Road structure health state recognition method based on graph neural network

Abstract

The invention discloses a road structure health state recognition method based on a graph neural network, and the method comprises the following steps: S1, carrying out the monitoring of a road structure through a sensor device, and collecting original data; s2, processing the collected original data; s3, constructing a graph structure; s4, performing graph convolution based on the improved R-GCN model, and optimizing the learning ability of the graph structure by combining a node reconstruction task of self-supervised learning; s5, proposing an optimization strategy for dynamically adjusting the number of graph convolutional layers, automatically adjusting the number of the graph convolutional layers according to the learning performance of the model in the training process, and outputting optimized node features; s6, the node features are input into an XGBoost algorithm, and classification and regression tasks of health states are carried out; according to the method, the improved R-GCN graph neural network and the XGBoost algorithm are fused, and high-precision and end-to-end health state classification and regression identification of the key components of the road structure are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring and intelligent transportation infrastructure, and in particular to a method for identifying the health status of highway structures based on graph neural networks. Background Art

[0002] In the field of transportation engineering, the long-term safety and reliability of highway structures have always been at the core of maintenance and management. Traditional structural health monitoring typically relies on devices such as accelerometers, strain gauges, displacement sensors, and temperature sensors to collect physical quantities such as vibration, deformation, and temperature in real time, and then evaluate them using time-domain or frequency-domain analysis methods. While this model is effective in reflecting the overall vibration characteristics of the structure, the lack of in-depth modeling of the physical connections between components often requires experts to manually extract features based on experience. Furthermore, the ability to process multi-source heterogeneous data is limited, resulting in insufficient diagnostic accuracy and generalization capabilities.

[0003] In recent years, machine learning methods have begun to be applied to highway structural health assessment. In particular, the XGBoost (Extended Gradient Boosting) tree in ensemble learning has become the preferred choice for health status classification and regression due to its strong ability to fit high-dimensional features and good resistance to overfitting. However, these methods typically only use the statistical characteristics of monitoring nodes as input, ignoring the structural topology and various physical connection types between nodes. This makes it difficult for the models to capture the internal correlations of the highway structure and cannot flexibly cope with feature loss caused by sensor noise or data loss.

[0004] The emergence of graph neural network (GNN) technology has brought new breakthroughs to the health monitoring of complex topological structures. Models based on graph convolutional networks (GCN) can jointly model nodes and their neighborhood information, achieving adaptive feature learning. Relational graph convolutional networks (R-GCN) go a step further and introduce independent weight matrices for different types of physical connections, which is suitable for differentiated modeling between multiple components. However, the existing R-GCN model has the problem of fixed network depth in practical applications: insufficient number of layers will cause the network to be unable to capture the interaction patterns of distant nodes, while too many layers can easily cause overfitting or gradient vanishing, reducing training efficiency and prediction stability.

[0005] Meanwhile, traditional self-supervised learning applications in graph networks are limited to general node or edge reconstruction tasks, lacking customized designs that closely integrate with the physical characteristics of highway structures. A single self-supervised objective often struggles to balance global topological information with fine-grained node feature reconstruction, making it difficult to achieve a more accurate representation of health status in an unlabeled environment.

[0006] In response to the above shortcomings, the present invention proposes an improved R-GCN model. While retaining R-GCN's use of independent weight mapping for multiple edge types, this model introduces a graph attention mechanism to highlight key nodes and important connections by calculating attention weights. At the same time, an adaptive convolutional layer mechanism is designed to dynamically increase or decrease the depth of the graph convolution layer according to model performance during training, effectively avoiding underfitting or overfitting caused by a fixed number of layers. In combination with a customized self-supervised node reconstruction task, a lightweight decoder is used to predict node features and minimize prediction errors, which together with the graph smoothing loss constitute a joint optimization objective, thereby further enhancing the model's ability to express local and global node features in unlabeled data scenarios. Finally, this method inputs the optimized node features into the XGBoost model, leveraging its advantages in classification and regression tasks to efficiently predict the health labels and health scores of key components of highway structures, providing an accurate and reliable basis for maintenance decisions. Summary of the Invention

[0007] One purpose of the present invention is to propose a method for identifying the health status of highway structures based on an improved graph neural network and XGBoost. The present invention makes full use of the improved R-GCN model (introducing a graph attention mechanism, adaptive convolutional layer depth adjustment, and self-supervised node reconstruction tasks) and the gradient boosting tree (XGBoost) algorithm, and describes in detail the complete intelligent process from sensor data acquisition and preprocessing, graph structure construction, R-GCN pre-training to dynamic depth fine-tuning, and then to XGBoost classification and regression evaluation. It has the advantages of strong feature expression ability, high model adaptability, and excellent recognition accuracy.

[0008] A method for identifying the health status of a highway structure based on a graph neural network according to an embodiment of the present invention includes the following steps:

[0009] S1. Use sensor equipment to monitor the highway structure and collect raw data;

[0010] S2. Clean, denoise, fill missing values ​​and standardize the collected raw data;

[0011] S3. Build a graph structure, where the nodes of the graph represent the components of the highway structure, and the edges between the nodes represent the physical connection relationships between the components;

[0012] S4. Perform graph convolution based on the R-GCN model, update the feature representation of each node, and optimize the learning ability of the graph structure by combining the node reconstruction task with self-supervised learning;

[0013] S5. Propose an optimization strategy for dynamically adjusting the number of graph convolutional layers, which automatically adjusts the number of graph convolutional layers according to the learning performance of the model during training.

[0014] S6. Input the node features into the XGBoost algorithm to perform health status classification and regression tasks.

[0015] Optionally, the S1 specifically includes:

[0016] S11. Collect real-time monitoring data of the highway structure by installing sensor equipment at key locations of the highway structure; the sensor equipment includes accelerometers, strain gauges, displacement sensors, and temperature sensors. The sensor equipment should be arranged according to the physical characteristics of the highway structure and key locations to ensure comprehensive and accurate data collection; the accelerometer is used to collect vibration data, the strain gauge is used to collect strain data, the displacement sensor is used to collect displacement data, and the temperature sensor is used to collect temperature data;

[0017] S12. The data collected by the sensor is recorded in time series to generate a time series data set, where each data point in the data set represents a measurement result at a certain time point and a certain sensor position.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Preprocess the raw data using data cleaning methods: Kalman filtering algorithm is used for denoising and linear interpolation is used for filling missing values;

[0020] S22, performing standardization on the preprocessed data, standardizing the output data of each sensor to a uniform scale to eliminate data deviations between different sensors;

[0021] S23. Time-align all collected data to ensure data synchronization, so that sensor data at different time steps can be processed and analyzed within the same time window.

[0022] Optionally, the S3 specifically includes:

[0023] S31. Based on the physical characteristics of the highway structure and the layout of sensors, a graph structure is constructed. The key components of each highway structure, including the pavement, bridge beams, support points, tunnel sections, ramps, piers, expansion joints, and cables, are considered as nodes of the graph. The edges between the nodes represent the physical connections between these components.

[0024] S32, define node features. The features of each node are composed of sensor data at the corresponding position, including vibration data X vib (t), strain data X strain (t) and temperature data X temp (t), X vib (t), X strain (t) and X temp(t) represents the vibration, strain and temperature data of the node at time step t respectively;

[0025] S33. Define the edges between the nodes of the graph. The type and strength of the edge are represented by the actual physical connection relationship. The type of edge can be based on the interaction between nodes, such as the connection between the road surface and the bridge, the connection between the tunnel section and the main highway, etc. The strength of the edge is determined by factors such as the stiffness or load-bearing capacity of the physical connection between the nodes. The strength of the edge is set to W. ij , where i and j are two nodes, W ij represents the physical connection strength between node i and node j;

[0026] S34. According to the physical characteristics of the structure and the location of data collection, establish the adjacency matrix A in the figure, where A ij Indicates whether there is a connection between node i and node j. If there is a connection, then A ij =1, otherwise A ij = 0, the adjacency matrix A is the basis of the graph convolution operation;

[0027] S35. Based on the characteristics of each node and the connection relationship between adjacent nodes, the topological structure of the graph is constructed to ensure that the graph can accurately represent the integrity and health status of the highway structure and provide basic data for subsequent graph neural network modeling.

[0028] Optionally, the S4 specifically includes:

[0029] S41. Set the initial characteristics of each node to where X i (t) represents the original feature vector of the i-th node at time step t;

[0030] S42. Under a fixed number of convolutional layers L, perform the following operations for each layer l = 0, ..., L-1: Calculate the relationship between node i and its neighbor nodes through the graph attention mechanism. The attention weight α between ij ; The adjacency matrix element A ij With attention weight α ij Multiply to weight the neighbor features; multiply the weighted neighbor features with the learnable weight matrix W of this layer (l) Multiply them together and pass through the activation function σ to get the output features of this layer, which is calculated according to the following formula:

[0031]

[0032] S43. After completing L layers of graph convolution, obtain the deep feature representation of each node

[0033]

[0034] S44, will Input to the lightweight self-supervised decoder g(·) to generate prediction features

[0035]

[0036] S45. Calculate the self-supervised reconstruction loss. The sum of the reconstruction errors of all nodes is:

[0037]

[0038] S46. Calculate the graph smoothing loss. The weighted sum of the feature differences between all nodes and their neighbors is:

[0039]

[0040] S47. Combine the two losses weighted by the proportional coefficient μ to form a joint loss:

[0041] L total =L GC +μ,L recon ;

[0042] S48, with a fixed number of layers L, minimize L through a one-time iteration total Update all trainable parameters of the improved R-GCN model.

[0043] Optionally, the improved R-GCN model updates and optimizes node feature representations in the following manner:

[0044] While retaining R-GCN, it uses independent weight matrices W for different types of edges. r On the basis of relationship mapping, the graph attention mechanism is introduced to calculate the attention weight α between each pair of neighbor nodes i and j ij , to dynamically adjust the importance of information transmission; dynamically increase or decrease the number of graph convolution layers through adaptive convolution layers to adapt to the complexity changes of each node and its neighborhood features; combined with the self-supervised node reconstruction task, a lightweight decoder g(·) is used to decode each convolution feature. Make predictions, generate And construct the reconstruction loss L recon ; At the same time, calculate the graph convolution smoothing loss L GC ; Combine the above two losses into a joint loss L according to the preset balance coefficient λ total =L GC +λ·L recon , and minimize L by one iteration total , synchronously update all trainable parameters to output the final optimized node features.

[0045] Optionally, the S5 specifically includes:

[0046] S51. Initialize the number of graph convolution layers to L (0) , and set the performance improvement lower limit threshold ε + and the upper threshold of performance fluctuation ε - , and the minimum number of layers L min ;

[0047] S52, after the tth round of fine-tuning training, calculate the current depth features Graph smoothing loss and self-supervised reconstruction loss And combined into a comprehensive loss according to the weight coefficient λ:

[0048]

[0049] S53. Calculate the relative change in comprehensive loss:

[0050]

[0051] S54, according to ΔL (t) The comparison result with the threshold dynamically adjusts the number of graph convolution layers:

[0052] When ΔL (t) <ε + When the performance improvement is insufficient, set L (t) =L (t-1) +1; when ΔL (t) >ε - When it is considered that overfitting or performance is stabilizing, set L (t) =max(L (t-1) -1,;L min ); otherwise keep L (t) =

[0053] L (t-1) ;

[0054] S55, with the updated number of layers L (t) Reconstruct the improved R-GCN model and regenerate node depth features and proceed to the next round of fine-tuning;

[0055] S56, repeating steps S52 to S55 until the number of training rounds reaches a preset value T or the number of layers does not change for multiple consecutive rounds;

[0056] S57. Output the final optimal number of graph convolution layers L and the corresponding optimized node feature set Used for subsequent health status identification.

[0057] Optionally, the S6 specifically includes:

[0058] S61, performing standardization processing on the optimized node features, adjusting each node feature to a standardized representation with a mean of zero and a variance of one;

[0059] S62. Aggregate all normalized node features in node order to form a feature input matrix, where each row of the feature input matrix corresponds to a normalized feature vector of each node;

[0060] S63. Build the XGBoost algorithm based on the gradient boosting tree, configure the number of iterations, learning rate, and regularization parameters of the model, and specify a loss function that supports both health status classification and continuous scoring.

[0061] S64. In each iteration, the XGBoost algorithm calculates the prediction error of all nodes based on the current ensemble results, fits new weak learners and adds them to the ensemble XGBoost algorithm to continuously reduce the overall error.

[0062] S65. After the iteration is completed, the XGBoost algorithm outputs the classification probability and continuous score for each node. The classification probability corresponds to a discrete health label after threshold determination, and the continuous score reflects the health of the node.

[0063] S66. Combining the health labels and continuous scores of each node, determine and output the final health status of each node as the health status identification result of this method.

[0064] The beneficial effects of the present invention are:

[0065] This paper first ensures high quality and consistency of data input by deploying multiple types of sensors at key locations on highway structures and systematically cleaning and standardizing the collected vibration, strain, displacement, and temperature data. Based on this, the paper then employs a relational graph convolutional network (R-GCN) model and integrates a graph attention mechanism during pre-training to dynamically weight information transfer between nodes, thereby more accurately capturing the impact of different physical connections on overall health.

[0066] Secondly, to overcome the underfitting or overfitting problems caused by the fixed number of layers in traditional graph neural networks, this paper introduces an adaptive convolutional layer depth adjustment strategy in the fine-tuning stage. By monitoring the changes in the joint loss in real time, the number of graph convolution layers can be increased or decreased independently. This enables the model to deeply explore the interaction patterns of distant nodes while avoiding training instability caused by an overly deep network.

[0067] This paper combines a customized self-supervised node reconstruction task with a lightweight decoder to predict node features and compare them with the original features. The resulting reconstruction loss and graph smoothing loss together constitute a joint optimization objective, significantly improving the model's ability to express local node details and overall generalization performance in an unlabeled environment.

[0068] Finally, the dynamically optimized node features are fed into the XGBoost classification and regression models. This not only inherits the excellent fit of the gradient boosting tree for classification and regression tasks in high-dimensional feature spaces, but also accelerates model convergence and improves the prediction accuracy of health labels and continuous scores. In summary, this invention implements a complete process from multi-source heterogeneous data fusion to end-to-end intelligent recognition, with the significant benefits of high robustness, high precision, and strong adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 This is an overall flow chart of a highway structure health status identification method based on graph neural network proposed in the present invention;

[0071] Figure 2 This is a schematic diagram of the structure of an improved R-GCN model proposed in this invention; DETAILED DESCRIPTION

[0072] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0073] refer to Figure 1-2 , a highway structure health status identification method based on graph neural network, including the following steps:

[0074] S1. Use sensor equipment to monitor the highway structure and collect raw data;

[0075] S2. Clean, denoise, fill missing values ​​and standardize the collected raw data;

[0076] S3. Build a graph structure, where the nodes of the graph represent the components of the highway structure, and the edges between the nodes represent the physical connection relationships between the components;

[0077] S4. Perform graph convolution based on the R-GCN model, update the feature representation of each node, and optimize the learning ability of the graph structure by combining the node reconstruction task with self-supervised learning;

[0078] S5. Propose an optimization strategy for dynamically adjusting the number of graph convolutional layers, which automatically adjusts the number of graph convolutional layers according to the learning performance of the model during training.

[0079] S6. Input the node features into the XGBoost algorithm to perform health status classification and regression tasks;

[0080] Real-time monitoring of key components such as road surfaces, bridges, and tunnels is carried out using devices such as accelerometers, strain gauges, displacement sensors, and temperature sensors. The raw data is then cleaned, denoised, missing value filled, and standardized to ensure input quality. A highway structure diagram is then constructed based on physical connection relationships, with each component as a node and corresponding edges defined. An improved R-GCN model is used to perform multi-layer graph convolution on this graph, and node feature learning is optimized in combination with a self-supervised node reconstruction task. During model fine-tuning, the depth of the graph convolution layer is dynamically increased or decreased according to changes in the joint loss to adapt to data complexity. Finally, the optimal node features obtained are input into the XGBoost algorithm to achieve accurate classification and continuous regression scoring of the health status of each node.

[0081] In this embodiment, S1 specifically includes:

[0082] S11. Collect real-time monitoring data of the highway structure by installing sensor equipment at key locations of the highway structure; the sensor equipment includes accelerometers, strain gauges, displacement sensors, and temperature sensors. The sensor equipment should be arranged according to the physical characteristics of the highway structure and key locations to ensure comprehensive and accurate data collection; the accelerometer is used to collect vibration data, the strain gauge is used to collect strain data, the displacement sensor is used to collect displacement data, and the temperature sensor is used to collect temperature data;

[0083] S12. The data collected by the sensor is recorded in time series to generate a time series data set, where each data point in the data set represents a measurement result at a certain time point and a certain sensor position;

[0084] In this embodiment, S2 specifically includes:

[0085] S21. Preprocess the raw data using data cleaning methods: Kalman filtering algorithm is used for denoising and linear interpolation is used for filling missing values;

[0086] S22, performing standardization on the preprocessed data, standardizing the output data of each sensor to a uniform scale to eliminate data deviations between different sensors;

[0087] S23. Time-align all collected data to ensure data synchronization, so that sensor data at different time steps can be processed and analyzed within the same time window.

[0088] In this embodiment, S3 specifically includes:

[0089] S31. Based on the physical characteristics of the highway structure and the layout of sensors, a graph structure is constructed. The key components of each highway structure, including the pavement, bridge beams, support points, tunnel sections, ramps, piers, expansion joints, and cables, are considered as nodes of the graph. The edges between the nodes represent the physical connections between these components.

[0090] S32, define node features. The features of each node are composed of sensor data at the corresponding position, including vibration data X vib (t), strain data X strain (t) and temperature data X temp (t), X vib (t), X strain (t) and X temp (t) represents the vibration, strain and temperature data of the node at time step t respectively;

[0091] S33. Define the edges between the nodes of the graph. The type and strength of the edge are represented by the actual physical connection relationship. The type of edge can be based on the interaction between nodes, such as the connection between the road surface and the bridge, the connection between the tunnel section and the main highway, etc. The strength of the edge is determined by factors such as the stiffness or load-bearing capacity of the physical connection between the nodes. The strength of the edge is set to W. ij , where i and j are two nodes, W ij represents the physical connection strength between node i and node j;

[0092] S34. According to the physical characteristics of the structure and the location of data collection, establish the adjacency matrix A in the figure, where A ij Indicates whether there is a connection between node i and node j. If there is a connection, then A ij =1, otherwise A ij = 0, the adjacency matrix A is the basis of the graph convolution operation;

[0093] S35. Based on the characteristics of each node and the connection relationship between adjacent nodes, the topological structure of the graph is constructed to ensure that the graph can accurately represent the integrity and health status of the highway structure and provide basic data for subsequent graph neural network modeling.

[0094] Based on the physical characteristics of the highway structure and the layout of sensors, key components such as the pavement, bridge beam sections, support points, tunnel sections, ramps, piers, expansion joints, and cables are mapped as nodes in the graph, and the edges between nodes are defined by the physical connection relationships between them. The characteristics of each node are composed of the vibration data, strain data, and temperature data of its corresponding position. The edge type and connection strength are set according to factors such as the interaction type and connection stiffness or bearing capacity between different components, and an adjacency matrix is ​​constructed accordingly, so that the corresponding entries in the matrix indicate the connectivity between nodes. Finally, the node characteristics and adjacency relationships are combined to form a graph topology that accurately reflects the integrity and health status of the highway structure, providing a solid foundation for subsequent graph neural network modeling and analysis.

[0095] In this embodiment, S3 specifically includes:

[0096] S41. Set the initial characteristics of each node to where X i (t) represents the original feature vector of the i-th node at time step t;

[0097] S42. Under a fixed number of convolutional layers L, perform the following operations for each layer l = 0, ..., L-1: Calculate the relationship between node i and its neighbor nodes through the graph attention mechanism. The attention weight α between ij ; The adjacency matrix element A ij With attention weight α ij Multiply to weight the neighbor features; multiply the weighted neighbor features with the learnable weight matrix W of this layer (l) Multiply them together and pass through the activation function σ to get the output features of this layer, which is calculated according to the following formula:

[0098]

[0099] S43. After completing L layers of graph convolution, obtain the deep feature representation of each node

[0100]

[0101] S44, will Input to the lightweight self-supervised decoder g(·) to generate prediction features

[0102]

[0103] S45. Calculate the self-supervised reconstruction loss. The sum of the reconstruction errors of all nodes is:

[0104]

[0105] S46. Calculate the graph smoothing loss. The weighted sum of the feature differences between all nodes and their neighbors is:

[0106]

[0107] S47. Combine the two losses weighted by the proportional coefficient μ to form a joint loss:

[0108] L total =L GC +μ,L recon ;

[0109] S48, with a fixed number of layers L, minimize L through a one-time iteration total Update all trainable parameters of the improved R-GCN model.

[0110] The present invention first converts the original feature vector of each node at time step t into As input, and under the preset fixed number of graph convolution layers L, the following operations are performed on each layer in sequence: the attention weight α between node i and its neighbor j is calculated through the graph attention mechanism ij , and add this weight to the adjacency matrix element A ij Multiply to weighted aggregate the features of all neighboring nodes; then combine the aggregation result with the learnable weight matrix W of this layer (l) Multiply and process with activation function σ to generate the output features of this layer; after completing all L layers of convolution, the depth features of each node are obtained Then, these deep features are input into the lightweight self-supervised decoder g(·) to generate prediction features The sum of the reconstruction errors of all nodes is calculated as the self-supervised reconstruction loss, and the weighted sum of the feature differences of all adjacent nodes is calculated as the graph smoothing loss. Finally, the two losses are merged into a joint loss according to the preset proportional coefficient λ, and all trainable parameters of the improved R-GCN model are updated by minimizing the joint loss in a one-time iteration with a fixed number of layers.

[0111] In this embodiment, the improved R-GCN model updates and optimizes node feature representations in the following ways:

[0112] While retaining R-GCN, it uses independent weight matrices W for different types of edges. r On the basis of relationship mapping, the graph attention mechanism is introduced to calculate the attention weight α between each pair of neighbor nodes i and j ij , to dynamically adjust the importance of information transmission; dynamically increase or decrease the number of graph convolution layers through adaptive convolution layers to adapt to the complexity changes of each node and its neighborhood features; combined with the self-supervised node reconstruction task, a lightweight decoder g(·) is used to decode each convolution feature. Make predictions, generate And construct the reconstruction loss L recon; At the same time, calculate the graph convolution smoothing loss L GC ; Combine the above two losses into a joint loss L according to the preset balance coefficient λ total =L GC +λ·L recon , and minimize L by one iteration total , synchronously update all trainable parameters to output the final optimized node features.

[0113] In this embodiment, the S5 specifically includes:

[0114] S51. Initialize the number of graph convolution layers to L (0) , and set the performance improvement lower limit threshold ε + and the upper threshold of performance fluctuation ε - , and the minimum number of layers L min ;

[0115] S52, after the tth round of fine-tuning training, calculate the current depth features Graph smoothing loss and self-supervised reconstruction loss And combined into a comprehensive loss according to the weight coefficient λ:

[0116]

[0117] S53. Calculate the relative change in comprehensive loss:

[0118]

[0119] S54, according to ΔL (t) The comparison result with the threshold dynamically adjusts the number of graph convolution layers:

[0120] When ΔL (t) <ε + When the performance improvement is insufficient, set L (t) =L (t-1) +1; when ΔL (t) >ε - When it is considered that overfitting or performance is stabilizing, set L (t) =max(L (t-1) -1,;L min ); otherwise keep L (t) =L (t-1) ;

[0121] S55, with the updated number of layers L (t) Reconstruct the improved R-GCN model and regenerate node depth features and proceed to the next round of fine-tuning;

[0122] S56, repeating steps S52 to S55 until the number of training rounds reaches a preset value T or the number of layers does not change for multiple consecutive rounds;

[0123] S57. Output the final optimal number of graph convolution layers L and the corresponding optimized node feature set Used for subsequent health status identification.

[0124] In this embodiment, S6 specifically includes:

[0125] S61, performing standardization processing on the optimized node features, adjusting each node feature to a standardized representation with a mean of zero and a variance of one;

[0126] S62. Aggregate all normalized node features in node order to form a feature input matrix, where each row of the feature input matrix corresponds to a normalized feature vector of each node;

[0127] S63. Build the XGBoost algorithm based on the gradient boosting tree, configure the number of iterations, learning rate, and regularization parameters of the model, and specify a loss function that supports both health status classification and continuous scoring.

[0128] S64. In each iteration, the XGBoost algorithm calculates the prediction error of all nodes based on the current ensemble results, fits new weak learners and adds them to the ensemble XGBoost algorithm to continuously reduce the overall error.

[0129] S65. After the iteration is completed, the XGBoost algorithm outputs the classification probability and continuous score for each node. The classification probability corresponds to a discrete health label after threshold determination, and the continuous score reflects the health of the node.

[0130] S66. Combining the health labels and continuous scores of each node, determine and output the final health status of each node as the health status identification result of this method.

[0131] The present invention cleans, standardizes and constructs a graph structure for real-time sensor data of key components of highway structures, regards road surfaces, bridge beam sections, etc. as graph nodes and defines edges with physical connections. Then, in an improved R-GCN model with a fixed number of layers, the graph attention mechanism is first used to weightedly aggregate neighbor features and pre-trained in combination with self-supervised reconstruction and graph smoothing loss. Subsequently, in the fine-tuning stage, the number of convolutional layers is dynamically increased or decreased according to the joint loss to optimize the node depth features. Finally, the optimal features are standardized and input into the XGBoost algorithm with configured hyperparameters. The health status classification and regression are completed by iteratively fitting weak learners, and the discrete health label and continuous score of each node are output, realizing accurate and efficient highway structure health status monitoring from data acquisition to end-to-end intelligent recognition.

[0132] Example 1:

[0133] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a highway bridge for structural health monitoring experiments. The total length of the bridge is about 3.2 kilometers, including multiple key components such as bridge deck, main beam, piers, expansion joints and cables. In the experiment, 10 typical monitoring nodes (numbered N1 to N10) were selected at the middle beam section of the bridge deck, the top of the pier, the bottom of the pier, the expansion joint connection and the cable anchor point. Accelerometers, strain gauges and temperature sensors were installed at each node to achieve synchronous collection of multi-source heterogeneous data of vibration, strain and temperature. To ensure the representativeness and integrity of the experimental data, the average daily traffic volume during this monitoring period was about 50,000 vehicles, the temperature difference between day and night was about 8°C, and there was no major maintenance work along the line, so the test environment conditions were typical and consistent.

[0134] During the data preprocessing phase, the collected raw time series data was first subjected to Kalman filtering to remove noise, followed by linear interpolation to fill in occasional missing values. Finally, the sensor data was normalized according to the time window to ensure that the vibration, strain, and temperature data at each node had the same dimension. Next, based on the bridge's structural topology and sensor distribution, a bridge graph structure consisting of 10 nodes and several edges was constructed. The edge types and strengths corresponded to beam-pier connections, pier-expansion joint connections, and cable-beam connections, respectively, to ensure the accuracy of the physical connection relationship model.

[0135] The improved R-GCN model of this invention was then used for a one-time pre-training of the graph structure. With a fixed convolutional depth of three layers, the model combined a graph attention mechanism and a self-supervised node reconstruction task for parallel optimization to generate preliminary features for each node. During the fine-tuning phase, the number of convolutional layers was dynamically increased or decreased based on the changes in the joint loss, ultimately determining the optimal number of layers to be five. During the experiment, 50 rounds of fine-tuning training were performed, and the joint loss decreased from an initial 1.25 to 0.12, indicating stable model convergence.

[0136] Finally, the optimal node features were normalized and fed into the XGBoost algorithm for health status assessment. XGBoost was configured with 100 iterations, a learning rate of 0.1, and regularization parameters γ = 0.1 and λ = 1, supporting both classification and regression tasks. The average prediction time for a single node during the entire identification process was only 14.7 milliseconds, meeting the requirements of real-time online monitoring.

[0137] Table 1 Highway structure health identification experimental data

[0138]

[0139] Table 1 shows experimental data for highway structural health identification. The table shows the actual health labels, model-predicted labels, health scores, single-node prediction time, and false alarm rate for 10 monitoring nodes. As can be seen from the table, the proposed method achieves an overall classification accuracy of 96%, with a false alarm rate of only 1.7%. The health scores closely match the actual damage level, with an average health score error of less than 0.05. These results fully demonstrate the efficiency, accuracy, and robustness of the proposed method under complex structural topologies and multi-source heterogeneous data, providing strong technical support for intelligent highway structural health maintenance.

[0140] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for identifying the health status of highway structures based on graph neural networks, characterized in that: The steps include: S1. Use sensor equipment to monitor the highway structure and collect raw data; S2. Clean, denoise, fill missing values ​​and standardize the collected raw data; S3. Build a graph structure, where the nodes of the graph represent the components of the highway structure, and the edges between the nodes represent the physical connection relationships between the components; S4. Perform graph convolution based on the improved R-GCN model, update the feature representation of each node, and optimize the learning ability of the graph structure by combining the node reconstruction task with self-supervised learning; S5. Propose an optimization strategy for dynamically adjusting the number of graph convolutional layers. According to the learning performance of the model during training, the number of graph convolutional layers is automatically adjusted and the optimized node features are output. S6. Input the node features into the XGBoost algorithm to perform health status classification and regression tasks.

2. A method for identifying the health status of highway structures based on graph neural networks according to claim 1, characterized in that: Said S1 specifically includes: S11. Collecting raw data of real-time monitoring of the highway structure by installing sensor equipment at key locations of the highway structure; the sensor equipment includes accelerometers, strain gauges, displacement sensors, and temperature sensors; S12. The raw data collected by the sensor is recorded in time series to generate a time series data set.

3. The method for identifying the health status of highway structures based on graph neural networks according to claim 1, characterized in that: The S2 specifically includes: S21. Preprocess the time series data set generated by the original data using data cleaning methods: the Kalman filter algorithm is used in the denoising process, and the linear interpolation method is used in the missing value filling process; S22, performing standardization on the pre-processed data, standardizing the output data of each sensor to a uniform scale; S23. Time-align all collected data.

4. The method for identifying the health status of highway structures based on graph neural networks according to claim 1, characterized in that: The S3 specifically includes: S31. Treat the key components of each highway structure, including the pavement, bridge beams, support points, tunnel sections, ramps, piers, expansion joints, and cables, as nodes in the graph. S32. The features of each node are composed of sensor data at the corresponding position, including vibration data X vib (t), strain data X strain (t) and temperature data X temp (t); S33. Define the edges between the nodes of the graph and set the strength of the edges to W ij , where i and j are two nodes, W ij represents the physical connection strength between node i and node j; S34. According to the physical characteristics of the structure and the location of data collection, establish the adjacency matrix A in the figure, where A ij Indicates whether there is a connection between node i and node j. If there is a connection, then A ij =1, otherwise A ij =0; S35. Construct a graph structure based on the characteristics of each node and the connection relationship between adjacent nodes.

5. The method for identifying the health status of highway structures based on graph neural networks according to claim 1, characterized in that: The S4 specifically includes: S41. Based on the graph structure, obtain a node set and an adjacency matrix A, and set the initial feature of each node to a vector composed of sensor data; S42. Under the preset fixed number of convolutional layers L, perform the following operations on each layer: for each node, first calculate the attention weights between the node and all neighboring nodes through the graph attention mechanism; combine the connection relationship of the corresponding position in the adjacency matrix with the above attention weights to aggregate the features of the neighboring nodes in a weighted form; multiply the aggregated result by the learnable parameter matrix of this layer, and then generate the output features of the node in this layer through the activation function; S43. After completing all L layers of convolution in sequence, the deep feature representation of each node is obtained; S44, inputting the deep features of each node into a lightweight self-supervised decoder to generate a corresponding prediction feature vector; S45. For all nodes, calculate the mean squared error between the predicted features and the actual depth features of the nodes, and sum the errors of all nodes to obtain the self-supervised reconstruction loss; S46. At the same time, for all nodes and their neighboring nodes, the mean square sum of adjacent depth feature differences is calculated to measure the feature smoothness after graph convolution and obtain the graph smoothing loss. S47, according to a pre-set proportional coefficient, weightedly combining the graph smoothing loss and the self-supervised reconstruction loss to form a joint loss; S48. While keeping the number of convolutional layers fixed, update all trainable parameters of the improved R-GCN model by minimizing the joint loss in one iteration.

6. The method for identifying the health status of highway structures based on graph neural networks according to claim 5, characterized in that: The improved R-GCN model updates and optimizes node feature representations in the following ways: First, R-GCN is used to apply relationship mapping of independent weight matrices to different types of edges; then, the graph attention mechanism is integrated into each convolution layer to dynamically allocate information transfer weights according to the relative importance of nodes; at the same time, through the adaptive convolution layer strategy, the convolution depth is automatically increased or decreased according to the performance feedback during the training process to match the feature complexity of different nodes and node neighborhoods; after the convolution, the node features are sent to a lightweight decoder to generate the reconstruction loss, and the reconstruction error between the features and the original features is calculated. The smoothing loss of the feature differences of all adjacent nodes is calculated, and the reconstruction loss and the smoothing loss are merged into a joint loss according to a preset ratio. The loss is minimized in a single iteration at a fixed depth, completing the synchronous update of all trainable parameters and outputting the final optimized node features.

7. The method for identifying the health status of highway structures based on graph neural networks according to claim 1, characterized in that: The S5 specifically includes: S51. Set the initial number of layers of the graph convolutional network, and predefine the lower threshold of performance improvement, the upper threshold of performance fluctuation, and the minimum number of layers allowed; S52. After each round of fine-tuning training, the graph smoothing loss and self-supervised reconstruction loss of this round are calculated separately, and the two are combined into a comprehensive loss value according to a preset ratio; S53. Compare the comprehensive loss of this round with the comprehensive loss of the previous round to evaluate the performance improvement of the improved R-GCN model at the current depth; S54. Based on the comparison result of the performance improvement and the preset threshold, the number of network layers is dynamically adjusted: if the improvement is lower than the lower threshold, the network depth is considered insufficient, and the number of layers is increased by one; if the improvement is higher than the upper threshold, it is considered that overfitting may occur or the performance is stable, and the number of layers is reduced by one; otherwise, the current number of layers remains unchanged; S55, reconstruct the improved R-GCN model with the updated number of layers, regenerate new node features, and enter the next round of fine-tuning; S56, repeat the above loss calculation, performance evaluation and layer number adjustment steps until the preset number of training rounds is reached; S57. After the training is completed, the final optimal number of graph convolution layers and the corresponding optimized node features are output.

8. The method for identifying the health status of highway structures based on graph neural networks according to claim 1, characterized in that: The S6 specifically includes: S61, performing standardization processing on the optimized node features, adjusting each node feature to a standardized representation with a mean of zero and a variance of one; S62. Aggregate all normalized node features in node order to form a feature input matrix, where each row of the feature input matrix corresponds to a normalized feature vector of each node; S63. Build the XGBoost algorithm based on the gradient boosting tree, configure the number of iterations, learning rate, and regularization parameters of the model, and specify a loss function that supports both health status classification and continuous scoring. S64. In each iteration, the XGBoost algorithm calculates the prediction errors of all nodes based on the current ensemble results, fits new weak learners, and adds the weak learners to the ensemble XGBoost algorithm. S65. After the iteration is completed, the XGBoost algorithm outputs the classification probability and continuous score for each node. The classification probability corresponds to a discrete health label after threshold determination, and the continuous score reflects the health of the node. S66. Combining the health labels and continuous scores of each node, determine and output the final health status of each node as the health status identification result of this method.

Images