Real-time analysis method and system for anti-seismic performance of bridge based on graph neural network

Abstract

The invention relates to the technical field of bridge anti-seismic performance analysis, and particularly discloses a bridge anti-seismic performance real-time analysis method and system based on a graph neural network, and the method comprises the following steps: 1, constructing a bridge inter-span structure graph representation: taking each span of a bridge as a node in the graph, structural elements such as connecting beams and supports between the spans are used as edges between the nodes to form a structural diagram representation of the bridge; each node comprises cross geometric attributes, material characteristics and seismic load information as input characteristics, 2, encoder processing: converting the input characteristics of each node into a higher-dimension hidden embedded layer vi through a linear single-layer perceptron (SLP) by using an encoder, and providing appropriate input for subsequent information transmission and structure response prediction; according to the method, the bridge structure is converted into graph representation, the calculation efficiency is improved, the complexity of traditional finite element analysis is reduced, the method can adapt to different types of bridge structures, and the adaptability of the model is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge seismic performance analysis, and particularly relates to a real-time analysis method and system for bridge seismic performance based on graph neural network. Background Art

[0002] Bridges are an important part of transportation infrastructure, undertaking a large number of transportation tasks and being a key link to ensure smooth road traffic and emergency rescue. With the frequent occurrence of natural disasters such as earthquakes, the seismic performance of bridges has increasingly become an important indicator that cannot be ignored in traffic network design and emergency management.

[0003] Traditional bridge seismic design and analysis usually adopt numerical calculation methods based on physical models such as finite element simulation (FEM). Although these methods perform well in terms of accuracy, due to their large computational amount and complex solution process, especially when the bridge is large in scale and complex in geometric shape, the time and computational resources consumed in the analysis process are extremely high. Therefore, how to improve the efficiency of bridge seismic performance analysis and reduce the computational cost has become an important challenge in bridge design and emergency response.

[0004] In recent years, the application of machine learning, especially deep learning technology, has gradually increased in the engineering field. With its powerful data fitting ability, remarkable achievements have been made in structural health monitoring, damage identification, and performance prediction. Neural networks, especially deep neural networks (DNNs), optimize model parameters through the backpropagation algorithm, can approximate various input-output relationships, and are widely used in tasks such as bridge material strength prediction, structural analysis, design optimization, and damage identification. However, traditional neural networks face the problem of being unable to handle structural geometric changes and topological variations. Once the geometric shape or structure type of the bridge changes, the trained neural network model often needs to be retrained, resulting in certain limitations in its practical application.

[0005] To address this problem, graph neural networks (GNNs), as an emerging deep learning method, have gradually attracted wide attention in the field of structural engineering due to their unique advantages in graph data processing. GNNs spread information between nodes through a message passing mechanism, can process data with irregular topological structures, and are particularly suitable for describing complex systems such as bridges composed of multiple nodes (bridge piers, bridge bearings, etc.) and edges (beams, support structures, etc.). Different from traditional neural networks, graph neural networks can learn the interactions between nodes and edges while maintaining the structural topological relationship, so they have stronger generalization ability and can better adapt to bridges of different types and geometric shapes.

[0006] Although the application prospects of graph neural networks in structural engineering are broad, existing graph neural network models still face some challenges. In bridge seismic analysis, the structural and load characteristics of bridges make the prediction of their seismic performance more complex. Traditional graph neural networks usually regard a bridge as an overall graph and adopt a fixed number of message passing layers and structure graph representation methods. Although certain prediction effects have been achieved in some standard structures, when facing more complex bridge structures or unseen structures, the generalization ability and prediction accuracy of the model are often affected. Summary of the Invention

[0007] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a real-time analysis method and system for bridge seismic performance based on graph neural networks, so as to solve the problem that the generalization ability and prediction accuracy of the model are often affected when facing more complex bridge structures or unseen structures.

[0008] To achieve the above purpose, the present invention provides the following technical solutions:

[0009] A real-time analysis method for bridge seismic performance based on graph neural networks includes the following steps:

[0010] Step 1, construct a structural graph representation of bridge spans: Each span of the bridge is used as a node in the graph, and structural elements such as connecting beams and bearings between spans are used as edges between nodes to form a structural graph representation of the bridge; each node contains the geometric attributes, material characteristics, and seismic load information of the span as input features.

[0011] Step 2, encoder processing: Use the encoder to convert the input features of each node into a higher-dimensional hidden embedding layer v i , providing a suitable input for subsequent information transmission and structural response prediction:

[0012]

[0013] Step 3, message passing layer processing: Use a graph neural network (GNN) as the message passing layer. In a single message passing layer, generate the information of each node through the information generation stage, and this information takes into account the current embedding of the node, the embedding of adjacent nodes, and the features of adjacent edges;

[0014]

[0015] Among them, i is the node, v i is the hidden embedding layer, j is the adjacent node, and e i,j is the adjacent edge;

[0016] Aggregate the received messages through an aggregation function and update the node embeddings through an update function, where both the information function and the update function are constructed by a linear single-layer perceptron (SLP). The generated information is aggregated by calculating the average value, and its calculation formula is:

[0017]

[0018] The construction formula of the linear single-layer perceptron (SLP) is:

[0019]

[0020] Introduce an adaptive message passing layer mechanism to dynamically adjust the number of message passing layers according to the number of spans and complexity of the bridge;

[0021] Step 4, Decoder processing: Use the decoder to convert the hidden embeddings of the nodes into the required outputs through a non-linear multi-layer perceptron (MLP), including displacements (in the X and Y directions), bending moments (in the Y and Z directions), and shear forces (in the Y and Z directions), where the decoding formula is:

[0022] MLP(x = SLP2(max(0, SLP1(x)))

[0023] f Response (v i ) = MLP Response (v i );

[0024] Step 5, Model training and optimization: Prepare a dataset containing bridge structural characteristics and seismic response labels, and optimize the model parameters by minimizing the loss function to achieve accurate prediction of the seismic performance of the bridge.

[0025] Preferably, the bridge inter-span structure diagram representation further includes the geometric relationship, material properties, and seismic load information of the inter-span as input features of the nodes.

[0026] Preferably, the adaptive message passing layer mechanism automatically adjusts the number of message passing layers according to the number of spans and complexity of the bridge to ensure that information can be fully propagated throughout the bridge structure.

[0027] Preferably, in the information generation stage of the message passing layer, by considering the interaction and connection relationship between nodes, information reflecting the behavior of bridge components, such as displacements, bending moments, and shear forces, is generated.

[0028] Preferably, the hidden layer in the decoder uses a rectified linear unit (ReLU) activation function to improve the expression ability of the model.

[0029] A real-time seismic performance analysis system for bridges based on graph neural networks, comprising:

[0030] Data input module: used to input the information of the intermediate structure of the bridge, including the geometric properties of the spans, material characteristics, and seismic load information;

[0031] Encoder module: used to convert the input node features into a hidden embedding layer of a higher dimension;

[0032] Message passing layer module: includes an information generation unit and a node update unit, used to achieve information interaction and feature update between nodes, and introduce an adaptive message passing layer number mechanism;

[0033] Decoder module: used to convert the hidden embedding of the node into the required output, including seismic responses such as displacement, bending moment, and shear force;

[0034] Model training and optimization module: used to prepare the dataset, train the model, and optimize the model parameters to achieve accurate prediction of the seismic performance of the bridge.

[0035] Preferably, the message passing layer module further includes an aggregation function and an update function, used to aggregate the received messages and update the embedding of the node.

[0036] Preferably, the data input module is also used to input the geometric relationship between the spans of the bridge, material properties, and seismic load information.

[0037] Preferably, the hidden layer in the decoder module uses a rectified linear unit (ReLU) activation function.

[0038] Preferably, the system further includes a model inference and seismic performance prediction module, used to generate seismic performance prediction results through the encoder, message passing layer, and decoder under a given bridge structure input.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] (1) Convert the finite element model of the bridge structure into a graph representation and analyze it in a graph neural network (GNN). By taking the spans of the bridge as the nodes of the graph and the structural elements between the spans as the edges of the graph, the overall structure of the bridge is mapped into a graph structure. This graph representation method is different from traditional finite element analysis. It expresses the relationship between the spans of the bridge structure and the interaction between nodes in the form of a graph, can effectively capture the complex dependencies between structural elements, convert the bridge structure into a graph representation, improve the computational efficiency, reduce the complexity of traditional finite element analysis, can adapt to different types of bridge structures, and enhance the adaptability of the model.

[0041] (2) By improving the traditional structure diagram representation method, especially in the representation of bridge span nodes, physical constraints are added. The introduction of pseudo-nodes enables the cross-nodes at each layer in the graph to more accurately reflect the behavioral consistency among different spans in the actual structure. This method effectively enhances the physical meaning of the graph neural network in bridge seismic performance analysis, improving the model's expressive ability and computational efficiency.

[0042] (3) Through an adaptive message passing mechanism to address the heterogeneity problem of bridge structures, by adaptively adjusting the number of message passing layers, the BridgeGNN model can automatically adjust the calculation depth according to the specific number of spans of the bridge, ensuring more efficient and accurate information propagation in the structure, significantly improving the model's computational efficiency, and being able to flexibly adjust according to different bridge types, effectively avoiding over-computation.

[0043] (4) Through the graph structure representation and message passing mechanism, the model can automatically learn the interactions between structures and their performance under seismic loads from the data. This model can adapt to different types of bridge structures and has strong generalization ability. Especially when facing new bridge designs or unseen structural configurations, it can quickly adapt and perform effective seismic performance prediction, greatly expanding the application scope of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is the flowchart of the present invention;

[0045] Figure 2 is the schematic diagram of the BrodgeGNN representation method based on bridge FEM node and edge attributes of the present invention;

[0046] Figure 3 is the schematic diagram of the bridge FEM graph representation matrix of the present invention;

[0047] Figure 4 is the schematic diagram of the method for determining the BridgeGNN edge attributes of the dust collection mechanism based on the mechanical model of key components of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0049] Embodiment 1:

[0050] Please refer to Figure 1 - Figure 4As shown in the figure, a real-time analysis method and system for the seismic performance of bridges based on graph neural networks includes the following steps:

[0051] Step 1: Construct a structural diagram representation of the bridge spans: Each span of the bridge is regarded as a node in the graph, and structural elements such as connecting beams and bearings between spans are regarded as edges between nodes to form a structural diagram representation of the bridge; each node contains the geometric attributes, material properties, and seismic load information of the span as input features;

[0052] Step 2: Encoder processing: Use the encoder to convert the input features of each node into a higher-dimensional hidden embedding layer v i , providing a suitable input for subsequent information transfer and structural response prediction:

[0053]

[0054] Step 3: Message passing layer processing: Use a graph neural network (GNN) as the message passing layer. In a single message passing layer, generate the information of each node through the information generation stage, which takes into account the current embedding of the node, the embeddings of adjacent nodes, and the features of adjacent edges;

[0055]

[0056] Among them, i is the node, v i is the hidden embedding layer, j is the adjacent node, and e i,j is the adjacent edge;

[0057] Aggregate the received messages through an aggregation function and update the embedding of the node through an update function. Among them, both the information function and the update function are constructed through a linear single-layer perceptron (SLP). The generated information is aggregated by calculating the average value, and its calculation formula is:

[0058]

[0059] The construction formula of the linear single-layer perceptron (SLP) is:

[0060]

[0061] Introduce an adaptive message passing layer number mechanism to dynamically adjust the number of message passing layers according to the number of spans and complexity of the bridge;

[0062] Step 4: Decoder processing: Use the decoder to convert the hidden embedding of the node into the required output through a non-linear multi-layer perceptron (MLP), including displacements (in the X and Y directions), moments (in the Y and Z directions), and shears (in the Y and Z directions). The decoding formula is:

[0063] MLP(x) = SLP2(max(0, SLP1(x)))

[0064] f Response (v i ) = MLP Response (v i );

[0065] Step 5, Model Training and Optimization: Prepare a dataset containing bridge structure characteristics and seismic response labels, and optimize the model parameters by minimizing the loss function to achieve accurate prediction of the seismic performance of the bridge.

[0066] Preferably, the bridge span structure diagram representation further includes the geometric relationship, material properties, and seismic load information of the span as the input features of the nodes.

[0067] As can be seen from the above, by taking each span of the bridge as a node in the graph, and the structural elements such as the connecting beams and bearings between the spans as the edges between the nodes, a structural diagram representation of the bridge is formed, and for each node, the geometric attributes, material characteristics, and seismic load information of the span are included as input features, realizing the graphical representation of the bridge structure. Thus, the complex bridge structure can be converted into an easily processed graph structure, facilitating subsequent graph neural network analysis. The graph representation method can intuitively display the relationship between the spans of the bridge and the interaction between the nodes, and at the same time can effectively capture the complex dependence relationships between the structural elements, providing a basis for subsequent analysis.

[0068] Embodiment 2:

[0069] Refer to Figure 1 - Figure 4 shown, Step 1, Construct the bridge span structure diagram representation: Take each span of the bridge as a node in the graph, and the structural elements such as the connecting beams and bearings between the spans as the edges between the nodes to form the structural diagram representation of the bridge; each node includes the geometric attributes, material characteristics, and seismic load information of the span as input features;

[0070] Step 2, Encoder Processing: Use the encoder to convert the input features of each node through a linear single-layer perceptron (SLP) into a higher-dimensional hidden embedding layer v i , providing a suitable input for subsequent information transfer and structural response prediction:

[0071]

[0072] Step 3, Message Passing Layer Processing: Use a graph neural network (GNN) as the message passing layer. In a single message passing layer, generate the information of each node through the information generation stage, and this information takes into account the current embedding of the node, the embeddings of adjacent nodes, and the features of adjacent edges;

[0073]

[0074] Among them, i is a node, v is a hidden embedding layer, j is an adjacent node, and e i,j is an adjacent edge;

[0075] Aggregate the received messages through an aggregation function and update the embedding of the node through an update function. Both the information function and the update function are constructed by a linear single-layer perceptron (SLP). The generated information is aggregated by calculating the average value, and its calculation formula is:

[0076]

[0077] The construction formula of the linear single-layer perceptron (SLP) is:

[0078]

[0079] Introduce an adaptive message passing layer number mechanism to dynamically adjust the number of message passing layers according to the number of spans and complexity of the bridge;

[0080] Step 4, decoder processing: Use the decoder to convert the hidden embedding of the node into the required output through a non-linear multi-layer perceptron (MLP), including displacements (in the X and Y directions), bending moments (in the Y and Z directions), and shear forces (in the Y and Z directions). The decoding formula is:

[0081] MLP(x) = SLP2(max(0, SLP1(x)))

[0082] f Response (v i ) = MLP Response (v i );

[0083] Step 5, model training and optimization: Prepare a dataset containing the characteristics of the bridge structure and seismic response labels, and optimize the model parameters by minimizing the loss function to achieve accurate prediction of the seismic performance of the bridge;

[0084] The adaptive message passing layer number mechanism automatically adjusts the number of message passing layers according to the number of spans and complexity of the bridge to ensure that information can be fully propagated throughout the bridge structure;

[0085] In the information generation stage of the message passing layer, by considering the interaction and connection relationship between nodes, information reflecting the behavior of bridge components, such as displacements, bending moments, and shear forces, is generated.

[0086] As can be seen from the above, by using a graph neural network as the message passing layer, in a single message passing layer, the information of each node is generated through the information generation stage, the received messages are aggregated through an aggregation function, and the embedding of the node is updated through an update function. At the same time, an adaptive message passing layer number mechanism is introduced to achieve information interaction and feature update between nodes, so as to capture the interaction between nodes and transfer structural features. Through message passing, the effective transfer and sharing of information between nodes are realized. Through the aggregation and update functions, the embedding state of the node is updated in real time, reflecting the changes in the structure. And the adaptive message passing layer number mechanism ensures the efficiency and accuracy of information propagation, avoids over-computation, can be flexibly adjusted according to different bridge types, and enhances the flexibility of the model.

[0087] Embodiment 3:

[0088] Reference Figure 1 - Figure 4 As shown, Step 1: Construct the structural diagram representation of the bridge span: Each span of the bridge is used as a node in the graph, and structural elements such as connecting beams and bearings between spans are used as the edges between nodes to form the structural diagram representation of the bridge; each node contains the geometric attributes, material properties, and seismic load information of the span as input features;

[0089] Step 2: Encoder processing: Use the encoder to convert the input features of each node into a higher-dimensional hidden embedding layer v i through a linear single-layer perceptron (SLP) to provide appropriate input for subsequent information transfer and structural response prediction:

[0090]

[0091] Step 3: Message passing layer processing: Use a graph neural network (GNN) as the message passing layer. In a single message passing layer, the information of each node is generated through the information generation stage, and this information takes into account the current embedding of the node, the embeddings of adjacent nodes, and the features of adjacent edges;

[0092]

[0093] where i is the node, v i is the hidden embedding layer, j is the adjacent node, and e i,j is the adjacent edge;

[0094] Aggregate the received messages through an aggregation function and update the embedding of the node through an update function. Among them, both the information function and the update function are constructed through a linear single-layer perceptron (SLP). The generated information is aggregated by calculating the average value, and its calculation formula is:

[0095]

[0096] The construction formula of the linear single-layer perceptron (SLP) is as follows:

[0097]

[0098] Introduce an adaptive message passing layer mechanism to dynamically adjust the number of message passing layers according to the span and complexity of the bridge;

[0099] Step 4, decoder processing: Use the decoder to convert the hidden embedding of the node into the required output through a non-linear multi-layer perceptron (MLP), including displacements (in the X and Y directions), bending moments (in the Y and Z directions), and shear forces (in the Y and Z directions). The decoding formula is as follows:

[0100] MLP(x) = SLP2(max(0, SLP1(x)))

[0101] f Response (v i ) = MLP Response (v i );

[0102] Step 5, model training and optimization: Prepare a dataset containing the characteristics of the bridge structure and seismic response labels, and optimize the model parameters by minimizing the loss function to achieve accurate prediction of the seismic performance of the bridge;

[0103] The hidden layer in the decoder uses the rectified linear unit (ReLU) activation function to improve the expressive ability of the model.

[0104] As can be seen from the above, by using the decoder to convert the hidden embedding of the node into the required output through a non-linear multi-layer perceptron, including displacements, bending moments, and shear forces, the prediction of the structural response is realized, thereby converting the hidden embedding of the node into specific seismic performance indicators, capable of predicting multiple seismic performance indicators, meeting the requirements of bridge seismic performance assessment. The output prediction through the non-linear multi-layer perceptron improves the prediction accuracy. By preparing a dataset containing the characteristics of the bridge structure and seismic response labels, and optimizing the model parameters by minimizing the loss function, the accurate prediction of the seismic performance of the bridge is achieved, thereby optimizing the model parameters, improving the prediction accuracy, enabling the model to accurately predict the seismic performance of the bridge. The trained model can adapt to different types of bridge structures and has strong generalization ability.

[0105] Furthermore, the design of the present invention can be widely applied to the following fields:

[0106] Bridge seismic design: Provide an efficient seismic performance analysis tool for bridge designers to help design bridge structures that meet seismic codes.

[0107] Bridge seismic inspection and assessment after earthquake: Conduct rapid and accurate seismic performance assessment of bridges to support post-disaster emergency response.

[0108] Bridge health monitoring: Continuously monitor bridges using sensor data to detect potential seismic performance problems and ensure the long-term safety of bridges.

[0109] Intelligent transportation system: In intelligent transportation management, real-time monitor the seismic performance of bridges to ensure traffic safety.

[0110] With the new structure graph representation and message passing mechanism, the BridgeGNN framework proposed by the present invention provides an efficient, accurate and flexible solution for real-time analysis of bridge seismic performance, which is particularly suitable for seismic performance assessment of complex bridge structures and provides important support for post-disaster emergency response. Its strong generalization ability and real-time analysis ability make it have broad application prospects in aspects such as bridge design, health monitoring and post-disaster assessment;

[0111] Use graph neural network to conduct real-time analysis of the seismic performance of bridge structures. During the implementation of the model, the inter-span structural characteristics of the bridge and the adaptive message passing mechanism are combined to provide efficient and accurate seismic performance prediction. First, it is necessary to model the bridge structure to be analyzed. During this modeling process, each span of the bridge is used as a node in the graph, and the structural elements between the spans (such as beams, bearings, etc.) are used as the edges between the nodes. Each inter-span node will be assigned a set of features, including the geometric dimensions of the span, material properties, seismic load information, etc.;

[0112] Among them, the features of each span node include geometric characteristics, such as the length, width, height, etc. of the span, material characteristics, such as the elastic modulus and density of the beam, and load information, including external excitations such as seismic loads. These features will be used as inputs and passed to the graph neural network to participate in subsequent feature processing and analysis;

[0113] During the construction of the graph, the bridge inter-span nodes are connected by edges to form a graph, and each node interacts with its adjacent nodes through the edges. The features of the edges include the connection method, material, force information, etc. between two spans;

[0114] In this implementation, the input features of each span node are linearly transformed through a single-layer perceptron (SLP) to obtain a hidden representation, that is, the high-dimensional features of the node.

[0115] Specifically, for the i-th node, the input feature is represented as x i , and after the linear mapping of the encoder, the hidden representation v i ;

[0116] In BridgeGNN, the number of message passing layers is adaptive and depends on the number of spans of the bridge. Specifically, bridges with more spans require more layers of message passing to ensure that information can be transmitted across the entire graph and updated effectively;

[0117] The task of the decoder is to convert the node features updated by the message passing layer into predicted seismic performance indicators. Each predicted indicator is generated by a separate multi-layer perceptron (MLP). Similarly, other seismic performance indicators (such as bending moment, shear force, etc.) are also predicted by the corresponding decoders.

[0118] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time analysis method for bridge seismic performance based on graph neural network, characterized in that: The steps include: Step 1: Construct a structural diagram representation of the bridge span: take each span of the bridge as a node in the diagram, and the connecting beams, supports and other structural elements between spans as the edges between nodes to form a structural diagram representation of the bridge; each node contains the geometric properties, material properties and seismic load information of the span as input features; Step 2: Encoder processing: The encoder is used to convert the input features of each node into a higher-dimensional hidden embedding layer v through a linear single-layer perceptron (SLP). i , providing suitable input for subsequent information transfer and structural response prediction: Step 3, message passing layer processing: Graph neural network (GNN) is used as the message passing layer. In a single message passing layer, the information of each node is generated through the information generation phase, which takes into account the current embedding of the node, the embedding of adjacent nodes, and the characteristics of adjacent edges; Among them, i is the node, v i is the hidden embedding layer, j is the adjacent node, e i,j are adjacent edges; The received messages are aggregated through the aggregation function, and the embedding of the nodes is updated through the update function. Both the information function and the update function are constructed through a linear single-layer perceptron (SLP). The generated information is aggregated by calculating the average value, and the calculation formula is: The linear single-layer perceptron (SLP) construction formula is: Introducing an adaptive message passing layer mechanism to dynamically adjust the number of message passing layers according to the span number and complexity of the bridge; Step 4, decoder processing: The decoder is used to convert the hidden embedding of the node into the required output through a nonlinear multi-layer perceptron (MLP), including displacement (X and Y directions), bending moment (Y and Z directions), and shear force (Y and Z directions). The decoding formula is: MLP(x)=SLP2(max(0,SLP1(x))) f Response (v i )=MLP Response (v i ); Step 5: Model training and optimization: Prepare a data set containing bridge structural characteristics and seismic response labels, optimize the model parameters by minimizing the loss function, and achieve accurate prediction of the seismic performance of the bridge.

2. The real-time analysis method of bridge seismic performance based on graph neural network according to claim 1 is characterized by: The bridge span structural diagram representation also includes the geometric relationship between spans, material properties and seismic load information as input features of the nodes.

3. The real-time analysis method of bridge seismic performance based on graph neural network according to claim 1 is characterized by: The adaptive message transmission layer mechanism automatically adjusts the number of message transmission layers according to the span number and complexity of the bridge to ensure that the information can be fully propagated to the entire bridge structure.

4. The real-time analysis method of bridge seismic performance based on graph neural network according to claim 1 is characterized by: The information generation stage in the message passing layer generates information reflecting the behavior of bridge components, such as displacement, bending moment and shear force, by considering the interaction and connection relationship between nodes.

5. The real-time analysis method of bridge seismic performance based on graph neural network according to claim 1 is characterized by: The hidden layer in the decoder uses a rectified linear unit (ReLU) activation function to improve the expressiveness of the model.

6. A real-time analysis system for bridge seismic performance based on graph neural network, characterized in that: include: Data input module: used to input the span structure information of the bridge, including the span geometry, material properties and seismic load information; Encoder module: used to convert the input node features into a higher-dimensional hidden embedding layer; Message transmission layer module: including information generation unit and node update unit, used to realize information interaction and feature update between nodes, and introduce adaptive message transmission layer number mechanism; Decoder module: used to convert the hidden embeddings of nodes into the required outputs, including seismic responses such as displacement, bending moment and shear force; Model training and optimization module: used to prepare data sets, train models, and optimize model parameters to achieve accurate prediction of bridge seismic performance.

7. The real-time analysis system for bridge seismic performance based on graph neural network according to claim 6 is characterized by: The message delivery layer module also includes an aggregation function and an update function for aggregating received messages and updating the embedding of nodes.

8. The real-time analysis system for bridge seismic performance based on graph neural network according to claim 6 is characterized by: The data input module is also used to input the span geometry relationship, material properties and earthquake load information of the bridge.

9. The real-time analysis system for bridge seismic performance based on graph neural network according to claim 6 is characterized by: The hidden layers in the decoder module use the rectified linear unit (ReLU) activation function.

10. The real-time analysis system for bridge seismic performance based on graph neural network according to claim 6 is characterized by: The system also includes a model reasoning and seismic performance prediction module, which is used to generate seismic performance prediction results through an encoder, a message passing layer and a decoder under a given bridge structure input.

Images