Bridge structure state assessment method based on health monitoring sequence and graph neural network

Abstract

The application discloses a bridge structure state evaluation method and system based on health monitoring sequences and a graph neural network, and the method comprises the following steps: constructing an adaptive bridge deck graph structure according to the characteristics of a bridge structure to be evaluated, inputting the node vehicle load state characteristics into a bridge mechanics analysis model for transient simulation, obtaining a load and structure response simulation dataset, and obtaining training input data through time window segmentation. A graph neural network model with an encoder-decoder architecture is constructed, and a graph attention mechanism and a time sequence feature extraction mechanism are integrated. The model is trained in stages, the encoder is first trained to obtain load inference capability, and then the parameters of the encoder are frozen to train the decoder to form a bridge proxy model. Finally, the model parameters are frozen, a real structure response monitoring sequence is input, a response prediction value is deduced, and whether the bridge structure state is abnormal is determined in real time by comparing the residual error with a health threshold. The bridge structure health state is rapidly, accurately and timely evaluated.

Description

Technical Field

[0001] This invention relates to the fields of structural health monitoring technology and signal processing technology, and more specifically, to a method and system for assessing the structural condition of bridges based on health monitoring sequences and graph neural networks. Background Technology

[0002] Urban small and medium-sized bridges are large in scale, diverse in type, span a wide range of construction periods, and vary greatly in safety standards, posing a significant challenge to the rapid assessment of their structural condition. Vehicle loads, as the primary load on urban small and medium-sized bridge structures, make their condition assessment a key focus of urban small and medium-sized bridge structural health monitoring.

[0003] Assuming the bridge structure conforms to a linear time-invariant system, if vehicle load conditions remain consistent, the response signals generated by a healthy structure will exhibit a certain statistical correlation. This correlation pattern can be used to reflect the structure's own state. When damage occurs to the bridge structure, its operating state will change, and the correlation characteristics in the original response data will also change accordingly. Under this premise, using a model trained under the bridge's healthy state to predict the structural response, the bridge's structural health state can be determined by the magnitude of the residual between the predicted and actual responses.

[0004] Bridge structural response sequences exhibit strong spatiotemporal correlations. Graph structures, due to their powerful representation capabilities and flexibility for non-Euclidean data, make graph neural networks the preferred surrogate model for extracting spatial correlations in structural response sequences. Each node in the graph iteratively updates its feature representation by aggregating the feature information of its neighbors. Through multiple iterations, the final representation obtained by each node captures the topological structure and feature information of its local neighborhood, thus characterizing the spatial correlation of response data in complex engineering structures. The Transformer, with its self-attention mechanism, can efficiently capture complex temporal dependencies in structural response sequences. This mechanism dynamically constructs global temporal dependencies by calculating the association weights between response data at any two time steps in the sequence, without relying on fixed time windows or recursive calculations, thereby comprehensively improving the ability to represent the dynamic behavior of structures. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of high latency, low accuracy, and insensitivity to damage response in traditional structural health monitoring methods, and to achieve rapid, accurate, and timely assessment of structural health status based on response sequences.

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for assessing the structural condition of bridges based on health monitoring sequences and graph neural networks, comprising: S1. For the bridge to be evaluated, construct a suitable bridge deck structure based on its structural characteristics; S2. Based on the mechanical analysis model of the bridge structure, the nodal vehicle load state characteristics of the bridge deck structure are input into the model for transient simulation analysis to obtain the simulation dataset corresponding to the vehicle load and structural response. The dataset is then divided into time windows to obtain the input data required for model training. S3. Construct a graph neural network model with an encoder-decoder architecture, wherein the encoder is used to infer the vehicle load state in future time periods based on historical structural response sequences, and the decoder is used to generate structural response sequences for corresponding time periods based on the inferred vehicle load state; the model integrates a graph attention mechanism to capture the spatial correlation of structural responses and integrates a temporal feature extraction mechanism to capture the temporal dependence of structural responses. S4. In the first stage, the encoder is trained based on the vehicle load-related data in the simulation dataset, so that the encoder has the ability to infer the load state; in the second stage, the encoder parameters are frozen, and the decoder is trained based on the structural response-related data in the simulation dataset, so that the decoder forms a surrogate model of the bridge structure and has the ability to generate response sequences. S5. Freeze the parameters of the trained graph neural network model, input the actual structural response monitoring sequence data of the bridge into the model, and use the decoder to infer the predicted value of the structural response in future periods; calculate the residual between the predicted response value and the actual monitoring value, and based on the preset residual threshold range under healthy conditions, determine in real time whether the bridge structural state has undergone abnormal changes.

[0007] Furthermore, in the bridge deck diagram structure in S1, nodes are evenly distributed and cover the entire bridge deck in the length and width directions. The number of nodes in the width direction is equal to the number of lanes, and the node spacing in the length direction is set to 3 to 10 meters. In the bridge deck diagram structure, adjacent nodes are connected by undirected edges to represent the interaction relationship between nodes.

[0008] Furthermore, the features of the nodes in S1 include the node's deflection response and vehicle load state. Based on the vehicle load sequence data and the vehicle's position on the bridge deck, a node vehicle load state feature is constructed for each moment, with a matrix size of [missing information]. For nodes and times where no vehicles are involved, the elements of the matrix are set to 0; in, The number of nodes along the bridge direction. To determine the number of transverse nodes, the bridge deck is evenly distributed along both the longitudinal and transverse directions. Each node is connected to its adjacent nodes in the same row and column by undirected edges. Furthermore, the time window in S2 uses a dimension of no less than 20.

[0009] Furthermore, the transient simulation analysis process in S2 is as follows: Based on the structural characteristics of the bridge to be evaluated and the node interconnection logic of the bridge deck structure, the vehicle load state characteristics of each node are transformed into the dynamic excitation input of the mechanical analysis model. By establishing the dynamic equilibrium equation of node displacement and load action, the time evolution law of structural response under instantaneous load action is characterized. The dynamic equilibrium equation is expressed as the sum of the second derivative of node deflection response with respect to time, the first derivative of node deflection response with respect to time, and node deflection response equals the node load. By combining the undirected edge relationships between adjacent nodes in the bridge deck structure, a response transfer model between nodes is constructed to map the mechanical relationships in the spatial dimension. The response transfer model between nodes is expressed as the deflection response of any node at a specific time being equal to the sum of the products of the deflection responses of all its adjacent nodes at that specific time and the corresponding structural mechanical transfer relationship function. The simulation process is constrained by the principle of energy conservation in structural dynamics to ensure that the dynamic matching of load and response conforms to the actual stress characteristics of the bridge structure. The principle of energy conservation states that the total energy of the system is equal to the sum of kinetic energy, potential energy and dissipated energy. By solving the above dynamic equilibrium equations, the inter-node response transfer model, and the equations corresponding to the energy conservation principle, the corresponding data of the load state and structural response of each node under the full time sequence are obtained, and time sequence data units are formed by dividing the data into a unified time window.

[0010] Furthermore, the encoder architecture in S3 includes an input feedforward layer, a graph attention convolutional layer, a temporal convolutional layer, and an output feedforward layer. The input feedforward layer performs feature upsizing on the input bridge deck node time series data; The graph attention convolutional layer contains a multi-layer graph attention network. By calculating the attention weights between the current node and its neighboring nodes, the features of the neighboring nodes are weighted and aggregated to extract the spatial topological features of the bridge structure. The temporal convolutional layer comprises multiple layers of one-dimensional dilated convolutional networks to perform convolution processing on the spatial topological features in order to extract long-range temporal evolution features. The output feedforward layer performs dimensionality reduction on the extracted spatiotemporal features and outputs the vehicle state prediction results for each node.

[0011] Furthermore, in the decoder of S3, the input first passes through an input feedforward layer for dimensionality increase, then through a three-layer GAT structure. The output of the GAT enters a Transformer decoder structure; the Transformer decoder receives two sets of inputs: and ; in, The node features output by the encoder after processing by the GAT layer have the following shape: ,in For batch size, For the number of nodes, ; It is obtained by linearly mapping the deflection data of the next 10 observed moments, and its shape is also... ; The Transformer decoder contains two standard multi-head cross-attention modules; each layer consists of a module that performs masked self-attention on the query sequence, a module that performs cross-attention between the query and memory, a feedforward neural network, and residual connections and layer normalization for each sub-layer.

[0012] Furthermore, the process of determining whether the bridge structural state has undergone abnormal changes in real time in S4 is as follows: By analyzing the residuals between the measured sequence data and the predicted sequence data of the bridge structure response under healthy conditions, and setting the mean of the reconstruction residuals of the model under healthy conditions to be denoted as 𝑚 and the standard deviation as 𝜎, a rapid assessment of the health status of the bridge structure is carried out using 𝑚±5𝜎 as the upper and lower thresholds.

[0013] As a second aspect of the present invention, a bridge structural condition assessment system based on health monitoring sequences and graph neural networks is also provided, comprising: Bridge deck diagram structure division unit is used to construct an appropriate bridge deck diagram structure based on the structural characteristics of the bridge to be evaluated. The dataset establishment unit is used for mechanical analysis models based on bridge structures. It inputs the nodal vehicle load state characteristics of the bridge deck structure into the model for transient simulation analysis, obtains the simulation dataset corresponding to vehicle load and structural response, and performs time window segmentation on the dataset to obtain the input data required for model training. The learning model building unit is used to construct a graph neural network model with an encoder-decoder architecture. The encoder is used to infer the vehicle load state in future time periods based on historical structural response sequences, and the decoder is used to generate structural response sequences for corresponding time periods based on the inferred vehicle load state. The model integrates a graph attention mechanism to capture the spatial correlation of structural responses and a temporal feature extraction mechanism to capture the temporal dependence of structural responses. The model training unit is used in the first stage to train the encoder based on the vehicle load-related data in the simulation dataset, so that the encoder has the ability to infer the load state; in the second stage, the encoder parameters are frozen and the decoder is trained based on the structural response-related data in the simulation dataset, so that the decoder forms a surrogate model of the bridge structure and has the ability to generate response sequences. The structural state diagnosis unit is used to freeze the parameters of the trained graph neural network model, input the actual structural response monitoring sequence data of the bridge into the model, and use the decoder to infer the predicted value of the structural response for future periods; calculate the residual between the predicted response value and the actual monitoring value, and determine in real time whether the bridge structural state has undergone abnormal changes based on the preset residual threshold range under healthy conditions.

[0014] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor to provide a bridge structural condition assessment method based on health monitoring sequences and graph neural networks as described in any one of the claims.

[0015] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The bridge structural state assessment method based on health monitoring sequences and graph neural networks of the present invention takes into account the actual stress state of the bridge while using a data-driven rapid deduction method of "response-response". It aligns the learned features with vehicle loads in the latent space and has interpretability in terms of the mechanical properties of civil engineering structures.

[0016] 2. The bridge structural status assessment method based on health monitoring sequences and graph neural networks of the present invention, through training based on structural response monitoring sequence data, uses a self-supervised training method to ensure that the diagnosis of structural health status does not depend on other conditions or constraints.

[0017] 3. The bridge structural condition assessment method based on health monitoring sequences and graph neural networks of this invention adopts a two-stage training strategy of pre-trained encoder + trained decoder, and is trained based on simulation datasets. Furthermore, the bridge deck diagram structural representation method is applicable to various types of bridges and has strong versatility. With the help of finite element data simulation, it can be quickly extended to other structures, realizing the transfer and application of the method. Attached Figure Description

[0018] Figure 1 This is a flowchart of a bridge structure status assessment method based on health monitoring sequences and graph neural networks according to an embodiment of the present invention. Figure 2 This is a flowchart of the bridge structural condition assessment method according to an embodiment of the present invention; Figure 3 This is a model architecture diagram of TC-GAT according to an embodiment of the present invention; Figure 4 This is a timing diagram of the damage detection state according to an embodiment of the present invention; Figure 5 This is a graph showing the relative error of response prediction in an embodiment of the present invention. Figure 6This is a system unit diagram of an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Example 1 Please refer to Figure 1 This embodiment 1 provides a bridge structural status assessment method based on health monitoring sequences and graph neural networks, including: S1. For the bridge to be evaluated, construct a suitable bridge deck structure based on its structural characteristics; S2. Based on the mechanical analysis model of the bridge structure, the nodal vehicle load state characteristics of the bridge deck structure are input into the model for transient simulation analysis to obtain the simulation dataset corresponding to the vehicle load and structural response. The dataset is then divided into time windows to obtain the input data required for model training. S3. Construct a graph neural network model with an encoder-decoder architecture, wherein the encoder is used to infer the vehicle load state in future time periods based on historical structural response sequences, and the decoder is used to generate structural response sequences for corresponding time periods based on the inferred vehicle load state; the model integrates a graph attention mechanism to capture the spatial correlation of structural responses and integrates a temporal feature extraction mechanism to capture the temporal dependence of structural responses. S4. In the first stage, the encoder is trained based on the vehicle load-related data in the simulation dataset, so that the encoder has the ability to infer the load state; in the second stage, the encoder parameters are frozen, and the decoder is trained based on the structural response-related data in the simulation dataset, so that the decoder forms a surrogate model of the bridge structure and has the ability to generate response sequences. S5. Freeze the parameters of the trained graph neural network model, input the actual structural response monitoring sequence data of the bridge into the model, and use the decoder to infer the predicted value of the structural response in future periods; calculate the residual between the predicted response value and the actual monitoring value, and based on the preset residual threshold range under healthy conditions, determine in real time whether the bridge structural state has undergone abnormal changes.

[0021] This embodiment 1 considers the structural response of bridge structures under vehicle loads and proposes the TC-GAT model. This model treats vehicle load information as a physical condition in the latent space, dividing the model into an encoder and a decoder. The encoder can infer the vehicle load conditions for future periods from existing bridge deck response sequences, while the decoder can generate the corresponding bridge deck response sequences based on the vehicle load conditions inferred by the encoder. Thus, the decoder acts as a surrogate model for the bridge structure, and its generated predicted structural response sequences can be used to assess the health status of the bridge structure in real time.

[0022] The method in this embodiment 1 introduces a bridge deck diagram structure to represent non-Euclidean data. The bridge deck topology is modeled by nodes and undirected edges, and the structural load and response sequence are embedded in the node features. Through a self-supervised training mode, based on finite element simulation data and bridge deck diagram structure representation, it can be quickly extended to the condition assessment of different types of bridge structures.

[0023] Next, please refer to Figure 2 This embodiment 1 further elaborates on the above steps.

[0024] (1) Bridge deck structure division First, for the bridge structure to be analyzed, its corresponding finite element model is constructed. The division of the bridge deck structure is based on the mesh division of the bridge deck in the bridge finite element model. The specific division can be determined according to the material, construction and number of lanes of the specific bridge structure. The nodes are evenly distributed in the length and width directions of the bridge deck, and the number of nodes in the width direction is equal to the number of lanes. If the number of lanes is small, more nodes can be divided in the width direction. The node spacing in the length direction is set to 3 to 10 meters. The difference in spacing between each node should not be too large to avoid affecting the model's learning of spatial relationships. Since each connected node in the bridge deck structure will affect each other, the edges in the graph structure are all undirected edges.

[0025] The bridge deck is evenly distributed along the longitudinal and transverse directions. There are nodes, among which The number of nodes along the bridge direction. The number of nodes in the transverse direction of the bridge is represented by an undirected edge connecting adjacent nodes in the same row and column. Node features include the node's deflection response and vehicle load status. Based on the vehicle load sequence data and the vehicle's position on the bridge deck, a node vehicle load status feature matrix is ​​constructed for each time step, with a matrix size of [missing value]. For nodes and times where no vehicles are involved, the elements of the matrix are set to 0.

[0026] (2) Dataset creation Based on the constructed finite element model of the bridge structure, existing nodal vehicle load state characteristics are input into the finite element model for transient analysis and simulation, obtaining data of similar magnitude. The structural response simulation dataset was used and segmented using a time window of dimension 20, which served as the input to the encoder.

[0027] In a preferred embodiment, the transient simulation analysis process is as follows: Based on the structural characteristics of the bridge under evaluation and the node interconnection logic of the bridge deck structure, the vehicle load state characteristics (including vehicle existence state and action sequence) of each node are transformed into dynamic excitation inputs for the mechanical analysis model. By establishing the dynamic equilibrium equations of node displacement and load action, the time evolution law of structural response under instantaneous load action is characterized. in, The time series function of nodal deflection response; Indicates nodal deflection response Regarding time The second derivative of the nodal deflection is the acceleration of the nodal deflection. Indicates nodal deflection response Regarding time The first derivative, i.e., the velocity of nodal deflection; For nodal load time sequence functions; By combining the undirected edge relationships between adjacent nodes in the bridge deck diagram structure, a response transfer model between nodes is constructed to map the mechanical relationships in the spatial dimension: in, To represent nodes In time Deflection response at time t, To represent nodes In time Deflection response at time t, For nodes The set of adjacent nodes, For nodes and The structural mechanics transfer relationship function is used to characterize the deflection response transfer law between adjacent nodes of a bridge structure under load. Its construction is based on the structural dynamics model and the principle of energy conservation. The simulation process is constrained by the principle of energy conservation in structural dynamics to ensure that the dynamic matching of load and response conforms to the actual stress characteristics of the bridge structure. in, Let be the total energy function. Let be the kinetic energy function. Let be the potential energy function. Let be the dissipation energy function; By solving the above series of equations, the corresponding data of the load state and structural response of each node under the full time sequence are obtained, and time sequence data units are formed by dividing the data into a unified time window.

[0028] Then, the measured sequence data of vehicle load is input into the finite element model to calculate the structural response sequence data corresponding to the nodes of the entire bridge. The number of structural response sensors for the entire bridge is usually less than the number of nodes. Therefore, during the pre-training of the encoder, only the nodes with structural response sensors have response data input into the model, while the node responses at other nodes are set to 0. The complete response sequence including all nodes can be used to train the decoder.

[0029] (3) Learning model establishment The core objective of decoupled representation learning is to divide the potential variable factors in data into independent, interpretable parts. Using this representation, relevant machine learning models can not only identify interpretable factors in the data structure but also achieve stronger transfer and generalization capabilities across different tasks, thus providing a robust technical path for bridge structure condition assessment. Traditional statistical analysis-based methods often fall short in terms of correlation accuracy and temporal accuracy of structural responses. Therefore, correlation modeling can be directly performed using structural response monitoring sequences, and physical constraints conforming to vehicle load characteristics can be introduced into the latent space of decoupled representation learning to guide the learning process, enabling the model to effectively extract latent representations corresponding to vehicle loads from the monitoring data. In other words, structural condition diagnosis and response correlation analysis can be uniformly understood as an "inference-generation" framework: first, the acting vehicle load is inferred from the structural response monitoring sequence, and then the corresponding response sequence is reconstructed based on this load. This generation process essentially constitutes a surrogate model for the bridge structure.

[0030] The TC-GAT model architecture in this implementation 1 consists of two parts: an encoder and a decoder; the encoder receives input... Bridge deck deflection sequence at these 20 moments Output Bridge vehicle load status at these 10 moments (Including vehicle position and weight); the decoder takes the input encoder output... The bridge deck vehicle load status at these 10 moments, output Bridge deck deflection sequence at these 10 moments The TC-GAT model architecture is as follows: Figure 3 As shown.

[0031] In this embodiment 1, both the encoder and decoder include a three-layer GAT (Graph Attention Network) structure. In addition, the encoder includes a two-layer one-dimensional dilated temporal convolutional layer, and the decoder includes a Transformer decoder structure to extract temporal features of the response sequence. Specifically, the encoder architecture consists of an input feedforward layer, a GAT convolutional layer, a temporal convolutional layer, and an output feedforward layer; the input feedforward layer, as a single-layer perceptron (SLP) structure, first receives data in dimension... Input, This represents the number of structural nodes in the bridge deck diagram. The time window size is set to 20. The feedforward layer increases the time dimension to 256, and then the input enters the GAT convolutional layer. The encoder's GAT convolutional layer contains a three-layer graph attention network to construct a hierarchical feature encoding structure. The first layer has an input dimension of 256 and an output dimension of 128, and uses a 4-head attention mechanism with a Dropout rate of 0.1 to enhance generalization ability. The second layer expands the input dimension to 512 and increases the output dimension to 256, and achieves feature fusion through a 2-head attention mechanism. The third layer maintains an input dimension of 512 and an output dimension of 256, and uses a single-head attention mechanism to focus on key topological relationships. On each floor In the middle, for nodes and neighbors Through a shared linear transformation (weight matrix) Transform the features and then calculate the attention score: in It is a learnable parameter vector. This indicates a concatenation operation, where the attention score is normalized using softmax to obtain the weight coefficients. : The representation of the current node is updated by summing the weighted features of adjacent nodes. The model's expressive power is enhanced through a multi-head attention mechanism, and its outputs are concatenated: A two-layer one-dimensional dilated temporal convolutional layer was designed after the GAT layer; firstly, the first convolutional layer receives 256-dimensional data from the output of the GAT layer, with the input channel being the number of nodes. Output channels expanded to It uses a convolutional kernel of size 3 with a dilation rate of 2; the output channels of the second convolutional layer are compressed back. To capture long-range temporal patterns, the same convolution parameters are maintained; each convolutional layer is followed by batch normalization (BNOR1d) and the GELU activation function to ensure gradient stability while introducing nonlinear modeling capabilities; finally, an input-output feedforward layer is used for dimensionality reduction, resulting in a final output with the shape of... The vehicle state prediction is used as the output of the encoder.

[0032] In the decoder, the data first goes through an input feedforward layer for dimensionality increase, then through a three-layer GAT structure. The output of the GAT enters a Transformer decoder structure; the Transformer decoder receives two sets of inputs: and ,in, The node features output by the encoder after processing by the GAT layer have the following shape: Where B is the batch size, N is the number of nodes, and d=256; It is obtained by linearly mapping the deflection data of the next 10 observed moments, and its shape is also... ; First, the input is fed into the Transformer decoder. sequence It will perform a linear mapping and combine it with learnable positional encoding. This assigns unique location information to each node: Position-encoded target sequence It will be used as the query input to the attention layer of the Transformer decoder.

[0033] The Transformer decoder contains two standard multi-head cross-attention modules; each layer consists of a module that performs masked self-attention on the query sequence, a module that performs cross-attention between the query and memory (both are 4-head attention), a feedforward neural network (FFN), and residual connections and layer normalization for each sub-layer.

[0034] In the decoder, the masked self-attention mechanism is responsible for modeling the temporal dependencies within the target sequence. To prevent the leakage of future information during training, a causal mask is used when calculating the attention score. The input is known. ,in For batch size, For the number of nodes, If the feature dimension is used, the calculation process of the masked self-attention mechanism is as follows: No. The output of each attention head is: in That is, causalmask, which is defined as follows: Causalmask ensures that each time step i can only focus on itself and the time steps preceding it. After the outputs of all attention heads have been computed, they are concatenated and linearly mapped to obtain the output: In the Transformer decoder, another attention mechanism—cross-attention—allows the GATT decoder to perform conditional predictions based on the encoder output (i.e., the predicted vehicle weight memory). The query is known... Memory represents the vehicle weight features output from the encoder after processing by the decoder's GAT layer. The query (Q), key (K), and value (V) matrices are calculated as follows: Attention score and output are calculated as follows: The final multi-head cross-attention output is: The cross-attention module enables the deflection prediction of each node to dynamically focus on vehicle weight information at certain key nodes and time steps, thereby improving the model's ability to model causal relationships.

[0035] In summary, the calculation process for the attention layer is as follows: in, All attention modules employ a 4-head multi-head attention mechanism, and the hidden layer dimension of the feedforward network is 2D. Finally, the attention layer outputs... The input is fed into a multilayer perceptron (MLP) to predict future deflection values. Finally, the dimension is restored to the predicted time step T by outputting a feedforward layer: (4) Model training To improve the model's generalization performance and convergence speed, the deflection data was processed. Normalize the data to transform it into a distribution with a mean of 0 and a standard deviation of 1. Perform maximum value normalization on the vehicle weight data to compress non-negative values ​​and prevent large values ​​from causing numerical instability during model training.

[0036] To enable the model to learn the normal distribution of the data during the pre-trained encoder stage, cope with interference, prevent overfitting, and improve robustness and generalization ability, a series of data augmentation processes were also performed on the response dataset. These included random masking, randomly setting 10% of the deflection data to zero, and injecting normally distributed noise with a standard deviation of 0.05.

[0037] Based on this, a loss function is used in the latent space between the encoder and decoder based on vehicle load data. Pre-trained encoder; loss function used based on encoder output and structural response dataset. Training the decoder.

[0038] Encoder predicts vehicle state matrix , For the predicted number of time points, Let be the total number of nodes in the bridge map structure. When a vehicle passes a node at a specific time, the corresponding element in the matrix represents the vehicle's weight; otherwise, the element is 0. The pre-training objective of the encoder is to minimize the vehicle matrix reconstruction error. in, and These represent the actual vehicle state matrix and the inferred vehicle state matrix, respectively. When dealing with large errors, use a loss function similar to the absolute error (MAE) to reduce the impact of outliers and avoid training instability that may result from excessively large gradients.

[0039] The decoder takes the vehicle state matrix output by the encoder as input. Inferring the nodal deflection response at the corresponding time point Therefore, the decoder structure can be viewed as a proxy model of the bridge structure. The decoder training objective is to minimize the response sequence reconstruction error. .

[0040] (5) Structural condition diagnosis All parameters are frozen, and future response values ​​are gradually extrapolated based on real structural response monitoring sequence data. The changes in the bridge structural state are inferred in real time based on the residual between the extrapolated values ​​and the actual values.

[0041] Specifically, by analyzing the residuals between the measured sequence data of the bridge structural response under healthy conditions and the sequence data predicted by TC-GAT, the mean of the reconstruction residuals of the model under healthy conditions is set to... The standard deviation is Then Using upper and lower thresholds, a rapid assessment of the structural health status of bridges is conducted.

[0042] The method in Example 1 performed well in a bridge damage identification experiment. The example simulates the decrease in stiffness caused by localized damage to the bridge structure by reducing the elastic modulus of the materials used in certain beam segments of the bridge finite element model. The deflection data used in the example totaled 30,000 s, divided into three damage time periods: 5001–6000 s, 15001–17000 s, and 25001–28000 s. The remaining time periods were considered healthy periods to verify whether the method described in this invention can effectively distinguish between structural damage and healthy states. The example set the mean of the reconstruction residuals of the model in the healthy state to R and the standard deviation to S. Using R ± 5 S as upper and lower thresholds, the bridge structural health status was assessed, achieving an accuracy of 99.8%. The results are as follows: Figure 4 as well as Figure 5 As shown.

[0043] The method in this embodiment 1 uses an index based on predicted residuals for structural state assessment. The steps include: dividing the bridge deck structure and constructing node features; establishing a simulation dataset of vehicle load-structural response; building a model, namely a graph neural network-based encoder-decoder decoupled representation learning model TC-GAT; first training the encoder part of TC-GAT based on the vehicle load simulation dataset, then freezing the encoder parameters, and training the decoder part of TC-GAT based on the structural response simulation dataset; performing structural state diagnosis, freezing all parameters, and gradually extrapolating future response values ​​based on real structural response monitoring sequence data; and inferring changes in the bridge structural state in real time based on the residuals between the extrapolated and actual values.

[0044] The method provided in Embodiment 1 is essentially a process of reconstructing structural response monitoring sequence data based on graph neural networks. It has two core steps: first, inferring vehicle loads from existing monitoring sequence data; and then regenerating the structural response sequence based on the inferred loads. This generation process serves as a surrogate model for the bridge structure, allowing for the assessment of the bridge components and overall operational status by using the residual between the surrogate model's predicted values ​​and the actual monitored values ​​as an indicator when the structural service state changes. The model proposed in this invention introduces a graph structure to represent the topological information of the bridge deck structure. Both structural loads and response sequences can be embedded with node features, greatly improving the model's representational power and versatility. This invention introduces vehicle loads as physical constraints in the latent space, providing interpretability in the mechanical properties of civil engineering structures. Simultaneously, through efficient modeling of the monitoring sequences, real-time prediction and rapid assessment of the structural response are achieved, effectively overcoming the shortcomings of traditional structural diagnostic methods in terms of interpretability, accuracy, and temporal resolution.

[0045] Example 2 Please refer to Figure 6 This embodiment 2 provides a bridge structural condition assessment system based on health monitoring sequences and graph neural networks, including: The structural division unit of the bridge deck is used to divide the details depending on the materials, construction and number of lanes of the specific bridge structure; the nodes are evenly distributed in the length and width directions of the bridge deck and cover the entire bridge deck. The node characteristics include deflection, vehicle presence and vehicle weight. The dataset creation unit is used to create a simulation dataset for vehicle load-structure response. The learning model building unit is used to build the encoder-decoder decoupled representation learning model TC-GAT based on graph neural networks; The model training unit is used to first train the encoder part of TC-GAT based on the vehicle load simulation dataset, then freeze the encoder part parameters of TC-GAT, and train the decoder part of TC-GAT based on the structural response simulation dataset. The structural state diagnosis unit is used to freeze all parameters, gradually extrapolate future response values ​​based on real structural response monitoring sequence data, and infer changes in the bridge structural state in real time based on the residual between the extrapolated values ​​and the actual values.

[0046] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement any step of a bridge structure status assessment method based on health monitoring sequences and graph neural networks.

[0047] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0048] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0049] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing the structural condition of bridges based on health monitoring sequences and graph neural networks, characterized in that, include: S1. For the bridge to be evaluated, construct a suitable bridge deck structure based on its structural characteristics; S2. Based on the mechanical analysis model of the bridge structure, the nodal vehicle load state characteristics of the bridge deck structure are input into the model for transient simulation analysis to obtain the simulation dataset corresponding to the vehicle load and structural response. The dataset is then divided into time windows to obtain the input data required for model training. S3. Construct a graph neural network model with an encoder-decoder architecture, wherein the encoder is used to infer the vehicle load state in future periods based on historical structural response sequences, and the decoder is used to generate structural response sequences for corresponding periods based on the inferred vehicle load state; the model integrates a graph attention mechanism to capture the spatial correlation of structural responses and integrates a temporal feature extraction mechanism to capture the temporal dependence of structural responses. S4. In the first stage, the encoder is trained based on the vehicle load-related data in the simulation dataset, so that the encoder has the ability to infer the load state. In the second stage, the encoder parameters are frozen, and the decoder is trained based on the structural response-related data in the simulation dataset, so that the decoder forms a proxy model of the bridge structure and has the ability to generate response sequences. S5. Freeze the parameters of the trained graph neural network model, input the actual structural response monitoring sequence data of the bridge into the model, and use the decoder to infer the predicted value of the structural response in future periods; The residual between the predicted response value and the actual monitored value is calculated, and based on the preset residual threshold range under healthy conditions, it is used to determine in real time whether the bridge structure status has undergone abnormal changes.

2. The bridge structural condition assessment method based on health monitoring sequences and graph neural networks according to claim 1, characterized in that, In the bridge deck diagram structure in S1, nodes are evenly distributed and cover the entire bridge deck in the length and width directions. The number of nodes in the width direction is equal to the number of lanes, and the node spacing in the length direction is set to 3 to 10 meters. In the bridge deck diagram structure, adjacent nodes are connected by undirected edges to represent the interaction relationship between nodes.

3. The bridge structural condition assessment method based on health monitoring sequences and graph neural networks according to claim 1, characterized in that, The features of the nodes in S1 include the node's deflection response and vehicle load status. Based on the vehicle load sequence data and the vehicle's position on the bridge deck, a node vehicle load status feature is constructed for each time step, with a matrix size of [missing information]. For nodes and times where no vehicles are involved, the elements of the matrix are set to 0; in, The number of nodes along the bridge direction. To determine the number of transverse nodes, the bridge deck is evenly distributed along both the longitudinal and transverse directions. Each node is connected to its adjacent nodes in the same row and column by undirected edges.

4. The bridge structural condition assessment method based on health monitoring sequences and graph neural networks according to claim 1, characterized in that, The time window in S2 uses a dimension of no less than 20.

5. The bridge structural condition assessment method based on health monitoring sequences and graph neural networks according to claim 1, characterized in that, The transient simulation analysis process in S2 is as follows: Based on the structural characteristics of the bridge to be evaluated and the node interconnection logic of the bridge deck structure, the vehicle load state characteristics of each node are transformed into the dynamic excitation input of the mechanical analysis model. By establishing the dynamic equilibrium equation of node displacement and load action, the time evolution law of structural response under instantaneous load action is characterized. The dynamic equilibrium equation is expressed as the sum of the second derivative of node deflection response with respect to time, the first derivative of node deflection response with respect to time, and node deflection response equals the node load. By combining the undirected edge relationships between adjacent nodes in the bridge deck structure, a response transfer model between nodes is constructed to map the mechanical relationships in the spatial dimension. The response transfer model between nodes is expressed as the deflection response of any node at a specific time being equal to the sum of the products of the deflection responses of all its adjacent nodes at that specific time and the corresponding structural mechanical transfer relationship function. The simulation process is constrained by the principle of energy conservation in structural dynamics to ensure that the dynamic matching of load and response conforms to the actual stress characteristics of the bridge structure. The principle of energy conservation states that the total energy of the system is equal to the sum of kinetic energy, potential energy and dissipated energy. By solving the above dynamic equilibrium equations, the inter-node response transfer model, and the equations corresponding to the energy conservation principle, the corresponding data of the load state and structural response of each node under the full time sequence are obtained, and time sequence data units are formed by dividing the data into a unified time window.

6. The bridge structural condition assessment method based on health monitoring sequences and graph neural networks according to claim 1, characterized in that, The encoder architecture in S3 includes an input feedforward layer, a graph attention convolutional layer, a temporal convolutional layer, and an output feedforward layer; the input feedforward layer performs feature upsizing on the temporal data of the input bridge surface nodes. The graph attention convolutional layer contains a multi-layer graph attention network. By calculating the attention weights between the current node and its neighboring nodes, the features of the neighboring nodes are weighted and aggregated to extract the spatial topological features of the bridge structure. The temporal convolutional layer comprises multiple layers of one-dimensional dilated convolutional networks to perform convolution processing on the spatial topological features in order to extract long-range temporal evolution features. The output feedforward layer performs dimensionality reduction on the extracted spatiotemporal features and outputs the vehicle state prediction results for each node.

7. The bridge structural condition assessment method based on health monitoring sequences and graph neural networks according to claim 1, characterized in that, In the decoder of S3, the input first passes through an input feedforward layer for dimensionality increase, then through a three-layer GAT structure. The output of the GAT enters a Transformer decoder structure; the Transformer decoder receives two sets of inputs: and ; in, The node features output by the encoder after processing by the GAT layer have the following shape: ,in For batch size, For the number of nodes, ; It is obtained by linearly mapping the deflection data of the next 10 observed moments, and its shape is also... ; The Transformer decoder contains two standard multi-head cross-attention modules; each layer consists of a module that performs masked self-attention on the query sequence, a module that performs cross-attention between the query and memory, a feedforward neural network, and residual connections and layer normalization for each sub-layer.

8. The bridge structural condition assessment method based on health monitoring sequences and graph neural networks according to claim 1, characterized in that, The process of determining whether abnormal changes have occurred in the bridge structural state in real time in S4 is as follows: By analyzing the residuals between the measured sequence data and the predicted sequence data of the bridge structure response under healthy conditions, and setting the mean of the reconstruction residuals of the model under healthy conditions to be denoted as 𝑚 and the standard deviation as 𝜎, a rapid assessment of the health status of the bridge structure is carried out using 𝑚±5𝜎 as the upper and lower thresholds.

9. A bridge structural condition assessment system based on health monitoring sequences and graph neural networks, characterized in that, include: Bridge deck diagram structure division unit is used to construct an appropriate bridge deck diagram structure based on the structural characteristics of the bridge to be evaluated. The dataset establishment unit is used for mechanical analysis models based on bridge structures. It inputs the nodal vehicle load state characteristics of the bridge deck structure into the model for transient simulation analysis, obtains the simulation dataset corresponding to vehicle load and structural response, and performs time window segmentation on the dataset to obtain the input data required for model training. The learning model building unit is used to construct a graph neural network model with an encoder-decoder architecture. The encoder is used to infer the vehicle load state in future time periods based on historical structural response sequences, and the decoder is used to generate structural response sequences for corresponding time periods based on the inferred vehicle load state. The model integrates a graph attention mechanism to capture the spatial correlation of structural responses and a temporal feature extraction mechanism to capture the temporal dependence of structural responses. The model training unit is used in the first stage to train the encoder based on the vehicle load-related data in the simulation dataset, so that the encoder has the ability to infer the load state. In the second stage, the encoder parameters are frozen, and the decoder is trained based on the structural response-related data in the simulation dataset, so that the decoder forms a proxy model of the bridge structure and has the ability to generate response sequences. The structural state diagnosis unit is used to freeze the parameters of the trained graph neural network model, input the actual structural response monitoring sequence data of the bridge into the model, and infer the predicted value of the structural response in future periods through the decoder. The residual between the predicted response value and the actual monitored value is calculated, and based on the preset residual threshold range under healthy conditions, it is used to determine in real time whether the bridge structure status has undergone abnormal changes.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor according to any one of claims 1-8, a method for assessing the state of a bridge structure based on health monitoring sequences and graph neural networks.

Images