A method and system for analyzing monitoring data of a large-span structure based on an internet of things

Abstract

This invention relates to the field of IoT monitoring technology and provides a method and system for analyzing monitoring data of large-span structures based on IoT. The method includes the following steps: acquiring the multimodal excitation-response spatiotemporal tensor of the target to be monitored; extracting topological feature barcodes by performing persistent cohomology calculations on the multimodal excitation-response spatiotemporal tensor, followed by nonlinear dimensionality reduction to obtain a topological damage fingerprint reflecting the stiffness distribution and damage evolution of the target; generating a global stiffness degradation field and damage probability map based on the topological damage fingerprint; and predicting the health status of the target within a preset time window based on the global stiffness degradation field and damage probability map. This invention, by integrating topological data analysis, constructing graph neural networks, and time-series prediction, can identify structural stiffness degradation early and accurately locate damage areas, offering advantages such as timely early warning and strong robustness.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) monitoring technology, and in particular relates to a method and system for analyzing monitoring data of large-span structures based on IoT. Background Technology

[0002] Long-span structures (such as cable-stayed bridges and suspension bridges) are prone to progressive damage in critical components during construction and long-term service due to environmental loads, material aging, and fatigue accumulation. To ensure structural safety, IoT-based structural health monitoring technology has been widely applied. Existing technologies typically use fixed sensor arrays (such as accelerometers, strain gauges, and cable force gauges) to collect structural response data, combined with drone inspections to obtain surface images, and then perform structural condition assessment through data fusion and feature extraction.

[0003] In recent years, topological data analysis methods such as continuous coherence have begun to be introduced into the field of structural damage identification (e.g., CN119942344A), and UAV formation path planning technology (e.g., CN121053570A) has also provided support for large-scale inspection.

[0004] However, the existing technologies still have the following shortcomings that need to be addressed: First, the application of continuous cohomology in the existing technologies mainly focuses on the extraction of topological features from static geometric point clouds, without deep integration with structural response data under dynamic excitation (such as equivalent dynamic stiffness and damping ratio), making it difficult to reflect the evolution trend of structural stiffness; Second, in the existing technologies, UAV inspections are mostly passive observations, lacking active excitation methods to stimulate the local dynamic response of the structure, resulting in insufficient sensitivity to the perception of early stiffness degradation. Summary of the Invention

[0005] The purpose of this invention is to provide a method for analyzing monitoring data of large-span structures based on the Internet of Things, in order to solve the above-mentioned technical problems.

[0006] This invention is implemented as follows: a data analysis method for monitoring large-span structures based on the Internet of Things, comprising the following steps: acquiring the multimodal excitation-response spatiotemporal tensor of the target to be monitored; the multimodal excitation-response spatiotemporal tensor includes the apparent image data of the target to be monitored, the dynamic response data of the structure, and the equivalent dynamic stiffness and damping ratio parameters of its local regions.

[0007] Persistent cohomology calculations are performed on the multimodal excitation-response spatiotemporal tensor to extract topological feature barcodes, followed by nonlinear dimensionality reduction to obtain a topological damage fingerprint that reflects the stiffness distribution and damage evolution of the target to be monitored.

[0008] Based on a preset graph neural network, a global stiffness degradation field and a damage probability map are generated according to the topological damage fingerprint.

[0009] Based on the global stiffness degradation field and damage probability map, the health status of the target to be monitored within a future preset time window is predicted.

[0010] Furthermore, the step of obtaining the multimodal excitation-response spatiotemporal tensor of the target to be monitored specifically includes: acquiring the apparent image data of the target to be monitored collected by the UAV formation, and performing alignment and tensor processing according to a unified spatiotemporal reference to obtain the processed apparent image data.

[0011] Acquire multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by UAV formation; the multimodal excitation-response data includes structural dynamic response data synchronously acquired by a fixed sensor array and equivalent dynamic stiffness and damping ratio parameters of local areas of the target under monitoring acquired synchronously.

[0012] The multimodal excitation-response data is fused with the processed appearance image data to generate a multimodal excitation-response spatiotemporal tensor.

[0013] Furthermore, the steps for acquiring multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by the UAV formation specifically include: planning the cooperative excitation path of each UAV in the UAV formation, so that different UAVs hover at different parts of the target under monitoring, and the hovering times are staggered; each UAV changes its rotor speed at its hovering point according to a preset pseudo-random sequence to generate broadband incoherent downwash airflow as an excitation signal, and at the same time records its own three-axis acceleration and angular velocity response through the inertial measurement unit on the UAV, and synchronously records the acceleration, strain and force response of the target under monitoring under the same excitation through a fixed sensor array; according to the calibration model of UAV rotor speed and downwash airflow velocity distribution, the downwash airflow excitation force spectrum of each hovering point is calculated, and then the frequency response function is calculated with the response signals of each sensor in the fixed sensor array, and the equivalent dynamic stiffness and damping ratio of the local area of ​​the target under monitoring is calculated by inversely using the coherent transfer function; the aerodynamic admittance distribution cloud map of the equivalent dynamic stiffness and damping ratio of all hovering points is obtained by spatial interpolation, which serves as multimodal excitation-response data.

[0014] Furthermore, the steps of performing persistent cohomology calculations on the multimodal excitation-response spatiotemporal tensor, extracting topological feature barcodes, and then performing nonlinear dimensionality reduction to obtain a topological damage fingerprint reflecting the stiffness distribution and damage evolution of the target under monitoring specifically include: treating the multimodal excitation-response spatiotemporal tensor as a scalar field function defined on the spatial grid of the target under monitoring; for each level set of the scalar field function, calculating its corresponding simplex, and tracking the generation and disappearance thresholds of topological features, including connected components, one-dimensional voids, and two-dimensional voids, that appear as the level set changes.

[0015] The generation threshold and the extinction threshold of each topological feature are combined to form an interval, and the set of all intervals constitutes the topological feature barcode.

[0016] After vectorizing the topological feature barcode, nonlinear dimensionality reduction is performed through a nonlinear manifold learning model based on Laplacian feature mapping to obtain the topological damage fingerprint.

[0017] Furthermore, the graph neural network uses the finite element nodes of the target to be monitored as graph nodes and the coherence between the structural dynamic response data at each graph node as edge weights to learn the stiffness coupling relationship between each graph node; based on the preset graph neural network, the step of generating a global stiffness degradation field and damage probability map according to the topological damage fingerprint specifically includes: establishing a finite element model of the target to be monitored, taking each finite element node as a graph node, and calculating the Pearson correlation coefficient between graph nodes based on the historical correlation of the structural dynamic response data at each graph node, as the initial edge weight of the graph neural network.

[0018] A graph neural network is constructed; the input of the graph neural network is the topological damage fingerprint of each graph node, its hidden layer uses graph convolution operation to aggregate the features of neighboring graph nodes, and its output is the stiffness degradation coefficient and damage probability of each graph node.

[0019] The graph neural network is trained using historical data under healthy conditions. The trained graph neural network outputs the stiffness degradation coefficient distribution map and damage probability distribution map of all finite element nodes of the entire target under monitoring, which are respectively used as the global stiffness degradation field and damage probability map.

[0020] Furthermore, the step of predicting the health status of the target under monitoring within a future preset time window based on the global stiffness degradation field and damage probability map specifically includes: First, flattening the global stiffness degradation field and damage probability map of each time step into a time-series state vector to form time-series data; then, constructing a long short-term memory network, taking the time-series data of the past M time steps as input, and outputting a prediction vector sequence for the next T time steps. This prediction vector sequence corresponds to the prediction result of the global stiffness degradation field and damage probability map within the future preset time window, and is used as the prediction result of the health status of the target under monitoring.

[0021] Another objective of this invention is to provide an IoT-based monitoring data analysis system for large-span structures, used to implement the aforementioned IoT-based monitoring data analysis method for large-span structures. Specifically, it includes a multimodal data acquisition module for acquiring the multimodal excitation-response spatiotemporal tensor of the target to be monitored; the multimodal excitation-response spatiotemporal tensor includes the apparent image data of the target to be monitored, the structural dynamic response data, and the equivalent dynamic stiffness and damping ratio parameters of its local regions.

[0022] The damage fingerprint extraction module is used to perform persistent cohomology calculation on the multimodal excitation-response spatiotemporal tensor, extract topological feature barcodes, and then perform nonlinear dimensionality reduction to obtain a topological damage fingerprint that reflects the stiffness distribution and damage evolution of the target to be monitored.

[0023] The stiffness field reconstruction module is used to generate a global stiffness degradation field and a damage probability map based on a preset graph neural network and the topological damage fingerprint.

[0024] The health status prediction module is used to predict the health status of the target to be monitored within a future preset time window based on the global stiffness degradation field and damage probability map.

[0025] Furthermore, the multimodal data acquisition module specifically includes: an image data acquisition unit, used to acquire the apparent image data of the target to be monitored collected by the UAV formation, and perform alignment and tensor quantization processing according to a unified spatiotemporal reference to obtain the processed apparent image data.

[0026] The physical data acquisition unit is used to acquire multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by UAV formation; the multimodal excitation-response data includes structural dynamic response data synchronously acquired by a fixed sensor array and equivalent dynamic stiffness and damping ratio parameters of local areas of the target under monitoring acquired synchronously.

[0027] The data fusion unit is used to fuse the multimodal excitation-response data with the processed appearance image data to generate a multimodal excitation-response spatiotemporal tensor.

[0028] Furthermore, the damage fingerprint extraction module specifically includes: a scalar field definition unit, used to regard the multimodal excitation-response spatiotemporal tensor as a scalar field function defined on the spatial grid of the target to be monitored; for each level set of the scalar field function, calculate its corresponding simplex, and track the generation threshold and disappearance threshold of topological features including connected components, one-dimensional holes, and two-dimensional holes that appear as the level set changes.

[0029] The topological feature generation unit is used to form an interval by combining the generation threshold and the extinction threshold of each topological feature, and the set of all intervals constitutes the topological feature barcode.

[0030] The topological feature dimensionality reduction unit is used to vectorize the topological feature barcode and then perform nonlinear dimensionality reduction through a nonlinear manifold learning model based on Laplacian feature mapping to obtain the topological damage fingerprint.

[0031] Furthermore, the stiffness field reconstruction module specifically includes: a finite element model establishment unit, used to establish a finite element model of the target to be monitored, taking each finite element node as a graph node, and calculating the Pearson correlation coefficient between graph nodes based on the historical correlation of the structural dynamic response data at each graph node, which serves as the initial edge weight of the graph neural network.

[0032] The graph neural network construction unit is used to construct a graph neural network. The input of the graph neural network is the topological damage fingerprint of each graph node. Its hidden layer uses graph convolution operation to aggregate the features of neighboring graph nodes. Its output is the stiffness degradation coefficient and damage probability of each graph node.

[0033] The network training and output unit is used to train the graph neural network using historical data under healthy conditions. The trained graph neural network outputs the stiffness degradation coefficient distribution map and damage probability distribution map of all finite element nodes of the entire target under monitoring, which serve as the global stiffness degradation field and damage probability map, respectively.

[0034] This invention provides a data analysis method for monitoring large-span structures based on the Internet of Things (IoT). By integrating topological data analysis, constructing graph neural networks, and time-series prediction, it can identify structural stiffness degradation at an early stage and accurately locate damage areas, offering advantages such as timely early warning and strong robustness. This invention can achieve ultra-sensitive identification of early-stage minor stiffness degradation in structures, robust damage feature extraction under strong noise interference, and high-precision inference of the stiffness field in areas without sensor deployment, significantly improving the sensitivity, noise resistance, positioning accuracy, and prediction reliability of structural health status monitoring. Attached Figure Description

[0035] Figure 1 A flowchart illustrating the IoT-based data analysis method for monitoring large-span structures provided in this embodiment of the invention.

[0036] Figure 2 This is a flowchart illustrating step S100 in the IoT-based large-span structure monitoring data analysis method provided in an embodiment of the present invention.

[0037] Figure 3 This is a flowchart illustrating step S200 in the IoT-based large-span structure monitoring data analysis method provided in this embodiment of the invention.

[0038] Figure 4 This is a flowchart illustrating step S300 in the IoT-based large-span structure monitoring data analysis method provided in an embodiment of the present invention.

[0039] Figure 5 This is a schematic diagram of the structure of the IoT-based large-span structure monitoring data analysis system provided in an embodiment of the present invention.

[0040] Figure 6 This is a schematic diagram of the structure of the multimodal data acquisition module provided in an embodiment of the present invention.

[0041] Figure 7 This is a schematic diagram of the damage fingerprint extraction module provided in an embodiment of the present invention.

[0042] Figure 8 This is a schematic diagram of the stiffness field reconstruction module provided in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0044] like Figure 1 As shown, in one embodiment of the present invention, a method for analyzing monitoring data of large-span structures based on the Internet of Things is provided, including the following steps: S100, obtaining the multimodal excitation-response spatiotemporal tensor of the target to be monitored; the multimodal excitation-response spatiotemporal tensor includes the apparent image data of the target to be monitored, the dynamic response data of the structure, and the equivalent dynamic stiffness and damping ratio parameters of its local region.

[0045] S200. Perform persistent cohomology calculation on the multimodal excitation-response spatiotemporal tensor, extract topological feature barcodes, and then perform nonlinear dimensionality reduction to obtain a topological damage fingerprint that reflects the stiffness distribution and damage evolution of the target to be monitored.

[0046] S300. Based on a preset graph neural network, a global stiffness degradation field and damage probability map are generated according to the topological damage fingerprint.

[0047] S400. Based on the global stiffness degradation field and damage probability map, predict the health status of the target to be monitored within a future preset time window.

[0048] In practical applications, the above method can be applied to the monitoring data analysis of long-span bridge structures, that is, the target to be monitored refers to the structure to be monitored in a long-span bridge, but it is not limited to this and is also applicable to monitoring scenarios of other structures.

[0049] like Figure 2 As shown, in a preferred embodiment of the present invention, the step of obtaining the multimodal excitation-response spatiotemporal tensor of the target to be monitored, namely step S100, specifically includes: S110, obtaining the apparent image data of the target to be monitored collected by the UAV formation, and performing alignment and tensor processing according to a unified spatiotemporal reference to obtain the processed apparent image data.

[0050] In practical applications, taking a long-span cable-stayed bridge as an example, the monitoring target is firstly a large-span cable-stayed bridge. First, surface image data is acquired from a drone swarm. The swarm can consist of three (or more) hexacopter drones, each equipped with a high-resolution industrial camera. Under calm or low-wind conditions, the drones cruise along a preset route, capturing high-definition images covering the entire bridge, including the underside of the main girder, the surface of the towers, the cable groups, and key connection nodes. This information serves as the surface image data. After acquiring the images, using the drone's built-in positioning system as the time reference and the bridge's design coordinate system as the spatial reference, the timestamp of the capture time and the drone's RTK positioning coordinates are extracted from each frame of the image. The image pixel coordinates are then mapped onto a three-dimensional spatial grid. Next, all images are tensorized according to the spatial grid: the RGB channel values ​​of all images within each spatial grid cell are fused using median fusion to form a third-order tensor, which is the processed surface image data.

[0051] S120. Acquire multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by UAV formation; the multimodal excitation-response data includes structural dynamic response data synchronously acquired by a fixed sensor array and equivalent dynamic stiffness and damping ratio parameters of local areas of the target under monitoring acquired synchronously.

[0052] It should be noted that broadband incoherent airflow excitation refers to the generation of downwash airflows that are temporally uncorrelated and have a wide spectral coverage (e.g., from around 10 Hz to hundreds of Hz) by multiple UAVs independently changing their rotor speeds according to mutually orthogonal pseudo-random sequences. This excitation can simultaneously excite multiple vibration modes of the structure, facilitating subsequent analysis of the structure's dynamic characteristics. Additionally, the fixed sensor array includes accelerometers, strain gauges, and force gauges pre-positioned in the main beam, pylon, and cable anchorage areas. The UAV formation hovers at preset key points (e.g., the middle section of the cable, the outer side of the middle column of the pylon, and the bottom surface of the main beam mid-span), generating excitation by modulating the rotor speeds according to a pseudo-random sequence. The acceleration and angular velocity of the UAVs' own IMUs, as well as the acceleration, strain, and cable force response of the fixed sensor array, are recorded simultaneously. Subsequently, the equivalent dynamic stiffness and damping ratio of the local area at each hovering point are calculated using the frequency response function and coherent transfer function. Equivalent dynamic stiffness reflects the structure's ability to resist deformation under dynamic excitation, while damping ratio reflects the structure's ability to dissipate vibrational energy. Both are extremely sensitive to early stiffness degradation (such as slight loss of cable force or microcracks in the anchorage zone). These parameters, together with the structural dynamic response data synchronously acquired by a fixed sensor array, constitute multimodal excitation-response data.

[0053] S130. The multimodal excitation-response data is fused with the processed appearance image data to generate a multimodal excitation-response spatiotemporal tensor.

[0054] Specifically, using a three-dimensional spatial grid as an index, at each grid point, the RGB values ​​of the processed apparent image data, the equivalent dynamic stiffness value, the damping ratio value corresponding to that grid point, and the acceleration amplitude, strain amplitude, cable force amplitude measured by the fixed sensor are stacked into a multi-dimensional feature vector. All the multi-dimensional feature vectors of the grid points are arranged in spatial order to form a multi-dimensional tensor containing spatial dimension, feature channel dimension, and time dimension, which is the multimodal excitation-response spatiotemporal tensor.

[0055] In a preferred embodiment of the present invention, the step of acquiring multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by a UAV formation specifically includes: planning the cooperative excitation path of each UAV in the UAV formation, so that different UAVs hover at different parts of the target under monitoring, and the hovering times are staggered; each UAV changes its rotor speed at its hovering point according to a preset pseudo-random sequence to generate broadband incoherent downwash airflow as an excitation signal, and at the same time records its own three-axis acceleration and angular velocity response through the inertial measurement unit on the UAV, and synchronously records the acceleration, strain and force response of the target under monitoring under the same excitation through a fixed sensor array; according to the calibration model of UAV rotor speed and downwash airflow velocity distribution, the downwash airflow excitation force spectrum of each hovering point is calculated, and then the frequency response function is calculated with the response signals of each sensor in the fixed sensor array, and the equivalent dynamic stiffness and damping ratio of the local area of ​​the target under monitoring is calculated by inversely using the coherent transfer function; the aerodynamic admittance distribution cloud map of the equivalent dynamic stiffness and damping ratio of all hovering points is obtained by spatial interpolation, which serves as multimodal excitation-response data.

[0056] In practical applications, based on the results of bridge finite element modal analysis, the locations most sensitive to low-order vibration modes are selected: for example, one drone hovers on the outer side of the middle tower column of the south tower to excite the tower's lateral bending mode; another hovers on the lower edge of the main beam mid-span to excite the main beam's first-order vertical bending mode; and a third hovers on the side of the middle section of the longest cable-stayed cable to excite the cable's local vibration mode. The hovering times of the three drones are staggered to avoid mutual airflow interference caused by simultaneous excitation. Simultaneously, the distance between the hovering centers of two adjacent drones is much greater than 10 times the rotor diameter, meeting the obstacle avoidance requirements.

[0057] Each UAV changes its rotor speed at its hovering point according to a preset pseudo-random sequence. The pseudo-random sequence, such as an m-sequence, has good autocorrelation properties. By fluctuating the rotor speed around the base speed according to this sequence, the resulting downwash velocity changes accordingly. The following calibration relationship exists between the UAV rotor speed and the downwash velocity distribution (i.e., the downwash excitation force spectrum): In the formula, The velocity of the downwash airflow. The rotor speed, Hovering height Radial distance, and This is a constant calibrated through ground experiments; the calibration formula reflects that the airflow velocity is greater at higher rotational speeds and lower hovering distances, and decays exponentially with radial distance. Different UAVs use mutually orthogonal pseudo-random sequences, making their excitation signals uncorrelated.

[0058] The inertial measurement unit (IMU) onboard the UAV records its triaxial acceleration and angular velocity responses at a high sampling rate. Simultaneously, a fixed sensor array (accelerometers, strain gauges, and cable force gauges) deployed on the bridge synchronously records the structure's response under this excitation at its respective sampling frequency. For each hovering point, the calculated downwash airflow excitation force spectrum and sensor response signals are used to calculate the frequency response function; specifically, an H1 estimator is employed. : In the formula, For the cross-power spectrum of the excitation and response signals, The self-power spectrum of the excitation signal, ω is the angular frequency.

[0059] Then coherent transfer function for: In the formula, The power spectrum of the response signal; the closer the value of the coherent transfer function is to 1, the more likely the response is caused by that excitation, and the higher the data quality. Typically, only [the following is retained]: The reliable frequency band; within the reliable frequency band, the equivalent dynamic stiffness Damping ratio Obtained through the half-power bandwidth method: In the formula, The natural angular frequency refers to the frequency at which a structure vibrates freely in an undamped state. In practice, it is usually determined by the peak value of the frequency response function. and The half-power point angular frequency refers to the angular frequency at two frequency points near the peak of the frequency response function when the power (square of the amplitude) drops to half of the peak power.

[0060] The equivalent dynamic stiffness and damping ratio calculated from all hovering points (e.g., setting multiple hovering points at each critical location to increase coverage) are used to obtain an aerodynamic admittance distribution cloud map across the entire bridge using existing radial basis function space interpolation. Radial basis function interpolation can smoothly connect discrete measurement points to form a continuous distribution map. This aerodynamic admittance distribution cloud map is a component of the multimodal excitation-response data.

[0061] In this embodiment of the invention, airflow interference from multiple UAVs is avoided by time staggering and spatial distance control; pseudo-random sequence modulation gives the downwash airflow wide bandwidth and incoherent characteristics, which can simultaneously excite multiple modes and facilitate the screening of reliable data through coherent transfer functions; the calculated equivalent dynamic stiffness and damping ratio are extremely sensitive to early small stiffness degradation.

[0062] like Figure 3 As shown, in a preferred embodiment of the present invention, the steps of performing persistent cohomology calculation on the multimodal excitation-response spatiotemporal tensor, extracting topological feature barcodes, and then performing nonlinear dimensionality reduction to obtain a topological damage fingerprint reflecting the stiffness distribution and damage evolution of the target to be monitored, namely step S200, specifically include: S210, regarding the multimodal excitation-response spatiotemporal tensor as a scalar field function defined on the spatial grid of the target to be monitored; for each level set of the scalar field function, calculating its corresponding simplex, and tracking the generation threshold and disappearance threshold of topological features including connected components, one-dimensional holes, and two-dimensional holes that appear with the change of the level set.

[0063] Specifically, persistent homology calculations are performed on the scalar field function. Persistent homology is a computational topology method used to analyze the generation and disappearance of topological features (connected components, loops, cavities) as a threshold parameter changes. The specific approach involves gradually increasing the threshold from its minimum to its maximum value. For each threshold... Take all scalar field functions that satisfy less than or equal to The points form a level set, and these points are used as vertices to construct a simplicial complex (i.e., a topological structure composed of points, line segments, and triangles). As the threshold increases, the level set expands, and new topological features continuously appear in the simplicial complex. The threshold at which each topological feature first appears is recorded as the generation threshold. And the threshold at which it finally disappears, i.e., the extinction threshold. For example, connected components: when the threshold is low enough, all points are disconnected, and each point is a connected component; as the threshold increases, adjacent points connect to form larger connected regions, and when two regions meet, one of the regions "disappears." One-dimensional voids (loops) are like closed paths around a damaged area in a bridge structure; two-dimensional voids (cavities) are like volumetric regions inside a closed curved surface. The generation and disappearance thresholds of these features collectively describe the topological structure of the scalar field function at different scales.

[0064] S220. The generation threshold and the extinction threshold of each topological feature are combined into an interval, and the set of all intervals constitutes the topological feature barcode.

[0065] Specifically, the generation threshold and the disappearance threshold of each topological feature are combined to form an interval. The set of all intervals is called a topological feature barcode. A topological feature barcode can be graphically represented as a line segment with the threshold on the horizontal axis and each feature representing a segment from birth to death. In a healthy state, a topological feature barcode typically contains several long-lived intervals (corresponding to large-scale structural regions); however, when local damage occurs, the damaged area will generate new short-lived intervals (reflecting abrupt changes in local stiffness).

[0066] S230. After vectorizing the topological feature barcode, nonlinear dimensionality reduction is performed through a nonlinear manifold learning model based on Laplacian feature mapping to obtain the topological damage fingerprint.

[0067] Since the number of intervals in a topological feature barcode is not fixed, it cannot be directly used as input to the model; therefore, it needs to be vectorized. This embodiment uses the persistence histogram method: calculate the length of each interval (i.e., persistence value), normalize all persistence values, divide them into multiple equally wide histograms, and count the number of intervals falling within each histogram to obtain a high-dimensional vector of fixed length, thus completing the vectorization. Then, input this high-dimensional vector into a nonlinear manifold learning model based on Laplacian eigenmaps for dimensionality reduction. The goal of this nonlinear manifold learning model is to preserve the local neighborhood relationships between high-dimensional data points while reducing dimensionality, i.e., to solve the following generalized eigenvalue problem: In the formula, Let W be the graph Laplacian matrix, W be the heat kernel weight matrix, D be the degree matrix, and V be the eigenvectors to be solved. The eigenvectors corresponding to the smallest d+1 non-zero eigenvalues ​​are taken to obtain the low-dimensional embedding coordinates, where d represents the dimension of the target after dimensionality reduction. In this embodiment, the high-dimensional vector is reduced to a low-dimensional embedding space of 2 to 5 dimensions, and these low-dimensional embedding coordinates are the topological damage fingerprint.

[0068] like Figure 4 As shown, in a preferred embodiment of the present invention, the graph neural network uses the finite element nodes of the target to be monitored as graph nodes and the coherence between the structural dynamic response data at each graph node as edge weights to learn the stiffness coupling relationship between each graph node; based on the preset graph neural network, the step of generating a global stiffness degradation field and damage probability map according to the topological damage fingerprint, i.e., step S300 specifically includes: S310, establishing a finite element model of the target to be monitored, taking each finite element node as a graph node, and calculating the Pearson correlation coefficient between graph nodes based on the historical correlation of the structural dynamic response data at each graph node, as the initial edge weight of the graph neural network.

[0069] In practical applications, taking a cable-stayed bridge as an example, a full bridge model is built using existing finite element software (such as ANSYS, ABAQUS, etc.). The main beam, towers, cables, piers, and other components are discretized into finite element elements, and the node corresponding to each element is a graph node of a graph neural network. This embodiment yields thousands of finite element nodes.

[0070] Based on the historical correlation of structural dynamic response data (such as acceleration, strain, and cable force time series over the past 30 days) at each graph node, the Pearson correlation coefficient between each pair of graph nodes is calculated. This coefficient measures the degree of linear correlation between two time series, ranging from -1 to 1, with a larger absolute value indicating a stronger correlation. The absolute value of the Pearson correlation coefficient is used as the initial edge weight. When the absolute value of the Pearson correlation coefficient is less than a preset threshold (e.g., 0.3), it is considered that there is no direct stiffness coupling relationship between the two graph nodes, and the edge weight is set to 0, i.e., they are not connected. This yields the initial adjacency matrix and edge weights of the graph neural network, reflecting the mechanical coupling strength between different parts of the structure in terms of dynamic response.

[0071] S320. Construct a graph neural network; the input of the graph neural network is the topological damage fingerprint of each graph node, its hidden layer uses graph convolution operation to aggregate the features of neighboring graph nodes, and its output is the stiffness degradation coefficient and damage probability of each graph node.

[0072] The architecture of the graph neural network can adopt the existing architecture of conventional graph neural networks, which includes several graph convolutional layers. For each graph node, the graph convolutional layer performs a weighted summation of its own features and the features of all its neighboring graph nodes (the weights are determined by the edge weights and the graph node degree, where the node degree refers to the sum of the edge weights of all directly connected neighboring graph nodes). After linear transformation and nonlinear activation, new features are output. After multiple layers of graph convolution, each graph node possesses comprehensive information about its local neighborhood, which can reflect the stiffness coupling role of the graph node in the overall structure. Finally, through a fully connected layer and a sigmoid activation function, two values ​​are output for each graph node: a stiffness degradation coefficient (valued between 0 and 1, where 0 represents no degradation and 1 represents complete failure) and a damage probability (valued between 0 and 1).

[0073] S330. The graph neural network is trained using historical data under healthy conditions. The trained graph neural network outputs the stiffness degradation coefficient distribution map and damage probability distribution map of all finite element nodes of the entire target under monitoring, which are respectively used as the global stiffness degradation field and damage probability map.

[0074] In practical applications, the training process of a graph neural network is as follows: The graph neural network is pre-trained using historical data under healthy conditions (such as data from the first year after a bridge's completion and acceptance, at which point the structure is considered undamaged). Since damage labels are lacking under healthy conditions, simulation enhancement can be used: small random perturbations are applied to the stiffness of each graph node in the finite element model to generate simulated damage samples, with the perturbed stiffness degradation coefficient serving as the monitoring signal. Network parameters are optimized by minimizing the error between the predicted values ​​and the simulated labels (including the mean square error of the stiffness degradation coefficient and the cross-entropy loss of the damage probability). After training, the current topological damage fingerprint is input into the trained graph neural network, and forward propagation yields the stiffness degradation coefficient and damage probability of each graph node. The prediction results of all graph nodes are plotted as a cloud map according to their spatial location, resulting in the global stiffness degradation field and damage probability map.

[0075] In a preferred embodiment of the present invention, the step of predicting the health status of the target to be monitored within a future preset time window based on the global stiffness degradation field and damage probability map, namely step S400, specifically includes: first, flattening the global stiffness degradation field and damage probability map of each time step into a time-series state vector to form time-series data; then, constructing a long short-term memory network, taking the time-series data of the past M time steps as input, and outputting a prediction vector sequence for the next T time steps. This prediction vector sequence corresponds to the prediction result of the global stiffness degradation field and damage probability map within the future preset time window, and serves as the prediction result of the health status of the target to be monitored.

[0076] In practical applications, the global stiffness degradation field contains the stiffness degradation coefficients of all finite element nodes of the target to be monitored. Each coefficient ranges from 0 to 1, where 0 represents complete health with no degradation and 1 represents complete failure. The damage probability map contains the damage probability of the same number of nodes, also ranging from 0 to 1, indicating the possibility of damage to that node. Taking a long-span cable-stayed bridge as an example, the finite element model contains J nodes. Therefore, the global stiffness degradation field at each time step is a list containing J values, and the damage probability map is also a list containing J values. These two lists are concatenated end-to-end according to node order to form a one-dimensional vector of length 2J, which is the time-series state vector. To eliminate the influence of different dimensions, each dimension of this vector is standardized: the original value of that dimension is subtracted from the mean value under healthy conditions, and then divided by the standard deviation under healthy conditions. This operation is repeated for all historical time steps to obtain a series of time-series state vectors, which are arranged in chronological order to form time-series data.

[0077] Long Short-Term Memory (LSTM) networks are deep learning models specifically designed for processing time-series data. Their core feature is the introduction of gating mechanisms to control the forgetting, storage, and output of information. Specifically, an LSTM unit contains three gates: the forget gate determines which information is discarded from the previous time step; the input gate determines which new information from the current time step is stored in long-term memory; and the output gate determines which information is selected from the current memory as the output. Through the synergistic effect of these three gates, LSTM can effectively capture long-term dependencies in time series data, avoiding the gradient vanishing or gradient exploding problems that are common in traditional recurrent neural networks.

[0078] In this embodiment, the input to the Long Short-Term Memory (LSTM) network is a sequence of temporal state vectors from the past M time steps. The value of M can be determined based on the characteristics and requirements of the actual monitoring data; for example, 28 corresponds to the past 7 days (4 time steps per day). The output of the LSTM network is a sequence of predicted vectors for the next T time steps, with each predicted vector having the same dimension as the input temporal state vector. The value of T is also determined based on the prediction requirements; for example, 12 corresponds to the next 3 days (4 time steps per day). The LSTM network typically has two hidden layers, each containing several LSTM units (e.g., 128). Finally, a fully connected layer maps the hidden states output by the LSTM to the target dimension, i.e., each component of the predicted vector.

[0079] During the training phase, historical monitoring data is used to train the Long Short-Term Memory (LSTM) network. Specifically, training samples are constructed from existing time-series data using a sliding window approach. The input to each training sample is a temporal state vector spanning M consecutive time steps, and the target output is the temporal state vector spanning T consecutive time steps after that window. For example, using monitoring data from the past 6 months can generate thousands of training samples. The loss function during training is the mean squared error between the predicted and actual outputs, which measures the average of the squared differences between each predicted and actual value. Gradient descent algorithms, such as the Adam optimizer, are used to continuously adjust the weight parameters in the network, gradually reducing the loss function until the network converges.

[0080] In the prediction phase, the latest M time-series state vector sequences are input into a pre-trained LSTM network. The Long Short-Term Memory (LSTM) network performs forward computation and automatically outputs a sequence of predicted vectors for the next T time steps. Each predicted vector is a one-dimensional array with the same length as the input time-series state vector (i.e., twice the number of nodes). Each predicted vector is inversely normalized (i.e., restored to its original dimensions using stored mean and standard deviation), and then split into two parts according to node order: the first half corresponds to the predicted stiffness degradation coefficient of each node, and the second half corresponds to the predicted damage probability of each node. These two parts are plotted as contour maps according to their spatial locations, yielding the predicted global stiffness degradation field and predicted damage probability map for each future time step. These prediction results represent the predicted health status of the target object within a preset future time window.

[0081] In addition, a damage probability threshold (e.g., 0.6) can be set. When the predicted damage probability of a node at a future time step exceeds this threshold, an early warning signal is output, along with the node's spatial location (e.g., latitude and longitude coordinates or bridge station number) and the expected time of occurrence. This allows maintenance personnel to schedule inspections or repairs in advance, shifting from reactive response to proactive early warning.

[0082] like Figure 5 As shown, in another embodiment of the present invention, an IoT-based large-span structure monitoring data analysis system is also provided to implement the above-mentioned IoT-based large-span structure monitoring data analysis method. Specifically, it includes: a multimodal data acquisition module 10, used to acquire the multimodal excitation-response spatiotemporal tensor of the target to be monitored; the multimodal excitation-response spatiotemporal tensor includes the apparent image data of the target to be monitored, the structural dynamic response data, and the equivalent dynamic stiffness and damping ratio parameters of its local region.

[0083] The damage fingerprint extraction module 20 is used to perform persistent cohomology calculation on the multimodal excitation-response spatiotemporal tensor, extract topological feature barcodes, and then perform nonlinear dimensionality reduction to obtain a topological damage fingerprint that reflects the stiffness distribution and damage evolution of the target to be monitored.

[0084] The stiffness field reconstruction module 30 is used to generate a global stiffness degradation field and a damage probability map based on a preset graph neural network and the topological damage fingerprint.

[0085] The health status prediction module 40 is used to predict the health status of the target to be monitored within a future preset time window based on the global stiffness degradation field and damage probability map.

[0086] like Figure 6As shown, in a preferred embodiment of the present invention, the multimodal data acquisition module 10 specifically includes: an image data acquisition unit 11, used to acquire the apparent image data of the target to be monitored collected by the UAV formation, and perform alignment and tensor quantization processing according to a unified spatiotemporal reference to obtain the processed apparent image data.

[0087] The physical data acquisition unit 12 is used to acquire multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by the UAV formation; the multimodal excitation-response data includes structural dynamic response data synchronously acquired by a fixed sensor array and equivalent dynamic stiffness and damping ratio parameters of the local area of ​​the target under monitoring acquired synchronously.

[0088] The data fusion unit 13 is used to fuse the multimodal excitation-response data with the processed appearance image data to generate a multimodal excitation-response spatiotemporal tensor.

[0089] like Figure 7 As shown, in a preferred embodiment of the present invention, the damage fingerprint extraction module 20 specifically includes: a scalar field definition unit 21, used to regard the multimodal excitation-response spatiotemporal tensor as a scalar field function defined on the spatial grid of the target to be monitored; for each level set of the scalar field function, calculate its corresponding simplex, and track the generation threshold and disappearance threshold of topological features including connected components, one-dimensional holes, and two-dimensional holes that appear as the level set changes.

[0090] The topological feature generation unit 22 is used to form an interval by combining the generation threshold and the extinction threshold of each topological feature, and the set of all intervals constitutes the topological feature barcode.

[0091] The topological feature dimensionality reduction unit 23 is used to vectorize the topological feature barcode and then perform nonlinear dimensionality reduction through a nonlinear manifold learning model based on Laplacian feature mapping to obtain the topological damage fingerprint.

[0092] like Figure 8 As shown, in a preferred embodiment of the present invention, the stiffness field reconstruction module 30 specifically includes: a finite element model establishment unit 31, used to establish a finite element model of the target to be monitored, taking each finite element node as a graph node, and calculating the Pearson correlation coefficient between graph nodes based on the historical correlation of the structural dynamic response data at each graph node, as the initial edge weight of the graph neural network.

[0093] Graph neural network building unit 32 is used to build a graph neural network; the input of the graph neural network is the topological damage fingerprint of each graph node, its hidden layer uses graph convolution operation to aggregate the features of neighboring graph nodes, and its output is the stiffness degradation coefficient and damage probability of each graph node.

[0094] The network training and output unit 33 is used to train the graph neural network using historical data under healthy conditions. The trained graph neural network outputs the stiffness degradation coefficient distribution map and damage probability distribution map of all finite element nodes of the entire target under monitoring, which are respectively used as the global stiffness degradation field and damage probability map.

[0095] It should be noted that the above modules and units can be implemented as a computer program, which can run on a computer device. The computer device's memory can store the computer program that makes up the modules or units, enabling the processor to execute the various steps of the above method.

[0096] It should be understood that although the steps in the flowcharts of the embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0097] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods.

[0098] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A method for analyzing monitoring data of large-span structures based on the Internet of Things, characterized in that, Includes the following steps: Acquire the multimodal excitation-response spatiotemporal tensor of the target to be monitored; the multimodal excitation-response spatiotemporal tensor includes the apparent image data of the target to be monitored, the structural dynamic response data, and the equivalent dynamic stiffness and damping ratio parameters of its local regions; Persistent cohomology calculation is performed on the multimodal excitation-response spatiotemporal tensor to extract topological feature barcodes, and then nonlinear dimensionality reduction is performed to obtain a topological damage fingerprint that reflects the stiffness distribution and damage evolution of the target to be monitored. Based on a preset graph neural network, a global stiffness degradation field and a damage probability map are generated according to the topological damage fingerprint. Based on the global stiffness degradation field and damage probability map, the health status of the target to be monitored within a future preset time window is predicted. The steps for obtaining the multimodal excitation-response spatiotemporal tensor of the target to be monitored specifically include: The apparent image data of the target to be monitored is acquired by the UAV formation and aligned and tensed according to a unified spatiotemporal reference to obtain the processed apparent image data. Acquire multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by UAV formation; the multimodal excitation-response data includes structural dynamic response data synchronously acquired by a fixed sensor array and equivalent dynamic stiffness and damping ratio parameters of local areas of the target under monitoring acquired synchronously; The multimodal excitation-response data is fused with the processed appearance image data to generate a multimodal excitation-response spatiotemporal tensor; The steps for acquiring multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by UAV formations specifically include: The coordinated excitation paths of each UAV in the UAV formation are planned so that different UAVs hover at different parts of the target to be monitored, with staggered hovering times. At its hovering point, each UAV changes its rotor speed according to a preset pseudo-random sequence to generate a broadband incoherent downwash airflow as an excitation signal. At the same time, the UAV's onboard inertial measurement unit records its own three-axis acceleration and angular velocity response, and a fixed sensor array synchronously records the acceleration, strain, and force response of the target to be monitored under the same excitation. Based on the calibration model of UAV rotor speed and downwash airflow velocity distribution, the downwash airflow excitation force spectrum at each hovering point is calculated. Then, the frequency response function is calculated by comparing it with the response signals of each sensor in the fixed sensor array. The equivalent dynamic stiffness and damping ratio of the local area of ​​the target to be monitored are calculated by back-calculating the coherent transfer function. The aerodynamic admittance distribution cloud map of the equivalent dynamic stiffness and damping ratio of all hovering points is obtained by spatial interpolation, which serves as multimodal excitation-response data.

2. The method for analyzing monitoring data of large-span structures based on the Internet of Things according to claim 1, characterized in that, The steps of performing persistent cohomology calculations on the multimodal excitation-response spatiotemporal tensor, extracting topological feature barcodes, and then performing nonlinear dimensionality reduction to obtain a topological damage fingerprint reflecting the stiffness distribution and damage evolution of the target under monitoring specifically include: The multimodal excitation-response spatiotemporal tensor is regarded as a scalar field function defined on the spatial grid of the target to be monitored; for each level set of the scalar field function, its corresponding simplex is calculated, and the generation threshold and disappearance threshold of topological features including connected components, one-dimensional holes, and two-dimensional holes that appear as the level set changes are tracked. The generation threshold and the extinction threshold of each topological feature are combined to form an interval, and the set of all intervals constitutes the topological feature barcode. After vectorizing the topological feature barcode, nonlinear dimensionality reduction is performed through a nonlinear manifold learning model based on Laplacian feature mapping to obtain the topological damage fingerprint.

3. The method for analyzing monitoring data of large-span structures based on the Internet of Things according to claim 1, characterized in that, The graph neural network uses the finite element nodes of the target to be monitored as graph nodes and the coherence between the structural dynamic response data at each graph node as edge weights to learn the stiffness coupling relationship between each graph node. The steps of generating a global stiffness degradation field and a damage probability map based on a preset graph neural network and the topological damage fingerprint specifically include: A finite element model of the target to be monitored is established, each finite element node is treated as a graph node, and the Pearson correlation coefficient between graph nodes is calculated based on the historical correlation of the structural dynamic response data at each graph node, which serves as the initial edge weight of the graph neural network. Construct a graph neural network; the input of the graph neural network is the topological damage fingerprint of each graph node, its hidden layer uses graph convolution operation to aggregate the features of neighboring graph nodes, and its output is the stiffness degradation coefficient and damage probability of each graph node. The graph neural network is trained using historical data under healthy conditions. The trained graph neural network outputs the stiffness degradation coefficient distribution map and damage probability distribution map of all finite element nodes of the entire target under monitoring, which are respectively used as the global stiffness degradation field and damage probability map.

4. The method for analyzing monitoring data of large-span structures based on the Internet of Things according to claim 1, characterized in that, The step of predicting the health status of the target to be monitored within a future preset time window based on the global stiffness degradation field and damage probability map specifically includes: First, the global stiffness degradation field and damage probability map at each time step are flattened into a time-series state vector to form time-series data. Then, a long short-term memory network is constructed, which takes the time-series data of the past M time steps as input and outputs a prediction vector sequence for the next T time steps. This prediction vector sequence corresponds to the prediction results of the global stiffness degradation field and damage probability map within the future preset time window, and serves as the health status prediction result of the target to be monitored.

5. A large-span structure monitoring data analysis system based on the Internet of Things (IoT), used to implement the large-span structure monitoring data analysis method based on the IoT as described in any one of claims 1-4, characterized in that, include: The multimodal data acquisition module is used to acquire the multimodal excitation-response spatiotemporal tensor of the target to be monitored; the multimodal excitation-response spatiotemporal tensor includes the apparent image data, structural dynamic response data and the equivalent dynamic stiffness and damping ratio parameters of the local region of the target to be monitored; The damage fingerprint extraction module is used to perform persistent cohomology calculation on the multimodal excitation-response spatiotemporal tensor, extract topological feature barcodes, and then perform nonlinear dimensionality reduction to obtain a topological damage fingerprint that reflects the stiffness distribution and damage evolution of the target to be monitored. The stiffness field reconstruction module is used to generate a global stiffness degradation field and a damage probability map based on a preset graph neural network and the topological damage fingerprint. The health status prediction module is used to predict the health status of the target to be monitored within a future preset time window based on the global stiffness degradation field and damage probability map.

6. The IoT-based monitoring data analysis system for large-span structures according to claim 5, characterized in that, The multimodal data acquisition module specifically includes: The image data acquisition unit is used to acquire the apparent image data of the target to be monitored collected by the UAV formation, and perform alignment and tensing processing according to a unified spatiotemporal reference to obtain the processed apparent image data. The physical data acquisition unit is used to acquire multimodal excitation-response data of the target under monitoring under broadband incoherent airflow excitation generated by UAV formation; the multimodal excitation-response data includes structural dynamic response data synchronously acquired by a fixed sensor array and equivalent dynamic stiffness and damping ratio parameters of local areas of the target under monitoring acquired synchronously. The data fusion unit is used to fuse the multimodal excitation-response data with the processed appearance image data to generate a multimodal excitation-response spatiotemporal tensor.

7. The IoT-based monitoring data analysis system for large-span structures according to claim 5, characterized in that, The damaged fingerprint extraction module specifically includes: The scalar field definition unit is used to treat the multimodal excitation-response spatiotemporal tensor as a scalar field function defined on the spatial grid of the target to be monitored; for each level set of the scalar field function, the corresponding simplex is calculated, and the generation threshold and disappearance threshold of topological features including connected components, one-dimensional holes, and two-dimensional holes that appear as the level set changes are tracked. The topological feature generation unit is used to form an interval by combining the generation threshold and the extinction threshold of each topological feature, and the set of all intervals constitutes the topological feature barcode. The topological feature dimensionality reduction unit is used to vectorize the topological feature barcode and then perform nonlinear dimensionality reduction through a nonlinear manifold learning model based on Laplacian feature mapping to obtain the topological damage fingerprint.

8. The IoT-based monitoring data analysis system for large-span structures according to claim 5, characterized in that, The stiffness field reconstruction module specifically includes: The finite element model building unit is used to build the finite element model of the target to be monitored. Each finite element node is treated as a graph node, and the Pearson correlation coefficient between graph nodes is calculated based on the historical correlation of the structural dynamic response data at each graph node, which serves as the initial edge weight of the graph neural network. The graph neural network construction unit is used to construct a graph neural network. The input of the graph neural network is the topological damage fingerprint of each graph node. Its hidden layer uses graph convolution operation to aggregate the features of neighboring graph nodes. Its output is the stiffness degradation coefficient and damage probability of each graph node. The network training and output unit is used to train the graph neural network using historical data under healthy conditions. The trained graph neural network outputs the stiffness degradation coefficient distribution map and damage probability distribution map of all finite element nodes of the entire target under monitoring, which serve as the global stiffness degradation field and damage probability map, respectively.

Images