Local-global state mapping and real-time early warning method for long-span spatial structure

Abstract

The application discloses a large-span space structure local-global state mapping and real-time early warning method and related devices, wherein the method comprises the following steps: arranging multiple sensor networks to collect time series monitoring data at preset local measuring points of a large-span space structure; combining design drawings, finite element models and historical records to construct a priori knowledge base; performing space-time alignment on the monitoring data and the knowledge base to form a sensor network graph structure integrating physical connection and dynamic correlation; the sensor network graph structure extracts node time series features through a GRU unit, and aggregates spatial correlation features through an ST-AGCN unit; predicting a global mechanical state including unmeasured area strain, displacement and internal force through a physical constraint decoder; and performing spatial clustering analysis on the long-term prediction field data by using kernel density estimation and local anomaly factor algorithm to identify abnormal areas of spatial aggregation mode. The application realizes accurate reconstruction of the global mechanical state of the large-span space structure and real-time damage early warning by using sparse local monitoring data.

Description

Technical Field

[0001] This invention relates to the fields of engineering structure monitoring and artificial intelligence technology, specifically to a method and device for local-to-global state mapping and real-time early warning of large-span spatial structures, as well as computing equipment. Background Technology

[0002] Large-span spatial structures, as a core feature of modern large-scale public buildings and infrastructure, are characterized by novel designs and complex systems. These structures are exceptionally sensitive to localized damage, material degradation, and changes in boundary conditions. Even minor initial defects can trigger catastrophic cascading failures under dynamic loads such as wind, snow, and earthquakes. Traditional methods of periodic manual inspections and reactive maintenance suffer from drawbacks such as being "invisible, inaccurate in measurement, and slow to react," failing to achieve early detection and warning of structural "sub-health" conditions.

[0003] Existing detection and early warning methods mainly include simulation-based methods and neural network model prediction methods. Among them, simulation-based methods based on high-fidelity finite element models and data assimilation continuously correct and update model parameters (such as stiffness and boundary conditions) through data assimilation techniques such as Kalman filtering and particle filtering, aiming to make the simulation output approximate the real response and thus assess the structural state. However, since high-fidelity models of large-span structures often have hundreds of thousands or even millions of degrees of freedom, this frequently leads to the disaster of dimensionality calculation. Deep learning-based neural network models are extremely dependent on training data. Obtaining a labeled dataset that covers all possible health states, damage modes, and environmental load conditions of the structure is almost impossible in engineering practice, resulting in extremely weak model generalization ability.

[0004] To address the aforementioned issues, this invention proposes a method for local-to-global state mapping and real-time early warning of large-span spatial structures. This method utilizes sparse local monitoring data to achieve accurate reconstruction of the global mechanical state of large-span spatial structures and real-time damage early warning. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method, device and computing equipment for local-global state mapping and real-time early warning of large-span spatial structures.

[0006] According to one aspect of the present invention, a method for local-to-global state mapping and real-time early warning of large-span spatial structures is provided, comprising: Multiple sensor networks are deployed at predetermined local measurement points in the long-span spatial structure to collect time-series monitoring data; based on the design drawings, finite element model files, and historical inspection records of the long-span spatial structure, a prior knowledge base including geometric topology, material properties, and boundary conditions is constructed; wherein, the time-series monitoring data includes static response data and dynamic response data; the predetermined local measurement points include supports, mid-span, member connections, and dynamic response sensitive areas; The time-series monitoring data is spatiotemporally aligned and associated with the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges. The initial graph structure is input into the independent GRU unit of the corresponding sensor node. The node-level temporal feature vector is output by the GRU unit and input into the ST-AGCN unit. The ST-AGCN unit aggregates the spatial features of multi-hop neighbor nodes and outputs the node-level spatial correlation feature vector. A physical constraint decoder is constructed based on the node-level spatial association feature vectors of all nodes and the prior knowledge base to output global mechanical state prediction values; wherein, the global mechanical state prediction values ​​include strain, displacement and internal force in the area where no sensors are deployed; the loss function of the physical constraint decoder includes strong constraint terms of the finite element equilibrium equation, mean square error of data prediction and weak modal correlation loss terms. Based on the global mechanical state prediction field of the large-span spatial structure during a preset health status monitoring period, the probability distribution of each key parameter is fitted using a kernel density estimation method; cluster analysis of the probability distribution of each key parameter is performed using a local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, potential damage is determined and an early warning message is triggered; wherein, the early warning message includes the parameter exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the recommended review level.

[0007] In one alternative approach, the acquisition of time-series monitoring data by arranging multiple sensor networks at predetermined local measurement points on a large-span spatial structure further includes: Based on the structural finite element model in the prior knowledge base, multi-round modal analysis and static sensitivity analysis were performed on the large-span spatial structure to identify the key mechanical response regions and high strain energy density regions under different load conditions. Based on the critical mechanical response region and the high strain energy density region, the preset local measuring points are divided into two categories: critical measuring points and auxiliary measuring points. The critical measuring points are deployed at the end connection regions of the main load-bearing components, the mid-span sections of large-span components, the abrupt change regions of support reaction forces, and the mode peak points of the first N main modes identified by finite element analysis. The auxiliary measuring points are deployed in the secondary sensitive regions of the structure. Fiber optic strain sensors, total station prism groups, and triaxial accelerometers are deployed at the key measuring points; low-power wireless strain sensors and accelerometers are deployed at the auxiliary measuring points.

[0008] In one alternative approach, based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a prior knowledge base including geometric topology, material properties, and boundary conditions is constructed, further comprising: The component type, spatial coordinates, cross-sectional dimensions, material grade and connection relationship are extracted from the design drawings using BIM analysis software to generate a preliminary geometric topology diagram. The discretized node coordinates, element connection relationships, material constitutive model parameters and boundary condition constraint information are extracted from the finite element model file. Based on the preliminary geometric topology diagram and the extraction results of the finite element model file, a structural digital foundation model is constructed. The structural digital foundation model uses physical components as entity nodes and physical connections between physical components as entity edges. The node attributes of the entity nodes include spatial coordinates, material properties, and cross-sectional properties, and the edge attributes of the entity edges include connection type and stiffness coefficient. Based on the historical inspection records, historical damage locations, damage types, repair measures, environmental load events, and performance degradation assessments are extracted. The extraction results of the historical inspection records are then appended to the corresponding entity nodes or entity edges in the structural digital generation basic model to form a structural knowledge graph with time-varying state dimensions. The prior knowledge base is obtained by encoding the prior physical rules of the structure and the dynamic database of boundary conditions based on the structural knowledge graph.

[0009] In one alternative approach, spatiotemporally aligning and associating the time-series monitoring data with the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges further includes: Using the global coordinate system of the structural digital twin basic model as the spatial reference and the main GNSS clock of the sensor acquisition system as the time reference, the spatial installation coordinates of each sensor node are bound to the geometric center or feature point of the nearest component in the prior knowledge base through surveying data, and each sensor node is mapped to the entity node or entity edge in the structural digital twin basic model. If the entity nodes mapped by two sensor nodes are directly connected by an entity edge in the basic digital twin model of the structure, a physical connection edge is established for the two sensor nodes. For any two sensor nodes, calculate the damping correlation coefficient of the acceleration time series within the current time window. When the damping correlation coefficient between the two nodes exceeds a preset threshold, establish a dynamic correlation edge. The sensor nodes and the adjacency matrix composed of physical connection edges and dynamic correlation edges are encapsulated into a spatiotemporal graph data structure to obtain the initial graph structure.

[0010] In one alternative embodiment, the GRU unit comprises three parallel processing branch layers: the first branch layer is a high-frequency vibration feature extraction branch layer, comprising a gated depthwise separable convolutional network layer consisting of multiple 1x3 deep convolutional layers and a gated linear network layer; the second branch layer is a quasi-static response feature extraction branch layer, comprising two standard GRU layers for capturing slow changing trends in strain and displacement data and short- to medium-term load responses; and the third branch layer is a long-term trend and anomaly perception branch, comprising a single-layer GRU with an attention mechanism for identifying long-term sudden anomaly patterns in the data. The ST-AGCN unit includes a multi-head spatial attention sublayer, a spatiotemporal graph convolution sublayer, and a gated temporal convolution sublayer. The multi-head spatial attention sublayer employs K independent attention heads. The spatiotemporal graph convolution sublayer, after obtaining an adaptive adjacency matrix, uses Chebyshev polynomial approximation graph convolution for information propagation. The gated temporal convolution sublayer, after spatial convolution, applies dilated causal convolution to the feature sequence of each node to extract temporal patterns and uses a gated activation unit to control the information flow.

[0011] In an optional approach, fitting the probability distribution of each key parameter using a kernel density estimation method, based on the predicted global mechanical state field of the large-span spatial structure during a preset health monitoring period, further includes: During a preset health status monitoring cycle, time-series monitoring data is input into the physical constraint decoder to obtain a continuous global mechanical state prediction field sequence. For each key parameter in the global mechanical state prediction field sequence, its complete time series value is extracted at each structural location. The health benchmark feature vector of the corresponding structural location is obtained based on the complete time series value. A health status benchmark database is constructed based on all structural locations and key parameters. For each key parameter in the health status benchmark database, at each structural location, a probability distribution model is established using the adaptive bandwidth kernel density estimation method. Based on the probability distribution model of each statistical characteristic at each structural location and the generated virtual sample set, the Mahalanobis distance from the sample point to its distribution center is calculated to obtain the probability distribution of the health status at each structural location.

[0012] In one alternative approach, clustering analysis of the probability distributions of key parameters using a local anomaly factor algorithm to identify anomalous regions that form spatial clusters further includes: The statistical characteristic values ​​of each structural location are compared with the health benchmark characteristic vectors of the corresponding locations in the health status benchmark database to obtain the characteristic deviation vector of each structural location. The local anomaly factor algorithm is used to quantify the degree of anomaly of the feature deviation vector of each structural location; for each structural location, its spatial neighborhood is defined according to the structural topological relationship in the prior knowledge base; the spatial neighborhood forms a multi-scale neighborhood set for each location based on the physical connection neighborhood and the spatial distance neighborhood; Based on the Euclidean distance of the feature deviation vector in the high-dimensional feature space, calculate the reachable distance from each location to its spatial neighborhood; based on the reachable distance from each location to its neighborhood and the degree of anomaly, calculate the local anomaly factor value. The local anomaly factor values ​​at each structural location are mapped to structural spatial coordinates to form a structural anomaly degree distribution map. Cluster analysis is performed on the high anomaly degree locations in the structural anomaly degree distribution map to obtain a set of spatially clustered anomaly regions. The set of anomaly regions includes the boundary location, number of structural locations, average anomaly degree of the region, and spatial distribution morphology information of each anomaly region.

[0013] In one alternative approach, arranging fiber optic strain sensors, total station prism groups, and triaxial accelerometers at the critical measuring points, and arranging low-power wireless strain sensors and accelerometers at the auxiliary measuring points, further includes: For each critical measuring point, a spatial fiber optic grating sensor array consisting of at least six sensors with a sensor spacing of less than or equal to 150 cm is installed in the end connection area of ​​its main load-bearing component; at least five fiber optic grating sensors are arranged at equal intervals of 200 cm along the height direction of the component at its mid-span section to monitor the cross-sectional strain gradient distribution and neutral axis offset; a quadrangular prism adaptive array consisting of four reflecting prisms is installed in the abrupt change area of ​​the support reaction force to obtain the three-dimensional displacement of the measuring point; at least a dual-channel broadband accelerometer group is deployed at each critical measuring point. For auxiliary measurement points, a low-power triaxial wireless accelerometer network is deployed. The accelerometer network uses a low frequency of 5Hz during the structural calm period and automatically increases to 50Hz sampling during periods of wind vibration or active pedestrian load.

[0014] According to another aspect of the present invention, a device for local-to-global state mapping and real-time early warning of large-span spatial structures is provided, comprising: The data acquisition and prior knowledge base construction module is used to deploy multiple sensor networks at preset local measuring points in a large-span spatial structure to collect time-series monitoring data. Based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a prior knowledge base including geometric topology, material properties, and boundary conditions is constructed. The time-series monitoring data includes static response data and dynamic response data. The preset local measuring points include supports, mid-span, member connections, and dynamic response sensitive areas. The initial graph structure construction module is used to perform spatiotemporal alignment and association between the time-series monitoring data and the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges. The spatial correlation feature vector acquisition module is used to input the initial graph structure into the independent GRU unit of the corresponding sensor node, output the node-level temporal feature vector through the GRU unit and input it into the ST-AGCN unit, and aggregate the spatial features of multi-hop neighboring nodes through the ST-AGCN unit to output the node-level spatial correlation feature vector. The global mechanical state prediction value acquisition module is used to construct a physical constraint decoder based on the node-level spatial correlation feature vectors of all nodes and the prior knowledge base to output the global mechanical state prediction value; wherein, the global mechanical state prediction value includes the strain, displacement and internal force of the area where no sensors are deployed; the loss function of the physical constraint decoder includes the strong constraint term of the finite element equilibrium equation, the mean square error of data prediction and the weak modal correlation loss term. The abnormal region identification and early warning module is used to fit the probability distribution of each key parameter based on the global mechanical state prediction field of the large-span spatial structure during a preset health status monitoring period, using a kernel density estimation method; to perform cluster analysis on the probability distribution of each key parameter using a local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, it is determined that potential damage has occurred and an early warning message is triggered; wherein, the early warning message includes the parameter exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the recommended review level.

[0015] According to another aspect of the present invention, a computing device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described large-span spatial structure local-global state mapping and real-time early warning method.

[0016] According to the solution provided by the present invention, multiple sensor networks are arranged at preset local measurement points in a large-span spatial structure to collect time-series monitoring data; based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a priori knowledge base including geometric topology, material properties, and boundary conditions is constructed; wherein, the time-series monitoring data includes static response data and dynamic response data; the preset local measurement points include supports, mid-span, member connections, and dynamic response sensitive areas; the time-series monitoring data is spatiotemporally aligned and correlated with the structural topology information in the priori knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges; the initial graph structure is input into the independent GRU units corresponding to the sensor nodes, and the node-level time-series feature vectors are output by the GRU units and input into the ST-AGCN unit, and the spatial features of multi-hop neighborhood nodes are aggregated by the ST-AGCN unit. The system outputs node-level spatial correlation feature vectors; based on the node-level spatial correlation feature vectors of all nodes and the prior knowledge base, a physical constraint decoder is constructed to output global mechanical state prediction values; wherein, the global mechanical state prediction values ​​include strain, displacement, and internal forces in areas where no sensors are deployed; the loss function of the physical constraint decoder includes strong constraint terms of the finite element equilibrium equations, mean square error of data prediction, and weak modal correlation loss terms; based on the global mechanical state prediction field of the large-span spatial structure during a preset health state monitoring period, the probability distribution of each key parameter is fitted using a kernel density estimation method; cluster analysis of the probability distribution of each key parameter is performed using a local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, potential damage is determined and an early warning message is triggered; wherein, the early warning message includes the parameters exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the suggested review level. This invention utilizes sparse local monitoring data to achieve accurate reconstruction of the global mechanical state of a large-span spatial structure and real-time damage early warning. Specifically, by using a spatiotemporal graph neural network (GRU-ST-AGCN) to simulate the transmission and diffusion processes in structural topological networks, the mechanical state of sensorless areas can be inferred from limited local observations. Prior knowledge is structured and input into the model in the form of a structural knowledge graph and a spatiotemporal graph. In the constraint decoder, the finite element equilibrium equations are added as strong constraints to the loss function, ensuring the interpretability of the output global prediction field and making it easily acceptable and verifiable by engineers. A probability distribution model of the health status baseline is established through kernel density estimation. The Local Anomaly Factor (LOF) algorithm quantifies the isolation degree of each location, and spatial clustering analysis identifies anomalous regions forming clusters in space. This approach is extremely sensitive to early damage; even if the change at each point does not exceed the traditional threshold, the pattern of collective shift can be effectively captured, while significantly suppressing false alarms caused by single-point noise.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This illustrates a flowchart of the local-to-global state mapping and real-time early warning method for large-span spatial structures according to an embodiment of the present invention. Figure 1 ; Figure 2 This illustrates a flowchart of the local-to-global state mapping and real-time early warning method for large-span spatial structures according to an embodiment of the present invention. Figure 2 ; Figure 3 A schematic diagram of sensor deployment according to an embodiment of the present invention is shown; Figure 4 A schematic diagram of an abnormal region forming a cluster in space according to an embodiment of the present invention is shown; Figure 5 A schematic diagram of the framework of the local-to-global state mapping and real-time early warning device for large-span spatial structures according to an embodiment of the present invention is shown; Figure 6 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. Detailed Implementation

[0019] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0020] Figure 1 , Figure 2 The diagram illustrates the flow chart of the local-to-global state mapping and real-time early warning method for large-span spatial structures according to an embodiment of the present invention. Figure 1 , two Specifically, such as Figure 1 , Figure 2 As shown, it includes the following steps: Step S101: Multiple sensor networks are deployed at preset local measurement points in the large-span spatial structure to collect time-series monitoring data; based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a priori knowledge base including geometric topology, material properties, and boundary conditions is constructed; wherein, the time-series monitoring data includes static response data and dynamic response data; the preset local measurement points include supports, mid-span, member connections, and dynamic response sensitive areas.

[0021] In this embodiment, a sensor network is deployed in key mechanical regions (such as supports, mid-span, and member connections) of a large-span spatial structure to achieve comprehensive monitoring of the structure's local response. A priori knowledge base is constructed using design drawings, finite element models, and historical inspection records. The structure's geometry, topology, material properties, and boundary conditions are digitized, allowing monitoring data to be integrated with the structural context. This enhances the interpretability of state mapping, reduces inconsistencies inherent in purely data-driven methods, and improves the accuracy of state prediction, making it particularly suitable for large-span spatial structures with complex structures and variable loads. Simultaneous acquisition of static response data (such as strain and displacement) and dynamic response data (such as acceleration) covers the structural behavior under different loads (such as static load, wind vibration, and earthquakes), providing a closer reflection of engineering realities. Synchronous tracking of the structure's short-term sudden responses and long-term slow evolution facilitates early detection of damage, performance degradation, and abnormal events. The selection of measurement points focuses on mechanically sensitive areas (such as dynamic response sensitive areas), reducing the cost of large-scale sensor deployment while ensuring the monitoring system captures the most critical structural behavioral characteristics.

[0022] In one alternative approach, the acquisition of time-series monitoring data by arranging multiple sensor networks at predetermined local measurement points on a large-span spatial structure further includes: Based on the structural finite element model in the prior knowledge base, multi-round modal analysis and static sensitivity analysis were performed on the large-span spatial structure to identify the key mechanical response regions and high strain energy density regions under different load conditions. Based on the critical mechanical response region and the high strain energy density region, the preset local measuring points are divided into two categories: critical measuring points and auxiliary measuring points. The critical measuring points are deployed at the end connection regions of the main load-bearing components, the mid-span sections of large-span components, the abrupt change regions of support reaction forces, and the mode peak points of the first N main modes identified by finite element analysis. The auxiliary measuring points are deployed in the secondary sensitive regions of the structure. Fiber optic strain sensors, total station prism groups, and triaxial accelerometers are deployed at the key measuring points; low-power wireless strain sensors and accelerometers are deployed at the auxiliary measuring points.

[0023] In this embodiment, based on the finite element model in the prior knowledge base, the location of measuring points is determined through modal analysis and static sensitivity analysis. This enables the accurate identification of key mechanical response regions and high strain energy density regions with the greatest impact on the overall structural response under different load conditions. A limited number of sensors (such as fiber optic gratings and total stations) are concentrated on key measuring points that best reflect the overall health of the structure, ensuring the representativeness and information density of the monitoring data. Simultaneously, lower-cost wireless sensors are used to supplement coverage at auxiliary measuring points, significantly reducing the construction and maintenance costs of a large-scale monitoring network while ensuring system effectiveness. The measuring point classification (key / auxiliary) is not a simple hierarchical division but a functional distinction based on the monitoring objectives. Key measuring points are deployed at specific locations that best reveal the static and dynamic characteristics of the structure, such as connection areas (reflecting internal force transmission), mid-span sections (reflecting overall deformation and bending), support areas (reflecting boundary constraints and reactions), and modal peak points (reflecting dynamic characteristics). For example, strain gradient monitoring at mid-span sections can be directly used to assess neutral axis offset and bending stiffness degradation of components, while acceleration data at modal peak points are most useful for identifying changes in frequency and damping ratio. At critical measuring points, fiber optic strain sensors (high accuracy, good long-term stability, and anti-interference) and total station prism groups (absolute displacement measurement) are used to obtain the most reliable static response benchmark data; simultaneously, triaxial accelerometers are used to capture dynamic responses. At auxiliary measuring points, low-power wireless sensors are employed to balance the monitoring coverage with deployment complexity. For instance, in a comprehensive sports stadium, the main roof cable-stayed dome structure is lightweight and sensitive to dynamic loads, requiring long-term health monitoring to ensure event safety and public health. Using a finite element model of the stadium for simulation, static analysis revealed significant stress concentration and high strain energy density at the connection nodes between cables and struts, the mid-span region of the external pressure ring beam, and the support ring beam connecting to the lower concrete grandstand. Modal analysis revealed that the first mode shape was a vertical translational motion of the entire roof, with the peak value located at the center of the dome; the second mode shape was an antisymmetric torsion, with the peak value located in two diagonal regions along the diagonal direction. Measuring points were placed at all key cable clamp nodes and the upper and lower ends of the struts to monitor cable force changes and node slippage. Three representative sections were selected in the span of the central tension ring and the external pressure ring beam, and strain sensor arrays were placed at each section. Measuring points were placed above the eight key supports of the peripheral support ring beam to monitor uneven settlement and horizontal displacement. Measuring points were placed at the center of the dome (first-order peak value) and in the two diagonal regions (second-order peak values). Auxiliary measuring points were selected and placed on the remaining cable segments and grid nodes at approximately 30-meter grid intervals. A dense fiber optic grating sensor network was deployed at all key measuring points to monitor cable strain and ring beam strain. Miniature prism targets were installed for scanning by an automated total station and for fixing high-sensitivity triaxial accelerometers. Solar-powered wireless accelerometer nodes were installed at the auxiliary measuring points, forming a vibration sensing network covering the entire roof.

[0024] In one alternative approach, arranging fiber optic strain sensors, total station prism groups, and triaxial accelerometers at the critical measuring points, and arranging low-power wireless strain sensors and accelerometers at the auxiliary measuring points, further includes: For each critical measuring point, a spatial fiber optic grating sensor array consisting of at least six sensors with a sensor spacing of less than or equal to 150 cm is installed in the end connection area of ​​its main load-bearing component; at least five fiber optic grating sensors are arranged at equal intervals of 200 cm along the height direction of the component at its mid-span section to monitor the cross-sectional strain gradient distribution and neutral axis offset; a quadrangular prism adaptive array consisting of four reflecting prisms is installed in the abrupt change area of ​​the support reaction force to obtain the three-dimensional displacement of the measuring point; at least a dual-channel broadband accelerometer group is deployed at each critical measuring point. For auxiliary measurement points, a low-power triaxial wireless accelerometer network is deployed. The accelerometer network uses a low frequency of 5Hz during the structural calm period and automatically increases to 50Hz sampling during periods of wind vibration or active pedestrian load.

[0025] In this embodiment, the dense array can monitor the stress concentration distribution and possible slip or plastic deformation development patterns in the connection area, thereby effectively identifying local damage such as bolt loosening and weld cracking. At least five sensors are arranged along the height of the mid-span section to directly monitor the cross-sectional strain gradient distribution and accurately calculate the neutral axis position and its offset, which is a key indicator for assessing the overall bending stiffness degradation and potential instability risk of the component. In critical displacement-sensitive areas such as supports, a quadrangular prism adaptive array consisting of four reflecting prisms is used, with a dual-channel broadband accelerometer group deployed at each critical measuring point. This achieves true three-dimensional attitude calculation and anti-interference displacement measurement. The quadrangular prism array effectively eliminates ranging errors caused by small prism rotations through redundant observation, calculating the precise three-dimensional translational and angular displacements of the measuring points. The dual-channel broadband accelerometer group provides multi-axial, broadband vibration information, not only capturing the structural main frequency but also monitoring high-frequency local vibrations and dynamic characteristic changes in modally dense areas, providing richer dynamic fingerprints for damage identification. Figure 3 As shown, a quadrangular prism adaptive array (four reflecting prisms forming a small square array) is arranged in the support area; five fiber grating sensors are equidistantly arranged along the height direction of the member at the mid-span section; more than six fiber grating sensors (forming a small cluster) are arranged in the connection area at the end of the member; auxiliary measuring points are distributed in the secondary sensitive areas (such as non-main load-bearing web members and secondary connections).

[0026] In one alternative approach, based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a prior knowledge base including geometric topology, material properties, and boundary conditions is constructed, further comprising: The component type, spatial coordinates, cross-sectional dimensions, material grade and connection relationship are extracted from the design drawings using BIM analysis software to generate a preliminary geometric topology diagram. The discretized node coordinates, element connection relationships, material constitutive model parameters and boundary condition constraint information are extracted from the finite element model file. Based on the preliminary geometric topology diagram and the extraction results of the finite element model file, a structural digital foundation model is constructed. The structural digital foundation model uses physical components as entity nodes and physical connections between physical components as entity edges. The node attributes of the entity nodes include spatial coordinates, material properties, and cross-sectional properties, and the edge attributes of the entity edges include connection type and stiffness coefficient. Based on the historical inspection records, historical damage locations, damage types, repair measures, environmental load events, and performance degradation assessments are extracted. The extraction results of the historical inspection records are then appended to the corresponding entity nodes or entity edges in the structural digital generation basic model to form a structural knowledge graph with time-varying state dimensions. The prior knowledge base is obtained by encoding the prior physical rules of the structure and the dynamic database of boundary conditions based on the structural knowledge graph.

[0027] In this embodiment, data from design (BIM drawings), simulation (finite element model), and operation and maintenance (historical inspection records) are integrated. BIM data provides macroscopic geometry and component information, finite element data provides discretized mechanical models and boundary conditions, and historical records include time-varying state information from actual service, improving the overall data quality and computational reliability of the system. Integrating historical inspection records (damage, repair, environmental events) into a knowledge base and associating them with specific components or connections allows understanding that a member was damaged and repaired in an event such as a typhoon, therefore its current stiffness may differ from the design value. It also reveals that a support area is prone to icing in winter, and boundary conditions may change seasonally, providing crucial information for accurately assessing the structure's current state and identifying cumulative damage and performance degradation trends. Encoding structural prior rules (such as equilibrium equations and constitutive relations) and a dynamic boundary condition database prevents physically impossible states (such as internal forces not satisfying equilibrium), significantly improving the reliability of state mapping and its engineering practical value.

[0028] For example, a stadium has been in service for over 10 years, undergoing numerous repairs and withstanding strong typhoons. Detailed information on over 5000 members and 2000 nodes was extracted from the stadium's BIM as-built model, including the member's type (e.g., Φ180×10mm round pipe), material (Q345C), and precise location and connection method in 3D space. Approximately 15000 nodes and 30000 elements used for calculations were analyzed from the design institute's ANSYS finite element model to obtain the precise elastic modulus of the steel (2.06×10⁻⁶).5 MPa) and the constraint status of all support nodes (e.g., fixed in one direction, free in another direction). Organize the venue's operation and maintenance records for the past 10 years, key records: ① In 2015, micro-cracks appeared in the grid node area above Gate 3 in the East Zone due to weld fatigue, and carbon fiber reinforcement was carried out; ② In 2020, Super Typhoon Hagupit passed through, and the measured wind speed on the roof reached level 9; ③ In 2022, routine inspection found that the preload of some high-strength bolts on the south side showed a decreasing trend. Create a 3D model of the BIM components as solids, and map the ANSYS element information onto these solids to form the correspondence between the computational model and the physical components. Assign attributes to each member (solid node): coordinates, material = Q345C, section = Φ180×10. Assign attributes to each welded node or bolted connection node (solid edge): connection type = rigid connection and estimate or obtain its stiffness coefficient from the calculation book. Locate the corresponding "Node Group Above Gate 3 in the East Zone" in the BIM model within the knowledge graph and add the following attribute: Historical Damage: [{Time: 2015, Type: Weld Fatigue Crack, Repair: Carbon Fiber Patch Reinforcement, Note: Stiffness in this area may have changed}]. Add a major environmental event attribute to the entire roof structure: [{Time: 2020-08-04, Event: Typhoon Hagupit, Peak Wind Speed: 35m / s}]. Add the following attribute to certain bolted connection nodes on the south side: Monitoring Record: [{Time: 2022-05, Status: Bolt Preload Decreased by 10%}]. Establish a boundary condition database containing design constraints for all supports. When the stadium's monitoring system is running, its constraint decoder actively calls the prior knowledge base when predicting internal forces across the entire field. For example, when calculating the area around Gate 3 in the East Zone, which had been reinforced, the updated stiffness parameters are used instead of rigidly using the design values. When analyzing vibration data anomalies on the south side, the knowledge base records the history of bolt preload reduction, making it more likely to associate the anomalies in this area with connection stiffness degradation and provide a higher confidence warning, which significantly improves the accuracy and support value of the warning.

[0029] Step S102: The time-series monitoring data is spatiotemporally aligned and associated with the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges.

[0030] In this embodiment, discretely distributed sensors are precisely spatially bound and temporally synchronized with a digital twin model representing the structural ontology in a priori knowledge base via a global coordinate system and a GNSS clock. This anchors data points to specific components, eliminating the ambiguity of data and location disconnect in traditional monitoring. Each sensor's time-series data is assigned a physical identity and spatiotemporal label, ensuring that all subsequent analyses (such as feature extraction and state prediction) are built within the geometric and topological framework of the real structure, improving the accuracy and traceability of state assessment.

[0031] In one alternative approach, spatiotemporally aligning and associating the time-series monitoring data with the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges further includes: Using the global coordinate system of the structural digital twin basic model as the spatial reference and the main GNSS clock of the sensor acquisition system as the time reference, the spatial installation coordinates of each sensor node are bound to the geometric center or feature point of the nearest component in the prior knowledge base through surveying data, and each sensor node is mapped to the entity node or entity edge in the structural digital twin basic model. If the entity nodes mapped by two sensor nodes are directly connected by an entity edge in the basic digital twin model of the structure, a physical connection edge is established for the two sensor nodes. For any two sensor nodes, calculate the damping correlation coefficient of the acceleration time series within the current time window. When the damping correlation coefficient between the two nodes exceeds a preset threshold, establish a dynamic correlation edge. The sensor nodes and the adjacency matrix composed of physical connection edges and dynamic correlation edges are encapsulated into a spatiotemporal graph data structure to obtain the initial graph structure.

[0032] In this embodiment, for example, the stadium's overall design coordinate system (with the origin at the center of the field) is used as the global coordinate system. All data acquisition boxes receive the same PPS signal via a GPS synchronization module, ensuring microsecond-level time synchronization. The coordinates of accelerometer A-12, measured by a total station, are (205.332, 178.456, 42.115). A search in the prior knowledge base reveals that the nearest structural entity to this point is an upper chord node on the main truss T-8, approximately one-third of the span from the south end support. A-12 is automatically bound to this node. Strain gauge S-05 and accelerometer A-21 are bound to the same node on the main truss T-8 and the end node of the directly connected secondary truss ST-3, respectively. Since the two structural entities are directly connected in the model, a fixed physical connection edge is established between S-05 and A-21. Analyzing acceleration data over a 10-minute window during a concert revealed a high vertical acceleration correlation coefficient of 0.88 (exceeding the threshold of 0.75) between sensor A-E01 at the eastern cantilever end of the roof and sensor A-W01 at the symmetrical western cantilever end, indicating that the roof underwent overall first-order symmetrical vertical vibration under crowd load. Therefore, a dynamic correlation edge was dynamically established between A-E01 and A-W01. Simultaneously, A-E01 showed a high correlation with another nearby point, A-E02, but a low correlation with the more distant point A-Center (the center point). The final structure diagram incorporates both a rigid skeleton edge determined by component connections and a dynamic collaborative edge revealed by this crowd load. When the ST-AGCN model runs on this graph, information can propagate from the main truss to the secondary truss via physical connection edges, and jump from the east cantilever end to the west cantilever end via dynamic correlation edges. This allows the model to more accurately reconstruct the full-field displacement and internal force distribution of the entire roof under crowd loads, even though strain gauges are not installed at the west cantilever end. In a similar activity a few months later, the correlation coefficients of A-E01 and A-W01 were found to have decreased significantly to 0.45, while the correlation coefficients of A-E01 and A-E02 were abnormally high. The changes in the graph structure edges suggested that the local dynamic characteristics of the east cantilever end may have changed (e.g., loose connections), causing it to decouple from the overall vibration mode and become more coupled with closer local components. An early warning was issued based on the graph structure anomalies. Manual inspection revealed a loss of preload in a critical bolt group at the east cantilever end, demonstrating the excellent sensitivity of this dynamic graph construction method for detecting early, localized damage.

[0033] Step S103: Input the initial graph structure into the independent GRU unit of the corresponding sensor node, output the node-level temporal feature vector through the GRU (Gated Recurrent Unit) unit and input it into the ST-AGCN unit, and output the node-level spatial correlation feature vector through the ST-AGCN unit to aggregate the spatial features of multi-hop neighboring nodes.

[0034] In this embodiment, the ST-AGCN unit, through its graph convolution (GCN) and spatiotemporal attention mechanism, can aggregate the spatial features of multi-hop neighborhood nodes. The final feature of a node is not only influenced by its directly adjacent nodes (one-hop neighborhood), but also by information propagated from more distant nodes (two-hop, three-hop, etc.) through the graph structure. In large-span spatial structures, a small displacement of a support may affect the distant mid-span area through the rod network. By aggregating multi-hop information, an implicit relationship can be established between local measurement points and distant unmeasured areas. At the same time, the model can dynamically focus on the importance of different neighborhood nodes, for example, paying more attention to the information of upwind nodes during wind-induced vibration.

[0035] In one alternative embodiment, the GRU unit comprises three parallel processing branch layers: the first branch layer is a high-frequency vibration feature extraction branch layer, comprising a gated depthwise separable convolutional network layer consisting of multiple 1x3 deep convolutional layers and a gated linear network layer; the second branch layer is a quasi-static response feature extraction branch layer, comprising two standard GRU layers for capturing slow changing trends in strain and displacement data and short- to medium-term load responses; and the third branch layer is a long-term trend and anomaly perception branch, comprising a single-layer GRU with an attention mechanism for identifying long-term sudden anomaly patterns in the data. The ST-AGCN unit includes a multi-head spatial attention sublayer, a spatiotemporal graph convolution sublayer, and a gated temporal convolution sublayer. The multi-head spatial attention sublayer employs K independent attention heads. The spatiotemporal graph convolution sublayer, after obtaining an adaptive adjacency matrix, uses Chebyshev polynomial approximation graph convolution for information propagation. The gated temporal convolution sublayer, after spatial convolution, applies dilated causal convolution to the feature sequence of each node to extract temporal patterns and uses a gated activation unit to control the information flow.

[0036] In this embodiment, early structural damage or performance degradation often manifests first as subtle changes in dynamic characteristics or anomalies in local correlation patterns, before progressing to observable displacement or strain exceeding limits. By using parallel branch layers and ST-AGCN units, the ability to capture early, weak anomalous signals is enhanced, thereby achieving earlier warnings.

[0037] Step S104: Construct a physical constraint decoder based on the node-level spatial association feature vectors of all nodes and the prior knowledge base to output global mechanical state prediction values; wherein, the global mechanical state prediction values ​​include strain, displacement and internal force in the area where no sensors are deployed; the loss function of the physical constraint decoder includes strong constraint terms of the finite element equilibrium equation, mean square error of data prediction and weak modal correlation loss terms.

[0038] In this embodiment, the strain, displacement and internal force of the entire large-span structure in the unsensitized area are predicted by the constraint decoder, which reduces the density and cost of sensor deployment. Users only need to deploy a small number of sensors in key parts to see the mechanical state cloud map of the entire structure, realizing the transition from discrete point monitoring to continuous field perception, which significantly improves the coverage and information output value of the monitoring system.

[0039] Step S105: Based on the global mechanical state prediction field of the large-span spatial structure during a preset health status monitoring period, the probability distribution of each key parameter is fitted using the kernel density estimation method; the probability distribution of each key parameter is clustered using the local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, potential damage is determined and an early warning message is triggered; wherein, the early warning message includes the parameter exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the recommended review level.

[0040] In this embodiment, the kernel density estimation (KDE) method is used to establish a probability distribution model of key parameters (such as maximum strain, displacement amplitude, and vibration frequency) at each structural location under a healthy state. This enables a probabilistic description and assessment of the structural state, not only determining whether a parameter exceeds the standard but also assessing the degree of anomalousness (measured by probability or statistical distance) deviating from the healthy baseline. This effectively reduces false alarms caused by normal environmental fluctuations (such as temperature changes) and improves the detection sensitivity for progressive damage, which is more in line with engineering practice because the healthy state itself has a normal fluctuation range. Through the Local Anomaly Factor (LOF) algorithm and spatial clustering analysis, anomalous regions that cluster in space are actively sought. This significantly reduces false alarms caused by accidental sensor malfunctions, local interference, or data noise. True structural damage (such as crack propagation or loosening of connections) typically affects a region, causing coordinated anomalous changes in state parameters at multiple locations within that region. Identifying these spatially clustered anomalous patterns indicates damage with far greater reliability than isolated anomalous points. Simultaneously, it can characterize the approximate outline and range of potential damage, providing target areas for subsequent manual verification.

[0041] In an optional approach, fitting the probability distribution of each key parameter using a kernel density estimation method, based on the predicted global mechanical state field of the large-span spatial structure during a preset health monitoring period, further includes: During a preset health status monitoring cycle, time-series monitoring data is input into the physical constraint decoder to obtain a continuous global mechanical state prediction field sequence. For each key parameter in the global mechanical state prediction field sequence, its complete time series value is extracted at each structural location. The health benchmark feature vector of the corresponding structural location is obtained based on the complete time series value. A health status benchmark database is constructed based on all structural locations and key parameters. For each key parameter in the health status benchmark database, at each structural location, a probability distribution model is established using the adaptive bandwidth kernel density estimation method. Based on the probability distribution model of each statistical characteristic at each structural location and the generated virtual sample set, the Mahalanobis distance from the sample point to its distribution center is calculated to obtain the probability distribution of the health status at each structural location.

[0042] In this embodiment, a unique health baseline is established for each key parameter at each structural location using statistical methods. This baseline reflects the normal response range of the specific structure under real environmental loads (wind, temperature, crowds, etc.), including realistic factors such as structural installation errors, minor variations in material properties, and environmental background noise. This makes subsequent anomaly judgment criteria more reasonable and lenient, effectively distinguishing between normal operating condition fluctuations and true structural performance degradation, and significantly reducing the false alarm rate. An adaptive bandwidth kernel density estimation method is used to establish a probability distribution model of the statistical characteristics of key parameters. This does not require pre-assuming that the data follows a specific distribution (such as a normal distribution), but rather lets the data speak for itself, enabling a more accurate fit to the true distribution shape of the health status data, especially for data that may have multi-peak, skewed, or heavy-tailed distributions (for example, displacement data may form two distribution peaks under windless and windy conditions). The probability distribution of the health status at each structural location (usually represented by a probability density function PDF or a cumulative distribution function CDF) can calculate the Mahalanobis distance or the probability of falling into the health distribution for any observed real-time feature vector. Operations and maintenance personnel can not only know that a certain location may be abnormal, but also know the degree of its abnormality (the size of the Mahalanobis distance) or the rarity of the occurrence of the abnormal state (probability value).

[0043] In one alternative approach, clustering analysis of the probability distributions of key parameters using a local anomaly factor algorithm to identify anomalous regions that form spatial clusters further includes: The statistical characteristic values ​​of each structural location are compared with the health benchmark characteristic vectors of the corresponding locations in the health status benchmark database to obtain the characteristic deviation vector of each structural location. The local anomaly factor algorithm is used to quantify the degree of anomaly of the feature deviation vector of each structural location; for each structural location, its spatial neighborhood is defined according to the structural topological relationship in the prior knowledge base; the spatial neighborhood forms a multi-scale neighborhood set for each location based on the physical connection neighborhood and the spatial distance neighborhood; Based on the Euclidean distance of the feature deviation vector in the high-dimensional feature space, calculate the reachable distance from each location to its spatial neighborhood; based on the reachable distance from each location to its neighborhood and the degree of anomaly, calculate the local anomaly factor value. The local anomaly factor values ​​at each structural location are mapped to structural spatial coordinates to form a structural anomaly degree distribution map. Cluster analysis is performed on the high anomaly degree locations in the structural anomaly degree distribution map to obtain a set of spatially clustered anomaly regions. The set of anomaly regions includes the boundary location, number of structural locations, average anomaly degree of the region, and spatial distribution morphology information of each anomaly region.

[0044] In this embodiment, instead of analyzing each measurement point data in isolation, a neighborhood is defined based on the structural topological relationships in the prior knowledge base. Areas with tightly connected physical connections (such as rods connected at the same node) are treated as logical units for analysis. A minor anomaly at a point may be judged as noise or local interference if its surrounding points (within the physical connection neighborhood) all behave normally; conversely, if a region around it shows coordinated anomalies, it is highly likely a signal of real damage. This significantly improves the ability to distinguish real damage from data noise / random errors and reduces the false alarm rate. The local anomaly factor algorithm identifies high-anomaly points and then uses spatial clustering analysis to form a set of anomalous regions. It focuses not on individual out-of-specification points, but on anomaly patterns that form clusters in space. This achieves the transformation from detecting point anomalies to identifying regional anomaly patterns. Real structural damage (such as crack propagation, corrosion areas, and connection group failures) usually has spatial continuity and manifests as an anomalous region. Identifying anomalous areas not only more reliably indicates the presence of damage but also allows for a preliminary assessment of the approximate extent, shape, and scale of the damage (e.g., linear damage may indicate cracks, while clumping damage may indicate localized corrosion or node failure), providing valuable clues for damage characterization. Maintenance personnel can visually see on the 3D structural model where anomalies are high, the size of the anomalous areas, and their shape. Structured regional information (e.g., "Average regional anomaly score of 2.5, containing 15 nodes, elliptical distribution") provides a quantitative basis for assessing the severity and potential impact of the damage, enabling the generation of richer early warning information (e.g., "Northeast corner support area, high confidence"). Figure 4As shown, each structural location (node ​​or unit) is color-coded with its Local Anomaly Factor (LOF) value, including normal (LOF ≈ 1), slight anomaly, and high anomaly (LOF > threshold). A continuous cluster of nodes forms a clear spatial aggregation in a certain region of the structure (near one support or the lower chord in the middle of the span); the superimposed structural coordinate grid marks the center coordinates of the anomaly region (e.g., X=24.5m, Y=12.3m, Z=8.1m); the small image in the lower right corner indicates that the sample points in this region are far from the healthy distribution center in the high-dimensional feature space (large Mahalanobis distance).

[0045] According to the solution provided by the present invention, multiple sensor networks are arranged at preset local measurement points in a large-span spatial structure to collect time-series monitoring data; based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a priori knowledge base including geometric topology, material properties, and boundary conditions is constructed; wherein, the time-series monitoring data includes static response data and dynamic response data; the preset local measurement points include supports, mid-span, member connections, and dynamic response sensitive areas; the time-series monitoring data is spatiotemporally aligned and correlated with the structural topology information in the priori knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges; the initial graph structure is input into the independent GRU units corresponding to the sensor nodes, and the node-level time-series feature vectors are output by the GRU units and input into the ST-AGCN unit, and the spatial features of multi-hop neighborhood nodes are aggregated by the ST-AGCN unit. The system outputs node-level spatial correlation feature vectors; based on the node-level spatial correlation feature vectors of all nodes and the prior knowledge base, a physical constraint decoder is constructed to output global mechanical state prediction values; wherein, the global mechanical state prediction values ​​include strain, displacement, and internal forces in areas where no sensors are deployed; the loss function of the physical constraint decoder includes strong constraint terms of the finite element equilibrium equations, mean square error of data prediction, and weak modal correlation loss terms; based on the global mechanical state prediction field of the large-span spatial structure during a preset health state monitoring period, the probability distribution of each key parameter is fitted using a kernel density estimation method; cluster analysis of the probability distribution of each key parameter is performed using a local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, potential damage is determined and an early warning message is triggered; wherein, the early warning message includes the parameters exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the suggested review level. This invention utilizes sparse local monitoring data to achieve accurate reconstruction of the global mechanical state of a large-span spatial structure and real-time damage early warning. Specifically, by using a spatiotemporal graph neural network (GRU-ST-AGCN) to simulate the transmission and diffusion processes in structural topological networks, the mechanical state of sensorless areas can be inferred from limited local observations. Prior knowledge is structured and input into the model in the form of a structural knowledge graph and a spatiotemporal graph. In the constraint decoder, the finite element equilibrium equations are added as strong constraints to the loss function, ensuring the interpretability of the output global prediction field and making it easily acceptable and verifiable by engineers. A probability distribution model of the health status baseline is established through kernel density estimation. The Local Anomaly Factor (LOF) algorithm quantifies the isolation degree of each location, and spatial clustering analysis identifies anomalous regions forming clusters in space. This approach is extremely sensitive to early damage; even if the change at each point does not exceed the traditional threshold, the pattern of collective shift can be effectively captured, while significantly suppressing false alarms caused by single-point noise.

[0046] Figure 5 A schematic diagram of the framework of a local-to-global state mapping and real-time early warning device for large-span spatial structures according to an embodiment of the present invention is shown. The local-to-global state mapping and real-time early warning device for large-span spatial structures includes: The data acquisition and prior knowledge base construction module 510 is used to deploy multiple sensor networks at preset local measuring points in a large-span spatial structure to collect time-series monitoring data; based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, it constructs a prior knowledge base including geometric topology, material properties, and boundary conditions; wherein, the time-series monitoring data includes static response data and dynamic response data; the preset local measuring points include supports, mid-span, member connections, and dynamic response sensitive areas; The initial graph structure construction module 520 is used to perform spatiotemporal alignment and association between the time-series monitoring data and the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges. The spatial correlation feature vector acquisition module 530 is used to input the initial graph structure into the independent GRU unit of the corresponding sensor node, output the node-level temporal feature vector through the GRU unit and input it into the ST-AGCN unit, and aggregate the spatial features of multi-hop neighboring nodes through the ST-AGCN unit to output the node-level spatial correlation feature vector. The global mechanical state prediction value acquisition module 540 is used to construct a physical constraint decoder based on the node-level spatial correlation feature vectors of all nodes and the prior knowledge base to output the global mechanical state prediction value; wherein, the global mechanical state prediction value includes the strain, displacement and internal force of the area where no sensors are deployed; the loss function of the physical constraint decoder includes the strong constraint term of the finite element equilibrium equation, the mean square error of data prediction and the loss term of weak modal correlation. The abnormal region identification and early warning module 550 is used to fit the probability distribution of each key parameter based on the global mechanical state prediction field of the large-span spatial structure during a preset health status monitoring period, using a kernel density estimation method; to perform cluster analysis on the probability distribution of each key parameter using a local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, it is determined that potential damage has occurred and an early warning message is triggered; wherein, the early warning message includes the parameter exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the recommended review level.

[0047] Figure 6 The diagram shows a structural schematic of an embodiment of the computing device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computing device.

[0048] like Figure 6As shown, the computing device may include: a processor 602, a communications interface 604, a memory 606, and a communications bus 608.

[0049] The processor 602, communication interface 604, and memory 606 communicate with each other via communication bus 608. Communication interface 604 is used to communicate with other network elements such as clients or other servers. Processor 602 executes program 610, specifically performing the relevant steps in the above-described embodiment of the local-to-global state mapping and real-time early warning method for large-span spatial structures.

[0050] Specifically, program 610 may include program code that includes computer operation instructions.

[0051] Processor 602 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computing device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0052] Memory 606 is used to store program 610. Memory 606 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0053] According to the solution provided by the present invention, multiple sensor networks are arranged at preset local measurement points in a large-span spatial structure to collect time-series monitoring data; based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a priori knowledge base including geometric topology, material properties, and boundary conditions is constructed; wherein, the time-series monitoring data includes static response data and dynamic response data; the preset local measurement points include supports, mid-span, member connections, and dynamic response sensitive areas; the time-series monitoring data is spatiotemporally aligned and correlated with the structural topology information in the priori knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges; the initial graph structure is input into the independent GRU units corresponding to the sensor nodes, and the node-level time-series feature vectors are output by the GRU units and input into the ST-AGCN unit, and the spatial features of multi-hop neighborhood nodes are aggregated by the ST-AGCN unit. The system outputs node-level spatial correlation feature vectors; based on the node-level spatial correlation feature vectors of all nodes and the prior knowledge base, a physical constraint decoder is constructed to output global mechanical state prediction values; wherein, the global mechanical state prediction values ​​include strain, displacement, and internal forces in areas where no sensors are deployed; the loss function of the physical constraint decoder includes strong constraint terms of the finite element equilibrium equations, mean square error of data prediction, and weak modal correlation loss terms; based on the global mechanical state prediction field of the large-span spatial structure during a preset health state monitoring period, the probability distribution of each key parameter is fitted using a kernel density estimation method; cluster analysis of the probability distribution of each key parameter is performed using a local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, potential damage is determined and an early warning message is triggered; wherein, the early warning message includes the parameters exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the suggested review level. This invention utilizes sparse local monitoring data to achieve accurate reconstruction of the global mechanical state of a large-span spatial structure and real-time damage early warning. Specifically, by using a spatiotemporal graph neural network (GRU-ST-AGCN) to simulate the transmission and diffusion processes in structural topological networks, the mechanical state of sensorless areas can be inferred from limited local observations. Prior knowledge is structured and input into the model in the form of a structural knowledge graph and a spatiotemporal graph. In the constraint decoder, the finite element equilibrium equations are added as strong constraints to the loss function, ensuring the interpretability of the output global prediction field and making it easily acceptable and verifiable by engineers. A probability distribution model of the health status baseline is established through kernel density estimation. The Local Anomaly Factor (LOF) algorithm quantifies the isolation degree of each location, and spatial clustering analysis identifies anomalous regions forming clusters in space. This approach is extremely sensitive to early damage; even if the change at each point does not exceed the traditional threshold, the pattern of collective shift can be effectively captured, while significantly suppressing false alarms caused by single-point noise.

[0054] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination of all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed can be employed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose. Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

Claims

1. A method for local-to-global state mapping and real-time early warning of large-span spatial structures, characterized in that, include: Multiple sensor networks are deployed at pre-defined local measurement points in a large-span spatial structure to collect time-series monitoring data; Based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a priori knowledge base including geometric topology, material properties, and boundary conditions is constructed; wherein, the time-series monitoring data includes static response data and dynamic response data; the preset local measurement points include supports, mid-span, member connections, and dynamic response sensitive areas; The time-series monitoring data is spatiotemporally aligned and associated with the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges. The initial graph structure is input into the independent GRU unit of the corresponding sensor node. The node-level temporal feature vector is output by the GRU unit and input into the ST-AGCN unit. The ST-AGCN unit aggregates the spatial features of multi-hop neighbor nodes and outputs the node-level spatial correlation feature vector. A physical constraint decoder is constructed based on the node-level spatial association feature vectors of all nodes and the prior knowledge base to output global mechanical state prediction values; wherein, the global mechanical state prediction values ​​include strain, displacement and internal force in the area where no sensors are deployed; the loss function of the physical constraint decoder includes strong constraint terms of the finite element equilibrium equation, mean square error of data prediction and weak modal correlation loss terms. Based on the global mechanical state prediction field of the large-span spatial structure during a preset health status monitoring period, the probability distribution of each key parameter is fitted using a kernel density estimation method; cluster analysis of the probability distribution of each key parameter is performed using a local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, potential damage is determined and an early warning message is triggered; wherein, the early warning message includes the parameter exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the recommended review level.

2. The method for local-global state mapping and real-time early warning of large-span spatial structures according to claim 1, characterized in that, Further time-series monitoring data is collected by arranging multiple sensor networks at pre-defined local measuring points in a large-span spatial structure, including: Based on the structural finite element model in the prior knowledge base, multi-round modal analysis and static sensitivity analysis were performed on the large-span spatial structure to identify the key mechanical response regions and high strain energy density regions under different load conditions. Based on the critical mechanical response region and the high strain energy density region, the preset local measuring points are divided into two categories: critical measuring points and auxiliary measuring points. The critical measuring points are deployed at the end connection regions of the main load-bearing components, the mid-span sections of large-span components, the abrupt change regions of support reaction forces, and the mode peak points of the first N main modes identified by finite element analysis. The auxiliary measuring points are deployed in the secondary sensitive regions of the structure. Fiber optic strain sensors, total station prism groups, and triaxial accelerometers are deployed at the key measuring points; low-power wireless strain sensors and accelerometers are deployed at the auxiliary measuring points.

3. The method for local-global state mapping and real-time early warning of large-span spatial structures according to claim 1, characterized in that, Based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a priori knowledge base is constructed, including geometric topology, material properties, and boundary conditions. This further includes: The component type, spatial coordinates, cross-sectional dimensions, material grade and connection relationship are extracted from the design drawings using BIM analysis software to generate a preliminary geometric topology diagram. The discretized node coordinates, element connection relationships, material constitutive model parameters and boundary condition constraint information are extracted from the finite element model file. Based on the preliminary geometric topology diagram and the extraction results of the finite element model file, a structural digital foundation model is constructed. The structural digital foundation model uses physical components as entity nodes and physical connections between physical components as entity edges. The node attributes of the entity nodes include spatial coordinates, material properties, and cross-sectional properties, and the edge attributes of the entity edges include connection type and stiffness coefficient. Based on the historical inspection records, historical damage locations, damage types, repair measures, environmental load events, and performance degradation assessments are extracted. The extraction results of the historical inspection records are then appended to the corresponding entity nodes or entity edges in the structural digital generation basic model to form a structural knowledge graph with time-varying state dimensions. The prior knowledge base is obtained by encoding the prior physical rules of the structure and the dynamic database of boundary conditions based on the structural knowledge graph.

4. The method for local-global state mapping and real-time early warning of large-span spatial structures according to claim 1, characterized in that, The process of spatiotemporally aligning and associating the time-series monitoring data with the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges further includes: Using the global coordinate system of the structural digital twin basic model as the spatial reference and the main GNSS clock of the sensor acquisition system as the time reference, the spatial installation coordinates of each sensor node are bound to the geometric center or feature point of the nearest component in the prior knowledge base through surveying data, and each sensor node is mapped to the entity node or entity edge in the structural digital twin basic model. If the entity nodes mapped by two sensor nodes are directly connected by an entity edge in the basic digital twin model of the structure, a physical connection edge is established for the two sensor nodes. For any two sensor nodes, calculate the damping correlation coefficient of the acceleration time series within the current time window. When the damping correlation coefficient between the two nodes exceeds a preset threshold, establish a dynamic correlation edge. The sensor nodes and the adjacency matrix composed of physical connection edges and dynamic correlation edges are encapsulated into a spatiotemporal graph data structure to obtain the initial graph structure.

5. The method for local-global state mapping and real-time early warning of large-span spatial structures according to claim 1, characterized in that: The GRU unit comprises three parallel processing branches. The first branch is a high-frequency vibration feature extraction branch, consisting of a gated depthwise separable convolutional network layer composed of multiple 1x3 deep convolutional layers and a gated linear network layer. The second branch is a quasi-static response feature extraction branch, consisting of two standard GRU layers, used to capture the slow changing trends of strain and displacement data and short-to-medium-term load responses. The third branch is a long-term trend and anomaly perception branch, consisting of a single-layer GRU with an attention mechanism, used to identify long-term sudden anomaly patterns in the data. The ST-AGCN unit includes a multi-head spatial attention sublayer, a spatiotemporal graph convolution sublayer, and a gated temporal convolution sublayer. The multi-head spatial attention sublayer employs K independent attention heads. The spatiotemporal graph convolution sublayer, after obtaining an adaptive adjacency matrix, uses Chebyshev polynomial approximation graph convolution for information propagation. The gated temporal convolution sublayer, after spatial convolution, applies dilated causal convolution to the feature sequence of each node to extract temporal patterns and uses a gated activation unit to control the information flow.

6. The method for local-global state mapping and real-time early warning of large-span spatial structures according to claim 1, characterized in that, Based on the predicted global mechanical state field of the large-span spatial structure during a preset health monitoring period, the probability distribution of each key parameter is further fitted using a kernel density estimation method, including: During a preset health status monitoring cycle, time-series monitoring data is input into the physical constraint decoder to obtain a continuous global mechanical state prediction field sequence. For each key parameter in the global mechanical state prediction field sequence, its complete time series value is extracted at each structural location. The health benchmark feature vector of the corresponding structural location is obtained based on the complete time series value. A health status benchmark database is constructed based on all structural locations and key parameters. For each key parameter in the health status benchmark database, at each structural location, a probability distribution model is established using the adaptive bandwidth kernel density estimation method. Based on the probability distribution model of each statistical characteristic at each structural location and the generated virtual sample set, the Mahalanobis distance from the sample point to its distribution center is calculated to obtain the probability distribution of the health status at each structural location.

7. The method for local-global state mapping and real-time early warning of large-span spatial structures according to claim 1, characterized in that, Cluster analysis of the probability distribution of each key parameter using the local anomaly factor algorithm further identifies anomalous regions that form spatial clusters, including: The statistical characteristic values ​​of each structural location are compared with the health benchmark characteristic vectors of the corresponding locations in the health status benchmark database to obtain the characteristic deviation vector of each structural location. The local anomaly factor algorithm is used to quantify the degree of anomaly of the feature deviation vector of each structural location; for each structural location, its spatial neighborhood is defined according to the structural topological relationship in the prior knowledge base; the spatial neighborhood forms a multi-scale neighborhood set for each location based on the physical connection neighborhood and the spatial distance neighborhood; Based on the Euclidean distance of the feature deviation vector in the high-dimensional feature space, calculate the reachable distance from each location to its spatial neighborhood; based on the reachable distance from each location to its neighborhood and the degree of anomaly, calculate the local anomaly factor value. The local anomaly factor values ​​at each structural location are mapped to structural spatial coordinates to form a structural anomaly degree distribution map. Cluster analysis is performed on the high anomaly degree locations in the structural anomaly degree distribution map to obtain a set of spatially clustered anomaly regions. The set of anomaly regions includes the boundary location, number of structural locations, average anomaly degree of the region, and spatial distribution morphology information of each anomaly region.

8. The method for local-global state mapping and real-time early warning of large-span spatial structures according to claim 2, characterized in that, The key measuring points are equipped with fiber optic strain sensors, total station prism groups, and triaxial accelerometers; the auxiliary measuring points are equipped with low-power wireless strain sensors and accelerometers, further including: For each critical measuring point, a spatial fiber optic grating sensor array consisting of at least six sensors with a sensor spacing of less than or equal to 150 cm is installed in the end connection area of ​​its main load-bearing component; at least five fiber optic grating sensors are arranged at equal intervals of 200 cm along the height direction of the component at its mid-span section to monitor the cross-sectional strain gradient distribution and neutral axis offset; a quadrangular prism adaptive array consisting of four reflecting prisms is installed in the abrupt change area of ​​the support reaction force to obtain the three-dimensional displacement of the measuring point; at least a dual-channel broadband accelerometer group is deployed at each critical measuring point. For auxiliary measurement points, a low-power triaxial wireless accelerometer network is deployed. The accelerometer network uses a low frequency of 5Hz during the structural calm period and automatically increases to 50Hz sampling during periods of wind vibration or active pedestrian load.

9. A local-to-global state mapping and real-time early warning device for large-span spatial structures, characterized in that, include: The data acquisition and prior knowledge base construction module is used to deploy multiple sensor networks at preset local measuring points in a large-span spatial structure to collect time-series monitoring data; Based on the design drawings, finite element model files, and historical inspection records of the large-span spatial structure, a priori knowledge base including geometric topology, material properties, and boundary conditions is constructed; wherein, the time-series monitoring data includes static response data and dynamic response data; the preset local measurement points include supports, mid-span, member connections, and dynamic response sensitive areas; The initial graph structure construction module is used to perform spatiotemporal alignment and association between the time-series monitoring data and the structural topology information in the prior knowledge base to form an initial graph structure with sensors as nodes and physical connections and dynamic correlations as composite edges. The spatial correlation feature vector acquisition module is used to input the initial graph structure into the independent GRU unit of the corresponding sensor node, output the node-level temporal feature vector through the GRU unit and input it into the ST-AGCN unit, and aggregate the spatial features of multi-hop neighboring nodes through the ST-AGCN unit to output the node-level spatial correlation feature vector. The global mechanical state prediction value acquisition module is used to construct a physical constraint decoder based on the node-level spatial correlation feature vectors of all nodes and the prior knowledge base to output the global mechanical state prediction value; wherein, the global mechanical state prediction value includes the strain, displacement and internal force of the area where no sensors are deployed; the loss function of the physical constraint decoder includes the strong constraint term of the finite element equilibrium equation, the mean square error of data prediction and the weak modal correlation loss term. The abnormal region identification and early warning module is used to fit the probability distribution of each key parameter based on the global mechanical state prediction field of the large-span spatial structure during a preset health status monitoring period, using a kernel density estimation method; to perform cluster analysis on the probability distribution of each key parameter using a local anomaly factor algorithm to identify abnormal regions that form clusters in space; when the size of the abnormal region exceeds a preset threshold, it is determined that potential damage has occurred and an early warning message is triggered; wherein, the early warning message includes the parameter exceeding the standard, the coordinates of the abnormal spatial location, the confidence level of the abnormal region, and the recommended review level.

10. A computing device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described large-span spatial structure local-global state mapping and real-time early warning method.

Images