Aircraft structure strength prediction method and system based on multi-field coupled graph neural network
By using a multi-field coupled graph neural network approach, we can directly process the complex geometry of aircraft and construct a multi-physics perceptual graph neural network model. This solves the problems of low efficiency in traditional finite element analysis and the difficulty of machine learning models in handling multi-physics coupled conditions, and enables efficient and accurate prediction of aircraft structural strength.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA AERO POLYTECH ESTAB
- Filing Date
- 2025-12-24
- Publication Date
- 2026-04-24
AI Technical Summary
Existing technologies are insufficient for efficiently and accurately predicting the strength of aircraft structures under multi-physics coupling conditions. Traditional finite element analysis is inefficient, and machine learning models are difficult to uniformly handle complex multi-physics coupling conditions.
By employing a multi-field coupled graph neural network approach, a multi-physics perceptual graph neural network model is constructed through two-level normalization processing of complex aircraft-scale geometry. This model integrates multiple operating parameters from various sources, including aerodynamics, thermal, and mechanical parameters, to autonomously perceive the differences in the importance of structural connections and achieve second-level prediction of aircraft structural stress fields and strength indicators.
It significantly improves the engineering practicality and accuracy of aircraft structure prediction, directly handles complex geometry, avoids errors, autonomously senses load transfer and interaction, is applicable to aircraft with different configurations, and the frame can be expanded to incorporate more physical field parameters.
Smart Images
Figure CN121920197A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer-aided design technology for aircraft structural strength, specifically to a method and system for predicting aircraft structural strength based on a multi-field coupled graph neural network. Background Technology
[0002] Aircraft strength design is a complex process involving strong coupling of multiple disciplines such as aerodynamics, structure, thermal, and load. During the strength assessment phase, engineers need to verify the structural strength of the aircraft under millions of possible operating conditions. Traditional methods mainly rely on finite element analysis, which, although highly accurate, often takes hours to days for a single complete analysis, limiting the efficiency of design iteration and optimization.
[0003] In recent years, machine learning, especially graph neural networks (GNNs), has provided new ideas for accelerating engineering simulation. GNNs excel at handling non-Euclidean data such as aircraft finite element meshes and have shown potential in fields such as aerodynamic field prediction. However, existing research mostly focuses on single physical fields, such as pure aerodynamics or structures. In actual flight, aircraft structures are subjected to an extremely complex combined environment: high-altitude low temperatures, cabin pressurization, aerodynamic heating, and maneuvering overloads all work together. For example, during cruise, the fuselage skin must withstand pressure differential loads of up to approximately 0.06 MPa, while the external temperature may drop to -55°C, while the cabin temperature is maintained at approximately 25°C. This huge temperature difference leads to thermal stress. In addition, the overloads generated by maneuvering flight couple with the above loads, forming the most severe load conditions. Currently, there is a lack of an intelligent method that can efficiently and accurately handle such strongly coupled multi-physics conditions and quickly predict the strength of aircraft structures.
[0004] This invention aims to overcome the limitations of traditional finite element analysis (FEM) methods in terms of computational efficiency and the difficulty of existing machine learning models in handling strength prediction of aircraft structures under multi-physics coupled conditions. It provides a method that integrates aerodynamic, thermal, and mechanical parameters from multiple sources and uses graph neural networks to achieve second-level prediction of aircraft structural stress fields and strength indicators, thereby significantly improving the efficiency of aircraft design, strength verification, and optimization processes. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention aims to provide an aircraft structural strength prediction method based on a multi-field coupled graph neural network. This method employs a two-level normalization approach to directly handle the complex geometry of aircraft scale, avoiding errors introduced by traditional simplification methods. Furthermore, it naturally captures load transfer and interactions between aircraft structural components through the information transmission mechanism of the graph neural network model. Multiple physical field conditions, such as ambient temperature, cabin temperature, pressure difference, aerodynamic loads, and inertial loads, are uniformly parameterized and embedded into the graph neural network model. This makes the prediction results closer to the real complex flight environment during aircraft missions, enabling the model to autonomously perceive and learn the differences in importance of different structural connections in load transfer. This achieves a deep integration of prior physical knowledge and data-driven learning, improving the model's accuracy in modeling force transmission paths in complex structures and significantly enhancing the engineering practicality and accuracy of aircraft structural prediction.
[0006] Specifically, on the one hand, the present invention provides a method for predicting the structural strength of an aircraft based on a multi-field coupled graph neural network, which includes the following steps: S1: Determine key parameters of the aircraft structure under multiphysics conditions, including ambient temperature. Aircraft cabin temperature Aircraft cabin pressure difference External pneumatic pressure Vertical overload factor fuselage shear force and fuselage bending moment , forming the working condition vector Perform parameterized simulation of aircraft structural loads to obtain a multiphysics finite element simulation stress dataset; S2: Construct a hierarchical graph of the aircraft structure's structural features; based on the operating condition vector output in step S1... The finite element analysis results were used to extract the node numbers, three-dimensional coordinates, and element connection relationships of the aircraft structure; the aircraft structure coordinates were normalized into the characteristic matrix of the aircraft structural component model. ; through the global normalized coordinates of two-level nodes and the encoding of aircraft structural component types The node feature vectors constitute the node feature matrix of the aircraft structural component model. ; Edge feature matrix of aircraft structural component model It consists of the edge feature vectors of the aircraft structural component model and the feature distances between the nodes of the aircraft structural component model; S3: Construct a multiphysics-aware graph neural network model for aircraft structural strength prediction, specifically including: using a multiphysics-aware graph neural network model encoder, fusing node features, and obtaining the initial state matrix of the nodes of the aircraft structural component model through linear transformation. A multi-physics sensing graph neural network model iterative processor is constructed to perform graph-structure weighted aggregation of node states and output a node state matrix of global aircraft structure information. The multiphysics perceptual graph neural network model decoder outputs node-predicted stress values, thus constructing the stress field of the aircraft structure. ; S4: Training and validating the multiphysics perceptual graph neural network model for predicting aircraft structural strength constructed in step S3, outputting the aircraft stress field, automatically obtaining the safety margin of key aircraft structural parts, and performing real-time assessment of aircraft structural strength.
[0007] Preferably, step S2 specifically includes: S21: Perform global centering and scaling of the aircraft structural model; obtain the global bounding box of the aircraft structural model; use the bounding box to obtain the global scale parameters of the aircraft structural model, and obtain the global normalized coordinates of the aircraft structural model nodes; S22: Perform local normalization processing at the aircraft structural component level; S23: Construct the node feature matrix of the aircraft structural component model to obtain the node feature matrix of the aircraft structural component model. Construct the edge feature matrix of the aircraft structural component model ; S24: Construct the global operating condition matrix for the aircraft structure .
[0008] Preferably, step S21 specifically includes: ; ; ; ; ; in, The coordinates are the global center coordinates of the aircraft structural model; The maximum value of the x-coordinate of the aircraft structural node; This represents the minimum value of the x-coordinate of an aircraft structural node. The maximum value of the ordinate of the aircraft structural node; This represents the minimum value of the ordinate of the aircraft structural node; The maximum value of the vertical coordinate of the aircraft structural node; For the minimum value of the vertical coordinate of the aircraft structural node; These are the global scale parameters for the aircraft structural model; To find the square function; For the first The globally normalized x-coordinate of each node; For the first The globally normalized ordinate of each node; For the first The globally normalized vertical coordinates of each node; For the first The initial x-coordinates of each node; For the first The initial y-coordinates of each node; For the first The initial vertical coordinates of each node; Number the nodes of the aircraft structural model.
[0009] Preferably, step S22 specifically includes: ; ; ; in, For the first The global quadratic normalized x-coordinate of each node; For the first The global quadratic normalized ordinate of each node; For the first The global quadratic normalized vertical coordinates of each node; The center of the node of the aircraft structural component; These are the global scale parameters for the nodes of aircraft structural components.
[0010] Preferably, the node feature matrix of the aircraft structural component model in step S23 The edge feature matrix of the aircraft structural component model Specifically: ; ; in, The first model of aircraft structural components Each node's feature matrix; Encoding the types of aircraft structural components; For the aircraft structural component model The node of the first The elements of the edge feature matrix; The abscissa feature vector of the aircraft structural component model; The ordinate edge feature vector of the aircraft structural component model; The vertical coordinate edge feature vector of the aircraft structural component model; The feature distance between nodes in the aircraft structural component model; These are numerically stable characteristic parameters.
[0011] Preferably, step S3 specifically includes: S31: Constructing a multiphysics-aware graph neural network model encoder: The global operating condition features are copied to each node, fused with the node features, and after linear transformation and activation, the initial state matrix of the nodes of the aircraft structural component model is obtained. ; S32: Construct an iterative processor for a multiphysics-aware graph neural network model; input the initial state matrix of the nodes in the aircraft structural component model. Each graph network layer of the iterative processor performs graph structure-weighted aggregation of node states and updates the states through parameterized nonlinear transformations, ultimately outputting a node state matrix containing global aircraft structure information. ; S33: Constructing a multi-physics perceptual graph neural network model decoder; analyzing the node state matrix of aircraft structural information. A linear transformation is performed to directly output the predicted stress value of each node, thus forming the stress field of the aircraft structure. .
[0012] Preferably, step S31 specifically includes: ; in, This is the initial state matrix of the nodes in the aircraft structural component model; This is a copy function that replicates the global status to each node. It is a non-linear activation function; For splicing operations; The encoder weights of the graph neural network model; For encoder bias in graph neural network models; This is the global operating condition matrix for the aircraft structure.
[0013] Preferably, step S32 specifically includes: ; ; ; in, For the first The aggregated information matrix of a layered graph network layer; It is the inverse of the weighted degree matrix; This is a weighted adjacency matrix for aircraft structural components; The node state matrix for aircraft structural information; For the first The weight matrix of the layer processor; For the first The bias vector of the layer processor; For the processor layer index; This represents the total number of processor layers. Weighted adjacency matrix elements for aircraft structural components; It is an exponential function; It is a linear transformation function; is the linear rectified activation function; is the learning weight generation function.
[0014] Preferably, step S33 specifically includes: ; in, For the stress field of the aircraft structure; The weights for the stress field state of the aircraft structure; This is an offset of the stress field state of the aircraft structure.
[0015] On the other hand, the present invention proposes a prediction system for an aircraft structure strength prediction method based on a multi-field coupled graph neural network, which includes: a key multi-physics condition analysis module for aircraft structure, a hierarchical graph structure feature acquisition module for aircraft structure, a multi-physics perception graph neural network model construction module, and an aircraft structure strength prediction output module. The aircraft structure key multiphysics field working condition analysis module is used to determine the key parameters of the multiphysics field working condition of the aircraft structure, perform parameterized simulation of the aircraft structure load, and obtain the multiphysics field finite element simulation stress dataset. The hierarchical diagram structural feature acquisition module for aircraft structures analyzes the node numbers, three-dimensional coordinates, and element connection relationships of the aircraft structure based on the finite element analysis results of the working condition vectors; it then normalizes the aircraft structure coordinates into a feature matrix for the aircraft structural component model. The nodal feature matrix of the aircraft structural component model The edge feature matrix of the aircraft structural component model; The multi-physics perceptual graph neural network model building module is a graph neural network model that includes an encoder, processor and decoder architecture. The multi-physics perceptual graph neural network model takes the node feature matrix of the aircraft structural component model, the edge feature matrix of the aircraft structural component model and the global working condition matrix of the aircraft structure as input, and the aircraft structure output stress matrix as output, to realize the stress analysis of the aircraft structure. The aircraft structural strength prediction output module is used to train and validate the multiphysics perceptual graph neural network model, output the aircraft stress field, automatically obtain the safety margin of key aircraft structural parts, and realize the real-time assessment of aircraft structural strength.
[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention unifies and parameterizes various physical field conditions such as ambient temperature, aircraft cabin temperature, pressure difference, aerodynamic load, and inertial load and embeds them into a graph neural network model, making the prediction results closer to the real complex flight environment when the aircraft performs its mission, and significantly improving the engineering practicality and accuracy of aircraft structure prediction.
[0017] (2) This invention directly processes the complex geometry of aircraft scale, adopts a two-level normalization method, avoids the errors caused by traditional simplification methods, and naturally captures the load transfer and interaction between aircraft structural components through the information transmission mechanism of graph neural network model.
[0018] (3) This invention innovatively proposes a method for constructing a dynamic weighted adjacency matrix based on physical information. By using a learnable function, edge features, such as connection direction and distance, are dynamically decoded into adjacency weights, enabling the model to autonomously perceive and learn the differences in importance of different structural connections, such as riveting, welding, and bolting, in load transfer. This achieves a deep integration of physical prior knowledge and data-driven learning, improving the model's accuracy in modeling force transmission paths of complex structures.
[0019] (4) This invention does not depend on a specific aircraft model. It can be applied to aircraft with different configurations by changing the training data. In addition, the frame can be easily expanded to incorporate more physical field parameters, such as humidity and dynamic load spectrum. Attached Figure Description
[0020] Figure 1 The control block diagram is shown for an aircraft structural strength prediction method based on a multi-field coupled graph neural network. Figure 2 This is a diagram of the encoder-processor-decoder three-layer architecture of the graph neural network model in this embodiment of the invention; Figure 3 This is a scatter plot comparing the predicted stress and the finite element simulation stress under all working conditions in the embodiments of the present invention. Figure 4 This is a comparison chart of actual computation time on a logarithmic scale in an embodiment of the present invention; Figure 5 This embodiment of the invention presents a comparison chart of the speedup factors of relative computing speed; Figure 6 A continuous stress cloud diagram of the wing sparsity structure in an embodiment of the present invention; Figure 7 The stress cloud diagram is a finite element analysis of the wing sparsity structure in an embodiment of the present invention. Detailed Implementation
[0021] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.
[0022] This invention proposes a method for predicting the structural strength of aircraft based on a multi-field coupled graph neural network, such as... Figure 1 As shown, the key parameters of the aircraft structure under multiphysics operating conditions are determined, and load parameterized finite element simulation is performed; a hierarchical graph structural feature vector of the aircraft structure is constructed, and the aircraft structure coordinates are normalized into the feature matrix of the aircraft structural component model; a multiphysics-aware graph neural network model for predicting the strength of the aircraft structure is constructed; the multiphysics-aware graph neural network model is trained and validated to achieve real-time assessment of the strength of the aircraft structure; specifically, the following steps are included: Step S1: Perform parametric simulation of aircraft structural loads and build an aircraft structural database; set key parameters for multiphysics operating conditions of the aircraft structure; thermophysical field parameters include: ambient temperature. Aircraft cabin temperature The internal and external pressure fields include: the pressure difference within the aircraft cabin. External pneumatic pressure The overall structural load of the aircraft includes: vertical overload factor. fuselage shear force fuselage bending moment The working condition vector is composed of: ; in, This is the working condition vector; Ambient temperature, unit: ; Temperature inside the aircraft cabin, in units of ; The pressure difference in the aircraft cabin, in units of... ; External pneumatic pressure, unit: ; This is the vertical overload factor; Shear force of the fuselage, unit: ; The unit is the fuselage bending moment. .
[0023] Within the range of multiphysics operating parameters, p types of multiphysics operating conditions are generated parametrically. Multiple load conditions are applied to the aircraft structural model, and the corresponding stress fields are calculated through finite element analysis to construct a training sample library for the aircraft structure. In this embodiment of the invention, for a certain type of commercial aircraft, parameterized simulation and data generation of aircraft structural loads under multiphysics operating conditions are performed to determine the multiphysics operating scenario within its typical flight envelope. Through multiphysics finite element simulation, the stress analysis results of the aircraft structure under 5 operating conditions are obtained as shown in Table 1.
[0024] Table 1. Stress Results from Multiphysics Finite Element Simulation Step S2: Construct a hierarchical graph structure feature vector for the aircraft structure; normalize the aircraft structure coordinates into a feature matrix for the aircraft structural component model. Based on the working condition vector output in step S1 The finite element analysis results were used to extract the node numbers, three-dimensional coordinates, and element connection relationships of the aircraft structure. Node features were generated, and the global bounding box was calculated based on the coordinates of all nodes. Centering and scaling were then performed to obtain the globally normalized coordinates. Based on the structural components to which the aircraft structural nodes belong, such as the aircraft wings, fuselage, and tail, calculate the bounding boxes at the aircraft structural component level and perform secondary normalization to obtain locally normalized coordinates; then concatenate the two and combine them with the aircraft structural component type code to form the feature vector of the aircraft structural node.
[0025] The hierarchical diagram of the aircraft structure features the structural characteristic vectors. The node numbers, node coordinates, and three-dimensional connection relationships of the mesh are output from the finite element software as shown in the table. The total number of nodes in the aircraft structure model is N=51589. In this case, 10 consecutive aircraft structure nodes are selected from the wing front spars for display. Table 2 shows the initial coordinates, initial coordinates, initial coordinates, and detailed information of the components to which the 10 nodes belong.
[0026] Table 2 Statistical Table of Aircraft Structural Node Information The connection relationship between the aircraft structural nodes is that adjacent nodes are connected by aircraft structural beam units, forming a linear topology as shown in Table 3, which is a statistical table of the specific numbers of the starting and ending nodes corresponding to the 9 edge numbers.
[0027] Table 3. Statistical Table of Linear Topology of Adjacent Nodes Step S21: Perform global centering and scaling of the aircraft structural model; translate the aircraft structural model to the center and scale it to a standard size range. Obtain the global bounding box of the aircraft structural model; find the maximum and minimum values of all aircraft structural node coordinates in the X, Y, and Z directions: The global center coordinates of the aircraft structural model are obtained as follows: ; in, The coordinates are the global center coordinates of the aircraft structural model; The maximum value of the x-coordinate of the aircraft structural node; This represents the minimum value of the x-coordinate of an aircraft structural node. The maximum value of the ordinate of the aircraft structural node; This represents the minimum value of the ordinate of the aircraft structural node; The maximum value of the vertical coordinate of the aircraft structural node; This represents the minimum value of the vertical coordinate of the aircraft structural node.
[0028] The actual values of the coordinate range of the aircraft structural model obtained in this embodiment of the invention, in meters; X: [-15.20, 38.50], Y: [-15.80, 15.80], Z: [-4.80, 5.20]; the global center of the aircraft structural model is: ; ; ; in, The x-coordinate of the global center of the aircraft structural model; The vertical coordinate of the global center of the aircraft structural model; The vertical coordinates are the global center of the aircraft structural model.
[0029] The global scale parameters of the aircraft structural model are obtained using the diagonal length of the bounding box: ; in, These are the global scale parameters for the aircraft structural model; To find the square function.
[0030] The global scale parameters of the aircraft structural model obtained in this embodiment of the invention are as follows: .
[0031] For each node of the aircraft structural model The initial coordinates are transformed as follows to obtain the nodes of the aircraft structural model. The globally normalized coordinates are: ; ; ; in, For the first The globally normalized x-coordinate of each node; For the first The globally normalized ordinate of each node; For the first The globally normalized vertical coordinates of each node; For the first The initial x-coordinates of each node; For the first The initial y-coordinates of each node; For the first The initial vertical coordinates of each node; Number the nodes of the aircraft structural model.
[0032] The aircraft structural model of this invention is placed in a region centered on the origin with a size of approximately Within the virtual sphere, the globally normalized coordinates of nodes are calculated. Taking node 10001 as an example: ; ; ; Table 4 shows the global normalized coordinates of the aircraft structural model nodes in this embodiment of the invention, including a detailed statistical table of the global normalized coordinates of 10 nodes.
[0033] Table 4. Statistics of Globally Normalized Coordinates of Nodes in Aircraft Structural Model Step S22: Perform local normalization at the aircraft structural component level to make the aircraft structural model unaffected by absolute dimensions, which is more conducive to the neural network model learning component features. Divide the aircraft structural components, obtaining them from the finite element model element properties or region partitioning based on the component to which a node belongs, such as: wings, fuselage, tail, etc. Calculate the component's local bounding box: For each component, calculate the coordinate range of its internal nodes. Component-level normalization: For nodes belonging to component p, normalize their global coordinates. Based on this, a second normalization is performed using the bounding box dimensions of the component itself, specifically: ; ; ; in, For the first The global quadratic normalized x-coordinate of each node; For the first The global quadratic normalized ordinate of each node; For the first The global quadratic normalized vertical coordinates of each node; The center of the node of the aircraft structural component; These are the global scale parameters for the nodes of aircraft structural components.
[0034] The aircraft wing component enclosure, obtained from the aircraft structural model, is specifically: X: [8.50, 38.50], Y: [-15.80, 15.80], Z: [-1.20, 5.20] wing component center: ; ; ; in, The x-coordinate of the center of the node of the aircraft wing component; The ordinate of the center of the node of the aircraft wing component; The vertical coordinates of the center of the node of the aircraft wing component.
[0035] The coordinates of the aircraft wing component dimensions in the globally normalized coordinate system according to the embodiments of the present invention are as follows: ; in, The global normalized x-coordinate of the aircraft wing component node; The globally normalized ordinate of the aircraft wing component node; The globally normalized vertical coordinates of the aircraft wing component nodes; These are the global scale parameters for the nodes of the aircraft wing components.
[0036] This embodiment of the invention uses node 10001 of the aircraft structural model as an example to calculate the second-order normalization: ; ; ; in, The global quadratic normalized x-coordinate of the 10001st node; The global quadratic normalized ordinate of the 10001st node; This represents the global quadratic normalized vertical coordinate of the 10001st node.
[0037] conduct The component type coding of the aircraft structure, This represents the number of types of aircraft structural components. In the embodiment, if the components are set to include four types of components: [nose, fuselage, wing, and tail], then the component type codes for these four types of components are as follows: nose: [1,0,0,0], fuselage: [0,1,0,0], wing: [0,0,1,0], and tail: [0,0,0,1]. Component type code: [nose, fuselage, wing, tail] = [0, 0, 1, 0]. It can also be divided into more categories, such as dividing the tail into horizontal stabilizer and vertical stabilizer, and the wing into left wing and right wing.
[0038] Step S23: Construct the node feature matrix of the aircraft structural component model as follows: ; Where k is the aircraft structural component type code; The node feature matrix of the aircraft structural component model, with a size of N is the total number of nodes; The dimension of the node feature vector is 10 in this embodiment of the invention.
[0039] The node feature matrix v of the aircraft structural component model is composed of node feature vectors, which are: The node feature matrix v of the aircraft structural component model is 10×10, specifically: The edge feature matrix E of the aircraft structural component model is constructed as follows: ; in, The edge feature matrix of the aircraft structural component model, size ; For the aircraft structural component model The node of the first The elements of the edge feature matrix; The abscissa feature vector of the aircraft structural component model; The ordinate edge feature vector of the aircraft structural component model; The vertical coordinate edge feature vector of the aircraft structural component model; The feature distance between nodes in the aircraft structural component model; These are numerically stable characteristic parameters, typically... This is used to ensure the stability of numerical calculations and prevent the denominator from being zero; The total number of edges; Let be the initial dimension of the edge vector, which is 3 in this embodiment of the invention.
[0040] This embodiment of the invention takes edge 1 of an aircraft structural component model, connecting nodes 10001 and 10002, as an example to illustrate how the abscissa edge feature vector of the aircraft structural component model is obtained, specifically: ; ; ; Feature distance between model nodes of aircraft structural components in this embodiment of the invention for: ; The edge feature matrix of the aircraft structural component model in the embodiment is obtained. The vector is: ; The calculation results of all edge features of the aircraft structural component model obtained in this embodiment of the invention are shown in Table 5. The edge feature matrix E of the aircraft structural component model is obtained according to the starting node and ending node of each of the nine edges.
[0041] Table 5. Statistical table of calculation results for all edge features of the aircraft structural component model. Step S24: Construct the global operating condition matrix of the aircraft structure; The global operating condition matrix of the aircraft structure is composed of multiphysics operating conditions, with a size of [missing information]. , This represents the total number of simulation conditions. The multiphysics parameter dimension is 7 in this invention, representing the global operating condition matrix of the aircraft structure. It is 5×7 in size and is constructed based on 5 typical operating condition parameters, specifically: U=[-25, 24, 0.04, 18, 2.5, 1200, 5000], #Working condition 1: Limit pull-up maneuver; [-69, 24, 0.062, 26, 1.0, 400, 1800], #Operating Condition 2: High-Altitude Cruise; [-34, 24, 0.045, 38, 1.5, 500, 4000], #Condition 3: High-speed dive; [-15, 24, 0.035, 12, -1.0, -200, -1200], # Condition 4: Negative maneuver; [45, 28, 0, 1.6, 2.2, 600, 1500]# Condition 5: High-load landing.
[0042] Construct the output stress matrix of the aircraft structure. The output stress matrix has a size of Output the stress matrix from the finite element software. The size is 5×10. The von Mises stress values obtained by finite element analysis are shown in Table 6, with units of MPa. This is a statistical table of the output stress simulation results of 10 nodes of the aircraft structure under 5 working conditions.
[0043] Table 6 Statistical Table of Simulation Results of Aircraft Structural Output Stress Step S3: Construct a multiphysics-aware graph neural network model for aircraft structural strength prediction; construct a graph neural network model with an encoder-processor-decoder architecture; the multiphysics-aware graph neural network model adopts an encoder-processor-decoder architecture, and the specific construction process is as follows: like Figure 2 The diagram shown is a three-layer architecture diagram of the encoder-processor-decoder of the graph neural network model in this embodiment of the invention; the input matrix obtained according to step S2 is the node feature matrix of the aircraft structural component model. The edge feature matrix of the aircraft structural component model Aircraft structure global operating condition matrix Aircraft structural output stress matrix The graph neural network model uses the operating condition vectors output in step S1 to form the global operating condition matrix of the aircraft structure. Node feature matrix of aircraft structural component model The edge feature matrix of the aircraft structural component model For input, perform the following steps in sequence: Step S31: Construct a multiphysics perceptual graph neural network model encoder; copy the global operating condition features to each node, fuse them with the node features, and obtain the initial state matrix of the nodes of the aircraft structural component model through linear transformation and activation. for: ; in, This is the initial state matrix of the nodes in the aircraft structural component model; This is a copy function that replicates the global status to each node. It is a non-linear activation function, as shown in the example. function; For splicing operations; The encoder weights of the graph neural network model; This refers to the encoder bias of a graph neural network model.
[0044] Step S32: Construct a multiphysics-aware graph neural network model iterative processor; initialize the node state matrix of the aircraft structural component model. The input iterative processor contains L sequentially connected graph network layers, each layer according to the formula... The node states are weighted and aggregated using a graph structure, and then updated using a parameterized nonlinear transformation. The final output is a node state matrix containing global aircraft structural information. Specifically: ; ; ; in, This represents the aggregated information matrix of the i-th layer of the graph network. It is the inverse of the weighted degree matrix, used to normalize the aggregated information; The weighted adjacency matrix of the aircraft structural components is derived from the edge feature matrix of the aircraft structural component model. generate; For the first The weight matrix of the layer processor; For the first The bias vector of the layer processor; For the processor layer index; This represents the total number of processor layers.
[0045] The weighted adjacency matrix of aircraft structural components set in the embodiments of the present invention Specifically: ; in, Weighted adjacency matrix elements for aircraft structural components; It is an exponential function, which makes the generated weights positive; It is a linear transformation function. This function introduces nonlinearity after linear transformation, enabling the network to fit complex physical mapping relationships; is a linear rectified activation function, meaning that this operation enables a node to aggregate information from its neighboring nodes according to the physical weight edge features, which is equivalent to physical local stress balance; is a learning weight generation function; Input for the function; This is the weight matrix in the linear transformation function; It is the bias vector in the linear transformation function; This is a function that maximizes the value of a function.
[0046] Step S33: Construct a multi-physics perceptual graph neural network model decoder; process the node state matrix of aircraft structural information. A linear transformation is performed to directly output the predicted stress value of each node, thus forming the stress field of the aircraft structure. for: ; in, For the stress field of the aircraft structure; The node state matrix for aircraft structural information; The weights for the stress field state of the aircraft structure; This is an offset of the stress field state of the aircraft structure.
[0047] Based on the above inputs and formulas, a graph neural network model is constructed using mature software deep learning toolkits, such as MATLAB.
[0048] This invention constructs a graph neural network model, obtaining a three-layer graph neural network model and completing data training. The settings are as follows: node feature dimension: Fv = 10; edge feature dimension: Fe = 3; global condition dimension: Fu = 7; number of processor layers: L = 4; hidden layer dimension: 256; learning rate: 0.001; training epochs: 300; batch size: 4. Scientific computing tools are used for model construction and training. Training set: 16 conditions including extreme pull-up, high-altitude cruise, high-speed dive, negative maneuver, etc. Test set: 4 conditions, including high-load landing, etc.
[0049] Step S4: Train and validate the multiphysics-aware graph neural network model for predicting aircraft structural strength constructed in step S3; use the multiphysics finite element simulation stress dataset constructed in step S1, with the error between the predicted stress field and the simulation results as the loss function, such as mean square error, to supervise the training of the graph neural network model; after training, validate the model's accuracy and generalization ability on an independent test set.
[0050] Step S41: Construct a loss function for predicting aircraft structural strength. for: ; in, The loss function for predicting the structural strength of aircraft; For graph neural network models of aircraft structural nodes The predicted stress value; The nodes obtained through finite element simulation The actual stress value; This represents the total number of all aircraft structural nodes. This refers to the set of training parameters in a graph neural network model, including the weights and biases in the encoder, processor layers, and decoder. wait.
[0051] Step S42: Backpropagation optimizes the multiphysics sensing graph neural network model; using the chain rule, the loss function for predicting aircraft structural strength is used. To begin, the gradient of the loss with respect to the training parameters in each graph neural network model is calculated in reverse, specifically: ; gradient It specifies how each parameter should be fine-tuned, increased, or decreased to most effectively reduce the total error. This process starts from the decoder and traces back layer by layer to the encoder.
[0052] Step S43: Update the parameters of the multiphysics-aware graph neural network model; use an optimizer, such as the most commonly used Adam; update all parameters based on the calculated gradients; the specific update formula is as follows: ; in, Update the training parameters in the graph neural network model; These are the initial values for the training parameters in the graph neural network model. The learning rate parameter controls the step size for each update and is one of the most important hyperparameters during training. In the Adam optimizer, It is a general value that can be used as a starting point. Adam, short for Adaptive Moment Estimation, is a commonly used deep learning optimization algorithm that designs an adaptive learning rate for different parameters by calculating the first and second moment estimates of the gradient.
[0053] Step S44: Obtain the updated and optimized multiphysics-aware graph neural network model from step S43 for rapid prediction and evaluation of aircraft structural strength: For any new set of multiphysics operating parameters, firstly, its operating vector F is formed, and then a graph structure is constructed together with the parameterized mesh of the aircraft to be evaluated, which is then input into the trained multiphysics-aware graph neural network model; the multiphysics-aware graph neural network model completes forward propagation within seconds and outputs the aircraft stress field; furthermore, the safety margin of key aircraft structural parts can be automatically obtained based on this stress field, realizing real-time evaluation of aircraft structural strength.
[0054] like Figure 3 The figure shows a scatter plot comparing the predicted stress and the finite element simulation stress under all working conditions in this embodiment of the invention. Each data point in the figure represents the stress value of a structural node under a specific working condition. The horizontal axis represents the simulated value, and the vertical axis represents the model's predicted value. The red dashed line represents the ideal fitting law. All data points are closely distributed on both sides of the ideal line, indicating a high degree of consistency between the predicted and simulated values.
[0055] The rapid prediction example of this invention requires evaluating the strength of the aircraft under a new operating condition: T_env = -40°C, T_cabin = 22°C, ΔP = 0.05MPa, n_z = 1.8 gust condition, P_aero = 30, F_shear = 800, M_bend = 3000; the location of the maximum stress in the prediction result is: the middle of the lower surface of the aircraft wing, σ_max = 218.7 MPa.
[0056] like Figure 4 The figure shown is a logarithmic comparison of the actual computation time in an embodiment of the present invention, comparing the actual computation time of the traditional finite element analysis method and the method of the present invention in a single aircraft strength analysis task. Figure 5The figure shown is a comparison of the speedup factors of relative calculation speed in an embodiment of the present invention, illustrating the speedup factors of relative calculation speed between the traditional finite element analysis method and the method of the present invention in a single aircraft strength analysis task.
[0057] like Figure 6 The figure shows a continuous stress cloud diagram of the wing sparsity structure according to an embodiment of the present invention. The diagram is a schematic diagram of the wing sparsity structure, demonstrating the continuous stress cloud diagram of the rapid prediction results of the stress field of the wing structure under a novel flight condition and gust load using the model of the present invention. Figure 7 The figure shown is a finite element analysis stress cloud diagram of the wing sparsity structure in an embodiment of the present invention. It demonstrates the comparison and verification of the traditional finite element analysis stress cloud diagram with the rapid prediction results of the model of the present invention on the stress field of the wing structure under a new flight condition and gust load.
[0058] The second aspect of this invention proposes an aircraft structural strength prediction system based on a multi-field coupled graph neural network method for predicting aircraft structural strength. The system includes: a key multi-physics operating condition analysis module for aircraft structure, a hierarchical graph structural feature acquisition module for aircraft structure, a multi-physics perceptual graph neural network model construction module, and an aircraft structural strength prediction output module.
[0059] The key multiphysics operating condition analysis module for aircraft structures is used to determine the key parameters of multiphysics operating conditions of aircraft structures, perform parameterized simulation of aircraft structure loads, and obtain multiphysics finite element simulation stress datasets.
[0060] The hierarchical diagram structural feature acquisition module for aircraft structures analyzes the node numbers, three-dimensional coordinates, and element connection relationships of the aircraft structure based on the finite element analysis results of the working condition vectors; it then normalizes the aircraft structure coordinates into a feature matrix for the aircraft structural component model. The nodal feature matrix of the aircraft structural component model And the edge feature matrix of the aircraft structural component model.
[0061] The multi-physics perceptual graph neural network model construction module adopts an encoder-processor-decoder architecture to construct a graph neural network model with an encoder-processor-decoder architecture. The multi-physics perceptual graph neural network model takes the node feature matrix of the aircraft structural component model, the edge feature matrix of the aircraft structural component model, and the global working condition matrix of the aircraft structure as inputs, and outputs the stress matrix of the aircraft structure; thus realizing the stress analysis and prediction of the aircraft structure.
[0062] The aircraft structural strength prediction output module is used to train and validate the multiphysics perceptual graph neural network model, output the aircraft stress field, automatically obtain the safety margin of key aircraft structural parts, and realize the real-time assessment of aircraft structural strength.
[0063] The beneficial effects of this invention are as follows: This invention proposes a method and system for predicting the structural strength of aircraft based on a multi-field coupled graph neural network. It unifies and parameterizes various physical field conditions such as ambient temperature, cabin temperature, pressure difference, aerodynamic loads, and inertial loads, embedding them into the graph neural network model. This makes the prediction results closer to the real complex flight environment during aircraft missions, improving the engineering practicality and accuracy of aircraft structural prediction. A two-level normalization method is used to directly handle complex aircraft-scale geometry, avoiding errors caused by traditional simplification methods. The information transmission mechanism of the graph neural network model naturally captures the load transfer and interaction between aircraft structural components. A dynamic weighted adjacency matrix construction method based on physical information dynamically decodes edge features, such as connection direction and distance, into adjacency weights through learnable functions. This allows the model to autonomously perceive and learn the differences in importance of different structural connections, such as riveting, welding, and bolting, in load transfer, achieving a deep integration of physical prior knowledge and data-driven learning, and improving the model's accuracy in modeling the force transmission path of complex structures.
[0064] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for predicting the structural strength of an aircraft based on a multi-field coupled graph neural network, characterized in that, It includes: S1: Determine key parameters of the aircraft structure under multiphysics conditions, including ambient temperature. Aircraft cabin temperature Aircraft cabin pressure difference External pneumatic pressure Vertical overload factor fuselage shear force and fuselage bending moment , forming the working condition vector Perform parameterized simulation of aircraft structural loads to obtain a multiphysics finite element simulation stress dataset; S2: Construct a hierarchical graph of the aircraft structure's structural features; based on the operating condition vector output in step S1... The finite element analysis results were used to extract the node numbers, three-dimensional coordinates, and element connection relationships of the aircraft structure; the aircraft structure coordinates were normalized into the characteristic matrix of the aircraft structural component model. ; through the global normalized coordinates of two-level nodes and the encoding of aircraft structural component types The node feature vectors constitute the node feature matrix of the aircraft structural component model. ; Edge feature matrix of aircraft structural component model It consists of the edge feature vectors of the aircraft structural component model and the feature distances between the nodes of the aircraft structural component model; S3: Construct a multiphysics-aware graph neural network model for aircraft structural strength prediction, specifically including: using a multiphysics-aware graph neural network model encoder, fusing node features, and obtaining the initial state matrix of the nodes of the aircraft structural component model through linear transformation. A multi-physics sensing graph neural network model iterative processor is constructed to perform graph-structure weighted aggregation of node states and output a node state matrix of global aircraft structure information. The multiphysics perceptual graph neural network model decoder outputs node-predicted stress values, thus constructing the stress field of the aircraft structure. ; S4: Training and validating the multiphysics perceptual graph neural network model for predicting aircraft structural strength constructed in step S3, outputting the aircraft stress field, automatically obtaining the safety margin of key aircraft structural parts, and performing real-time assessment of aircraft structural strength.
2. The aircraft structural strength prediction method based on a multi-field coupled graph neural network according to claim 1, characterized in that: Step S2 is as follows: S21: Perform global centering and scaling of the aircraft structural model; obtain the global bounding box of the aircraft structural model; use the bounding box to obtain the global scale parameters of the aircraft structural model, and obtain the global normalized coordinates of the aircraft structural model nodes; S22: Perform local normalization processing at the aircraft structural component level; S23: Construct the node feature matrix of the aircraft structural component model to obtain the node feature matrix of the aircraft structural component model. Construct the edge feature matrix of the aircraft structural component model ; S24: Construct the global operating condition matrix for the aircraft structure .
3. The aircraft structural strength prediction method based on a multi-field coupled graph neural network according to claim 2, characterized in that: Step S21 is as follows: ； ； ； ； ； in, The coordinates are the global center coordinates of the aircraft structural model; The maximum value of the x-coordinate of the aircraft structural node; This represents the minimum value of the x-coordinate of an aircraft structural node. The maximum value of the ordinate of the aircraft structural node; This represents the minimum value of the ordinate of the aircraft structural node; The maximum value of the vertical coordinate of the aircraft structural node; For the minimum value of the vertical coordinate of the aircraft structural node; These are the global scale parameters for the aircraft structural model; To find the square function; For the first The globally normalized x-coordinate of each node; For the first The globally normalized ordinate of each node; For the first The globally normalized vertical coordinates of each node; For the first The initial x-coordinates of each node; For the first The initial y-coordinates of each node; For the first The initial vertical coordinates of each node; Number the nodes of the aircraft structural model.
4. The aircraft structural strength prediction method based on a multi-field coupled graph neural network according to claim 2, characterized in that: Step S22 is as follows: ； ； ； in, For the first The global quadratic normalized x-coordinate of each node; For the first The global quadratic normalized ordinate of each node; For the first The global quadratic normalized vertical coordinates of each node; The center of the node of the aircraft structural component; These are the global scale parameters for the nodes of aircraft structural components.
5. The aircraft structural strength prediction method based on a multi-field coupled graph neural network according to claim 2, characterized in that: The nodal feature matrix of the aircraft structural component model in step S23 The edge feature matrix of the aircraft structural component model Specifically: ； ； in, The first model of aircraft structural components Each node's feature matrix; Encoding the types of aircraft structural components; For the aircraft structural component model The node of the first The elements of the edge feature matrix; The abscissa feature vector of the aircraft structural component model; The ordinate edge feature vector of the aircraft structural component model; The vertical coordinate edge feature vector of the aircraft structural component model; The feature distance between nodes in the aircraft structural component model; These are numerically stable characteristic parameters.
6. The aircraft structural strength prediction method based on a multi-field coupled graph neural network according to claim 1, characterized in that: Step S3 is as follows: S31: Constructing a multiphysics-aware graph neural network model encoder: The global operating condition features are copied to each node, fused with the node features, and after linear transformation and activation, the initial state matrix of the nodes of the aircraft structural component model is obtained. ; S32: Construct an iterative processor for a multiphysics-aware graph neural network model; input the initial state matrix of the nodes in the aircraft structural component model. Each graph network layer of the iterative processor performs graph structure weighted aggregation on the node states and updates the states through parameterized nonlinear transformations, ultimately outputting a node state matrix containing global aircraft structure information. ; S33: Constructing a multi-physics perceptual graph neural network model decoder; analyzing the node state matrix of aircraft structural information. A linear transformation is performed to directly output the predicted stress value of each node, thus forming the stress field of the aircraft structure. .
7. The aircraft structural strength prediction method based on a multi-field coupled graph neural network according to claim 6, characterized in that: Step S31 is as follows: ； in, This is the initial state matrix of the nodes in the aircraft structural component model; This is a copy function that replicates the global status to each node. It is a non-linear activation function; For splicing operations; The encoder weights of the graph neural network model; For encoder bias in graph neural network models; This is the global operating condition matrix for the aircraft structure.
8. The aircraft structural strength prediction method based on a multi-field coupled graph neural network according to claim 6, characterized in that: Step S32 is as follows: ； ； ； in, For the first The aggregated information matrix of a layered graph network layer; It is the inverse of the weighted degree matrix; This is a weighted adjacency matrix for aircraft structural components; The node state matrix for aircraft structural information; For the first The weight matrix of the layer processor; For the first The bias vector of the layer processor; For the processor layer index; This represents the total number of processor layers. Weighted adjacency matrix elements for aircraft structural components; It is an exponential function; It is a linear transformation function; is the linear rectified activation function; is the learning weight generation function.
9. The aircraft structural strength prediction method based on a multi-field coupled graph neural network according to claim 6, characterized in that: Step S33 is as follows: ； in, For the stress field of the aircraft structure; The weights for the stress field state of the aircraft structure; This is an offset of the stress field state of the aircraft structure.
10. A prediction system for the aircraft structural strength prediction method based on a multi-field coupled graph neural network as described in any one of claims 1-9, characterized in that, It includes: a key multiphysics operating condition analysis module for aircraft structures, a hierarchical graph structural feature acquisition module for aircraft structures, a multiphysics perception graph neural network model construction module, and an aircraft structure strength prediction output module; The aircraft structure key multiphysics field working condition analysis module is used to determine the key parameters of the multiphysics field working condition of the aircraft structure, perform parameterized simulation of the aircraft structure load, and obtain the multiphysics field finite element simulation stress dataset. The hierarchical diagram structural feature acquisition module for aircraft structures analyzes the node numbers, three-dimensional coordinates, and element connection relationships of the aircraft structure based on the finite element analysis results of the working condition vectors; it then normalizes the aircraft structure coordinates into a feature matrix for the aircraft structural component model. The nodal feature matrix of the aircraft structural component model The edge feature matrix of the aircraft structural component model; The multi-physics perceptual graph neural network model building module is a graph neural network model that includes an encoder, processor and decoder architecture. The multi-physics perceptual graph neural network model takes the node feature matrix of the aircraft structural component model, the edge feature matrix of the aircraft structural component model and the global working condition matrix of the aircraft structure as input, and the aircraft structure output stress matrix as output, to realize the stress analysis of the aircraft structure. The aircraft structural strength prediction output module is used to train and validate the multiphysics perceptual graph neural network model, output the aircraft stress field, automatically obtain the safety margin of key aircraft structural parts, and realize the real-time assessment of aircraft structural strength.