Structural health monitoring method, device, equipment and medium

Abstract

The invention discloses a structure health monitoring method, device and equipment and a medium, and relates to the field of civil engineering structure health monitoring, and the method comprises the steps: obtaining the physical layout information of each monitoring point in an engineering structure and a multi-source monitoring signal of the monitoring point; constructing a graph structure based on the physical layout information; the graph structure comprises nodes and edges, the nodes are used for representing components with monitoring points in the engineering structure, and the edges are used for representing relationships among the components; constructing the multi-source monitoring signal into a graph signal, and based on a graph structure, performing spectral analysis and filtering processing on the graph signal through a graph signal processing model to obtain a processed graph signal; inputting the graph structure and the processed graph signal into a trained graph neural network model to obtain a health state monitoring result; the health state monitoring result comprises a damage probability value and a state classification result of each node. According to the invention, the accuracy of health monitoring of the engineering structure can be greatly improved.

Description

Technical Field

[0001] This application relates to the field of structural health monitoring technology in civil engineering, and in particular to a structural health monitoring method, device, equipment and medium. Background Technology

[0002] In modern civil engineering, the safe and stable operation of large and complex structures such as bridges, high-rise buildings, and large stadiums is crucial. Structural Health Monitoring (SHM), as a key technology for ensuring structural safety, aims to monitor key parameters such as stress, vibration, displacement, and temperature of the structure in a long-term, real-time, and continuous manner by deploying various monitoring devices, such as strain sensors, acceleration sensors, and displacement sensors, at critical locations within the structure. This allows for accurate assessment of the structure's current health status and timely warnings when abnormal damage or performance degradation occurs, providing a scientific basis for structural maintenance, repair, and safety decisions, effectively preventing serious safety accidents, and extending the structure's service life. With the continuous development of monitoring technology and the increasing demands for accuracy in structural monitoring, the scale and complexity of structural monitoring data have experienced explosive growth. Current structural monitoring data exhibits significant characteristics such as high dimensionality, heterogeneity, time-varying nature, and complex spatial coupling. Therefore, effective health monitoring of civil engineering structures is of paramount importance to ensure their safety and stability.

[0003] Currently, the relevant technologies use time-series signal analysis methods for structural status monitoring. However, this approach is difficult to fully characterize the potential spatial relationships and evolutionary patterns within the structure, resulting in low accuracy in health monitoring. Summary of the Invention

[0004] The purpose of this application is to provide a structural health monitoring method, device, equipment, and medium to solve the problem of low accuracy in civil structure health monitoring.

[0005] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a structural health monitoring method, comprising: Acquire the physical layout information of each monitoring point in the engineering structure and the multi-source monitoring signals of the monitoring points; Based on the physical layout information, a graph structure is constructed; the graph structure includes nodes and edges, the nodes are used to represent the components with monitoring points in the engineering structure, and the edges are used to represent the relationships between the components. The multi-source monitoring signals are constructed into a graph signal. Based on the graph structure, the graph signal is subjected to spectral analysis and filtering through a graph signal processing model to obtain the processed graph signal. The graph structure and the processed graph signal are input into the trained graph neural network model to obtain the health status monitoring results; the health status monitoring results include: the damage probability value and status classification result of each node.

[0006] Secondly, this application provides a structural health monitoring device, the device comprising: The acquisition module is used to acquire the physical layout information of each monitoring point in the engineering structure and the multi-source monitoring signals of the monitoring points; A construction module is used to construct a graph structure based on the physical layout information; the graph structure includes nodes and edges, the nodes are used to represent components with monitoring points in the engineering structure, and the edges are used to represent the relationships between the components. The processing module is used to construct the multi-source monitoring signals into a graph signal, and based on the graph structure, perform spectrum analysis and filtering on the graph signal through a graph signal processing model to obtain the processed graph signal; The monitoring module is used to input the graph structure and the processed graph signal into the trained graph neural network model to obtain health status monitoring results; the health status monitoring results include: the damage probability value and status classification result of each node.

[0007] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the structural health monitoring method described in any one of the above.

[0008] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the structural health monitoring method described in any one of the above.

[0009] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a structural health monitoring method, apparatus, device, and medium. The method includes: acquiring physical layout information of various monitoring points in an engineering structure and multi-source monitoring signals of the monitoring points; constructing a graph structure based on the physical layout information; the graph structure includes nodes and edges, where nodes represent components with monitoring points in the engineering structure, and edges represent the relationships between the components; constructing the multi-source monitoring signals into graph signals; and performing spectral analysis and filtering on the graph signals based on the graph structure using a graph signal processing model to obtain processed graph signals; inputting the graph structure and processed graph signals into a trained graph neural network model to obtain health status monitoring results; the health status monitoring results include: the damage probability value and state classification result of each node.

[0010] Compared with existing technologies, this solution acquires physical layout information of monitoring points and multi-source monitoring signals, enabling the data range to cover the physical location of components and multi-dimensional state signals. This not only ensures the comprehensiveness of data sources but also provides complete input information for subsequent modeling. Based on the physical layout information, a graph structure is constructed, overcoming the limitation of traditional time-series analysis focusing only on the time dimension. This transforms the spatial relationships of engineering structures into quantifiable graph topologies, achieving accurate characterization of potential spatial relationships within the structure. After constructing the multi-source monitoring signals into graph signals, spectral analysis and filtering are performed using a graph signal processing model. This effectively separates effective structural features (low-frequency global trends) and noise interference (high-frequency local disturbances) from the signals, improving the purity and effectiveness of input features and obtaining effectively processed graph signals. Finally, the graph structure and processed graph signals are input into a trained graph neural network model. Utilizing the graph neural network's joint learning capability of topological relationships and node features, the model can accurately output the damage probability value and state classification result for each node, significantly improving the accuracy of health monitoring. This solves the core problem of traditional time-series analysis's inability to characterize spatial relationships and correlation evolution, leading to low monitoring accuracy. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a structural schematic diagram of the application environment of a structural health monitoring method according to an embodiment of this application; Figure 2 A schematic flowchart illustrating a structural health monitoring method provided in an embodiment of this application; Figure 3 A schematic flowchart illustrating a method for constructing a graph structure according to an embodiment of this application; Figure 4 This is a schematic diagram of the functional modules of a structural health monitoring device provided in one embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] In related technologies, time-series signal analysis methods are used for structural status monitoring. However, this approach is difficult to fully characterize the potential spatial relationships and evolutionary patterns within the structure, resulting in low accuracy of health monitoring.

[0016] Based on the above-mentioned deficiencies, this application provides a method for structural health monitoring. Compared with existing technologies, this solution acquires physical layout information of monitoring points and multi-source monitoring signals, enabling the data range to cover the physical location of components and multi-dimensional state signals. This not only ensures the comprehensiveness of data sources but also provides complete input information for subsequent modeling. Based on the physical layout information, a graph structure is constructed, overcoming the limitation of traditional time-series analysis focusing only on the time dimension. This transforms the spatial relationships of engineering structures into quantifiable graph topologies, achieving accurate characterization of potential spatial relationships within the structure. After constructing the multi-source monitoring signals into graph signals, spectral analysis and filtering are performed using a graph signal processing model. This effectively separates effective structural features (low-frequency global trends) and noise interference (high-frequency local disturbances) from the signals, improving the purity and effectiveness of input features and obtaining effectively processed graph signals. Finally, the graph structure and processed graph signals are input into a trained graph neural network model. Utilizing the graph neural network's joint learning capability of topological relationships and node features, the model can accurately output the damage probability value and state classification result for each node, significantly improving the accuracy of health monitoring. This solves the core problem of traditional time-series analysis's inability to characterize spatial relationships and correlation evolution, leading to low monitoring accuracy.

[0017] The structural health monitoring method provided in this application embodiment can be applied to, for example... Figure 1The application environment of the structural health monitoring method shown is as follows. This application environment includes a terminal 102, a server 104, and a data storage system. The terminal 102 communicates with the server 104 via a network. The data storage system can store the physical layout information of each monitoring point in the engineering structure and the multi-source monitoring signals of each monitoring point, acquired by the server 104. The data storage system can be set up independently, integrated into the server 104, or placed in the cloud or on another server. The terminal 102 can send the acquired physical layout information and multi-source monitoring signals of each monitoring point to the server 104. After receiving the physical layout information and multi-source monitoring signals of each monitoring point, the server 104 processes them using a graph signal processing model and a graph neural network model to obtain the health status monitoring results. Furthermore, in some embodiments, the structural health monitoring method can also be implemented independently by the server 104 or the terminal 102. For example, the terminal 102 can directly obtain the health status monitoring results based on the physical layout information and the multi-source monitoring signals of each monitoring point using a graph signal processing model and a graph neural network model.

[0018] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0019] This structural health monitoring method is applicable to health monitoring data analysis scenarios for bridge structures, large building structures, utility tunnels, or underground structures.

[0020] In one exemplary embodiment, such as Figure 2 As shown, a structural health monitoring method is provided. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S201 to S204. Wherein: Step S201: Obtain the physical layout information of each monitoring point in the engineering structure and the multi-source monitoring signals of the monitoring points.

[0021] It should be noted that the aforementioned engineering structure can be a civil engineering structure requiring health status monitoring. The physical layout information refers to the spatial arrangement and physical connection information of multiple sensors in the structural monitoring system within the engineering structure. The aforementioned multi-source monitoring signals refer to the signals obtained by each sensor monitoring the engineering structure. These multiple sensors can include various types, such as temperature sensors, strain sensors, acceleration sensors, and environmental monitoring sensors. The engineering structure can include various different components, such as main beams, tower columns, and bridges.

[0022] The physical deployment information of each monitoring point and the multi-source monitoring signal of each monitoring point can be obtained from external devices, imported from blockchain or database, or obtained by spatial analysis and data collection of each monitoring point. This embodiment does not limit the method of obtaining the physical deployment information of each monitoring point and the multi-source monitoring signal of the monitoring point.

[0023] For example, by setting corresponding sensors at various detection points of the engineering structure, the engineering structure can be detected in real time, thereby obtaining multi-source monitoring signals from each detection point. The aforementioned physical layout information can be determined in advance through spatial analysis of the engineering structure, for example, by first converting the engineering structure into a spatial model and then analyzing the spatial model.

[0024] By acquiring the physical layout information of each monitoring point and the multi-source monitoring signals of each monitoring point in this step, we can more comprehensively consider the spatial structure and signal characteristics, and provide comprehensive and good data guidance information for subsequent health status.

[0025] Step S202: Based on the physical layout information, construct a graph structure; the graph structure includes nodes and edges, nodes are used to represent components with monitoring points in the engineering structure, and edges are used to represent the relationships between the components.

[0026] It is understandable that by analyzing the multiple sensors corresponding to each monitoring point in the structural monitoring system, the component graph structure G = (V, E) is generated, where V is the set of nodes, representing the components with monitoring points in the engineering structure, and E is the set of edges, representing the structural connections, distance similarities, or mechanical coupling relationships between the nodes. The components with monitoring points can be the key structural units in the engineering structure.

[0027] In one embodiment, this application also provides a specific implementation method for constructing a graph structure based on physical deployment information; please refer to [link to relevant documentation]. Figure 3 As shown, the method includes: Step S301: Abstract the components with monitoring points in the engineering structure into nodes.

[0028] Step S302: Obtain the relationship information between each node from the physical layout information, and establish edges based on the relationship information; the relationship information includes at least one of the following: physical connection relationship, spatial adjacency relationship, mechanical coupling relationship, and disease propagation relationship.

[0029] Step S303: Determine the parameter information of the edges and nodes, and assign a corresponding weight value to each edge according to the parameter information; the parameter information includes: edge distance, stiffness coefficient, and mutual information.

[0030] Step S304: Based on the weight values, edges, and nodes, a graph structure is formed.

[0031] Specifically, the process of constructing the component diagram can be executed through a structural diagram modeling module. Its purpose is to establish a relationship diagram between various components in the engineering structure, which serves as input to a graph neural network model. First, the input data of the structural diagram modeling module is acquired and processed through the model to obtain the graph structure. This input data can include: structural design drawings, BIM model, sensor deployment diagram, and a list of structural components. The list of structural components can include main beams, bridge towers, cables, connection nodes, etc. During processing, the input data is first analyzed, abstracting each component in the engineering structure with deployed monitoring points into nodes, such as main beam segments (one node per segment), cables, upper and lower bridge tower sections, and bridge deck expansion joints. This step realizes the transformation from physical entities to graph model elements, decomposing the complex structural system into computable basic units. Each node is associated with the monitoring signals and attribute characteristics of the corresponding component, providing a foundation for subsequent analysis.

[0032] Based on this, the multi-dimensional relationships between nodes are extracted through physical layout information and transformed into edges, covering physical connection relationships, spatial adjacency relationships, mechanical coupling relationships, and disease propagation relationships. This comprehensively captures the implicit and explicit connections within the structure, assigning different edge type codes to each edge type. Physical connection relationships include, for example, beams and columns forming a direct connection through rigid connections; spatial adjacency relationships include adjacent but not directly connected components; mechanical coupling relationships include, for example, the interaction of components along the load transfer path; and disease propagation relationships include, for example, the potential path for cracks or corrosion to extend from one component to related components.

[0033] Further, edge weights are assigned by combining parameters such as distance, stiffness coefficient, and mutual information. For example, closer edges have higher weights (spatial adjacency), and larger stiffness coefficients have higher weights (mechanical coupling). This allows the numerical value of an edge to intuitively reflect the tightness of the relationship or the strength of its influence. Weights can include static edge weights and dynamic edge weights. Static edge weights refer to edge weights that are fixed constants, such as fixed values ​​set based on the inherent physical properties of the component. Dynamic edge weights, on the other hand, are dynamically updated over time, such as time-varying relationship strengths calculated based on real-time monitoring data. This adapts to the needs of characterizing stable structural associations and dynamically evolving associations, respectively. The structural graph modeling module outputs a graph structure, which can be encapsulated into GraphData objects using mainstream graph learning frameworks such as DGL or PyG for subsequent calculations. It also supports heterogeneous graphs with various node / edge types as input, adapting to the modeling needs of different components and different relationships in complex engineering structures.

[0034] The establishment of edges based on relational information includes at least one of the following: If there is a physical connection between two nodes, an edge is created between the two nodes; if the spatial adjacency between two nodes is less than a first preset threshold, an edge is created between the two nodes; if the mechanical coupling between two nodes is greater than a second preset threshold, an edge is created between the two nodes; if the disease propagation relationship between two nodes meets preset conditions, an edge is created between the two nodes.

[0035] In the process of constructing edges, multi-dimensional judgment rules based on physical meaning and engineering characteristics can be used: For any two nodes, if there is an actual physical connection between them, such as the direct connection between a beam and a support in a bridge, or the rigid connection between a column and a floor slab in a building, then an edge is directly constructed for them to characterize the basic topological skeleton of the structure; if the spatial adjacency between them, such as the spatial distance between sensors being less than a first preset threshold, which can be determined according to the structural type and safety standards, indicates that the components are closely adjacent in space and may have potential mutual influence (such as vibration transmission, temperature conduction), then an edge is constructed to capture this spatial relationship. If the mechanical coupling relationship between the two components exceeds a second preset threshold (e.g., the load transfer coefficient or vibration mode coupling degree calculated by finite element analysis exceeds the second preset threshold, which can be set based on the structural mechanics model and safety factor), it indicates that there is a significant mechanical interaction between the components, and an edge needs to be established to reflect this key mechanical relationship. If the disease propagation relationship between the two components meets preset conditions (e.g., rules set based on historical disease data and structural mechanisms, such as cracks easily propagating from concrete beams to adjacent concrete columns, or corrosion easily spreading between components of the same material and with similar humidity), then an edge is established to characterize the potential path of disease propagation. By establishing edges through this series of rules, the resulting graph structure can comprehensively and accurately map the multi-dimensional relationships between components in the engineering structure, providing a realistic topological basis for subsequent health monitoring and analysis.

[0036] In this embodiment, the graph structure formed based on physical layout information not only retains the physical topological essence of the structure, but also achieves a fine characterization of relationships through weight quantization, providing a structured input that fits the actual engineering practice for subsequent graph signal processing and graph neural network analysis.

[0037] Step S203: Construct the multi-source monitoring signals into a graph signal. Based on the graph structure, perform spectral analysis and filtering on the graph signal using a graph signal processing model to obtain the processed graph signal.

[0038] In this step, the graph signal processing model can be GSP. The aforementioned multi-source monitoring signals can include time-series signals from various nodes of the engineering structure. These time-series signals can include, for example, acceleration signals, strain signals, displacement signals, and cable force signals. After acquiring the multi-source monitoring signals, a graph signal X∈R is constructed based on the time-series signals from each node within the multi-source monitoring signals. N×T In this matrix, N represents the number of nodes, and T represents the number of time series. Each row of the matrix corresponds to the signal sequence of a node, and each column corresponds to the monitoring data of all nodes at the same time. This realizes the transformation of the three-dimensional information of "node-time-signal" into a structured matrix, allowing a one-to-one correspondence between the signal and the node of the graph structure. Subsequently, the graph Fourier transform is used to process X. With the help of the graph Fourier basis obtained by graph Laplace matrix decomposition, the time-domain signal of the node domain is projected into the spectral domain. Through spectral analysis, the frequency components related to damage sensitivity can be accurately located. For example, overall structural damage is often reflected in the low-frequency spectrum, which reflects the global trend, while the micro-cracks and loosening of local components will show abnormal fluctuations in the high-frequency spectrum. At the same time, in the spectral domain, high-frequency noise can be filtered out, effective low-frequency or specific frequency band components can be retained, and redundant spectral coefficients can be discarded to complete the noise reduction and compression of the original signal. This not only improves the signal quality but also reduces the computational complexity of the subsequent model, laying the foundation for high-quality features to be input into the graph neural network.

[0039] Specifically, the multi-source monitoring signals are constructed into a graph signal. Based on the graph structure, the graph signal is subjected to spectral analysis and filtering through the graph Fourier transform in the graph signal processing model to obtain the processed graph signal, including: A graph signal is constructed based on the time-series signals of each node in the multi-source monitoring signal. Rows in the graph signal represent nodes, and columns represent time-series signals. The adjacency matrix, degree matrix, and identity matrix are determined according to the graph structure, and a graph Laplacian matrix is ​​constructed based on these matrices. Eigenvalue decomposition is performed on the graph Laplacian matrix based on the graph signal to obtain eigenvector matrices. Based on the eigenvector matrix, the graph signal is converted from the node domain to the frequency domain using a graph Fourier transform to obtain a frequency domain signal. The frequency domain signal is then filtered to obtain the processed graph signal. Filtering includes noise reduction and / or dimensionality reduction.

[0040] Specifically, in the graph signal processing flow, the construction of the graph Laplacian matrix is ​​fundamental to the connection structure topology and signal analysis. Based on the node characteristics in the engineering structure, we determine A: the adjacency matrix, representing the connection relationships between nodes; D: the degree matrix, i.e., the diagonal matrix; and I: the identity matrix. Based on the adjacency matrix A, the degree matrix D, and the diagonal matrix I, we construct the graph Laplacian matrix. This Laplacian matrix can include an unnormalized Laplacian matrix, a symmetric normalized Laplacian matrix, or a random walk Laplacian matrix. Among these, the unnormalized Laplacian matrix... Symmetric normalized Laplace matrix Alternatively, the random walk Laplace matrix can be used: the unnormalized form directly describes the node connectivity through the difference between the degree matrix D and the adjacency matrix A, suitable for structures with small differences in node degree; the symmetric normalized form uses the inverse square root of the degree matrix to symmetrically scale the adjacency matrix, stabilizing the eigenvalue range between [0, 2], making it more suitable for complex structures with large differences in node degree, such as bridges containing different types of components like main beams, supports, and cables; the random walk form is based on the definition of transition probabilities between nodes, suitable for analyzing dynamic evolution processes such as disease propagation. Here, A is the adjacency matrix of nodes in the graph structure, D is the degree matrix of nodes in the graph structure, and I is the diagonal matrix of nodes in the graph structure.

[0041] Eigenvalue decomposition of the graph Laplacian matrix based on graph signals is the core of spectral domain analysis, which involves... The graph is represented by the eigenvector matrix U, which forms the graph's Fourier basis. Each column corresponds to a natural vibration mode of the graph. Low-frequency eigenvectors correspond to small eigenvalues, exhibiting a globally smooth distribution and reflecting the overall coordinated change trend of the structure. High-frequency eigenvectors correspond to large eigenvalues, exhibiting localized, intense fluctuations and sensitively capturing the differences between components. The eigenvalue matrix Λ represents the elements on the diagonal, i.e., the frequency components. Their magnitude quantifies the "activity level" of the corresponding mode, providing a basis for subsequent filtering.

[0042] Based on the eigenvector matrix U and the graph signal X in the node domain, the graph signal X in the node domain is projected into the frequency domain using the Graph Fourier Transform (GFT). This indicates the realization of the transformation from the "space-time" domain to the "frequency-time" domain: in the frequency domain signal... In the middle, low-frequency coefficients correspond to the overall health trend of the structure, such as the vibration mode changes of the entire bridge, while high-frequency coefficients correspond to local disturbances, such as the abnormal vibration of a cable or the sudden strain change of a support, providing mathematical tools for the precise location of damage characteristics.

[0043] After obtaining the frequency domain signal, the signal is further purified through spectral domain filtering and noise reduction. Low-pass filtering retains low-frequency components and can filter out high-frequency interference such as sensor noise and environmental interference, highlighting the overall structural state characteristics. Band-pass filtering can selectively extract signals within a specific frequency range, such as specific frequency bands related to typical damage. By retaining the first k spectral coefficients (corresponding to the k smallest eigenvalues) and reconstructing the signal, the processed graphical signal is obtained. This approach can both reduce data dimensionality (reduce redundant features) and preserve key structural information through linear combinations of low-frequency bases, where U K The first K eigenvectors selected for the eigenvector matrix U To project the original image signal X onto U K Zhang Cheng's low-dimensional subspace is used to extract features related to core frequencies. This is the processed image signal.

[0044] The final output of the processed image signal By integrating the topological correlations of the structure with the purified physical signal features, noise interference is removed while key damage-related patterns are preserved. This provides high-quality input for subsequent node state prediction in Graph Neural Networks (GNNs), enabling GNNs to more efficiently learn the mapping relationship between "structural topology - signal features - health state." N represents the number of nodes, and d represents the number of feature dimensions retained after frequency domain denoising or dimensionality reduction, such as retaining the first k spectral coefficients. Optionally, an automatic adjustment of the frequency domain filtering window size based on the graph signal noise level can be implemented to achieve adaptive graph signal denoising.

[0045] In this embodiment, a graph signal processing model is used to perform spectral analysis and filtering on the graph signal. This effectively reduces noise, compresses, and purifies the features of multi-source monitoring signals of engineering structures, retaining key information sensitive to the overall trend of the structure and local damage. This provides high-quality input for the subsequent graph neural network to learn the mapping relationship between "structure topology-signal features-health status", thereby achieving more accurate structural health monitoring and damage identification, and improving the accuracy and efficiency of engineering structure status assessment.

[0046] Step S204: Input the graph structure and the processed graph signal into the trained graph neural network model to obtain the health status monitoring results; the health status monitoring results include: the damage probability value and status classification result of each node.

[0047] The aforementioned graph neural network model can include multiple graph convolutional layers, graph attention layers, and graph classification layers. Residual connections and layer normalization mechanisms can be introduced into each graph convolutional layer to improve network stability and training depth. The graph neural network model is used to monitor the health status of engineering structures based on graph structure and processed graph signals, aggregating spatial features using inter-node relationships, and obtaining the damage probability value and status classification result for each node. The damage probability value of each node characterizes the probability of loss at that node, and this probability value can range from 0 to 1. The status classification result can include at least one of the following: a. Component-level status score, ranging from 0 to 1; b. Health level classification (intact, minor, moderate, severe); c. Risk level and warning indicator; d. Recommended maintenance and reinforcement strategies. Risk levels can include: normal, attention level, and high risk.

[0048] The graph structure and the processed graph signal are input into a trained graph neural network model to obtain health status monitoring results, including: Obtain node attribute information for each node; perform dimensional unification and feature concatenation on the node attribute information, processed graph signal, and graph structure to obtain concatenated features; input the concatenated features into each graph convolutional layer, and sequentially perform activation function, residual connection, and normalization processing in each graph convolutional layer to extract structural features; structural features include global structural features and local structural features; perform attention allocation processing on the structural features according to their importance through a graph attention layer to obtain attention features; perform classification processing on the attention features through a graph classification layer to obtain health status monitoring results.

[0049] Specifically, after acquiring the graph structure and processed graph signal, they are input into a graph neural network model to learn the state evolution patterns of nodes in the structural space, thereby achieving damage state scoring or label prediction. First, node attribute information is acquired, including component type, design life, and ontology vectors. The graph structure, processed graph signal, and node attribute information are then processed along a unified dimension and concatenated to obtain the concatenated features. The aforementioned graph neural network model can include a GCN, which may include three convolutional layers. Each convolutional layer uses an activation function, residual connections (Residual), and a normalized LayerNorm. The concatenated features are then sequentially processed through activation functions (such as ReLU, which introduces nonlinearity to enhance the model's expressive power and capture complex structural relationships), residual connections (to alleviate gradient vanishing, preserve shallow features, and allow deep networks to learn effectively), and normalization (such as layer normalization, which stabilizes the training process and accelerates convergence), thereby fully extracting structural features. The structural features here include global structural features and local structural features. Global structural features are used to reflect the overall topological correlation and overall vibration modes of the structure, as well as other macroscopic laws. Local structural features are used to focus on the microscopic details such as the local connections between components and the sudden changes in local strain.

[0050] Next, these structural features are input into the graph attention layer. This layer allocates attention based on the importance of the features; for example, local strain features more closely related to damage are given higher weights to highlight key information and obtain attention features. Finally, the attention features are fed into the graph classification layer, where a classifier (such as a fully connected layer with Softmax) performs classification processing, ultimately outputting the health status monitoring results of the engineering structure, achieving an accurate assessment of the structural health condition.

[0051] This embodiment utilizes the activation function, residual connections, and normalization processing of graph convolutional layers to fully and efficiently extract global and local features of the structure, ensuring comprehensive feature extraction while addressing the challenges of training deep networks. Through attention allocation in the graph attention layer, key features are precisely focused, strengthening the weight of effective information. Following classification by the graph classification layer, the final output of engineering structure health status monitoring results is accurate, significantly improving the accuracy and relevance of monitoring and providing a reliable basis for structural maintenance and damage early warning. Furthermore, it integrates graph signal processing (GSP) and graph neural network (GNN) frameworks, for the first time combining graph Fourier transform (GFT) with a graph neural network model for multi-sensor time-series data processing and status assessment of civil structures. Traditional SHM methods either rely on frequency domain signal analysis or data-driven models, while the proposed solution bridges the gap between frequency domain feature extraction and deep structural learning, improving the robustness and interpretability of damage identification. Furthermore, a spectral domain sensitive component extraction and noise reduction mechanism is employed. Through GFT transformation, the graph signal spectrum is analyzed to automatically identify the frequency bands most sensitive to damage in the structural response and perform targeted filtering and dimensionality reduction. Compared to traditional time-domain / frequency-domain filtering methods, graph spectral domain filtering is more suitable for irregular node connections and local signal disturbances, improving the precision and stability of data processing. Simultaneously, a GNN model design oriented towards component-level state recognition is adopted, refining the graph neural network to the component-level output. This not only outputs the node damage probability but also correlates causal paths, topological influence chains, and state evolution trends in the graph. This achieves a leap from "sensor-level monitoring" to "component-level intelligent diagnosis," providing stronger engineering guidance and maintenance decision support.

[0052] Optionally, this application also provides a method for constructing a graph neural network model, the method comprising: The process involves acquiring structural layout information, domain rule data, and sample structure monitoring signals for a sample engineering structure. These signals correspond to state annotation results. Based on the structural layout information, a sample graph structure is constructed, and the sample structure monitoring signals are subjected to spectral analysis and filtering using a graph signal processing model to obtain processed sample graph signals. The processed sample graph signals and sample graph structures are divided into training and validation sets. The training set is input into an initial graph neural network model to obtain the output results. A cross-entropy loss term is constructed based on the output results and state annotation results. Domain rule data is transformed into logical rule terms, and a logical rule loss term is constructed based on these logical rule terms. The cross-entropy loss term and the logical rule loss term are combined to construct a loss function. The parameters in the initial graph neural network model are iteratively trained and optimized according to minimizing the loss function to obtain the trained model. The validation set is input into the trained model for validation processing to obtain the final graph neural network model.

[0053] It should be noted that in the process of building a graph neural network model, it is necessary to first collect basic information of the sample engineering structure, including structural layout information reflecting the position and connection relationship of components, domain rule data, and sample structure monitoring signals collected by sensors. These sample structure monitoring signals include vibration and strain data, and each sample monitoring signal can be manually labeled, with corresponding state labeling results, providing data input and supervision basis for model training. State labeling results include, for example, "intact" and "minor damage".

[0054] The aforementioned domain rule data can include industry standard data, expert rule data, and component semantic data. Industry standard data includes bridge codes and inspection procedures, while expert rule data is based on long-term maintenance experience, such as maintenance experience, empirical formulas, and component semantics. Component semantics can be categorized as decreasing in influence from main beam to support to pier. By integrating ontology rules with expert prior knowledge, the module aims to deeply combine industry knowledge, expert experience, and model training to improve the reliability of structural health monitoring.

[0055] During implementation, the domain rule data is first represented in the form of RDF triples of a knowledge graph, such as "<Lasso 3#><Service Years>". <18> The data is clearly linked to components, attributes, and values ​​by specifying "<Cable 3#><Stress Fluctuation><23%>". Then, the domain knowledge data is converted into logical formulas or regularization formulas. For example, a logical formula might state, "If the cable's service life is >15 years and stress fluctuation >20%, then the risk of damage is high." This is then integrated into the model training process. It can be used as part of the loss function to guide model learning, making model predictions more aligned with expert rules, or as a post-processing rule to verify and correct model output. By incorporating domain knowledge data into the model training process, it prevents the model from arbitrarily fitting incorrect patterns due to data bias. Furthermore, for rare components made of special materials or in extreme structural abnormalities after strong earthquakes, the model can better generalize with the guidance of prior knowledge, improving its ability to judge the health status of such scenarios.

[0056] In this embodiment, a sample graph structure is constructed based on the structural layout information. At the same time, a graph signal processing model is used to perform spectral analysis and filtering on the sample monitoring signal, such as separating low-frequency global trends from high-frequency local disturbances and removing noise contamination, so as to obtain a processed sample graph signal that can accurately reflect the structural state. Subsequently, the processed sample graph signal and the sample graph structure are divided into a training set and a validation set according to the proportion. The training set is used for model parameter learning, and the validation set is used to evaluate the model's generalization ability.

[0057] During the model training phase, the initial training parameters of the graph neural network are configured first. After configuration, the training set is input into the initial graph neural network model to obtain the model's predicted output of the structural state. A cross-entropy loss term is constructed based on the predicted output and the actual state labeling results to measure the fit between the model's predictions and the data labels. Simultaneously, the domain rule data is transformed into computable logical rule terms, and a logical rule loss term is constructed to penalize predictions that violate the domain rules. The two losses are combined proportionally to form the final loss function. The parameters of the initial model are iteratively adjusted using optimization algorithms such as gradient descent to continuously reduce the loss function, gradually improving the model's prediction accuracy and rule compliance, resulting in the pre-trained model. Finally, the validation set is input into the trained model for validation. The performance of the validation set (such as accuracy and F1 score) is used to determine whether the model meets the requirements. If it meets the requirements, it is determined to be the final usable graph neural network model. If it does not meet the requirements, the data processing method, model structure, or loss function weights need to be adjusted and retrained to ensure that the model has both data fitting ability and domain rule adaptability in engineering structural health monitoring tasks.

[0058] After inputting the training set into the initial graph neural network model, different types of input features may have different dimensions and distributions. These need to be mapped to the same dimensional space through linear transformation or normalization. These input features include: processed sample graph signals, type embeddings (Type_Embedding), and ontology vectors (Ontology_Vector). Type_Embedding represents the type information of nodes. For example, in bridge monitoring, nodes may represent different components such as "beams," "supports," and "cables." Embedding transforms category information into learnable vectors. Ontology_Vector represents the conceptual relationships in the encoded domain knowledge system, enabling the model to utilize prior knowledge, such as the relationship between "cables" and "anchors."

[0059] Taking three graph convolutional layers as an example, three graph convolutional layers (GraphConv) are used to capture more complex graph structure features, but they are also prone to overfitting and gradient vanishing problems. Each graph convolutional layer can use the ReLU activation function, adding a residual connection and a LayerNorm. The residual connection is used to solve the training difficulties of deep networks and help gradient backpropagation. The LayerNorm is used to normalize the output of each layer, stabilize the numerical distribution during training, prevent excessive numerical fluctuations, and accelerate convergence. The processing of the graph convolutional layer can be represented by the following formula: ; in, It is an adjacency matrix with self-loops, allowing the node's own information to participate in the propagation; It is an adjacency matrix The corresponding degree matrix, It is a symmetric normalized adjacency matrix, which ensures the stability of feature scale during propagation; It is the first The learnable weight matrix of the layer; The activation function (which can be ReLU here) introduces non-linearity. Residual connections are used to add the input of each layer to the output of each layer, alleviating the vanishing gradient problem in deep networks and helping the model learn more complex features. For the first The output of the layer, For the first The node features of the layer's input.

[0060] The output layer of the initial graph neural network model is used to perform node-level predictions, including node-level classification and regression tasks. For example, predicting the "damage level" (discrete categories such as intact, minor, severe, etc.) can be done using the Softmax activation function, outputting the probability of each node belonging to each category. For regression tasks such as predicting "damage probability scores," which include continuous values ​​from 0 to 100, the output layer does not use an activation function and directly outputs the continuous numerical value. The purpose of the output layer is to map the high-order features learned by the network to the task's target space to obtain the final prediction result.

[0061] The loss function described above is used to avoid the model relying solely on data fitting. It improves the rationality and reliability of predictions through domain rules, and is particularly suitable for scenarios with limited labeled data or noisy data. It can be expressed by the following formula: ; in, This is the cross-entropy loss term, used to measure the difference between the model's predictions and the labeled data, ensuring that the model learns patterns from the data; The balance coefficient controls the constraint strength of the ontology rules; For logical rule loss terms, domain knowledge data is transformed into a computable loss term to constrain model predictions to conform to prior knowledge. For example, domain knowledge data might state: if cable tension fluctuation > 15% and service life > 20 years, then the condition cannot be rated as "intact".

[0062] Training strategies are designed to address the characteristics of graph data. For supervised training, labeled nodes (such as structural components with known damage) can be used to calculate the loss function, and network parameters can be updated through backpropagation. DropEdge and node sampling techniques can also be introduced to stabilize training. DropEdge randomly discards some edge connections to reduce the risk of overfitting and enhance the model's robustness to graph structural perturbations. Node sampling, in large-scale graphs, samples only a subset of nodes and their neighbors for computation in each training iteration (such as the sampling strategy in GraphSAGE), reducing computational costs. An early stopping mechanism is employed to monitor the validation set loss; training stops when the loss no longer decreases after several consecutive rounds to prevent overfitting.

[0063] This embodiment utilizes a graph neural network for structural health monitoring, fully leveraging the graph structure's ability to characterize topological relationships such as spatial connections and mechanical couplings among components. It efficiently fuses processed graph signal features (including global structural trends and local details) through graph convolution and graph attention mechanisms. Simultaneously, it combines a hybrid loss function of cross-entropy loss and logical rule loss, balancing data fitting accuracy with domain knowledge constraints. This avoids model misfitting due to data noise or sparsity while improving generalization ability for rare components and extreme structural states. Ultimately, it achieves accurate prediction of the health status (damage probability, classification results) of each node in the structure, providing reliable decision support for preventative maintenance and damage early warning of engineering structures, significantly improving the accuracy, relevance, and engineering practicality of health monitoring. Furthermore, it proposes a graph modeling approach that integrates structural geometric connections, sensor data associations, and component attribute semantics into a single graph structure, forming a modeling framework of heterogeneous graphs, multi-type edges, and multi-source attributes. This overcomes the limitations of previous modeling methods that only considered physical structures or single-class data, significantly improving the graph model's ability to fit complex structural behaviors. Furthermore, logical rules and prior ontology knowledge are introduced for training guidance. Engineering experience, ontology knowledge, and logical rules (such as "over-aged components + severe fluctuations = high risk") are encoded as regularization terms and participate in neural network training. This solves the "black box" problem of deep models and improves the interpretability, reliability, and engineering trustworthiness of the model results.

[0064] Optionally, after obtaining the health status monitoring results, the above method further includes: Convert the health status monitoring results into a graphical format; the graphical format includes at least one of the following: model diagram, bar chart, or line graph; display the graphical format through a visualization interface, and provide alarm prompts for data in the health status monitoring results that meet the alarm conditions.

[0065] Specifically, in the application phase of structural health monitoring results, the health status monitoring results output by the model can be converted into intuitive graphical formats. At least one of these formats can be selected based on actual needs: model diagrams, bar charts, and line graphs. Model diagrams visually present the structural topology, using different colors or markers to indicate the health status of each component, making the overall health distribution of the structure clear at a glance. For example, red indicates components with high damage risk, and green indicates intact components. Bar charts can compare the damage probability values ​​of different components, clearly showing the differences in risk levels among them. Line graphs can dynamically present the trend of the health status changes of the same component at different times, assisting in the analysis of damage development patterns. These graphs are then displayed through a visual interface, allowing staff to view them in real time and quickly grasp the structural health status. Visual interfaces include web-based monitoring platforms and local monitoring software interfaces. Simultaneously, the system automatically compares the health status monitoring results with preset alarm conditions. For example, if the damage probability exceeds 80%, the status is judged as "severe damage." Once data meeting the alarm conditions is detected, an alarm is immediately triggered through interface pop-ups, audio-visual prompts, and SMS notifications, ensuring that staff detect structural anomalies immediately and take timely repair and maintenance measures to effectively reduce structural safety risks.

[0066] This application provides a structural health monitoring method. Compared with existing technologies, this method acquires the physical layout information of monitoring points and multi-source monitoring signals, enabling the data range to cover the physical location of components and multi-dimensional state signals. This not only ensures the comprehensiveness of data sources but also provides complete input information for subsequent modeling. Based on the physical layout information, a graph structure is constructed, breaking through the limitation of traditional time-series analysis focusing only on the time dimension. The spatial relationships of engineering structures are transformed into quantifiable graph topologies, achieving accurate characterization of potential spatial relationships within the structure. After constructing multi-source monitoring signals into graph signals, a graph signal processing module is used... The graph model performs spectral analysis and filtering, effectively separating effective structural features (low-frequency global trends) and noise interference (high-frequency local disturbances) from the signal, improving the purity and effectiveness of the input features, and obtaining an effectively processed graph signal. Finally, the graph structure and the processed graph signal are input into a trained graph neural network model. By utilizing the graph neural network's joint learning ability on topological associations and node features, the model can accurately output the damage probability value and state classification result of each node, significantly improving the accuracy of health monitoring. This solves the core problem that traditional time series analysis is unable to characterize spatial relationships and correlation evolution, leading to low monitoring accuracy.

[0067] For example, taking a cable-stayed bridge as a civil engineering structure, health monitoring of the cable-stayed bridge structure can be carried out. Multiple monitoring points can be set up at key locations of the cable-stayed bridge, and one or more types of sensors can be deployed at each monitoring point for long-term monitoring of structural response. For example, strain sensors and acceleration sensors can be installed on several main beams, tower columns, and stay cables to acquire data such as strain and vibration. Environmental and temperature sensors can be added as needed to assist in calibrating the structural response. The sensor placement covers typical cross-sections and key components of the bridge to reflect local and global stress and deformation characteristics. These key components include, for example, the mid-span of the main beams, the top of the towers, and the cable-beam connections. The monitoring data collected by each sensor at different time periods can form a time-series signal, whose feature dimensions can be selected as needed, such as recording the strain or acceleration amplitude of a single channel at each node, or multi-channel information (such as acceleration in three directions). After each sensor collects data, the collected data undergoes synchronization and preprocessing to form a spatiotemporally aligned graph signal data set. Preprocessing may include operations such as drift removal and standardization.

[0068] The physical layout information of the various sensors in the cable-stayed bridge is represented as a graph structure. In this graph structure, nodes correspond to sensors at each monitoring point, and node attributes include the sensor's measurement characteristics, such as strain values ​​or vibration amplitudes at a specific moment. Edges in the graph structure represent the structural relationships and signal correlations between sensor nodes. Edges can be established based on the spatial location of the sensors, the type of structural member they are located in, and the correlation degree of the observed signals. Preferably, the correlation coefficient of sensor signals or the dynamic time warping (DTW) distance can be used to quantify the similarity between signals from any two nodes. If two sensor responses are highly correlated, for example, located on the same main beam and affected by the same vehicle load, their strain curves will have a high correlation coefficient, and an edge will be connected in the graph structure with a large weight; conversely, if the correlation is weak, no connection will be made or a small weight will be assigned. To highlight the most important topological relationships, each node can retain only a few edges (such as the first three) with the highest weights. The graph structure constructed in this way not only considers the geographical proximity of each sensor but also reflects functional relationships. For example, sensors located on the bridge deck and under the main beam may be geographically close, but their strain changes show opposite trends (the bridge deck is under tension while the main beam is under compression), thus their correlation is low and they may not be directly connected in the graph. The final graph model has the same number of nodes as the number of effective sensors, and the number and weight distribution of edges reflect the interaction relationships between the various parts of the cable-stayed bridge structure.

[0069] After constructing the graph structure, graph signal processing and feature extraction are performed. For the constructed sensor graph structure, graph signal processing techniques are applied to extract structural state features. On one hand, the time-series signal acquired by the sensor is transformed from the node domain to the spectral domain of the graph using Graph Fourier Transform (GFT) to obtain the graph signal. Specifically, the adjacency matrix and degree matrix are first determined based on the graph structure, and the graph Laplacian matrix L is calculated. Let L be the graph matrix L, where D is the degree matrix and A is the adjacency matrix. Then, perform eigenvalue decomposition on the graph Laplacian matrix L. The eigenvector matrix V and its corresponding eigenvalue diagonal matrix Λ are obtained. These eigenvector matrices serve as the "graphical Fourier basis," and the eigenvalue diagonal matrix corresponds to the "frequency" of the graphical signal. Then, the measurement signals from all sensors at a given moment are used to construct the graphical signal vector s, whose corresponding GFT coefficients (spectral domain representation) are... =V T s. , where V represents the eigenvector matrix. Let s be the transpose of the eigenvalue matrix and s be the graph signal. By analyzing the amplitude distribution of the GFT coefficients under different eigenvalues, the spectral components contained in the structural response can be determined, which is equivalent to identifying the mode shape components and spatial frequencies of the signal under the graph topology. On the other hand, filtering can be applied to the graph spectral coefficients to remove noise interference. A graph domain low-pass filter h(Λ) can be designed to filter the signal above the cutoff frequency. The signal components corresponding to the higher-order eigenvalues ​​are weakened or filtered out, retaining only the low-frequency graphical signal. The low-frequency graphical signal represents a pattern of smooth, consistent changes in the responses of adjacent sensors on the bridge structure, which can be considered the normal global vibration pattern of the structure. The high-frequency graphical signal characterizes local anomalies with significant differences in the responses of neighboring nodes, which may be caused by noise or local damage. By adjusting the parameters of the low-pass filter in the spectral domain (such as selecting the cutoff eigenvalue λc or the filter order), the relationship between noise removal and signal distortion can be balanced. After filtering, an inverse graphical Fourier transform s′=V is performed on the spectral coefficients. The signal is restored to the nodal domain, where V is the eigenvector matrix. s represents the GFT coefficients after spectral domain filtering (the filtered spectral domain signal), s′ represents the output of the inverse graphical Fourier transform, and s′ represents the reconstructed nodal domain graphical signal. By multiplying the filtered spectral coefficients with the graphical Fourier basis, the purified frequency features in the spectral domain are transformed into physical signals in the nodal domain, ultimately obtaining denoised sensor measurement signals or signals focusing on key information (such as structural vibration signals after removing environmental interference, strain signals highlighting damage characteristics). These signals can be directly used for subsequent structural health status analysis or as input to graphical neural networks, and the resulting denoised signal sequence more clearly reflects the normal response characteristics of the structure. Furthermore, the graphical signal processing model can also extract global features such as mode shapes: for example, by comparing the graphical mode shapes corresponding to different order feature vectors with the finite element modal analysis results, the main modal forms of bridges can be identified, such as the first-order vertical bending mode shape and torsional mode shape. Once the graphical signal modes corresponding to these mode shapes change significantly, it may mean changes in global parameters such as structural stiffness and damping, which can serve as a basis for damage early warning.

[0070] After acquiring the graph structure and its signal features, a graph neural network is introduced to model and infer the structural health status. A graph convolution-based deep learning model, such as a Graph Convolutional Network (GCN) or a Message Passing Neural Network (MPNN), is preferred. These models achieve efficient modeling of graph data by performing message passing and feature aggregation on the neighborhood of graph nodes. In this embodiment, a two-layer graph convolutional network is selected as an example: the first layer extracts the primary feature representation of each sensor node and its neighbors, and the second layer further aggregates global information to output the state discrimination result. The network input is the sensor graph and its node feature matrix, such as the statistical characteristics of each sensor within a certain time window (maximum strain, mean square vibration amplitude, etc.) or the instantaneous signal after GSP filtering. Through graph convolution calculation, each node will share information with its neighboring nodes. The output layer of the graph convolutional neural network model can vary depending on the monitoring task: if used for strain or displacement prediction, it outputs the response estimate of each node; if used for structural state classification, it outputs the health status category of the entire structure or the damage probability distribution of each component.

[0071] The graph convolutional neural network (GNN) model is trained for different tasks. Training data can be labeled samples from historical monitoring or high-reliability numerical simulation datasets (digital twin technology). For example, to achieve damage localization and quantitative assessment of cable-stayed bridges, the tensile force changes under different weakening degrees of cable sections can be simulated using a finite element model. The multi-condition sensor graph data obtained from the simulation is used for supervised learning. When damage identification is the goal, the training labels include the structural health status corresponding to each sample, such as "undamaged" or "Nth cable fracture x%". A supervised learning strategy can be used during model training, adjusting the GNN weights using existing labels to make its output match the real state as closely as possible. Through continuous iterative training, the model gradually learns to correlate changes in multiple sensor signals in the graph space to determine the structural state. After training, the real-time monitored graph signals are input into the trained GNN model to perform inference and obtain the current state identification result of the cable-stayed bridge structure. For example, the model outputs the node anomaly score distribution of a cable-stayed bridge's sensor network, showing that the node corresponding to a specific cable has the highest anomaly score, indicating that the cable may be damaged. Alternatively, the output classification result can indicate "structurally healthy" or "damage present," and provide the probability value of the damaged area.

[0072] After outputting the state identification results through the model, these results can be presented to users intuitively through a visualization interface. For example, the 3D model or engineering drawing of the bridge can be overlaid with the monitoring results: if a cable is determined to be damaged, it can be highlighted and the estimated degree of damage can be marked, such as the percentage reduction in cross-sectional area; for the overall state classification results, safety indicators or risk warning levels can be given on the interface. The measured data and predicted data from the sensors can also be plotted synchronously, such as comparing the original strain signal of a key section with the GNN predicted signal using curves to verify the model's accuracy. Simultaneously, the modal vibration modes extracted from the graph signal processing can be visualized: for example, displaying a graph of the bridge's vibration modes or a bar chart showing the energy proportion of each mode in the spectrum, allowing engineers to understand the structural dynamic characteristics and their changes. Through the above methods, this embodiment achieves a complete closed loop from data acquisition to result presentation, enabling timely and accurate identification of the health status of cable-stayed bridge structures and location of potential damage.

[0073] Optionally, for the graph signal processing module, the application process and details of Graph Fourier Transform (GFT) in signal denoising and spectrum analysis of cable-stayed bridge sensor networks can be explained. Specifically, assuming that when a vehicle load event occurs on a cable-stayed bridge, the sensor strain response values ​​collected at a certain moment constitute a graph signal s=[s1,s2,…,s…]. N ] TWhere N is the number of sensor nodes and T is the number of time-series signals. For example, at this moment, a heavy-duty vehicle is crossing the bridge, causing the strain of the main beam sensors near the wheel path to increase, while the strain change in the far region is smaller. Abnormal peak values ​​may appear in the sensors at individual damaged locations (such as where there are cracks or loose cables).

[0074] In the process of performing GFT and filtering analysis on the above graph signals, the graph Laplacian matrix can be constructed first: based on the topology and edge weights of the sensor network, the graph Laplacian matrix L = D − A is formed. Here, A is the adjacency matrix established based on sensor correlation, and D is the node degree matrix (diagonal elements D... ii =∑ j A ij (where i and j represent rows and columns, respectively). For a cable-stayed bridge sensor network, the structure of A reflects the connection relationships between various measuring points on the bridge, such as the correlation strength between nodes of the cable-tower-beam system. A normalized Laplace operator is used. Similar spectral decomposition results can also be obtained, where I is the identity matrix.

[0075] Then, eigenvalue decomposition and graphical Fourier transform are performed. First, eigenvalue decomposition is performed on the graphical Laplacian matrix L to obtain N eigen pairs. ,satisfy , For eigenvalues, Let these be eigenvectors. Arrange all eigenvectors in ascending order of their eigenvalues ​​to form an orthogonal matrix. The corresponding eigenvalues ​​form a diagonal matrix. Since the structure graph is usually connected, the eigenvalue sequence starts from 0 (λ0=0). Eigenvectors corresponding to smaller eigenvalues ​​change slowly on the graph (global mode), while eigenvectors corresponding to larger eigenvalues ​​oscillate more violently locally (local mode). The original graph signal is projected onto the eigenvector basis, and then calculated... This yields the graphical Fourier coefficients, and each scalar in the coefficients... Essentially, the original signal is in the first... 1st order feature mode (corresponding feature vector) The process of converting the signal from the "node space" to the "spectral domain" involves analyzing the intensity of the component components. By analyzing the distribution of the graphical Fourier coefficients with characteristic frequencies, the spectral characteristics of the structural response caused by the load event can be obtained: for example, if the vehicle load mainly excites the first-order vertical vibration mode of the bridge, then... The amplitude is significant at the corresponding low-order frequencies; if a local slack in a cable causes an anomaly, the strain signal will exhibit localized characteristics, making the high-frequency components... Increase.

[0076] To eliminate the interference of measurement noise and local anomalies on the signal, in the spectral domain... Filtering is performed. A low-pass filter function g(λ) is designed, which assigns a value close to 1 for the characteristic frequency and a value close to 0 for the high-frequency range. A simple choice is to set the cutoff frequency λc, when... The Fourier coefficients corresponding to >λc are calculated according to Process the signal, or retain the original value (or attenuate it according to a smoothing function). Commonly used filtering methods in this field also include window function filtering and Chebyshev polynomial approximation filtering; appropriate filter types and parameters can be selected as needed. The spectral domain signal obtained from the graphical Fourier transform... After applying filter g, the filtered spectral coefficients are obtained. Each spectral coefficient component The filter's response at the corresponding frequency Compared with the original spectral coefficients Multiply them to get the result. The first of the Graph Laplace matrix The eigenvalue (corresponding to the eigenvalue) The cutoff frequency is selected because the main energy of the bridge structure response is usually concentrated in the lower-order modes (such as the first few vibration modes), while the high-frequency components are mostly caused by measurement noise or local abrupt changes. When using a signal, the known modal frequency range can be referenced, retaining only the first few modal frequencies as signal components. Applying a low-pass filter to the signal can effectively reduce noise and improve the accuracy of subsequent predictions. This embodiment selects a cutoff frequency. The first three non-zero eigenvalues ​​are retained to cover the first vertical bend, the second vertical bend, and the first torsional mode of the cable-stayed bridge, while filtering out higher-frequency local disturbances.

[0077] Further, signal reconstruction and result analysis are performed, specifically by performing an inverse graphical Fourier transform on the filtered spectral coefficient vector to restore it to the nodal domain: s′= The denoised sensor response values ​​are obtained, representing the inverse graphical Fourier transform (IGFT), which is used to reconstruct the filtered signal in the spectral domain back to the node domain. U is the graphical Fourier basis matrix (composed of the eigenvectors of the graphical Laplace matrix), serving as a bridge for the transformation from the spectral domain to the node domain. The Fourier coefficients are filtered in the spectral domain, retaining only the effective frequency components, such as removing low-frequency components of noise or signals in specific frequency bands. s′ is the reconstructed node-domain signal, preserving key features related to structural health (such as global vibration modes and local damage signals) while filtering out interference noise. It can be directly used for subsequent structural state analysis or as input features for graph neural networks. Compared to the original signal s, s′ significantly reduces the influence of random noise and local anomalies, exhibiting a smoother spatial distribution. In engineering, the signals before and after filtering can be compared, for example, plotting the strain values ​​of adjacent sensors on a cross-section: the original data curve is jagged and undulating, while the reconstructed curve is smoother and more continuous, indicating that noise is suppressed. Furthermore, modal spectrum analysis can be further performed in the spectral domain: based on the graph Fourier coefficients... By analyzing the energy proportions in different λ segments, we can determine which mode dominates the structural vibration. For example, if the coefficient corresponding to the first eigenvalue has the highest proportion, it indicates that the event mainly excited the first-order mode. If higher-order components increase significantly, it may mean that there is a local anomaly or damage (because healthy structures rarely excite high-frequency local modes). Combining the results of finite element modal analysis, the mode shapes identified by the graph signal spectrum analysis can be compared with the normal state benchmark: when the overall stiffness of the bridge decreases, the lower-order eigenvalues ​​will decrease and the mode shapes will change; the corresponding component changes in the graph can be used as identification indicators. When a component (such as a cable-stayed bridge) is locally damaged, the signal of the relevant nodes of that cable will generate high-frequency components, reflected in the increased energy in the high λ segment of the graph Fourier spectrum. Through the above spectrum analysis method, maintenance personnel can extract the global modal characteristics and local anomaly characteristics of the structure from the sensor network data: global characteristics are used to assess the overall dynamic performance and health status of the cable-stayed bridge, while local high-frequency anomalies are used to locate potential damage sites.

[0078] In this embodiment, the graph signal processing module transforms the structural monitoring signal from the node domain to the spectral domain. Using graph Fourier transform, it separates low-frequency components reflecting global trends from high-frequency components capturing local anomalies. Then, targeted filtering (such as retaining key frequency bands and filtering out noise) purifies the signal, effectively reducing noise to eliminate environmental interference and focusing on damage-related characteristic frequencies. Simultaneously, spectral coefficient filtering achieves data compression, significantly reducing the computational load of subsequent models. This processing method fully integrates structural topology and physical signal characteristics, making the extracted features more closely aligned with engineering practice. It provides high-quality input for models such as graph neural networks, ultimately improving the accuracy and efficiency of structural health monitoring. It exhibits significant advantages, particularly in identifying subtle local damage and distinguishing complex interference signals. By combining GFT-based graph signal processing strategies with data-driven methods such as graph neural networks, the accuracy and reliability of cable-stayed bridge structural health monitoring can be improved in engineering practice.

[0079] Based on the same inventive concept, this application also provides a structural health monitoring device for implementing the above-mentioned structural health monitoring method. The solution provided by this device is similar to the solution described in the above-described method; therefore, the specific limitations of the one or more structural health monitoring device embodiments provided below can be found in the limitations of the structural health monitoring method described above, and will not be repeated here.

[0080] In one exemplary embodiment, such as Figure 4 As shown, a structural health monitoring device is provided, the device comprising: The acquisition module 510 is used to acquire the physical layout information of each monitoring point in the engineering structure and the multi-source monitoring signals of the monitoring points. Module 520 is used to construct a graph structure based on physical layout information. The graph structure includes nodes and edges. Nodes are used to represent components with monitoring points in the engineering structure, and edges are used to represent the relationships between the components. Processing module 530 is used to construct a graph signal from multi-source monitoring signals, and based on the graph structure, perform spectrum analysis and filtering on the graph signal through a graph signal processing model to obtain the processed graph signal. The monitoring module 540 is used to input the graph structure and the processed graph signal into the trained graph neural network model to obtain the health status monitoring results. The health status monitoring results include the damage probability value and status classification result of each node.

[0081] As an optional implementation, module 520 is specifically used for: Abstract the components with monitoring points in the engineering structure into nodes; Obtain the relationship information between each node from the physical layout information, and establish edges based on the relationship information; the relationship information includes at least one of the following: physical connection relationship, spatial adjacency relationship, mechanical coupling relationship, and disease propagation relationship; Determine the parameter information of edges and nodes, and assign corresponding weight values ​​to each edge based on the parameter information; the parameter information includes: edge distance, stiffness coefficient, and mutual information; A graph structure is formed based on weight values, edges, and nodes.

[0082] As an optional implementation, the construction module 520 is also used for: If there is a physical connection between two nodes, create an edge between the two nodes; If the spatial adjacency between two nodes is less than the first preset threshold, an edge is established between the two nodes; If the mechanical coupling relationship between two nodes is greater than the second preset threshold, an edge is established between the two nodes. If the disease propagation relationship between two nodes meets the preset conditions, an edge is established between the two nodes.

[0083] As an optional implementation, the processing module 530 is specifically used for: A graph signal is constructed based on the time series signals of each node in the multi-source monitoring signal; the rows in the graph signal represent each node, and the columns in the graph signal represent the time series signals. Determine the adjacency matrix, degree matrix, and identity matrix based on the graph structure, and construct the graph Laplacian matrix based on the adjacency matrix, degree matrix, and identity matrix; Based on the graph signal, the graph Laplacian matrix is ​​subjected to eigenvalue decomposition to obtain the eigenvector matrix; Based on the eigenvector matrix, the graph signal is converted from the node domain to the frequency domain through graph Fourier transform to obtain the frequency domain signal; The frequency domain signal is filtered to obtain the processed graph signal; the filtering includes noise reduction and / or dimensionality reduction.

[0084] As an optional implementation, the monitoring module 540 is specifically used for: Obtain the node attribute information of each node; The node attribute information, the processed graph signal, and the graph structure are subjected to dimensional unification and feature splicing to obtain spliced ​​features. The concatenated features are input into each graph convolutional layer, and activation functions, residual connections, and normalization are performed sequentially in each graph convolutional layer to extract structural features. The structural features include global structural features and local structural features. The structural features are processed by assigning attention to them according to their importance through a graph attention layer to obtain attention features. Attention features are classified using a graph classification layer to obtain health status monitoring results.

[0085] As an alternative implementation, the graph neural network model is constructed through the following steps: Acquire structural layout information, domain rule data, and sample structure monitoring signals of the sample engineering structure; the sample structure monitoring signals correspond to state annotation results. Based on the structural layout information, a sample graph structure is constructed, and the sample structure monitoring signal is subjected to spectrum analysis and filtering through a graph signal processing model to obtain the processed sample graph signal. The processed sample image signals and sample image structures are divided into training sets and validation sets; The training set is input into the initial graph neural network model to obtain the output result; Based on the output results and state labeling results, construct the cross-entropy loss term; Transform domain rule data into logical rule items, and construct logical rule loss items based on logical rule items; The cross-entropy loss term and the logical rule loss term are combined to construct a loss function. The parameters in the initial graph neural network model are iteratively trained and optimized according to minimizing the loss function to obtain the trained model. The validation set is input into the trained model for validation processing, resulting in a graph neural network model.

[0086] As an optional implementation, the above-described apparatus is further used for: Convert health status monitoring results into graphical formats; the graphical formats include at least one of the following: model diagram, bar chart, and line graph. The graph format is displayed through a visual interface, and alarm prompts are given for data in the health status monitoring results that meet the alarm conditions.

[0087] The structural health monitoring device provided in this application acquires physical layout information of monitoring points and multi-source monitoring signals, enabling the data range to cover the physical location of components and multi-dimensional state signals. This not only ensures the comprehensiveness of data sources but also provides complete input information for subsequent modeling. Based on the physical layout information, a graph structure is constructed, overcoming the limitation of traditional time-series analysis that only focuses on the time dimension. This transforms the spatial relationships of engineering structures into quantifiable graph topologies, achieving accurate characterization of potential spatial relationships within the structure. After constructing the multi-source monitoring signals into graph signals, they are processed through a graph signal processing model. Spectral analysis and filtering effectively separate effective structural features (low-frequency global trends) and noise interference (high-frequency local disturbances) in the signal, improving the purity and effectiveness of the input features and obtaining an effectively processed graph signal. Finally, the graph structure and the processed graph signal are input into a trained graph neural network model. By utilizing the graph neural network's joint learning ability on topological associations and node features, the damage probability value and state classification result of each node can be accurately output, significantly improving the accuracy of health monitoring. This solves the core problem that traditional time series analysis is unable to characterize spatial relationships and correlation evolution, resulting in low monitoring accuracy.

[0088] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores video tag processing data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a structural health monitoring method.

[0089] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0090] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0091] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0092] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0093] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0094] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0095] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0097] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for monitoring structural health, characterized in that, The structural health monitoring method includes: Acquire the physical layout information of each monitoring point in the engineering structure and the multi-source monitoring signals of the monitoring points; Based on the physical layout information, a graph structure is constructed; the graph structure includes nodes and edges, the nodes are used to represent the components with monitoring points in the engineering structure, and the edges are used to represent the relationships between the components. The multi-source monitoring signals are constructed into a graph signal. Based on the graph structure, the graph signal is subjected to spectral analysis and filtering through a graph signal processing model to obtain the processed graph signal. The graph structure and the processed graph signal are input into the trained graph neural network model to obtain the health status monitoring results; the health status monitoring results include: the damage probability value and status classification result of each node.

2. The structural health monitoring method according to claim 1, characterized in that, Based on the physical deployment information, a graph structure is constructed, including: The components with monitoring points in the engineering structure are abstracted as nodes; The relationship information between each node is obtained from the physical layout information, and the edge is established based on the relationship information; the relationship information includes at least one of the following: physical connection relationship, spatial adjacency relationship, mechanical coupling relationship, and disease propagation relationship; The parameter information of the edges and the nodes is determined, and a corresponding weight value is assigned to each edge according to the parameter information; the parameter information includes: edge distance, stiffness coefficient, and mutual information; The graph structure is formed based on the weight values, the edges, and the nodes.

3. The structural health monitoring method according to claim 2, characterized in that, The edges are established based on the relationship information, including at least one of the following: If there is a physical connection between two nodes, an edge is established between the two nodes; If the spatial adjacency between two nodes is less than a first preset threshold, an edge is established between the two nodes; If the mechanical coupling relationship between two nodes is greater than a second preset threshold, an edge is established between the two nodes. If the disease propagation relationship between two nodes meets the preset conditions, an edge is established between the two nodes.

4. The structural health monitoring method according to claim 1, characterized in that, The multi-source monitoring signals are constructed into a graph signal. Based on the graph structure, the graph signal is subjected to spectral analysis and filtering using the graph Fourier transform in the graph signal processing model to obtain the processed graph signal, including: Based on the time series signals of each node in the multi-source monitoring signals, a graph signal is constructed; the rows in the graph signal represent each node, and the columns in the graph signal represent the time series signals. The adjacency matrix, degree matrix, and identity matrix are determined based on the graph structure, and the graph Laplacian matrix is ​​constructed based on the adjacency matrix, degree matrix, and identity matrix. Based on the graph signal, the graph Laplacian matrix is ​​subjected to eigenvalue decomposition to obtain the eigenvector matrix; Based on the eigenvector matrix, the graph signal is converted from the node domain to the frequency domain through the graph Fourier transform to obtain the frequency domain signal; The frequency domain signal is filtered to obtain the processed graph signal; the filtering includes: noise reduction, and / or, dimensionality reduction.

5. The structural health monitoring method according to claim 1, characterized in that, The graph neural network model includes: multiple graph convolutional layers, graph attention layers, and graph classification layers; The graph structure and the processed graph signal are input into a trained graph neural network model to obtain health status monitoring results, including: Obtain the node attribute information of each node; The node attribute information, the processed graph signal, and the graph structure are subjected to dimensional unification and feature concatenation to obtain the concatenated features. The spliced ​​features are input into each of the graph convolutional layers, and activation functions, residual connections, and normalization are sequentially performed in each graph convolutional layer to extract structural features; the structural features include: global structural features and local structural features; The structural features are processed by assigning attention to them according to their importance through the graph attention layer to obtain attention features; The attention features are classified through the graph classification layer to obtain the health status monitoring results.

6. The structural health monitoring method according to claim 1, characterized in that, The graph neural network model is constructed through the following steps: Acquire structural layout information, domain rule data, and sample structure monitoring signals of the sample engineering structure; the sample structure monitoring signals correspond to state annotation results. Based on the structural layout information, a sample graph structure is constructed, and the sample graph monitoring signal is subjected to spectrum analysis and filtering processing through a graph signal processing model to obtain the processed sample graph signal. The processed sample image signal and the sample image structure are divided into a training set and a validation set; The training set is input into the initial graph neural network model to obtain the output result; Based on the output results and the state labeling results, a cross-entropy loss term is constructed; The domain rule data is transformed into logical rule terms, and a logical rule loss term is constructed based on the logical rule terms. The cross-entropy loss term and the logical rule loss term are combined to construct a loss function. The parameters in the initial graph neural network model are iteratively trained and optimized according to minimizing the loss function to obtain the trained model. The validation set is input into the trained model for validation processing to obtain the graph neural network model.

7. The structural health monitoring method according to claim 1, characterized in that, After obtaining the health status monitoring results, the method further includes: The health status monitoring results are converted into a graphical format; the graphical format includes at least one of the following: model diagram, bar chart, and line graph. The graphical format is displayed through a visual interface, and alarm prompts are given for data in the health status monitoring results that meet the alarm conditions.

8. A structural health monitoring device, characterized in that, The structural health monitoring device includes: The acquisition module is used to acquire the physical layout information of each monitoring point in the engineering structure and the multi-source monitoring signals of the monitoring points; A construction module is used to construct a graph structure based on the physical layout information; the graph structure includes nodes and edges, the nodes are used to represent the components with monitoring points in the engineering structure, and the edges are used to represent the relationships between the components. The processing module is used to construct the multi-source monitoring signals into a graph signal, and based on the graph structure, perform spectrum analysis and filtering on the graph signal through a graph signal processing model to obtain the processed graph signal; The monitoring module is used to input the graph structure and the processed graph signal into the trained graph neural network model to obtain health status monitoring results; the health status monitoring results include: the damage probability value and status classification result of each node.

9. A computer device, comprising: The memory and processor contain a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the structural health monitoring method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the structural health monitoring method according to any one of claims 1-7.

Images