Graph neural network-based bridge cable response adaptive prediction and anomaly identification method and system

Abstract

The invention provides a bridge cable response adaptive prediction and anomaly recognition method and system based on a graph neural network. The method comprises the steps that node feature vectors of all monitoring points are obtained, the node feature vectors serve as graph nodes, and all the node feature vectors form a graph node set; and calculating an edge weight between the graph nodes according to the node feature vectors, and judging whether the graph nodes are connected or not according to a preset physical connection matrix. And if the judgment result is yes and the edge weight is greater than the weight threshold, adding the corresponding edge weight into an edge set, and constructing a graph structure according to the graph node set and the edge set. And inputting the graph structure into the preset graph neural network for prediction to obtain the prediction response data and the side attention weight, and obtaining an anomaly recognition result according to the prediction response data, the preset actual response data and the side attention weight, thereby improving the accuracy of bridge cable prediction and anomaly recognition.

Description

Technical Field

[0001] This invention belongs to the field of bridge anomaly identification, and particularly relates to a method and system for adaptive prediction and anomaly identification of bridge cable response based on graph neural networks. Background Technology

[0002] With the continuous increase in the number of long-span bridges and the increasing complexity of their service environments, bridge structures exhibit significant multi-factor coupling characteristics during long-term operation. This is particularly true in cable-stayed and suspension bridges, where the cable-stayed system is simultaneously affected by multiple factors such as temperature changes, wind loads, structural deformation, and vehicle actions, resulting in response behavior that is clearly nonlinear, time-varying, and structure-dependent. Current bridge structural health monitoring systems typically collect environmental parameters and structural response data by deploying various sensors, and then perform anomaly analysis and assessment based on this data. However, existing methods often rely on finite element models, simplified mechanical analysis, or empirical rules for inference. These methods are highly dependent on the accuracy of structural modeling and parameter calibration, making them difficult to apply consistently under actual operational conditions, leading to low accuracy in predicting and identifying anomalies in bridge cable-stayed systems. Therefore, improving the accuracy of predicting and identifying anomalies in bridge cable-stayed systems has become an urgent technical problem to be solved. Summary of the Invention

[0003] The purpose of this invention is to design an adaptive prediction and anomaly identification method and system for bridge cable response based on graph neural networks, which can improve the accuracy of bridge cable prediction and anomaly identification.

[0004] To achieve the above objectives, a method for adaptive prediction and anomaly identification of bridge suspension cable response based on graph neural networks is provided in a first aspect of the present invention, the method comprising: Obtain the node feature vector of each monitoring point; wherein, the node feature vector is used as a graph node, and all the node feature vectors constitute a graph node set; The edge weights between each graph node are calculated based on the node feature vectors, and the connection between each graph node is determined based on the preset physical connection matrix. If the judgment result is yes and the edge weight is greater than the weight threshold, then the corresponding edge weight is added to the edge set, and the graph structure is constructed based on the graph node set and the edge set; The graph structure is input into a preset graph neural network for prediction to obtain predicted response data and edge attention weights. Anomaly identification results are obtained based on the predicted response data, the preset actual response data, and the edge attention weights.

[0005] Further, obtaining the node feature vector of each monitoring point includes: Obtain the raw data for each monitoring point; The mean, standard deviation, effective data ratio, and time embedding vector are calculated based on the original data. The mean, standard deviation, effective data ratio, and time embedding vector are concatenated to obtain the node feature vector for each monitoring point.

[0006] Further, the node feature vector includes mean, standard deviation, and effective data ratio; calculating the edge weights between each graph node based on the node feature vector includes: The similarity between each graph node is calculated based on the mean and the standard deviation. The data integrity between each graph node is calculated based on the effective data ratio and a preset adjustable coefficient. The edge weight is obtained by adding the similarity and the data completeness.

[0007] Further, obtaining the anomaly identification result based on the predicted response data, the preset actual response data, and the side attention weights includes: Calculate the cable residual index based on the predicted response data and the actual response data; Calculate the path change index based on the edge attention weights; The anomaly identification result is obtained by searching the cable residual index and the path change index in a preset index result mapping table.

[0008] Further, the preset response data includes predicted cable force and predicted root mean square acceleration, and the actual response data includes measured cable force and measured root mean square acceleration. The calculation of the cable residual index based on the predicted response data and the actual response data includes: Subtract the measured cable force from the predicted cable force and take the absolute value to obtain the first data; Subtract the measured root mean square acceleration from the predicted root mean square acceleration and take the absolute value to obtain the second data; The cable residual index is obtained by adding the first data and the second data.

[0009] Further, the step of calculating the path change index based on the edge attention weights includes: Subtract the edge attention weight of the current time window from the edge attention weight of the previous time window to obtain the third data. Add the third data of the cable node to the third data of all neighboring nodes to obtain the path change index.

[0010] Furthermore, the graph neural network includes a graph attention module and a temporal modeling module. The step of inputting the graph structure into a preset graph neural network for prediction to obtain predicted response data and edge attention weights includes: The graph structure is input into the graph attention module for weighted aggregation to obtain node feature encodings; The node feature encoding is input into the time modeling module for prediction to obtain the predicted response data and the edge attention weights.

[0011] A second aspect of the present invention provides a bridge cable response adaptive prediction and anomaly identification system based on graph neural networks, the system comprising: An acquisition unit is used to acquire the node feature vector of each monitoring point; wherein, the node feature vector is used as a graph node, and all the node feature vectors constitute a graph node set; The calculation unit is used to calculate the edge weight between each graph node based on the node feature vector, and to determine whether each graph node is connected based on the preset physical connection matrix. If the determination result is yes and the edge weight is greater than the weight threshold, the construction unit is used to add the corresponding edge weight to the edge set and construct the graph structure according to the graph node set and the edge set. The prediction unit is used to input the graph structure into a preset graph neural network for prediction, obtain predicted response data and edge attention weights, and obtain anomaly identification results based on the predicted response data, preset actual response data and edge attention weights.

[0012] In a third aspect of the invention, an electronic device is provided, the electronic device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the method described in the first aspect above.

[0013] In a fourth aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0014] The beneficial technical effects of the present invention are at least as follows: To address the problems of multi-factor coupling, incomplete data, and difficulty in identifying anomaly mechanisms in bridge cable-stayed systems under actual operating conditions, leading to inaccurate anomaly identification, this invention provides a method and system for adaptive prediction and anomaly identification of bridge cable-stayed responses based on graph neural networks. The core of this method lies in the structured modeling of multi-source monitoring data, explicitly incorporating the action paths formed by physical connections and response transmission between monitoring points in the bridge into the modeling process. This allows the prediction process to reflect the propagation mechanism of the cable-stayed response within the structure. Based on this, the prediction results are not only used to obtain numerical estimates of key response quantities but also combined with the dynamic characterization of the structural action paths within the model to identify changes in the dominant response mechanism in the cable-stayed system, thereby distinguishing between sudden anomalies and trend-based deterioration. By unifying cable residual indices and path change indices into the anomaly discrimination logic, a shift from "numerical deviation detection" to "structural mechanism identification" is achieved, giving the anomaly identification results clear engineering semantics and interpretability. This method can adapt to the response evolution characteristics of bridges under different operating stages and environmental conditions, providing more stable, reliable and structurally significant technical support for the condition assessment and operation and maintenance decisions of cable-stayed systems. Attached Figure Description

[0015] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0016] Figure 1 This is a flowchart of the adaptive prediction and anomaly identification method for bridge cable response based on graph neural networks provided in the embodiments of this application.

[0017] Figure 2 This is a schematic diagram of the structure of the bridge cable response adaptive prediction and anomaly identification system based on graph neural network provided in the embodiments of this application. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] Please refer to Figure 1 , Figure 1 This is a flowchart of the adaptive prediction and anomaly identification method for bridge suspension cable response based on graph neural networks provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S104.

[0020] Step S101: Obtain the node feature vector of each monitoring point; wherein, the node feature vector is used as a graph node, and all node feature vectors constitute a graph node set. Step S102: Calculate the edge weights between each graph node based on the node feature vectors, and determine whether each graph node is connected based on the preset physical connection matrix. Step S103: If the judgment result is yes and the edge weight is greater than the weight threshold, then add the corresponding edge weight to the edge set, and construct the graph structure based on the graph node set and the edge set. Step S104: Input the graph structure into the preset graph neural network for prediction to obtain the predicted response data and edge attention weights. Based on the predicted response data, the preset actual response data, and the edge attention weights, obtain the anomaly identification result.

[0021] In step S101 of some embodiments, the bridge cable-stayed structure health monitoring system typically deploys multiple types of sensors to collect environmental parameters (such as wind speed and temperature) and structural response parameters (such as displacement, acceleration, and cable force). Because these sensors come from different manufacturers and use different interface standards, their data formats are inconsistent, their sampling frequencies differ, and some channels experience intermittent data loss or drift. To ensure that the subsequent graph neural network can stably process this heterogeneous data, the raw data collected from each monitoring point first needs to undergo unified structured processing. This step employs a sliding time window mechanism to extract the statistical characteristics of each monitoring point over a past period and introduces a time embedding vector to represent periodic changes. Simultaneously, a data integrity metric is introduced to construct a credibility score.

[0022] Specifically, the raw data includes five categories: temperature Wind speed Main beam displacement Bridge tower displacement Historical values ​​of Hesoli ,in This indicates continuous sampling time points. Temperature was collected using thermocouple sensors deployed in the middle of the main girder; wind speed was measured using three-dimensional ultrasonic anemometers deployed laterally on the bridge deck, with the crosswind component retained; the main girder displacement was obtained by acquiring vertical displacement using LVDT sensors at key locations on both sides; the bridge tower displacement was calculated by combining a triaxial accelerometer and an inclinometer at the tower top to obtain horizontal displacement; historical cable force values ​​were obtained from fiber optic force measurement systems deployed at the cable anchor heads. All data were recorded on the monitoring platform at 5-minute intervals.

[0023] Set the sliding time window length to ,For example This indicates that 12 sampling points from the past hour are used to extract the statistical characteristics of each type of signal within that time window. Within each time window, the mean is calculated for each monitoring point. Standard deviation Valid data ratio and time embedding vector The mean Standard deviation Valid data ratio and time embedding vector By concatenating the data, we obtain the node feature vector for each monitoring point. As shown below: ; in, Indicates the first time window The mean of the raw data collected from each monitoring point is used to represent the status level of that monitoring point within the current window; For the first time window The standard deviation of the raw data collected from each monitoring point describes the degree of response fluctuation of that monitoring point within that time period; It is the effective data ratio, defined as the number of data points actually recorded in the time window divided by K, with a value range of [0,1], used to reflect the integrity of the data at that node; It is a time embedding vector of length d, which uses a periodic function encoding method to represent the time position within a day or week. For example, sin(2πt / 288) is used to represent the position of the current time window within the daily cycle, where is the time step number.

[0024] Using the node feature vectors as graph nodes, the node feature vectors of all monitored points constitute a graph node set, as shown below: ; in, This represents the number of nodes in the diagram, corresponding to all monitoring points with structural semantics in the monitoring system, including key nodes of the main beam, nodes at the top of the bridge tower, and nodes at the cable anchor heads.

[0025] In one example, assuming the mean horizontal displacement of the top node of the bridge tower is 0.003 and the standard deviation is 0.0012 within a certain time window, and 10 out of 12 data points are valid data points, then... =0.83. The temporal embedding vector of this node. This indicates that the current time window falls within the daytime high-frequency load zone. All this information is then pieced together to form... This is for use in subsequent graph structure construction.

[0026] In steps S102 to S103 of some embodiments, the aim is to construct a graphical structure of the bridge cable-stayed structure state. , where graph nodes The set includes nodal feature vectors of key monitoring points such as main beams, bridge towers, and cable anchors. Edge set This step requires the construction of a structure that reflects the physical action paths, response coupling patterns, and environmental driving characteristics within the structure. The goal of this step is to combine static structural connections with dynamic state dependencies to construct a structural topology that can be input into a graph neural network. This ensures that subsequent modeling possesses structural adaptability and anomaly interpretability while also considering the physical semantics of the bridge. Since the cable-stayed system in a bridge structure is significantly affected by the coupling effect of environmental factors (such as temperature and wind speed) and structural deformation, the graph structure must not only represent static connections but also be dynamically weighted to reflect the strength of the propagation and impact of monitoring data along different physical paths.

[0027] Specifically, considering that the response of the suspension cable may be driven by multiple sources, such as temperature changes causing beam-cable thermal expansion stress coupling, and wind speed excitation causing tower top displacement changes that affect cable vibration, a connection scoring mechanism based on multi-factor weighted similarity is designed when constructing edge weights. A structural stability regularization term is introduced to ensure the rationality of the graph structure is maintained even when the structure is disturbed or the data is incomplete. Specifically, the similarity between each graph node is calculated based on the mean and standard deviation; the data completeness between each graph node is calculated based on the effective data ratio and a preset adjustable coefficient; and the similarity and data completeness are added together to obtain the edge weights. Let node 1 and node 2 be two monitoring points with structural connections, and construct their edge weights. The expression is as follows: ; in, and They are nodes and nodes The mean within the time window, and They are nodes and nodes The standard deviation of each response represents its volatility. It is a minimal constant used for numerical stability. and They are nodes and nodes The effective data ratio. The first item is a similarity metric, used to measure whether the response states of two nodes are consistent; the second item is a data integrity coupling term, reflecting whether the two connected nodes have high observation quality. This is an adjustable coefficient that controls the influence of the integrity regularization term on the edge weights, typically selected between 0.1 and 0.3. This design has strong physical interpretability: if both nodes respond smoothly and have sufficient observations, the edge weights are higher; otherwise, the information propagation capability of the path is automatically weakened.

[0028] In the specific construction of edge sets First, a fixed physical connection matrix is ​​constructed based on the bridge structural layout diagram. This is used to identify which nodes have explicit structural coupling relationships. If nodes and If a physical connection exists, then Otherwise, it is 0. Then for all satisfying... For each pair of nodes, calculate its edge weight. If its value is higher than the weight threshold If so, then it is included in the edge set, that is: ; in, Typically, a weight of around 0.1 is used to eliminate "lazy edges" that have weak state correlation but are structurally connectable, preventing the graph neural network from capturing meaningless signal paths during training. This edge construction method enables the graph structure to dynamically adapt to changes in the bridge's operating state. For example, if an abnormal increase in wind speed at a certain time leads to increased fluctuations in the tower top displacement, the edge weight connecting the wind speed node and the tower top node will be adjusted accordingly. This will increase, further affecting the subsequent cable force prediction mechanism. The completed graph structure. Include Each node is connected to several weighted edges, and each edge reflects both the physical connection semantics and the response intensity information related to the current time window state. This graph structure will be used as input to the graph neural network in the next step to support subsequent structural response prediction tasks.

[0029] In step S104 of some embodiments, after completing the structural diagram... After the initial construction, the goal of this step is to predict the critical response quantities of the cable and obtain predicted response data, including predicted cable force. With the predicted root mean square acceleration The response of bridge cable-stayed systems is highly dependent on the structural topology, environmental excitations, and time-varying node states. Traditional black-box regression models or time-series models struggle to effectively incorporate the internal coupling relationships within the structure and cannot dynamically adapt to changes in the structural path. Therefore, this step designs a graph neural network modeling method that integrates structural awareness. This method fully utilizes the node attributes and edge weights constructed in the first two steps to achieve accurate prediction and mechanistic explanation of the cable response.

[0030] The input for this step is a graph structure. , where the graph node set This indicates the observation status of each monitoring point in the structure within the current sliding time window, specifically including the mean value within the time window. Standard deviation Valid data ratio and time embedding vector Edge set For structural connection paths with physical meaning, edge weights This combines state similarity and observation quality to represent nodes. For nodes The state propagation strength. The graph structure constructed from these inputs possesses the triple characteristics of real physical topology, data-driven weights, and temporal context, providing a high-quality information flow carrier for subsequent model predictions.

[0031] Considering the complex factors that bridges face during operation, such as sensor malfunctions, severe local state disturbances, and fluctuations in physical path stability, an edge-node dual attention mechanism and a structural consistency control term are introduced into the graph neural network. This allows the graph neural network to adaptively adjust information weights according to the actual state while maintaining respect for structural constraints. The graph attention module uses the graph attention mechanism to weighted aggregate the neighbor information of each node, constructing the aggregated node representation, i.e., node feature encoding. The formula is as follows: ; in, Represents a node The neighborhood group, Neighboring nodes The node feature vectors, The edge attention weight represents the degree of influence of a node on another node in the current state. , This is a trainable linear transformation weight matrix. The first term in this expression represents the propagation of state information, and the second term introduces the edge weights. The weighted terms form a state-structure joint propagation path, which interpretably integrates the structural path strength and state coupling.

[0032] During training, to prevent the model from over-relying on high-similarity paths and causing structural path degradation, a regularization term based on structural uncertainty is introduced to regulate edge stability. Its function is to suppress the model at weak edges (i.e., This results in an unreasonably high level of attentional focus on shorter paths. Specifically: ; in, These are the edge attention weights calculated internally by the model. For edge weights, To prevent division by zero by extremely small constants, this regularization term acts as a structural consistency constraint in the training objective; that is, if the edge weight of a certain edge... If the value is too low (indicating that the path does not have a stable physical response in the current state), then giving too much attention to that side will be penalized, thus forcing the model to restrict the flow of information to paths with stable physical structures.

[0033] Finally, after two layers of graph convolution and attention aggregation, the node feature encodings of all nodes are obtained. The cable-related nodes will be fed into a shared temporal modeling module, which consists of a two-layer bidirectional long short-term memory network (Bi-LSTM). The input sequence is the node feature encoding of the most recent time window. Output the predicted cable force at the current moment. With the predicted root mean square acceleration This process integrates spatial structure information with temporal evolution characteristics to form the final cable response prediction result. It should be noted that cable-related nodes mainly refer to the nodes where sensors are deployed at the cable anchor head, because the prediction target is the root mean square of the cable force and acceleration; the nodes corresponding to these output values ​​are the cable response nodes. Therefore, only these nodes need to be analyzed. The input is fed into the shared temporal modeling module for sequence modeling and response prediction. Other nodes (such as main beam nodes and bridge tower nodes) participate in the information propagation and neighbor feature aggregation of the previous graph structure, but the prediction output of this step is not on these nodes, so they are not sent to the temporal modeling module.

[0034] Predicted response data and represent the root mean square of the predicted cable force and predicted acceleration at the end of the current sliding time window, respectively. Throughout the modeling process, all inputs are derived from the structured node data and edge structure constructed in previous steps. The prediction task is accomplished through physical information perception and structural path optimization. This model structure fully utilizes the characteristics of significant structural paths, easily distorted data, and strong periodic effects in bridge cable-stayed systems. The introduced structural regularization term and edge weight fusion mechanism constitute the core innovation of this step, enabling the prediction to possess not only engineering credibility but also path interpretability.

[0035] Please see Figure 2 In some embodiments, step S101 may include, but is not limited to, steps S201 to S203: Furthermore, based on the prediction results of graph neural networks and the attention distribution within the graph, an anomaly identification method for bridge cable-stayed systems is constructed, ultimately achieving refined judgment and output of the cable's operational status. Since cables may be subject to multiple sources of disturbance during actual operation, such as temperature, wind load, and vehicle load, and the structure itself has time-varying characteristics, anomalies are not merely instantaneous deviations in values, but rather evolutions in the action path, dominant factors, and even response patterns. Therefore, judgment cannot be simply based on the prediction residual threshold; instead, dynamic attention information of the structural path is introduced to identify "deep anomalies" caused by changes in the structural coupling mechanism.

[0036] The input data is entirely derived from the output of the previous prediction step. Specifically, it includes the predicted response data of the cable nodes. And the edge attention weights of the graph neural network output in the last layer. Edge attention weights This is derived from the adaptive learning results of the contribution of each path during the graph convolution process.

[0037] First, it is necessary to obtain the actual response data of the cable nodes within the current time window. , respectively, the measured cable force =and the measured root mean square acceleration Cable tension data is collected by fiber Bragg grating sensors deployed at the cable anchor points, with a sampling period of once every 5 minutes. Acceleration signals are collected by high-sensitivity MEMS accelerometers installed in the middle of the cable, and the RMS value is calculated after time-domain filtering. All data is uniformly formatted by the edge gateway and sent to the processing system for alignment with the predicted values.

[0038] To reflect the deviation between prediction and reality, a cable residual index is defined for the current time window. The cable residual index integrates errors from both cable force and acceleration dimensions. The calculation method involves subtracting the measured cable force from the predicted cable force and taking the absolute value to obtain the first data; subtracting the measured root mean square acceleration from the predicted root mean square acceleration and taking the absolute value to obtain the second data; and finally, adding the first and second data to obtain the cable residual index. The formula is shown below: ; in, For cable residual index, To measure the cable force, This is the measured root mean square acceleration. Predicting cable force, This expression is used to predict the root mean square of acceleration. It ensures that structural anomalies are captured in both mechanical and dynamic response directions, avoiding model bias towards certain types of anomalies (such as low-frequency structural subsidence) or certain types of sensors (such as accelerometers). For example, if a sudden drop in bridge temperature leads to a decrease in cable force and an increase in vibration frequency, the two residuals may differ significantly, but the abnormal trend can still be identified through the cable residual index.

[0039] However, relying solely on cable residual indices is insufficient to determine the underlying causes of anomalies; therefore, it is necessary to consider the attention evolution characteristics of the structural path. This involves extracting the set of edge attention weights for all edges of the cable nodes within the current time window. and the previous time window edge attention weight set Comparison to construct path change indicators The third data is obtained by subtracting the edge attention weight of the current time window from the edge attention weight of the previous time window. This third data is then added to the third data of the cable node and all its neighboring nodes to obtain the path change index. The formula is shown below: ; in, As a path change indicator, The set of neighboring nodes of the cable node. The edge attention weights for the current time window. The edge attention weights for the previous time window are learned through a graph attention mechanism, representing the neighbor nodes. For cable nodes The predicted contribution. In practice, the system retains the edge attention weights for each sliding time window as a historical record of the structural path dynamics. If A significant change in the dominant path of the cable response, such as a sudden shift from "temperature-dominated" to "wind load-dominated," or a significant decrease in the importance of the main beam nodes, usually indicates that the structural response mechanism has deviated from its normal operating mode and carries a high engineering risk.

[0040] In practical engineering, a mapping table of index results is pre-constructed based on the cable residual index. and path change indicators The anomaly identification results are obtained by searching the indicator result mapping table. For example, If it significantly exceeds the historical mean standard deviation, and Simultaneously, a drastic change is identified as a "sudden response bias," such as when cable vibrations are severe during a strong wind attack, but the model still makes predictions based on temperature paths. By comparing it with the historical average rate of change, we can determine... Does it change drastically? For example, under normal structural conditions, what are the path change indicators? The average value is 0.12, and the standard deviation is 0.04. If within the current time window... More than 3 standard deviations above the mean (i.e. If the value is greater than 0.24, the indicator is considered to have experienced a drastic jump. In practical engineering cases, such as during a sudden strong wind, the edge attention weight from the wind speed node to the top node of the bridge tower... If the value suddenly increases by more than 0.3 (e.g., from 0.15 to 0.48), while the weight of the previously dominant path (e.g., temperature → main beam → cable) decreases significantly, then... The calculated value increased significantly, which was determined to be a sudden change in the dominant structural path, belonging to the category of sudden response bias.

[0041] like Medium but The gradual increase over multiple consecutive time windows indicates a slow change in the structural path, often accompanied by prestress decay or structural relaxation, which is identified as "response mechanism drift." That is, when... It fluctuates within the normal range (e.g., 0.08–0.12), but The value gradually increased from 0.10 to 0.22 over multiple consecutive time windows (e.g., 5 sliding windows, approximately 25 minutes), meaning the average increase per window was greater than 0.02, indicating response mechanism drift. In a real-world scenario of gradual temperature decrease, the temperature slowly dropped from 28°C to 18°C. This changed the path of action on the main beam displacement and cable forces, causing the system's focus to gradually shift from the temperature node to the wind speed node, manifested as... It continues to decrease along the original path and continues to increase along the new path.

[0042] when High but "Basically stable" indicates that abrupt changes in cable node observations without altering the dominant path are mostly false alarms caused by local sensor interference, and are classified as "intermittent data fluctuations." Such scenarios are typically caused by local sensor anomalies or temporary communication failures. For example, the measured cable force at a certain time period... A sudden, abnormal jump (such as a sudden drop from 88kN to 72kN, followed by a rapid rebound) leads to... It suddenly rises to 0.25, while the edge attention weights within that time window... No significant changes occurred, and the calculated values ​​were... The value stabilized at around 0.11, differing from the previous time window by only 0.01. In this case, the possibility of a change in the dominant structural path can be ruled out, and it is presumed to be a short-term sensor distortion, which the system classifies as an occasional data fluctuation.

[0043] The system's final output is the anomaly detection result. , It also includes the current anomaly level (normal, warning, alarm) and anomaly type label (sudden, trend, measurement point), and provides the main attention change path in this round of identification, such as "wind speed → tower top → cable channel reinforcement" or "main beam → cable path fracture". The anomaly identification result is sent to the structural health management module of the bridge operation and maintenance system to issue an alert or trigger further structural inspection.

[0044] Steps S101 to S104 of this embodiment involve obtaining the node feature vector of each monitoring point. The node feature vectors are used as graph nodes, and all node feature vectors constitute a graph node set. Edge weights between each graph node are calculated based on the node feature vectors, and a preset physical connection matrix is ​​used to determine whether each graph node is connected. If the determination result is yes and the edge weight is greater than a weight threshold, the corresponding edge weight is added to the edge set, and a graph structure is constructed based on the graph node set and the edge set. The graph structure is input into a preset graph neural network for prediction, obtaining predicted response data and edge attention weights. Anomaly identification results are obtained based on the predicted response data, preset actual response data, and edge attention weights, thus improving the accuracy of bridge cable prediction and anomaly identification.

[0045] Please see Figure 2 This application also provides a graph neural network-based adaptive prediction and anomaly identification system for bridge cable response, which can realize the above-mentioned graph neural network-based adaptive prediction and anomaly identification method for bridge cable response. The system includes: The acquisition unit 201 is used to acquire the node feature vector of each monitoring point; wherein, the node feature vector is used as a graph node, and all node feature vectors constitute a graph node set. The calculation unit 202 is used to calculate the edge weight between each graph node based on the node feature vector, and to determine whether each graph node is connected based on the preset physical connection matrix. Construction unit 203: If the judgment result is yes and the edge weight is greater than the weight threshold, the construction unit is used to add the corresponding edge weight to the edge set and construct the graph structure according to the graph node set and the edge set. The prediction unit 204 is used to input the graph structure into a preset graph neural network for prediction, obtain the predicted response data and edge attention weights, and obtain the anomaly identification result based on the predicted response data, the preset actual response data and the edge attention weights.

[0046] The specific implementation of the bridge cable response adaptive prediction and anomaly identification system based on graph neural network is basically the same as the specific implementation of the bridge cable response adaptive prediction and anomaly identification method based on graph neural network described above, and will not be repeated here.

[0047] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for adaptive prediction and anomaly identification of bridge suspension cable response based on graph neural networks, characterized in that, The method includes: Obtain the node feature vector of each monitoring point; wherein, the node feature vector is used as a graph node, and all the node feature vectors constitute a graph node set; The edge weights between each graph node are calculated based on the node feature vectors, and the connection between each graph node is determined based on the preset physical connection matrix. If the judgment result is yes and the edge weight is greater than the weight threshold, then the corresponding edge weight is added to the edge set, and the graph structure is constructed based on the graph node set and the edge set; The graph structure is input into a preset graph neural network for prediction to obtain predicted response data and edge attention weights. Anomaly identification results are obtained based on the predicted response data, the preset actual response data, and the edge attention weights.

2. The method for adaptive prediction and anomaly identification of bridge suspension cable response based on graph neural networks according to claim 1, characterized in that, The process of obtaining the node feature vector for each monitoring point includes: Obtain the raw data for each monitoring point; The mean, standard deviation, effective data ratio, and time embedding vector are calculated based on the original data. The mean, standard deviation, effective data ratio, and time embedding vector are concatenated to obtain the node feature vector for each monitoring point.

3. The method for adaptive prediction and anomaly identification of bridge suspension cable response based on graph neural networks according to claim 1, characterized in that, The node feature vector includes the mean, standard deviation, and effective data ratio. The step of calculating the edge weights between each graph node based on the node feature vector includes: The similarity between each graph node is calculated based on the mean and the standard deviation. The data integrity between each graph node is calculated based on the effective data ratio and a preset adjustable coefficient. The edge weight is obtained by adding the similarity and the data completeness.

4. The method for adaptive prediction and anomaly identification of bridge suspension cable response based on graph neural networks according to claim 1, characterized in that, The step of obtaining the anomaly identification result based on the predicted response data, the preset actual response data, and the side attention weights includes: Calculate the cable residual index based on the predicted response data and the actual response data; Calculate the path change index based on the edge attention weights; The anomaly identification result is obtained by searching the cable residual index and the path change index in a preset index result mapping table.

5. The method for adaptive prediction and anomaly identification of bridge suspension cable response based on graph neural networks according to claim 4, characterized in that, The preset response data includes predicted cable force and predicted root mean square acceleration, and the actual response data includes measured cable force and measured root mean square acceleration. The calculation of cable residual indices based on the predicted response data and the actual response data includes: Subtract the measured cable force from the predicted cable force and take the absolute value to obtain the first data; Subtract the measured root mean square acceleration from the predicted root mean square acceleration and take the absolute value to obtain the second data; The cable residual index is obtained by adding the first data and the second data.

6. The method for adaptive prediction and anomaly identification of bridge suspension cable response based on graph neural network according to claim 4, characterized in that, The step of calculating the path change index based on the edge attention weights includes: Subtract the edge attention weight of the current time window from the edge attention weight of the previous time window to obtain the third data. Add the third data of the cable node to the third data of all neighboring nodes to obtain the path change index.

7. The method for adaptive prediction and anomaly identification of bridge suspension cable response based on graph neural networks according to claim 1, characterized in that, The graph neural network includes a graph attention module and a temporal modeling module. The step of inputting the graph structure into a preset graph neural network for prediction to obtain predicted response data and edge attention weights includes: The graph structure is input into the graph attention module for weighted aggregation to obtain node feature encodings; The node feature encoding is input into the time modeling module for prediction to obtain the predicted response data and the edge attention weights.

8. A bridge cable response adaptive prediction and anomaly identification system based on graph neural networks, characterized in that, The system includes: An acquisition unit is used to acquire the node feature vector of each monitoring point; wherein, the node feature vector is used as a graph node, and all the node feature vectors constitute a graph node set; The calculation unit is used to calculate the edge weight between each graph node based on the node feature vector, and to determine whether each graph node is connected based on the preset physical connection matrix. If the determination result is yes and the edge weight is greater than the weight threshold, the construction unit is used to add the corresponding edge weight to the edge set and construct the graph structure according to the graph node set and the edge set. The prediction unit is used to input the graph structure into a preset graph neural network for prediction, obtain predicted response data and edge attention weights, and obtain anomaly identification results based on the predicted response data, preset actual response data and edge attention weights.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the adaptive prediction and anomaly identification method for bridge cable response based on graph neural networks as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the adaptive prediction and anomaly identification method for bridge cable response based on graph neural networks as described in any one of claims 1 to 7.

Images