obtain an initial grid. Finally, edge constraint algorithm, Laplace smoothing algorithm and edge folding optimization algorithm are used to further optimize the grid quality, and finally the output body grid is obtained. The embodiment introduces a deep learning model to autonomously generate a grid, which greatly reduces the time cost of manually dividing the grid by researchers, and helps to improve the iteration and optimization speed of aircraft design.

Claims

1. An autonomous generation method of an aerodynamic shape conformal mesh for an aircraft, characterized in that, The method comprises the following steps: Step S1, obtaining a body-fitted mesh data for forming a body-fitted mesh of a three-dimensional aerodynamic shape of an aircraft: an input control parameter set of an aerodynamic shape of the aircraft and an output three-dimensional coordinate data set of each node of the body-fitted mesh; Step S2, preprocessing the body-fitted mesh data set to obtain body-fitted mesh features; Step S3, inputting the body-fitted mesh features into a mesh generation model to obtain the body-fitted mesh of the aircraft; The mesh generation model comprises a mesh initial model, a post-processing module and a mesh reconstructor connected in sequence; The mesh initial model is used to generate a mesh tensor; The post-processing module is used to restore the mesh tensor to obtain a mesh coordinate point set, and to obtain face index information by using Delaunay triangulation on the mesh coordinate point set; The mesh reconstructor receives the mesh coordinate point set and the face index information, and outputs the body-fitted mesh of the aircraft according to the mesh coordinate point set and the face index information; Wherein: The mesh initial model is a mesh initial model trained according to the body-fitted mesh feature data set of a known three-dimensional aerodynamic shape body-fitted mesh of an aircraft.

2. The method of claim 1, wherein, The preprocessing of the body-fitted mesh data is: using a maximum and minimum normalizer to fit the input control parameter set and the output three-dimensional coordinate data set respectively to obtain input normalized data and output normalized data as the body-fitted mesh feature data set, and the normalization calculation formula is: ; where: and are the maximum and minimum values of the single feature, respectively, and for the input, the maximum and minimum values of each control parameter, respectively; and for the output, the maximum and minimum values of the coordinates in the three-dimensional coordinate space, the single feature being the feature of the volume mesh.

3. The method of claim 1, wherein, The inverse normalization formula is: ； wherein: and are the maximum and minimum values of the single feature, respectively; In the process of restoring the output coordinate points, abnormal point filtering is performed: the local density of all mesh points based on the near neighbor distance is calculated, the outlier points with a density lower than a threshold value are removed, and the missing points are supplemented by interpolation.

4. The method of claim 1, wherein, The Delaunay triangulation is based on the coordinates after inverse normalization, and a triangular mesh satisfying the empty circumscribed sphere criterion is generated: the Bowyer-Watson algorithm is used to insert mesh points point by point, and then the face index information is obtained, and the body-fitted mesh of the aircraft is obtained after boundary constraint in the mesh reconstructor.

5. The method of claim 1, wherein, The method further comprises: after obtaining the body-fitted mesh of the aircraft, using Laplace smoothing to update the mesh, and the formula is: ； Wherein: denotes the updated new value of the embedding vector of node i; original value of the embedding vector representing node i; denotes the learning rate, which controls the magnitude of each update, = 0.2; N(i) denotes the set of neighbor nodes of node i, j is a node in this set; represents the embedding vector of node j; denotes the size of the set of neighbor nodes of node i, and the number of neighbor nodes.

6. The method of claim 1, wherein, The training process of the mesh initial model is: dividing the body-fitted mesh feature data set of a known three-dimensional aerodynamic shape body-fitted mesh of an aircraft into a training set and a test set, setting a total mesh generation loss, repeatedly inputting the mesh feature data of the training set and the test set, and forming the mesh initial model after the total mesh generation loss is less than a set value.

7. The method of claim 6, wherein, The total mesh generation loss is composed of coordinate regression error and geometric smoothness constraint, and the calculation formula is: ； ； ； Wherein: L is the total loss; MSE is the mean square error; is the curvature regularization term, which is based on the average curvature of the mesh edges calculated from the predicted coordinates, to suppress surface distortion; and is an edge normal vector of the two endpoints; and are the end point coordinates; is set to 0.1 ; represents the total number of grid nodes generated; is the grid node coordinate value; represent the model prediction grid node coordinate values.

8. The method of claim 6, wherein, The sample number of the body-fitted mesh feature data set of the known three-dimensional aerodynamic shape body-fitted mesh of the aircraft is at least 2000; the total mesh generation loss is 1%, if the total mesh generation loss cannot meet the requirements, the training parameters and model parameters are optimized, and the model is trained multiple times until the model accuracy meets the requirements.

Images