Industrial structure mechanics simulation and prediction system based on enhanced graph attention network

Abstract

The application relates to the technical field of engineering simulation, and discloses an industrial structure mechanics simulation prediction system based on an enhanced graph attention network, which comprises a finite element data mapping and enhancement module, a prediction module, an optimization module and a deployment module. The system reconstructs a finite element grid into a graph structure tensor, captures geometric mutation characteristics by using high-dimensional feature projection and adaptive feature recalibration, and transmits deep physical information through dense residual connection. In the training stage, a physical constraint loss function containing attention area weighting and structure smoothness is introduced; in the reasoning stage, a mixed precision mode is adaptively switched according to hardware attributes, and large graph blocking and fusion calculation based on overlapping boundaries are performed. The application effectively solves the contradiction between calculation efficiency and numerical accuracy in complex industrial structure mechanics response prediction, and realizes millisecond-level high-fidelity online simulation.

Description

Technical Field

[0001] This invention relates to the field of engineering simulation technology, specifically to an industrial structural mechanics simulation and prediction system based on enhanced graph attention networks. Background Technology

[0003] Currently, industry primarily relies on the Finite Element Method (FEM) for structural mechanics calculations. While this method, based on rigorous solutions to physical equations, can provide high-fidelity simulation results, its computational process essentially involves solving extremely large systems of partial differential equations, resulting in extremely high computational complexity. For complex industrial models containing millions of degrees of freedom, a single simulation can often take tens of minutes or even hours. This high time cost makes the FEM method unsuitable for meeting the millisecond-level real-time feedback requirements of digital twin systems, thus hindering online monitoring and real-time control.

[0004] To address the computational efficiency issue, deep learning-based surrogate model techniques have emerged. These methods attempt to fit the mapping relationship between loads and physical fields by training neural networks. However, existing deep learning simulation methods still face technical bottlenecks when dealing with complex industrial structures.

[0005] Industrial structural components typically possess highly irregular geometries and complex topologies (e.g., numerous chamfers, bolt holes, and variable cross-sections), making their data a typical example of non-Euclidean data. Traditional convolutional neural networks (CNNs) require voxelizing or projecting the model onto a regular grid, inevitably leading to the loss of geometric details and discretization errors. While graph neural networks (GNNs) can directly process unstructured grids, conventional graph convolutional networks are prone to over-smoothing when dealing with deep feature extraction, causing the network to fail to distinguish the feature differences between adjacent nodes. Specifically, the model struggles to capture peak features in stress concentration areas (such as flange roots and keyways), and predictions tend to be averaged, thus severely underestimating the structural failure risk.

[0006] Most existing data-driven models are based on pure regression algorithms and lack embedded constraints on physical mechanisms. Because they do not include prior knowledge of continuum mechanics, the physical fields predicted by these models often lack smoothness in spatial distribution, easily generating numerical oscillations or noise between adjacent nodes that do not conform to physical laws. Furthermore, these black-box models are extremely sensitive to the distribution of input data; when small perturbations occur in actual operating conditions, the output results may fluctuate drastically, lacking the robustness required for industrial applications.

[0007] As the accuracy of industrial models increases, the number of grid nodes grows exponentially, posing a significant challenge to the GPU memory resources of computing hardware. Existing deep learning frameworks typically require loading the entire graph data at once during inference, often leading to memory overflow (OOM) issues when processing models with hundreds of millions of degrees of freedom. Furthermore, they lack adaptive adjustment mechanisms for different computing power environments, making efficient deployment on edge computing devices or ordinary workstations difficult.

[0008] In summary, developing a real-time simulation and prediction system that can maintain high fidelity at the finite element level while adapting to complex geometric topologies and possessing physical consistency is a key technical challenge that urgently needs to be addressed in the field of industrial digital twins. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides an industrial structural mechanics simulation and prediction system based on an enhanced graph attention network. This system solves the problems of traditional finite element methods being too time-consuming to meet the real-time digital twin requirements in the mechanical response analysis of complex industrial structural components, while conventional deep learning models suffer from inaccurate prediction of extreme values ​​in stress concentration areas and uneven distribution of the output physical field.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] This invention constructs an industrial structural mechanics simulation and prediction system based on an enhanced graph attention network. Logically, the system first establishes a bridge from the physical space to the feature space through mapping and enhancement modules. Unlike conventional processing that treats structures as images, this system analyzes the topological connectivity of finite element mesh data, reconstructing it into a graph structure tensor in non-Euclidean space. In this process, the system not only extracts node geometric information, material property information, and load information as node feature vectors, but also standardizes heterogeneous physical quantities based on global statistical features through dynamic normalization units, eliminating the order-of-magnitude differences between different dimensions (such as geometric coordinates and load values) and solving the gradient imbalance problem during deep learning model training. Furthermore, to address the weak generalization ability of the model with small sample sizes, the system introduces a load perturbation enhancement mechanism, applying Gaussian-distributed random noise to the load vector during the training phase to actively expand the diversity of the sample space.

[0012] At the core prediction level, this system deploys a prediction module. This module abandons the traditional fully connected or convolutional architecture, instead employing a multi-layered stacked graph neural network. Internally, it integrates a high-dimensional feature projection unit, which maps initial physical features to a high-dimensional latent space through a multilayer perceptron to decouple complex nonlinear mechanical relationships. To accurately capture stress concentration phenomena in key areas such as bolt holes and chamfers in structures like wind turbine main shafts, the system utilizes an adaptive feature recalibration unit to implement an attention mechanism. Based on the intensity of load transmission across different topologies, it dynamically calculates and allocates the mutual influence weights between nodes. Simultaneously, to address the over-smoothing phenomenon that often occurs in deep graph networks, the system designs dense residual connection units to establish skip paths between network layers, ensuring that deep networks retain the original geometric structural features, thereby supporting high-precision mechanical response prediction field output.

[0013] To ensure that the prediction results conform to physical laws, this system incorporates an optimization module. This module introduces a hybrid loss function that incorporates physical mechanisms during the model training phase. In addition to the conventional data fitting error, the system specifically constructs a structural smoothness constraint term. This term, based on the principle of graph Laplace smoothing, calculates the difference between the predicted values ​​of a node and its neighboring nodes, forcing the model output to maintain a continuous and smooth spatial distribution, thereby effectively suppressing non-physical numerical noise and oscillations. Furthermore, the system combines a region-of-interest weighting strategy with an anti-interference robustness constraint term. By assigning higher weights to regions of geometric abrupt changes and restricting the norm of model parameters, the system's robustness under extreme conditions and complex boundary conditions is improved.

[0014] At the engineering deployment level, this system integrates a deployment module. Addressing the diverse hardware environments of industrial sites, the system identifies computing resources through a hardware abstraction layer and automatically switches between mixed-precision and single-precision inference modes using a precision switching unit, maximizing computational efficiency while ensuring accuracy. For ultra-large-scale node data exceeding the single-card memory limit, the system employs a large-graph block inference and boundary fusion strategy, dividing the large graph into overlapping subgraphs for separate computation before weighted fusion, achieving full-field prediction of models with hundreds of millions of degrees of freedom under limited hardware resources.

[0015] This invention provides an industrial structural mechanics simulation and prediction system based on an enhanced graph attention network. It has the following beneficial effects:

[0016] 1. This invention transforms unstructured meshes into graph tensors through a finite metadata graph structure mapping and prediction module. By utilizing adaptive feature recalibration and dense residual connection mechanisms, it can explicitly model the non-Euclidean topological relationships between nodes. This enables the system to accurately capture load transfer paths and stress concentration features when dealing with complex structures such as wind turbine main shafts with geometrical abrupt changes such as bolt holes and chamfers. This avoids the feature oversmoothing problem in deep network training and improves the prediction accuracy of mechanical response in key areas.

[0017] 2. This invention introduces structural smoothness constraints and anti-interference robustness constraints based on the graph Laplacian operator in the optimization module, internalizing the smoothness assumption of continuous medium mechanics as the training objective. This mechanism forces the spatial distribution of the model output to maintain physical continuity, effectively eliminating the non-physical numerical oscillations common in pure data-driven models. At the same time, in conjunction with the load perturbation enhancement strategy, it ensures that the model can maintain the stability of the output even under small input fluctuations or extreme working conditions.

[0018] 3. This invention utilizes a deployment module combined with hardware abstraction layer technology to automatically switch between mixed-precision and single-precision computing based on underlying computing resources, balancing speed and accuracy. In particular, through large-graph block inference and boundary fusion technology, the system successfully overcomes the limitation of single-card memory on large-scale node data, enabling industrial structure simulation models with hundreds of millions of degrees of freedom to achieve low-cost, high-efficiency online inference on ordinary computing devices. Attached Figure Description

[0019] Figure 1 This is a diagram showing the overall module architecture of the system according to an embodiment of the present invention;

[0020] Figure 2 This is a flowchart illustrating the overall prediction method according to an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram comparing the predicted stress field of the wind turbine main shaft with the actual value in an embodiment of the present invention;

[0022] Figure 4 This is a comparison chart of predicted and actual values ​​in an embodiment of the present invention;

[0023] Figure 5 This is a schematic diagram of the absolute error distribution in an embodiment of the present invention.

[0024] Among them, 100 is the mapping and enhancement module; 200 is the prediction module; 300 is the optimization module; and 400 is the deployment module. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] See attached document Figure 1 The present invention provides an industrial structural mechanics simulation and prediction system based on an enhanced graph attention network, which may include: a mapping and enhancement module 100, a prediction module 200, an optimization module 300, and a deployment module 400.

[0027] The industrial structural mechanics simulation and prediction system based on enhanced graph attention networks is deployed in a computing device containing a processor and memory. The memory stores computer program instructions and finite element mesh data of the industrial structural components to be predicted. The processor executes the computer program instructions to instantiate and run the mapping and enhancement module 100, the prediction module 200, the optimization module 300, and the deployment module 400.

[0028] The computing device has an internal system bus. The mapping and enhancement module 100, the prediction module 200, the optimization module 300, and the deployment module 400 establish data communication connections and complete data interaction through the system bus.

[0029] The mapping and enhancement module 100 reads the finite element mesh data from the memory through a data interface. The mapping and enhancement module 100 parses the nodal geometry information, material property information, and working condition load information in the finite element mesh data and extracts them as nodal feature vectors.

[0030] The mapping and enhancement module 100 standardizes the node feature vectors. The mapping and enhancement module 100 then reconstructs the processed node feature vectors into a graph structure tensor containing the node feature vectors and edge index relationships.

[0031] The overall mathematical expression of a graph structure tensor is:

[0032] ;

[0033] in, Represents the reconstructed graph structure tensor. This represents the set of node eigenvectors in the graph structure tensor. The set of edge index relationships in the graph structure tensor.

[0034] The prediction module 200 receives the graph structure tensor output by the mapping and enhancement module 100. The prediction module 200 performs feature extraction and propagation through internal multi-layered stacked graph neural network layers.

[0035] The prediction module 200 dynamically calculates the mutual influence weights between nodes using an attention mechanism. Based on the calculated mutual influence weights, the prediction module 200 performs feature aggregation and outputs the mechanical response prediction field of the industrial structural component to be predicted.

[0036] The mapping calculation form of the mechanical response prediction field is expressed as:

[0037] ;

[0038] in, The mechanical response prediction field represents the output. The mapping function represents the multi-layered stacked graph neural network layers in the prediction module 200. This represents the set of network parameters for a graph neural network layer.

[0039] The optimization module 300 is activated and runs during the model training phase. The optimization module 300 reads the mechanical response prediction field output by the prediction module 200 and the real physical field label data stored in the memory, and constructs a hybrid loss function.

[0040] The hybrid loss function includes a data fitting term and a physical consistency constraint term. The optimization module 300 calculates the gradient by minimizing the hybrid loss function and adjusts the network parameters of the prediction module 200 based on the backpropagation algorithm.

[0041] The basic mathematical form of the hybrid loss function is:

[0042] ;

[0043] in, Represents a mixed loss function. Represents the data fitting term, Represents physical consistency constraints. The weight coefficients representing the data fitting terms. The weighting coefficients represent the physical consistency constraints.

[0044] The deployment module 400 is used to control the operating mode of the system based on the underlying hardware computing resources during the inference phase. The deployment module 400 reads the processor's hardware computing power status and memory resource status through the underlying system interface and automatically configures the inference calculation precision.

[0045] Based on the configured inference calculation accuracy, the deployment module 400 calls the prediction module 200, which has completed parameter adjustments, to generate the final stress field and displacement field data, and writes the generated data to the memory or outputs it to the display terminal through the output interface.

[0046] See attached document Figure 1 and attached Figure 2 The mapping and enhancement module 100 further includes a topology reconstruction unit, a feature construction unit, a dynamic normalization unit, and a load perturbation enhancement unit.

[0047] The topology reconstruction unit reads the finite element mesh data file of the industrial structural component to be predicted through a file parsing interface. The finite element mesh data file contains a node list and an element connection list. The node list records the index numbers of all discrete nodes. The element connection list records the set of node indices contained in each finite element mesh element.

[0048] The topology reconstruction unit traverses each finite element mesh element in the element connection list. For the currently traversed finite element mesh element, the topology reconstruction unit extracts the indices of all nodes contained in that element. The topology reconstruction unit establishes an undirected connection edge between any two nodes belonging to the same finite element mesh element.

[0049] The topology reconstruction unit generates a sparse adjacency matrix or edge index list based on all established undirected connections to represent the topological relationships in the graph structure tensor. For graphs containing... A mesh model with 10 nodes, the graph structure is defined as follows: .

[0050] For each node in the node set, the feature building element extracts the corresponding physical property data from the finite element mesh data. The physical property data includes geometric coordinate data, material property data, and boundary condition load data.

[0051] The feature construction unit concatenates the extracted physical attribute data into an initial node feature vector. The process of constructing the node feature vector is represented as follows:

[0052] ;

[0053] in, Representing the The initial node feature vectors of each node. Representing the The three-dimensional spatial geometric coordinate vector of each node , Representing the Material property vectors for each node (including Young's modulus and Poisson's ratio). Representative applied in the Three-dimensional load component vectors at each node , This represents a vector concatenation operation.

[0054] The dynamic normalization unit standardizes each dimension of the node feature vector. The dynamic normalization unit pre-traverses the training dataset, calculating the global statistics for each dimension of the feature vector.

[0055] The calculation formula for standardization is as follows:

[0056] ;

[0057] in, Representing the The node at the th Normalized eigenvalues ​​of dimension Representing the The node at the th The original feature values ​​of the dimension. Represents the first in the training dataset The global mean of the dimensional features. Represents the first in the training dataset The global standard deviation of the dimensional features. This represents a tiny constant used to prevent division by zero errors.

[0058] The dynamic normalization unit also stores the global mean and global standard deviation in the system parameter configuration file. At the system output, the dynamic normalization unit reads the prediction result vector and performs an inverse normalization operation using the global mean and global standard deviation to restore the dimensionless prediction values ​​to stress or displacement values ​​with physical units.

[0059] The load perturbation augmentation unit is activated only during the model training phase. It receives the node feature vectors output by the feature building unit. The load perturbation augmentation unit identifies the load component portion of the node feature vectors.

[0060] The load disturbance enhancement unit applies random noise to the load component to generate enhanced samples. These enhanced samples are used to expand the distribution diversity of the training data and improve the model's robustness to operating condition fluctuations.

[0061] The calculation logic for enhanced load disturbance is expressed as follows:

[0062] ;

[0063] in, Represents the enhanced load vector. Represents the original load vector. The representative amplitude disturbance coefficient follows a pattern with a mean of 0 and a variance of . Gaussian distribution , The vector represents the directional perturbation vector, which contains three random small components that follow a uniform distribution.

[0064] The final output of the mapping and enhancement module 100 includes normalized feature vectors. With edge index list The graph structure tensor is passed to the prediction module 200.

[0065] See attached document Figure 1 and attached Figure 2 The prediction module 200 is configured as a deep neural network structure, which contains multiple stacked graph neural network layers. Each graph neural network layer specifically includes a high-dimensional feature projection unit, an adaptive feature recalibration unit, a feature aggregation and update unit, and a dense residual connection unit.

[0066] The high-dimensional feature projection unit is located at the input of each graph neural network layer. The high-dimensional feature projection unit receives node feature vectors from the previous layer or the initial input. Since the node feature vectors of the initial input have low dimensionality and heterogeneous physical meanings (such as coordinates, modulus, load), they are difficult to use directly to capture complex nonlinear mechanical relationships. The high-dimensional feature projection unit uses a multilayer perceptron (MLP) to perform dimensionality-up mapping on the input features.

[0067] The high-dimensional feature projection unit maps low-dimensional node feature vectors to a high-dimensional latent feature space through linear transformations and nonlinear activation functions, thereby expanding the expressive power of the features. The mathematical expression of this mapping process is as follows:

[0068] ;

[0069] in, Representing the In the layer network The high-dimensional latent feature vector of each node Representing the The input feature vector of the layer, Representing the The learnable weight matrix of the layer, Representing the Layer bias vector, This represents a non-linear activation function (such as ELU or ReLU).

[0070] Specifically, to preserve the nonlinear features of low-dimensional physical information while increasing the feature dimensionality, the high-dimensional feature projection unit employs a two-layer bottleneck structure to construct a multilayer perceptron. This structure consists of a first fully connected layer, a GELU activation layer, and a second fully connected layer connected sequentially. Assume the dimension of the input feature vector is... (e.g., 8-dimensional), the first fully connected layer maps it to the extended dimension. (For example The second fully connected layer compresses and maps it to the output latent dimension. (For example ).

[0071] Compared to the ordinary ReLU function, the GELU (Gaussian Error Linear Unit) activation function used here allows small negative values ​​to pass through, which can effectively alleviate the neuron death problem when processing sparse and finite metadata. Its mathematical expression is:

[0072] ;

[0073] The adaptive feature recalibration unit performs attention computation on every connecting edge in the graph structure tensor. The adaptive feature recalibration unit considers not only the features of the nodes themselves but also the interaction strength between connected nodes. For any edge connecting the source node and the target node, the adaptive feature recalibration unit first calculates the correlation coefficient between the two in the high-dimensional feature space, which characterizes the degree of mechanical influence between the source and target nodes.

[0074] The adaptive feature recalibration unit uses a single-layer feedforward neural network as the core of the attention mechanism. It uses the LeakyReLU activation function to process the correlation coefficients and uses the Softmax function to normalize all the correlation coefficients in the first-order neighborhood of the target node, thereby obtaining the final attention weights.

[0075] The calculation process for attention weights is as follows:

[0076] ;

[0077] ;

[0078] in, Representing the Layer nodes With nodes The non-normalized correlation coefficient between them Representing the The shared weight matrix used for feature transformation in the layer. Representing the The attention mechanism parameter vector of the layer, This represents a vector concatenation operation. Representative node The set of first-order adjacent nodes (including nodes) itself), Represents the normalized attention weights. This represents the vector transpose operation.

[0079] The feature aggregation update unit performs a weighted aggregation operation based on the calculated attention weights. The unit iterates through all neighboring nodes of the target node, multiplying the feature vectors of each neighboring node by their corresponding attention weights and then summing the results. This process simulates the transmission and diffusion of mechanical loads on the mesh topology; neighboring nodes with larger weights have a greater impact on the stress / displacement state of the target node.

[0080] The aggregated feature vector output by the feature aggregation update unit is calculated as follows:

[0081] ;

[0082] in, Representing the The updated feature vector is generated after the layer undergoes graph attention aggregation.

[0083] In a preferred embodiment of the present invention, in order to enhance the model's ability to capture features of different subspaces, the adaptive feature recalibration unit adopts a multi-head attention mechanism.

[0084] Specifically, the adaptive feature recalibration unit is executed independently. The above attention calculation process ( For the number of attention heads, for example Each calculation uses an independent weight matrix and attention parameters.

[0085] The system will The output features of each attention head are concatenated or averaged to obtain the final aggregated feature vector:

[0086] ;

[0087] Through a multi-head mechanism, the system can simultaneously focus on the differences in geometric positional relationships and physical properties (such as material stiffness) between nodes, thereby improving the robustness of feature extraction.

[0088] Dense residual connection units are configured at the output of each layer of the graph neural network. As the number of network layers increases, in order to prevent gradient vanishing and loss of high-frequency structural information caused by overly smooth features in deep networks, dense residual connection units establish skip connection paths across layers.

[0089] The dense residual connection unit adds the high-dimensional feature vector input to the current layer to the aggregated feature vector output by the feature aggregation update unit element-wise. This operation forces the model to learn the residual mapping, allowing the deep network to retain the original input structure information.

[0090] The mathematical logic representation of dense residual connections is as follows:

[0091] ;

[0092] in, Representing the The input feature vector of a layered graphical neural network, i.e., the first layer... The final output of the layer. When When it is the last layer, the vector is decoded by the fully connected layer and output as the final mechanical response prediction (stress field or displacement field).

[0093] In the initial stage of model training, to break the symmetry of network parameters and maintain the variance stability of gradients during deep propagation, this invention does not use all-zero initialization or random Gaussian initialization. For all weight matrices in the prediction module 200, the system uses a Xavier uniform distribution for initialization.

[0094] The weighting parameters are sampled uniformly from the following intervals:

[0095] ;

[0096] in, The dimension of the input features for this layer. Let be the dimension of the output features of this layer. This initialization strategy ensures that the variance of the output features of each layer remains consistent during forward propagation, thereby accelerating the convergence of the physical constraint loss function.

[0097] See attached document Figure 1 and attached Figure 2 The optimization module 300 is used to construct and calculate the hybrid loss function during the model training phase and update the network parameters of the prediction module 200 through the backpropagation algorithm.

[0098] The hybrid loss function constructed by optimization module 300 consists of three core parts: a data fitting term based on the region of interest, a structural smoothness constraint term, and an anti-interference robustness constraint term.

[0099] The optimization module 300 first calculates the data fitting terms. To address the issue of a small number of nodes with high importance in critical areas (such as stress concentration zones) in industrial structural components, the optimization module 300 adopts a region-of-interest weighting strategy.

[0100] The optimization module 300 identifies geometrically abrupt regions (including chamfers and hole edges) or high-stress-risk regions marked based on historical simulation data in the mesh model during the preprocessing stage.

[0101] Specifically, the optimization module 300 automatically identifies regions of geometric abrupt changes based on the local curvature characteristics of the mesh nodes. For any node... The system calculates the discrete variance of the normal vectors of all its adjacent elements. Define nodes. Geometric mutation index as follows:

[0102] ;

[0103] in, Representative shared node The set of all surface mesh elements, Representative Unit The unit normal vector, Representative node The average normal vector at that location. When When the curvature exceeds a preset threshold (e.g., 0.1), the system determines that the node belongs to a region of geometric abrupt change (such as a chamfer or edge) and assigns its corresponding weight coefficient. Set to a high weight value (e.g.) This allows for adaptive attention to complex geometric features.

[0104] The system assigns higher weight coefficients to nodes within the region and basic weight coefficients to nodes in other ordinary regions, thus constructing a weight vector.

[0105] The formula for calculating the data fitting term is as follows:

[0106] ;

[0107] in, This represents the total number of nodes participating in the training. Representing the Predicted mechanical response values ​​for each node. Representing the The true label values ​​of each node (from finite element simulation true values) Represents the L1 norm (absolute error). Representing the The loss weight coefficients of each node, when the node When located in the area of ​​interest, .

[0108] Optimization module 300 calculates the structural smoothness constraint sub-term. This structural smoothness constraint sub-term is based on the graph Laplace smoothing principle and aims to force the physical field output by the model to satisfy the local continuity assumption of continuum mechanics, thereby suppressing non-physical high-frequency oscillations (noise) in the prediction results.

[0109] The optimization module 300 calculates the Euclidean distance between each node and the predicted mean of all its first-order neighbors. The smaller the Euclidean distance, the smoother the local physics field.

[0110] The formula for calculating the structural smoothness constraint sub-item is as follows:

[0111] ;

[0112] in, Representative node The set of first-order adjacent nodes, Represents the number of neighboring nodes. Representing neighbor nodes The predicted value, This represents the square of the L2 norm.

[0113] Optimization module 300 calculates the robustness constraint term. This robustness constraint term employs a regularization strategy, which limits the overfitting response of the model to small input perturbations (noise introduced by the load perturbation enhancement unit) by constraining the numerical amplitude of the model weight parameters, thus ensuring the generalization stability of the model under extreme conditions.

[0114] The formula for calculating the robustness constraint term is as follows:

[0115] ;

[0116] in, Represents the regularization coefficient. Representing the The weight matrix of a layered graph neural network. This represents the square of the Frobenius norm.

[0117] See attached document Figure 1 and attached Figure 2 The deployment module 400 specifically includes a hardware abstraction layer unit, a precision switching unit, a large image block inference unit, and a boundary fusion unit.

[0118] The Hardware Abstraction Layer (HAL) unit is activated during system startup or inference task initialization. The HAL unit queries the physical hardware attributes of the current runtime environment by calling the driver interface of the underlying computing device (such as CUDA Driver API or OpenCLAPI). These physical hardware attributes include the GPU device's compute capability version, the number of stream processors, the tensor core support status, and the currently available video memory (VRAM).

[0119] The precision switching unit dynamically determines the numerical precision mode for inference calculations based on the hardware attributes returned by the hardware abstraction layer unit. When the hardware device has a tensor acceleration operation unit and the available video memory capacity is greater than a preset threshold, the precision switching unit activates the mixed precision inference mode.

[0120] In mixed-precision inference mode, the precision switching unit converts the input graph tensor data from single-precision floating-point format (FP32) to half-precision floating-point format (FP16 or BF16). The prediction module 200 uses the half-precision format to perform matrix multiplication and convolution operations, thereby reducing memory usage and increasing computational throughput. After the computation is complete, the precision switching unit converts the output prediction results back to single-precision floating-point format to ensure the accuracy requirements of the physics field values.

[0121] When the hardware does not support tensor acceleration operations or detects a shortage of video memory resources, the precision switching unit locks the system in single-precision floating-point inference mode and uses FP32 format for calculation throughout the process, prioritizing the numerical stability of the calculation process.

[0122] The large graph partitioning inference unit is used to process ultra-large-scale industrial structure models that exceed the memory limitations of a single graphics card. When the number of finite element mesh nodes to be predicted exceeds the hardware's capacity limit, the large graph partitioning inference unit uses a graph partitioning algorithm (such as the METIS algorithm or spectral clustering algorithm) to divide the complete global graph structure into multiple subgraphs.

[0123] The specific execution logic of the graph partitioning algorithm is not random cutting, but based on the joint optimization objective of minimizing the edge weight and balancing the node load.

[0124] System construction objective function The aim is to find the optimal segmentation scheme. This minimizes the number of cut edges (i.e., edges connecting different subgraphs) while restricting each subgraph through constraints. The difference in the number of nodes in the system does not exceed a preset balance factor (e.g., 5%).

[0125] The objective function is expressed as:

[0126] ;

[0127] in, Edge weights (corresponding to the physical connection stiffness between nodes), For indicator functions, when node and nodes The value is 1 when it is divided into different subgraphs.

[0128] By minimizing this objective function, the system ensures that the most physically connected regions (such as the entire bearing housing) are preserved in the same subgraph, thereby minimizing the loss of physical information caused by forced segmentation.

[0129] To avoid feature extraction bias caused by missing neighbor information when processing subgraph boundary nodes in graph neural networks, the large graph block inference unit constructs overlapping boundary regions during the segmentation process. For each subgraph, the large graph block inference unit not only includes nodes in the core region but also extends outwards. Rank neighborhood ( Corresponding to the number of stacked layers in a graph neural network, the boundary nodes belonging to adjacent subgraphs are included in the computation scope of the current subgraph, forming an overlapping subgraph with halo nodes.

[0130] The boundary fusion unit is used to merge the inference results of various subgraphs to recover the global physics field. The boundary fusion unit performs inference calculations for each overlapping subgraph separately to obtain local prediction results. For nodes located within the overlapping boundary region, since they are repeatedly calculated by multiple subgraphs, the boundary fusion unit uses a weighted average strategy to fuse the prediction values ​​of different subgraphs.

[0131] The weighted average fusion strategy calculates the weight based on the topological location confidence of nodes in each subgraph. Generally, the closer a node is to the center of the subgraph, the more complete its neighborhood information and the higher its prediction confidence; the closer a node is to the edge of the subgraph, the lower its weight.

[0132] The calculation formula for boundary merging is as follows:

[0133] ;

[0134] in, Representative node The final predicted mechanical response value after fusion Represents the containing node The set of all subgraphs, Representative node In the The weighting coefficients in each subgraph (these coefficients are usually related to the nodes) To the The topological distance between the centers of each subgraph is proportional to the total subgraph center. Representing the Subgraph for nodes The predicted output value.

[0135] In this embodiment, the weighting coefficient A smooth decay strategy based on the Sigmoid function is used for calculation. The calculation formula is as follows:

[0136] ;

[0137] in, Representative node To the Topological distance (hop count) between the centers of each subgraph The radius threshold representing the core security area, This represents the decay rate constant. This function ensures that the weight of nodes located at the center of the subgraph is close to 1, while the weight of nodes located at the edge of the overlapping boundary decays smoothly to 0, thus mathematically eliminating the step effect at the subgraph splicing point.

[0138] Through this fusion operation, the system effectively eliminates numerical jumps at the subgraph splicing points, ensuring the smooth continuity of the global physical field.

[0139] To further illustrate the application effects and technical advantages of the industrial structural mechanics simulation and prediction system based on enhanced graph attention networks provided by this invention in practical engineering, the following detailed explanation is provided with specific industrial component examples. The foregoing embodiments have analyzed the various functional modules of the system and their internal logic in detail. This embodiment will demonstrate how these modules work together to solve the contradiction between accuracy and efficiency faced by complex industrial structural components in real-time simulation.

[0140] See attached document Figure 3 and attached Figure 4 In this embodiment, the main shaft of a large wind turbine generator set is selected as the industrial structural component to be predicted. As a key transmission component connecting the hub and the gearbox, the wind turbine main shaft has the characteristics of complex geometric topology (including flanges, variable cross-section shafts, multiple chamfers and bolt holes), variable load conditions (bearing random wind loads and torques), and extremely high safety margin requirements.

[0141] The specific implementation steps for predicting the mechanical response of wind turbine main shafts using the system of this invention are as follows:

[0142] Step S1: Data parsing and graph structure reconstruction

[0143] The mapping and enhancement module 100 reads the finite element mesh file of the wind turbine main shaft through a data interface. This finite element mesh file contains complex tetrahedral element connection relationships. The topology reconstruction element constructs a sparse adjacency matrix, and the feature construction element extracts the node spatial coordinates, material properties (42CrMo steel), and dynamic wind load vectors to construct a feature matrix.

[0144] Step S2: Enhanced Graph Attention Feature Inference

[0145] The prediction module 200 receives the graph structure tensor and captures the stress concentration features at the flange root and bearing housing transition fillets through high-dimensional feature projection and adaptive feature recalibration. Dense residual connection units ensure that the deep network can effectively transmit the overall geometric structure information of the spindle.

[0146] Step S3: Multi-precision inference and physics output

[0147] The deployment module 400 identifies the hardware environment, automatically performs inference calculations, and the boundary fusion unit outputs the final stress field prediction results.

[0148] Experimental verification and quantitative analysis:

[0149] To verify the effectiveness of the technical solution of this invention, a high-fidelity simulation dataset of wind turbine main shafts under different working conditions and loads was constructed and compared with the ground truth calculated by commercial finite element software.

[0150] Comparison of experimental results indicators:

[0151] The statistical analysis results based on the test set data are shown in the table below:

[0152] Evaluation indicators Numerical results illustrate Mean Absolute Error (MAE) 0.84MPa The average prediction bias across all nodes is at an extremely low level. Coefficient of determination (R²Score) 0.992 The degree of fit between the predicted result and the true value is characterized, with a value close to 1 indicating an excellent fit. Maximum stress relative error 4.07% The prediction deviation at the most dangerous point of the structure was used to verify the ability to capture stress concentration zones.

[0153] Explanation of experimental results charts:

[0154] See attached document Figure 3 The figure shows, from left to right: the predicted stress field, the actual stress field, and the absolute error field. Figure 3 The cloud map distribution shows that the predicted stress field and the actual stress field are highly consistent in color distribution trends. Especially in the high-stress areas shown in dark red (such as the bearing housing steps), the morphological characteristics of the two are extremely similar. The absolute error field on the right shows that the error is close to 0 in most areas (dark blue / purple areas), with only a few minor error fluctuations at the edges of geometrical abrupt changes.

[0155] See attached document Figure 4 and attached Figure 5 , attached Figure 4 This is a scatter plot of correlation. The horizontal axis represents the ground truth, and the vertical axis represents the predicted value. The scatter points are closely clustered around each other. The reference line distribution did not show any obvious dispersion.

[0156] Appendix Figure 5 This is a histogram of the absolute error distribution. Statistical data shows that the prediction errors of the vast majority of grid nodes are concentrated in an extremely low range, and the error distribution exhibits a zero-mean long-tail characteristic.

[0157] Detailed verification of the maximum stress point (data support) is based on the system's monitoring data of extreme points for the samples:

[0158] The actual maximum stress is approximately 270.56 MPa (based on the true value from finite element simulation).

[0159] The predicted maximum stress is approximately 281.57 MPa (based on model inference). Calculations show that the relative error of this system's prediction for this critical structural failure risk point is only 4.07%. This result is within the acceptable range of engineering tolerances, demonstrating the reliability of this invention in critical area safety assessment and effectively overcoming the technical shortcomings of traditional deep learning models, which are prone to peak clipping or numerical oscillations in extreme value prediction.

[0160] In summary, by combining graph attention mechanism with physical mechanism constraints, this invention achieves high-precision mechanical performance prediction of complex structural components such as wind turbine main shafts while ensuring millisecond-level inference speed. The error index fully meets the accuracy requirements of industrial-grade rapid design iteration and digital twin monitoring.

Claims

1. An industrial structural mechanics simulation and prediction system based on enhanced graph attention networks, characterized in that, include: The mapping and enhancement module (100) is used to receive finite element mesh data of industrial structural components to be predicted, parse the node geometric information, material property information and working condition load information in the finite element mesh data and extract them as node feature vectors, perform standardization processing on the node feature vectors, and reconstruct the processed node feature vectors into a graph structure tensor containing node feature vectors and edge index relationships. The prediction module (200) is used to receive the graph structure tensor, extract and transmit features through a multi-layer stacked graph neural network layer, dynamically calculate the mutual influence weights between nodes using an attention mechanism, and output the mechanical response prediction field of the industrial structural component to be predicted. An optimization module (300) is used to construct a hybrid loss function during the model training phase. The hybrid loss function includes a data fitting term and a physical consistency constraint term. The network parameters of the prediction module (200) are adjusted by minimizing the hybrid loss function. The deployment module (400) is used to automatically configure the inference calculation accuracy based on the underlying hardware computing resources during the inference phase, and to generate the final stress field and displacement field data based on the prediction module (200).

2. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 1, characterized in that, The mapping and enhancement module (100) includes: A topology reconstruction unit is used to traverse the finite element mesh elements contained in the finite element mesh data, identify the nodes contained in each finite element mesh element, and establish a connection edge between the two nodes when any two nodes belong to the same finite element mesh element, thereby constructing an adjacency matrix of the graph structure and determining the edge index relationship. The feature construction unit is used to extract the spatial three-dimensional coordinates, Young's modulus, Poisson's ratio and three-dimensional load components of each node from the finite element mesh data, and to concatenate the extracted data to form the node feature vector. The dynamic normalization unit is used to calculate the global mean and global standard deviation of each dimension of features in the training dataset, and to standardize the node feature vector using the global mean and global standard deviation. At the same time, the system output uses the global mean and global standard deviation to perform inverse standardization recovery on the prediction result.

3. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 1, characterized in that, The prediction module (200) include: A high-dimensional feature projection unit is used to nonlinearly map the initial node feature vectors to a high-dimensional latent feature space using a multilayer perceptron. An adaptive feature recalibration unit is used to calculate the correlation coefficient between the connected source node and the target node in the high-dimensional latent feature space for each connection edge in the graph structure, and to convert the correlation coefficient into attention weights using a nonlinear activation function and a normalization function. The attention weights characterize the strength of the load transfer between different topologies. The feature aggregation and update unit is used to perform a weighted summation of the feature vectors of all neighboring nodes in the first-order neighborhood of the target node according to the attention weight, so as to update the feature representation of the target node.

4. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 3, characterized in that, The prediction module (200) further includes: The dense residual connection unit is used to establish skip connection paths between each layer of the graph neural network, and directly adds the input feature vector of the current layer to the aggregated feature vector output by the feature aggregation update unit as the input of the next network layer.

5. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 1, characterized in that, The mapping and enhancement module (100) is also provided with a load disturbance enhancement unit; The load perturbation enhancement unit is used to apply random amplitude noise and random direction perturbation vectors that conform to a Gaussian distribution to the input node load vector during the training process, thereby generating enhanced training samples.

6. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 1, characterized in that, In the optimization module (300), the data fitting term adopts a region-of-interest weighting strategy, which specifically includes: Identify geometric abrupt change regions or high stress risk regions in the industrial structural components, assign loss weight coefficients greater than a preset threshold to nodes located in the geometric abrupt change regions or high stress risk regions, and calculate the weighted average absolute error between the predicted value and the actual value.

7. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 1, characterized in that, In the optimization module (300), the physical consistency constraint term includes a structural smoothness constraint sub-term; The structural smoothness constraint sub-item is constructed based on the graph Laplace smoothing principle and is used to calculate the Euclidean distance between the predicted physical quantity of the current node and the mean of the predicted physical quantities of all its neighboring nodes, and the Euclidean distance is included as a penalty term in the hybrid loss function.

8. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 1, characterized in that, In the optimization module (300), the physical consistency constraint term also includes an anti-interference robustness constraint sub-term; The robustness constraint sub-item for anti-interference adopts a regularization constraint strategy. By limiting the L2 norm of the model weight parameters, it constrains the response amplitude of the model to small input perturbations, ensuring prediction stability under extreme conditions.

9. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 1, characterized in that, The deployment module (400) includes: The hardware abstraction layer unit is used to identify the hardware device type and video memory capacity of the current operating environment; The precision switching unit, when it detects that the hardware device supports tensor acceleration operations and the video memory capacity is sufficient, calls the mixed precision inference mode to convert the input data into half-precision floating-point format for calculation, and converts the output result back to single-precision floating-point format. When it is detected that the hardware device does not support tensor acceleration operations, it is locked in single-precision floating-point inference mode.

10. The industrial structural mechanics simulation and prediction system based on enhanced graph attention network according to claim 9, characterized in that, The deployment module (400) also includes: The large graph segmentation inference unit is used to segment the complete graph structure into multiple subgraphs with overlapping boundary regions when processing large-scale node data that exceeds the video memory capacity limit. The boundary fusion unit is used to perform reasoning on each of the subgraphs separately, and to fuse the prediction results of different subgraphs using a weighted average strategy for nodes in the overlapping boundary region.

Images