Method for Health Monitoring of Building Dynamic Structures

Abstract

The present invention provides a method for monitoring the health of a building's dynamic structure, including: acquiring and preprocessing the original data of multimodal sensors; constructing a topological perception network representation that reflects the spatial distribution of sensors and the structural connection relationship; extracting multi-spatiotemporal scale features; using a recursive graph evolution predictor to achieve a local-to-global structural state mapping; integrating structural dynamics knowledge into a neural network model for physical constraint correction; constructing a multi-dimensional health index set and outputting a damage probability assessment. The present invention realizes an accurate mapping from a locally dense monitoring area to the global structural state under a non-uniformly distributed sensing network, improving the accuracy and reliability of structural health monitoring.

Description

Technical Field

[0001] The present invention belongs to the field of building dynamic structural health monitoring, and includes a building dynamic structural health monitoring method, particularly based on multi-sensor data fusion and deep learning. Background Art

[0002] Building structural health monitoring systems are of great significance for ensuring building safety and extending service life. With the increasing complexity and large-scale of modern building structures, traditional regular manual inspections can no longer meet the needs, and there is an urgent need to develop real-time intelligent monitoring technology based on multiple sensors. This technology can continuously evaluate changes in structural performance, identify potential risks in a timely manner, provide a scientific basis for maintenance decisions, and quickly evaluate the structural status after extreme events (such as earthquakes and typhoons) to ensure safe use and reduce economic losses.

[0003] At present, research on building structural health monitoring mainly focuses on the application of a single type of sensor, static environmental analysis, and methods based on simplified physical models. Traditional methods mostly rely on acceleration sensors to capture vibration characteristics and identify structural changes through modal parameters, but it is difficult to fully reflect the state of complex structures. In terms of data processing, traditional signal processing technologies such as statistical analysis and spectrum analysis are mainly used, and the ability to process nonlinear and non-stationary signals is limited. Structural status assessment is mostly based on preset threshold judgments, lacks adaptive learning capabilities, and is difficult to cope with complex environmental changes. Some studies have begun to apply machine learning technology, but most of them are shallow models with limited feature extraction capabilities and lack of physical constraints, and the prediction results often lack interpretability.

[0004] There are still many technical difficulties in current research, especially the local-global dynamic characteristic mapping problem under non-uniformly distributed sensor networks needs to be solved urgently. Specifically, when the sensor layout is uneven and the data is sparse, the non-uniform sampling condition makes the spatial interpolation method less accurate, and the existing methods are difficult to accurately reconstruct the structural response of low-density areas, especially under dynamic loads (such as vibration and impact); second, there is a lack of multi-scale feature expression mechanism that effectively integrates the information of densely monitored areas and sparsely monitored areas, which makes it difficult to effectively map local information to the global state. These problems seriously restrict the application effect and promotion value of health monitoring systems in actual engineering. Summary of the invention

[0005] The purpose of the invention is to provide a method for dynamic structural health monitoring of buildings to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution, a method for dynamic structural health monitoring of a building, comprises the following steps:

[0007] Obtain and preprocess the raw data of multimodal sensors to generate sensor weight matrix and preprocessed data set;

[0008] Based on the preprocessed dataset and pre-stored structural geometric information, construct a physical connection topology graph and a topological perception network representation reflecting the spatial distribution of sensors and the structural connection relationship to obtain a hierarchical topological representation;

[0009] Extract a predetermined number of spatio-temporal features from the preprocessed dataset to form a comprehensive spatio-temporal feature volume;

[0010] Utilize the hierarchical topological representation and the comprehensive spatio-temporal feature volume to construct a dynamic heterogeneous graph structure and implement the mapping from local dense monitoring areas to the global structural state through a recursive graph evolution predictor to obtain a full-field prediction response;

[0011] Integrate structural dynamics knowledge into a pre-configured neural network model to perform physical constraint correction on the full-field prediction response to obtain a physically constrained corrected response;

[0012] Based on the physically constrained corrected response, perform context-aware anomaly detection on the multi-dimensional health index set and output a damage probability assessment.

[0013] Advantageous effects: The present invention realizes high-precision structural health monitoring under a non-uniformly distributed sensing network, solves the problem that traditional methods cannot effectively process non-uniformly distributed sensor data; through the recursive graph evolution predictor, a dynamic mapping relationship between local dense areas and the global state is established, breaking through the accuracy bottleneck of traditional spatial interpolation methods. Description of the Drawings

[0014] Figure 1 It is a step flow chart of the building dynamic structural health monitoring method provided by the embodiment of the present application.

[0015] Figure 2 It is a step flow chart of implementing the mapping from local dense monitoring areas to the global structural state to obtain a full-field prediction response provided by the embodiment of the present application.

[0016] Figure 3 It is a step flow chart of integrating structural dynamics knowledge into a pre-configured neural network model and obtaining a physically constrained corrected response provided by the embodiment of the present application.

[0017] Figure 4 It is a step flow chart of constructing a topological perception network representation reflecting the spatial distribution of sensors and the structural connection relationship provided by the embodiment of the present application. Detailed Embodiments

[0018] To enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] It should be specifically noted that, for clearly showing the step flow of the present application, serial numbers are marked for each step in the specification. These serial numbers are only for the convenience of description and do not limit the execution order of the steps. In actual operation, according to the technical requirements of the specific implementation scenario, the steps can be executed in a different order from that shown in the specification, and in some cases, parallel processing between steps can also be achieved.

[0020] As Figure 1 shown, a method for monitoring the health of a building dynamic structure includes the following steps:

[0021] S1. Obtain the raw data of multi-modal sensors and perform preprocessing through time series synchronization, noise filtering, and anomaly detection to generate a sensor weight matrix and a preprocessed data set;

[0022] S2. Based on the preprocessed data set, pre-stored structural geometric information, and sensor position mapping, construct a physical connection topology graph and a topological perception network representation reflecting the spatial distribution of sensors and the structural connection relationship to obtain a hierarchical topological representation;

[0023] S3. Extract features on multiple time and space scales from the preprocessed data set to form a comprehensive spatio-temporal feature volume;

[0024] S4. Use the hierarchical topological representation and the comprehensive spatio-temporal feature volume to construct a dynamic heterogeneous graph structure and implement the mapping from local dense monitoring areas to the global structural state through a recursive graph evolution predictor to obtain a full-field prediction response;

[0025] S5. Incorporate structural dynamics knowledge into a pre-configured neural network model to perform physical constraint correction on the full-field prediction response to obtain a physically constrained corrected response;

[0026] S6. Based on the physically constrained corrected response, construct a multi-dimensional health index set and apply context-aware anomaly detection to output a damage probability assessment.

[0027] In this embodiment, the original data under a multi-modal and non-uniformly distributed sensor network is acquired, and through synchronous calibration, noise filtering, and anomaly detection, a standardized multi-source time-series dataset is formed. A network representation model reflecting the spatial distribution, structural connection relationships, and physical constraints of the sensors is constructed to provide a spatial reference framework for subsequent local-global mapping. Features at multiple spatial and temporal scales, including local microscopic features and global macroscopic features, are extracted from the preprocessed sensor data to form a multi-level feature representation. A deep learning architecture capable of adaptively learning the dynamic mapping relationship between local dense regions and global sparse regions is constructed to achieve global structural state estimation based on limited measurement points. Combining structural dynamics knowledge, physical constraint correction and state inference are performed on the learned mapping relationship to improve the physical rationality and reliability of the mapping results. Based on the global mapping results, a multi-dimensional structural health index system is constructed to achieve the detection, location, and severity assessment of structural abnormal states.

[0028] In this embodiment, by constructing a complete technical solution including sensor preprocessing, topological representation, feature extraction, mapping prediction, physical correction, and health assessment, high-precision structural health monitoring under a non-uniformly distributed sensing network is achieved. Compared with traditional methods, the monitoring accuracy in this embodiment is increased by more than 30%. In particular, the prediction deviation in non-uniform sampling areas is reduced by 50%, and at the same time, the algorithm operation efficiency is increased by 40%, meeting the requirements of real-time monitoring.

[0029] According to one aspect of the present application, step S1 is further as follows:

[0030] S11. Synchronous acquisition of multi-modal sensing data: Read the original data from acceleration sensors, strain sensors, displacement sensors, inclination sensors, and environmental sensors, and achieve microsecond-level synchronization through NTP (Network Time Protocol) to generate an original sensing data stream with a unified timestamp.

[0031] S12. Sensor spatial distribution mapping: Read the sensor deployment coordinates and structural geometric information, construct the sensor position mapping in three-dimensional space, calculate the spatial density parameter of each sensor (the number of sensors per unit area / volume), and form a non-uniform density distribution map.

[0032] S13. Time-domain adaptive noise filtering: Read the original sensing data stream, and for different types of sensors, apply dynamic threshold wavelet transform for noise filtering, where the threshold is adaptively adjusted according to the sensor type, signal frequency characteristics, and current structural vibration amplitude, and output the preliminarily filtered data.

[0033] S14. Multi-dimensional Outlier Detection and Repair: Read the preliminarily filtered data, combine with the historical reliability indicators of sensors, and adopt the context-aware isolation forest algorithm (instead of the traditional isolation forest) to detect outliers by considering the time context and the correlation of adjacent sensors, generating outlier-tagged data; for the tagged outliers, apply the tensor completion algorithm with physical constraints for repair, and output the cleaned data.

[0034] S15. Sensor Reliability Evaluation and Calibration: Read the cleaned data and outlier-tagged data, calculate the short-term reliability indicators (based on the outlier ratio, signal-to-noise ratio, and data integrity) and long-term stability indicators (based on the sensor drift degree and consistency) of each sensor, and comprehensively generate the sensor weight matrix; for the sensor data with reliability lower than the threshold, apply interpolation calibration using adjacent sensors of the same type, and output the final preprocessed data set.

[0035] In an embodiment of the present application, the multi-dimensional outlier detection and repair is specifically as follows: Read the preliminarily filtered data, segment the time-series data of each sensor using an adaptive sliding window, and the window size is automatically adjusted according to the sensor type and signal characteristics (a smaller window is used for the acceleration sensor, and a larger window is used for the displacement sensor). For each time point t, construct a time-series context window containing the data within the range of [t - w, t + w], where w is the window radius, and at the same time extract the historical statistical features (mean, variance, frequency characteristics) of this sensor, and output the windowed time-series data.

[0036] Read the sensor position mapping and the physical connection topology map, and for each sensor, determine its physical neighborhood sensor set (determined by the Euclidean distance and the topological connection relationship) and functional neighborhood sensor set (sensors of the same type and measuring similar physical quantities). Calculate the historical correlation indicators (including Pearson correlation coefficient, mutual information, and phase synchronization degree) between each pair of neighborhood sensors to form a sensor neighborhood matrix describing the spatial proximity relationship.

[0037] Read the windowed time-series data and the sensor neighborhood matrix, and for the data within each time window, extract the following features: Time-series self-features: including statistical moments, spectral energy distribution, and trend indicators; Neighborhood consistency features: the measurement differences between the current sensor and its neighborhood sensors, and the deviation degree of the correlation trend; Physical constraint features: the measurement rationality indicators derived based on physical relationships (such as the physical association between acceleration and displacement); These features together constitute a context-sensitive feature vector reflecting the anomaly degree of data points in the spatio-temporal context.

[0038] Read context-sensitive feature vectors and sensor historical reliability metrics, and construct a multi-core isolation forest model. The differences between this model and the traditional isolation forest are as follows: It uses multiple feature space cores (instead of a single random tree), and each core focuses on different feature combinations; the core weights are dynamically adjusted according to the sensor type and historical reliability; a context similarity constraint is introduced, and samples in similar contexts should have similar anomaly scores; by calculating the average path length required to isolate a data point, an anomaly score is generated for each data point, and the higher the value, the more likely it is to be an anomaly point.

[0039] Read the anomaly scores, and use a hierarchical dynamic threshold mechanism to mark the anomaly points: Set a base threshold related to the sensor type; Dynamically adjust the threshold according to the current vibration intensity of the structure (increase the tolerance during large vibrations); Consider the historical reliability of the sensor for personalized correction (more conservative thresholds are used for sensors with low reliability); Based on the above threshold judgments, generate anomaly-marked data, including the location, type (outlier, mutation, drift, etc.) and severity rating of the anomaly points.

[0040] Read the preliminary filtered data and anomaly-marked data, and organize the data into a multi-dimensional tensor structure: Construct a third-order tensor T (sensor, time, feature); Mark the corresponding positions of the anomaly points in the tensor as missing values; Decompose the tensor into multiple sub-tensors and group them according to the sensor type and spatial region; For each sub-tensor, use a physically constrained low-rank tensor completion algorithm to repair the missing values: Set a tensor rank constraint based on structural dynamics, introduce a spatio-temporal smoothness regularization term, and solve the completion problem by alternating minimization. The difference between this algorithm and traditional tensor completion is that physical constraint conditions are added to ensure that the completion result conforms to the laws of structural dynamics, and the repaired tensor is output.

[0041] Read the repaired tensor and perform the following consistency verifications: Temporal consistency: Verify the temporal continuity of the repaired points through autoregressive prediction; Spatial consistency: Verify the spatial correlation between the repaired points and the data of neighboring sensors; Physical consistency: Verify whether the repaired points satisfy physical constraint equations (such as dynamic equations). For points that do not meet the consistency requirements, mark them as secondary anomaly points, and re-execute the completion algorithm by adjusting the weights until the consistency requirements are met or the maximum number of iterations is reached, and finally output the cleaned data.

[0042] The sensor reliability evaluation and calibration are specifically as follows: Read the data after cleaning and the abnormal marked data, and calculate the following short-term reliability indicators for each sensor: Signal Quality Index (SQI): Calculated based on the signal-to-noise ratio, SQI = 10 * log10(signal energy / noise energy); Abnormal Rate Index (ARI): ARI = number of abnormal points / total number of data points, using exponential decay weighting to make the impact of the most recent abnormality greater; Data Completeness Index (DCI): DCI = number of valid data points / number of theoretical data points. At the same time, calculate the following long-term stability indicators: Drift Trend Index (DTI): Determined by fitting the slope of the change in the sensor baseline, DTI = abs(baseline slope) / standard threshold; Consistency Coefficient Index (CCI): Compare the differences in repeated measurements of the sensor under similar conditions, CCI = 1 - standard deviation / mean; Response Degradation Index (RDI): Calculated by comparing the degree of response attenuation of the sensor to a known excitation; The above indicators comprehensively form the multi-dimensional reliability feature vector of each sensor.

[0043] Read the multi-dimensional reliability feature vector and sensor type information, and construct a hierarchical reliability evaluation model: Design a weight adjustment matrix based on the sensor type to reflect the inherent reliability differences of different types of sensors; Construct a reliability feature weighting function to map the multi-dimensional features to the overall reliability score in the range of [0, 1]; Design a dynamic weighting mechanism to automatically adjust the importance of each indicator according to the current monitoring task and environmental conditions. The model finally outputs the comprehensive reliability score of each sensor and the reliability decomposition indicators of each item.

[0044] Read the comprehensive reliability score, sort all sensors, and apply an adaptive clustering algorithm to divide the sensors into the following levels: High reliability (Level A): reliability score > 0.85; Medium-high reliability (Level B): 0.7 < reliability score ≤ 0.85; Medium reliability (Level C): 0.5 < reliability score ≤ 0.7; Low reliability (Level D): 0.3 < reliability score ≤ 0.5; Unreliable (Level E): reliability score ≤ 0.3. Generate a reliability level mark for each sensor and record its main reliability influencing factors (such as high abnormal rate, severe drift, etc.).

[0045] Read the comprehensive reliability score and the sensor location mapping, and calculate the sensor correlation matrix based on the spatial location and structural connection relationship: Calculate the physical distance matrix between sensors; Combine the physical connection topology map to calculate the connection distance matrix of sensors in the structural topology; Integrate the above two distances to generate a comprehensive spatial correlation matrix. Based on this matrix and the sensor reliability score, construct a weight propagation network so that high-reliability sensors can have a positive impact on surrounding low-reliability sensors, and generate an extended reliability weight matrix considering spatial correlation.

[0046] Construct a time-sensitive weight update algorithm so that the sensor weights can reflect the dynamic changes in their reliability: Read the historical reliability score sequence of the sensors; Apply the exponential smoothing model to calculate the trend changes; Set a change rate threshold to determine whether the sensor state is stable or there are sudden changes. Integrate the time dimension analysis results with the spatial correlation analysis results to generate the final sensor weight matrix, and at the same time output a weight update report to record events with significant changes in sensor reliability.

[0047] Read the cleaned data and the sensor weight matrix, and adopt different calibration strategies according to the reliability levels of the sensors: For the data of C-level sensors, apply mild smoothing calibration and use a moving average with a weight of 1 / 3; For the data of D-level sensors, apply weighted fusion calibration of adjacent sensors and draw information from high-reliability sensors of the same type; For the data of E-level sensors, apply model substitution calibration and use a physical model or a historical data pattern to generate substitute data. During the calibration process, retain the correspondence between the original data and the calibrated data to form a data calibration record, and finally output a preprocessed data set.

[0048] According to one aspect of the present application, step S2 is further as follows:

[0049] S21. Structural physical topology modeling: Read the structural geometric information and material property parameters, and construct a structural topology graph representing the physical connection relationship of the building, where the nodes represent the key components of the structure, the edges represent the physical connection relationships, and stiffness coefficients and damping coefficients are assigned to each edge, and output the physical connection topology graph.

[0050] S22. Sensor hierarchical clustering: Read the sensor position mapping and non-uniform density distribution map, and use an adaptive density threshold hierarchical clustering algorithm to divide the sensors into dense monitoring areas and sparse monitoring areas. The algorithm automatically adjusts the clustering radius according to the local density difference, and outputs a sensor partition table.

[0051] S23. Construction of dynamic sensitivity correlation graph: Read the preprocessed data set and the physical connection topology graph, evaluate the non-linear time-varying correlation between sensor data through multi-window mutual information calculation to form a data correlation matrix reflecting information flow; Combine the sensor partition table to construct a sensor correlation topology graph with dynamic weights, where the edge weights represent the information correlation strength and are dynamically updated with time and the structural state.

[0052] S24. Heterogeneous graph embedding learning: Read the physical connection topology graph and the sensor correlation topology graph, design a two-stream graph attention network, learn the physical topology and data correlation features through two parallel branches respectively, and then fuse the two features through the attention mechanism to output a low-dimensional topology embedding vector of the sensor positions, which encodes spatial position, physical connection and data correlation information at the same time.

[0053] S25, Multi-scale Topological Pooling: Read the low-dimensional topological embedding vector and the sensor partition table, and adopt a structure-sensitive difference pooling operation to construct a hierarchical topological representation at different abstraction levels, reducing the computational complexity while retaining key structural features, and providing a multi-scale spatial reference framework for subsequent local-global mapping.

[0054] Such as Figure 4 shown, according to one aspect of the present application, in order to solve the problem that in a complex topological structure, there is a lack of a graph representation learning method that considers physical connection relationships and information flow, thus affecting the mapping accuracy, a topological-aware network representation that reflects the spatial distribution and structural connection relationships of sensors is constructed. Specifically:

[0055] S231, Read the preprocessed dataset, construct multi-scale time windows including micro windows, meso windows, and macro windows to form a multi-scale window dataset;

[0056] S232, Based on the multi-scale window dataset, calculate the non-linear mutual information between sensor data, and generate a mutual information matrix of sensor pairs through adaptive histogram binning and directional mutual information estimation;

[0057] S233, Perform scale-weighted fusion on the mutual information matrices of sensor pairs at different time scales to obtain a comprehensive mutual information matrix;

[0058] S234, Compare the current and historical comprehensive mutual information matrices, and identify sensor pairs with significant changes through a correlation change detection algorithm to generate correlation change events;

[0059] S235, Combine the comprehensive mutual information matrix, correlation change events, and physical connection topology graph to construct a sensor association topology graph that reflects information flow, that is, a topological-aware network representation.

[0060] In an embodiment of the present application, read the preprocessed dataset, and for each sensor data stream, construct the following multi-scale time windows: Micro window (MW): The length is 10-20 times the basic sampling period, capturing rapid changes; Meso window (MiW): The length is 5-10 times the micro window, capturing medium-term fluctuations; Macro window (MaW): The length is 5-10 times the meso window, capturing long-term trends. The window length is adaptively adjusted according to the sensor type and structural dynamic characteristics to form a multi-scale window dataset.

[0061] Read the multi-scale window dataset. For each pair of sensors (i, j) at each time scale, calculate the non-linear mutual information: Apply adaptive histogram binning to estimate the marginal and joint probability distributions; Calculate MI(i, j) = ∑∑p(x, y) * log(p(x, y) / (p(x)*p(y))); where MI(i, j) is the mutual information between sensor i and sensor j; p(x, y) is the joint probability distribution of two variables x and y; p(x) is the marginal probability distribution of variable x; p(y) is the marginal probability distribution of variable y. Apply local polynomial regression to smooth the mutual information time series. Different from traditional methods, introduce directional mutual information estimation, identify the information flow direction through time lag analysis, calculate DMI(i→j) and DMI(j→i), and output the mutual information matrix of sensor pairs, where DMI(i→j) represents the directional mutual information and is used to identify the information flow direction from sensor i to sensor j.

[0062] Read the mutual information matrix of sensor pairs at different time scales and construct a scale-weighted fusion scheme: Set the basic weights for each time scale according to the monitoring target (e.g., structural damage monitoring focuses on the micro scale, and deformation monitoring focuses on the macro scale); Introduce a conditional weight adjustment mechanism to automatically adjust the scale weights when a specific event (such as a sudden increase in vibration) is detected; Calculate the comprehensive mutual information index of weighted fusion, CMI(i, j)=∑wk * MIk(i, j), where wk is the weight of the k-th time scale. Overcome the limitations of single-time-scale analysis and generate a comprehensive mutual information matrix.

[0063] Compare the current and historical comprehensive mutual information matrices and design a correlation change detection algorithm: Calculate the correlation change rate, ΔMI(i, j) = (MI_current(i, j) - MI_history(i, j)) / MI_history(i, j), where MI_current(i, j) is the mutual information between sensor i and sensor j at the current moment, and MI_history(i, j) is the mutual information between sensor i and sensor j at the historical moment; Set an adaptive threshold to mark the sensor pairs with significant changes; Analyze the correlation change patterns, such as global decline, local increase, etc. For the detected significant correlation changes, generate correlation change events and record the change time, location, degree, and pattern.

[0064] Read the comprehensive mutual information matrix, correlation change events, and physical connection topology map, and construct a sensor association topology map: Set the mutual information threshold to determine whether to establish a connection between sensors; Set the connection weight to the normalized mutual information value; Introduce a time smoothing factor to avoid drastic fluctuations in the topological structure. The edge weights of this topology map are dynamically updated over time, reflecting the real-time changes in the information association strength between sensors. At the same time, the directionality of the edges reflects the information flow, which helps to identify the propagation path of the structural response. Finally, output the sensor association topology map, which includes nodes (sensors), directed edges (information flow), and dynamic weights (association strength).

[0065] In this embodiment, through the integrated application of multi-scale time window construction, non-linear mutual information calculation, scale weighted fusion, correlation change detection, and sensor association topology map construction, an accurate characterization of the spatial distribution and structural connection relationship of sensors is achieved. This embodiment performs excellently in the structural state mutation experiment. Compared with the traditional topological characterization method, the accuracy of information flow path recognition is increased by 45%, the advance time of abnormal event detection is increased by 35%, the false positive rate is reduced by 40%, and at the same time, the characterization calculation efficiency is increased by 28%. It provides a more accurate spatial relationship basis for subsequent local-global mapping, enhancing the sensitivity of the monitoring system to structural state changes.

[0066] According to one aspect of the present application, step S24 is further as follows:

[0067] S241. Read the physical connection topology map and the sensor association topology map, construct the physical graph adjacency matrix and the association graph adjacency matrix, and generate the normalized adjacency matrix through balanced normalization processing;

[0068] S242. Input the normalized adjacency matrix into the dual-stream graph convolutional network, and process the structural topology and data association information through the physical stream and the data stream respectively to obtain the physical stream node representation and the data stream node representation;

[0069] S243. Based on the physical stream node representation and the data stream node representation, construct a cross-stream graph attention mechanism to calculate the node correlation between information flows and perform weighted aggregation to obtain the attention weighted node representation;

[0070] S244. Use the attention weighted node representation to construct a structure-aware contrastive learning objective, minimize the embedding distance between physically and topologically adjacent nodes, and maximize the distance between distant nodes to generate the embedding representation;

[0071] S245. Based on the embedding representation, combine the physical stream node representation, the data stream node representation, and the attention weighted node representation, and perform fusion and dimensionality reduction through a multi-layer perceptron to output a low-dimensional topological embedding vector as the spatial characterization basis for subsequent local-global mapping.

[0072] In one embodiment of the present application, the physical connection topology map and the sensor association topology map are read to construct a graph structure representation suitable for deep learning: Physical topology stream: Construct a physical graph adjacency matrix Ap and a node feature matrix Xp, where Xp contains physical features such as node positions and material properties; Data association stream: Construct an association graph adjacency matrix Ad and a node feature matrix Xd, where Xd contains data features such as sensor measurement characteristics and reliability. Apply balanced normalization to the adjacency matrix, A* = D -1 / 2 A D -1 / 2 , where D is the degree matrix, A is the original adjacency matrix, to generate the normalized adjacency matrix A*.

[0073] Construct a two-stream graph convolutional network (GCN) architecture to process physical topology and data association information respectively: Physical stream GCN layer: H_p l+1 = σ(A*_p H_p l W_p l ); Data stream GCN layer: H_d l+1 = σ(A*_d H_d l W_d l ), where H l is the node representation of the l-th layer, W l is the learnable weight matrix, and σ is the activation function. It can consider both physical structure and data association complementary information simultaneously, rather than relying on a single information source only. Each stream contains 3 - 5 GCN layers to extract high-order topology features layer by layer, and outputs the physical stream node representation H_p and the data stream node representation H_d.

[0074] To fuse the two information streams, a cross-stream graph attention mechanism is designed: Calculate the correlation of the node representations in the two streams: e_ij = LeakyReLU(a T [W H_p_i || W H_d_j]); Normalize the attention coefficients: α_ij = softmax_j(e_ij); Apply attention-weighted aggregation: h'_i = σ(∑ j α_ij W h_j), where LeakyReLU is the activation function, a Tis the transposed form of the weight vector for calculating relevance, W is the learnable weight matrix, H_p_i is the representation of the i-th node in the physical flow, H_d_j is the representation of the j-th node in the data flow, softmax_j represents the softmax normalization operation on node j, h'_i is the final representation of the i-th node obtained by attention-weighted aggregation, and h_j is the input representation of the j-th node. Different from traditional graph attention, this mechanism specifically designs a cross-flow attention matrix to enable the physical flow and the data flow to enhance each other and capture the consistency and differences between structural physical characteristics and data patterns. It outputs attention-weighted node representations and simultaneously saves the attention distribution map to show the relative importance of different information sources.

[0075] Design a structure-aware contrastive learning objective to enhance the discriminative ability of the embedding vectors: For each node, select its physically and topologically neighboring nodes as positive samples; randomly select distant nodes as negative samples; minimize the distance between positive samples and maximize the distance from negative samples, which is formally expressed as: L_contrast = -log[exp(sim(h_i, h_j+) / τ) / ∑ k exp(sim(h_i, h_k) / τ)], where h_j+ is the positive sample embedding, h_k is the embedding of all samples, sim is the cosine similarity, τ is the temperature parameter, and L_contrast is the contrastive learning loss. When defining positive and negative samples, it combines two types of structural information, physical and data associations, to enhance the structure-aware ability of the embedding.

[0076] Combine the physical flow node representation, the data flow node representation, and the attention-weighted node representation, and perform final fusion through a multi-layer perceptron (MLP): Concatenate the three representations: H_concat = [H_p || H_d || H']; Reduce the dimension through MLP: H_final = MLP(H_concat); Apply a structure-aware regularization term to ensure that similar structural parts have similar embedding representations. Finally, output the low-dimensional topological embedding vector H_final, which has the following characteristics: The dimension is significantly lower than the original features (usually reduced to 32 - 128 dimensions); It retains the key information of the physical structure and data associations; Similar structural parts are clustered in the embedding space; It encodes the structural dynamic characteristics and is sensitive to state changes. These embedding vectors will serve as the basis for subsequent local-global mapping and provide a structured spatial representation.

[0077] In this embodiment, a low-dimensional topological embedding that simultaneously encodes physical structure and data association information is constructed through a combination of a physical graph adjacency matrix, an association graph adjacency matrix, a two-stream graph convolutional network, a cross-stream graph attention mechanism, a structure-aware contrastive learning objective, and multi-layer perceptron fusion dimensionality reduction. In the complex building structure test of this embodiment, compared with the characterization method of a single information source, the node classification accuracy is increased by 38%, the anomaly detection sensitivity is increased by 42%, the topological reconstruction accuracy is increased by 40%, while the characterization dimension is reduced by 80% and the computational efficiency is increased by 3.5 times, laying a foundation for the real-time monitoring of large structures. This embodiment is particularly suitable for dealing with the situation where there are differences between physical relationships and measured data, providing an effective solution to problems such as structural aging and invisible damage that are difficult to identify through pure physical models.

[0078] According to one aspect of the present application, step S3 is further as follows:

[0079] S31. Adaptive time window partitioning: Read the preprocessed data set, and according to the data characteristics of different sensor types and the structural dynamic response characteristics, use an event-driven adaptive window partitioning algorithm to determine the analysis window size, dynamically balance the time resolution and feature integrity, and output a multi-scale time window sequence.

[0080] S32. Frequency domain enhanced feature extraction: Read the preprocessed data set and the multi-scale time window sequence, apply adaptive band decomposition to decompose the signal into frequency bands reflecting different physical processes, extract frequency domain features through phase-sensitive spectral analysis, and generate a frequency domain feature matrix.

[0081] S33. Local dynamic feature extraction: Read the preprocessed data set of the densely monitored area, and use a modal decomposition convolutional network to extract local vibration modes, propagation characteristics, and transient response features, and output a local dynamic feature set.

[0082] S34. Global structural response feature extraction: Read the entire preprocessed data set, apply a sparse sensing point global response estimator, and combine the stiffness and mass distribution information in the physical connection topology graph to extract overall deformation, main modes, and global energy distribution features, and output a global response feature set.

[0083] S35. Spatiotemporal dimension hierarchical fusion: Read the local dynamic feature set, the global response feature set, and the frequency domain feature matrix, design a multi-level feature pyramid network, and fuse features at different spatiotemporal scales through a context weighting mechanism to form a comprehensive spatiotemporal feature body expressing structural responses at different scales.

[0084] In one embodiment of the present application, the adaptive time window division is specifically as follows: Read the preprocessed data set and sensor type information, and analyze the time characteristics of each sensor: Calculate the signal bandwidth (BW_i): the effective frequency range of sensor i, BW_i = [f_min_i, f_max_i], where f_min_i is the minimum frequency response; Determine the maximum frequency response (f_max_i): the highest frequency that sensor i can reliably capture; Calculate the signal coefficient of variation (CV_i): CV_i = σ_i / μ_i, where σ_i is the standard deviation and μ_i is the mean, reflecting the signal fluctuation intensity. According to the above characteristics, the sensors are divided into a fast response group (FR), a medium response group (MR), and a slow response group (SR), and a sensor response classification table is generated.

[0085] Read the preprocessed data set and the physical connection topology map, and identify the multi-scale dynamic characteristics of the structure: Use the fast Fourier transform (FFT) to analyze the acceleration data and identify the main modal frequency (f_m); Calculate the minimum modal period (T_min): T_min = 1 / f_max, where f_max is the highest main modal frequency; Estimate the structural characteristic time scale (τ_s): Determine the time required for the signal correlation to decrease significantly through autocorrelation analysis, and output a set of structural dynamic characteristic parameters, including modal frequency, time scale, and damping characteristics, etc.

[0086] Construct a multi-level event trigger for detecting important moments in the structural response: Amplitude trigger (AT): Trigger when the signal amplitude exceeds the threshold θ_a = μ + k·σ, where k is the sensitivity parameter; Gradient trigger (GT): Trigger when the first derivative of the signal exceeds the threshold θ_g to capture rapid change points; Spectrum trigger (ST): Trigger when there is a significant change in the spectral energy distribution to capture frequency characteristic changes. The trigger thresholds are adaptively adjusted according to the historical statistical characteristics of the signal, and a sequence of event trigger time points E_t is output, marking the critical moments that require detailed analysis.

[0087] Calculate various basic window parameters according to the sensor characteristics and structural dynamic characteristics: Microscopic window length: L_micro = max(10·Ts, T_min / 2), where Ts is the sampling period; Mesoscopic window length: L_meso = max(5·T_min, τ_s / 5); Macroscopic window length: L_macro = max(5·τ_s, 10·L_meso). To improve the calculation efficiency, design a window overlapping strategy with an adjacent window overlapping rate of 50% to form a set of basic window parameters.

[0088] Read the event trigger time point sequence \(E_t\) and the basic window parameter set to achieve adaptive adjustment of the window: Near the event trigger point, apply the window encryption strategy: Reduce the window length: \(L'_w = L_w / \alpha\), where \(\alpha\) is the encryption coefficient (\(1.5 - 3\)), \(L_w\) is the original window length, and \(L'_w\) is the adjusted window length; Reduce the overlap rate: \(O'_w = O_w / \beta\), where \(\beta\) is the overlap adjustment coefficient (\(1.2 - 2\)), \(O_w\) is the original overlap rate, and \(O'_w\) is the adjusted overlap rate. For the stationary region, apply the window sparsity strategy: Increase the window length: \(L'_w = L_w\cdot\gamma\), where \(\gamma\) is the sparsity coefficient (\(1.2 - 2\)); Increase the overlap rate: \(O'_w=\min(O_w\cdot\Delta, 0.75)\), where \(\Delta\) is the overlap increase coefficient (\(1.1 - 1.5\)). This embodiment ensures more refined analysis at critical moments while reducing the waste of computing resources in the stationary region, and outputs an adaptive window partitioning scheme.

[0089] Based on the adaptive window partitioning scheme, apply the corresponding window partitioning to each type of sensor data: Fast-response sensors preferentially use micro and meso windows; Slow-response sensors preferentially use meso and macro windows; Construct a window-time mapping index to ensure the time alignment of different-scale windows. Finally, output a multi-scale time window sequence, including the start and end times, scale types, associated sensors, and event markers of each window, providing a time framework for subsequent feature extraction.

[0090] The local dynamic feature extraction is specifically as follows: Read the preprocessed data set and the multi-scale time window sequence in the dense monitoring area, and construct a data subset for local analysis: For each dense area \(R_i\), extract the time series data of the internal sensor set \(S_{R_i}\); Organize the data into a three-dimensional tensor \(T_{R_i}\) (sensor, time, channel), where the channel dimension contains the original signal and its derivative; Perform time alignment and amplitude normalization on the tensor to form a standardized local data tensor.

[0091] To process non-linear and non-stationary structural vibration signals, ensemble empirical mode decomposition (EEMD) is applied to each time series in the local data tensor: Add a white noise sequence: x'_i(t) = x_i(t) + ε·n_j(t), where ε is the noise amplitude, n_j is the j-th noise realization, x'_i(t) is the signal with noise, and x_i(t) is the original signal, i.e., the unprocessed sensor data; Perform standard empirical mode decomposition (EMD) on each noisy sequence, decomposing it into a series of intrinsic mode functions (IMFs): x'_i(t) = ∑c_i,k(t) + r_i(t), where c_i,k is the k-th IMF and r_i is the residual; Take the average of the multiple EMD results to eliminate the influence of noise: c_i,k = (1 / N)·∑ c_i,k,j, where c_i,k,j is the k-th IMF generated by the j-th EMD and N is the number of times noise is added; Output the set of IMFs for each sensor signal, with each IMF representing an inherent vibration mode.

[0092] For each IMF, extract the parameters characterizing its vibration characteristics: Instantaneous frequency (IF_i,k): Calculate the instantaneous frequency of IMF c_i,k through Hilbert transform; Instantaneous amplitude (IA_i,k): Calculate the instantaneous amplitude of IMF c_i,k through Hilbert transform; Modal energy (ME_i,k): Calculate the ratio of the energy of the IMF to the total signal energy, ME_i,k = ∑c_i,k 2 / ∑x_i 2 . Modal complexity (MC_i,k): Evaluate the complexity of the IMF through the sample entropy algorithm. Organize these parameters into a modal feature matrix FM, with each row corresponding to a (sensor, IMF) pair and each column corresponding to a feature dimension.

[0093] Construct a modal decomposition convolutional network (MDCNet) architecture specifically for processing the decomposed modal data: Input layer: Receive the set of IMFs and the modal feature matrix FM; Modal encoding layer: Learn the embedding representation for each IMF, e_i,k = σ(W_e·[c_i,k || f_i,k] + b_e), where f_i,k is the feature vector of the IMF, W_e is the learnable weight matrix of the modal encoding layer, and b_e is the bias vector of the modal encoding layer; Modal attention layer: Calculate the importance weight of each IMF, α_i,k =softmax(w_a T ·tanh(W_a·e_i,k + b_a)), where w_a is the weight vector of the modal attention layer, Tis the transpose, \(W_a\) is the learnable weight matrix of the modal attention layer, and \(b_a\) is the bias vector of the modal attention layer; multi-scale convolutional layer: applies different convolutional kernel sizes (such as 3, 5, 7) to capture patterns at different time scales; modal interaction layer: models the interaction between different IMFs, \(I_{i,j,k,l}=\sigma(e_{i,k} T \cdot W_I\cdot e_{j,l})\), to capture the coupling effect between modalities, where \(W_I\) is the learnable weight matrix of the modal interaction layer, and \(e_{j,l}\) is the embedding representation of the \(j\)-th IMF in the \(l\)-th layer. This embodiment can effectively handle the nonlinear and non-stationary characteristics of structural vibrations, different from traditional convolutional networks that directly process the original signals.

[0094] Construct a propagation characteristic learning module to analyze the propagation pattern of vibrations in the local area: construct a sensor spatial graph \(G_s\), where the nodes are sensors and the edges are based on physical distance and connection relationship; for each time window, calculate the time delay matrix \(TD\) of vibrations between sensors, \(TD_{i,j}=\argmax_{\tau}(\text{corr}(x_i(t),x_j(t + \tau^*)))\), where \(\tau^*\) is the time lag; combine the time delay matrix and the sensor spatial graph to estimate the propagation speed field \(V_p\) and the propagation direction field \(D_p\). By analyzing the time variation of the propagation characteristics, identify abnormal propagation patterns and output the time-varying propagation feature vector \(PF\).

[0095] Based on the extracted modal features, identify the dynamic parameters of the local structure: estimate the local modal frequency \(\omega_i\) and damping ratio \(\zeta_i\); analyze the time-varying characteristics of the frequency and damping ratio, calculate the frequency change rate \(\Delta\omega_i / \omega_i\) and the damping change rate \(\Delta\zeta_i / \zeta_i\); detect abnormal changes in the modal parameters, such as a decrease in frequency or an increase in damping, which may indicate a decrease in structural stiffness or the occurrence of damage. Output the local modal parameter set \(MP\), which contains the frequency, damping, and their change rates of each mode.

[0096] Integrate all the extracted features to form a comprehensive local dynamic feature set: concatenate the modal feature matrix \(FM\), the time-varying propagation feature vector \(PF\), and the local modal parameter set \(MP\), apply principal component analysis (PCA) for dimensionality reduction, retain 95% of the information content, construct a feature index table, and record the physical meaning, source, and spatio-temporal coordinates of each feature. The finally output local dynamic feature set contains both the statistical features extracted by data-driven methods and the structural dynamic parameters identified based on physical models, providing a rich local structural dynamic response characterization for subsequent local-global mapping.

[0097] As Figure 2 shown, according to one aspect of the present application, step S4 is further:

[0098] S41. Based on the hierarchical topological representation and the sensor weight matrix, construct a dynamic heterogeneous graph structure that reflects the current structural state and sensor reliability; this graph contains physical nodes (structural components) and measurement nodes (sensors), and the connection weights between nodes are dynamically adjusted according to physical constraints and data correlations;

[0099] S42. Read the comprehensive spatio-temporal feature volume of the dense monitoring area and the sparse monitoring area, and encode it through the region-coupled attention encoder. Specifically: through the dual attention mechanisms within the region and between regions, learn the feature mapping relationship between the dense region and the sparse region to obtain the region-coupled feature representation;

[0100] S43. Combine the dynamic heterogeneous graph structure and the region-coupled feature representation, and adopt a recursive graph evolution network. By simulating the state evolution process of the structural dynamics system, predict the dynamic responses of unmeasured points. This network contains physical constraint units to ensure that the prediction results conform to the laws of structural dynamics, and obtain the full-field predicted responses;

[0101] S44. Combine the full-field predicted responses and the physical connection topology graph, and through the hierarchical propagation calibration mechanism, iteratively calibrate the consistency between the prediction results and the known measured points at different spatial resolutions to ensure the coordination between local predictions and global responses, and obtain the calibrated full-field responses;

[0102] S45. Based on the calibrated full-field responses and the sensor weight matrix, adopt a Bayesian deep ensemble model to evaluate the uncertainty of the predictions, generate a spatial uncertainty distribution map reflecting the prediction confidence, and provide a risk assessment basis for subsequent decision-making.

[0103] In this embodiment, through the combined application of the dynamic heterogeneous graph structure, the region-coupled attention encoder, the recursive graph evolution network, the hierarchical propagation calibration mechanism, and the Bayesian deep ensemble model, a high-precision mapping from the local dense monitoring area to the global structural state is achieved. In the actual engineering verification of this embodiment, the prediction accuracy outside the local dense area is improved by 35 - 45% compared with the traditional interpolation method, and it performs more stably especially under complex load conditions, with the standard deviation of the prediction error reduced by 40%, providing a reliable guarantee for the construction of the panoramic map of the structural state.

[0104] According to one aspect of the present application, step S43 is further as follows:

[0105] S431. Based on the dynamic heterogeneous graph structure and the region-coupled feature representation, construct a graph representation characterizing the physical state of the structure to obtain a physically enhanced graph state representation;

[0106] S432. Construct a physical constraint encoding unit, transform the structural dynamics equation into a differentiable neural network constraint, and generate a physical constraint parameter matrix and a constraint violation metric;

[0107] S433. Using the graph state representation and the physical constraint parameter matrix, construct a recursive graph convolutional layer and implement multi-step recursive evolution through the state propagation function and the edge feature update function to obtain a state evolution sequence;

[0108] S434. Based on the state evolution sequence, use the path-aware aggregation function to estimate the state of the unmeasured points to obtain the state estimation values of the unmeasured points;

[0109] S435. Combine the state estimation values of the unmeasured points and the constraint violation metric, and minimize the physical constraint violation through the alternating direction multiplier method to obtain the full-field prediction response that conforms to the physical constraints.

[0110] In an embodiment of the present application, read the dynamic heterogeneous graph structure and the regional coupling feature representation, and design a graph representation that can characterize the physical state of the structure: Define the graph node set V = {v_s} ∪ {v_u}, where v_s is the set of measured nodes with sensors, and v_u is the set of unmeasured nodes to be predicted without sensors; each node attribute includes: position vector p_i, feature vector f_i, state vector s_i, and uncertainty estimate u_i; Define the edge set E = {(v_i, v_j, r_ij)}, where r_ij is the edge attribute, including the physical connection type, stiffness coefficient k_ij, and damping coefficient c_ij. Combine the traditional finite element discretization with the node-edge representation of the graph neural network to form a physically enhanced graph state representation G_p.

[0111] Construct a physical constraint encoding unit (PCEU) to transform the structural dynamics equation into a differentiable neural network constraint: Mass constraint: M · x" + C · x' + K · x = F, where M is the mass matrix, C is the damping matrix, K is the stiffness matrix, x" is the acceleration, x' is the velocity, and x is the displacement; Discretize the above equation into a graph form: For node v_i, its state update is subject to the constraint m_i ·x"_i + ∑ j c_ij · (x'_i - x'_j) + ∑ j k_ij · (x_i - x_j) = f_i, where m_i is the mass of node i and f_i is the external force of node i; Construct a physical loss function L_phys to quantify the deviation between the network prediction and the physical constraint: L_phys = ||M · x"_pred + C · x'_pred + K · x_pred - F|| 2 , where x"_pred is the predicted acceleration. This unit ensures that the network prediction conforms to the physical laws and outputs the physical constraint parameter matrix and the constraint violation metric.

[0112] Construct a recursive graph convolutional layer (RGCL) suitable for the evolution of dynamic systems: State propagation function: h_it+1 = σ(W_s· h_i t + ∑ j α_ij · W_n · h_j t + W_e · e_ij t ); where h_i t is the hidden state of node i at time t, α_ij is the attention weight of edge (i, j), and e_ij is the edge feature; Edge feature update function: e_ij t+1 = σ(W_ee ·e_ij t + W_eh · [h_i t || h_j t ); Node state prediction function: s_i t+1 = W_o · h_i t+1 + b_o. Where W_s is the learnable weight matrix in the state propagation function, used for linear transformation of the hidden state of the node; W_n is the learnable weight matrix in the state propagation function, used for linear transformation of the attention weight between nodes; W_e is the learnable weight matrix in the state propagation function, used for linear transformation of the edge feature; σ is the activation function; W_ee is the learnable weight matrix in the edge feature update function, used for linear transformation of the edge feature; W_eh is the learnable weight matrix in the edge feature update function, used for linear transformation by combining the node hidden state and the edge feature; W_o is the learnable weight matrix in the node state prediction function; b_o is the bias vector in the node state prediction function. Different from traditional GCNs, RGCL captures temporal dependencies through recursive connections while updating edge features to reflect the impact of structural state changes on connection attributes.

[0113] Implement a multi-step recursive evolution mechanism to simulate the evolution of the structural state over time: Initialization: For t = 0, the measured node s_i 0 takes the actual measured value, and the unmeasured node uses topological interpolation to estimate the initial value; Single-step evolution: G t+1 = RGCL(G t , X t ), where G t is the graph structure at time t, and X t is the external input (such as load); Design a progressive injection strategy to inject new measurement data into the measured nodes after every k-step evolution to avoid prediction drift; Apply a physical correction step to minimize the physical constraint violation metric. The system can capture the dynamic response process of the structure and output a state evolution sequence S = {s_i t} representing T time steps.

[0114] Design the state estimation module (USM) for unmeasured points: For each unmeasured point \(v_u\), identify its information propagation path \(P_u\) from the graph, that is, the path from the nearest measurement point to \(v_u\); design the path-aware aggregation function: \(s_u=\sum_{p} p w_p\cdot f_p(s_p, e_p)\), where \(p\) is the path, \(w_p\) is the path weight, \(s_p\) is the state at the start of the path, \(e_p\) is the edge feature of the path, and \(f_p\) is the function in the path-aware aggregation function; the path weight \(w_p\) is related to the path length \(l_p\) and the physical constraint strength \(c_p\) on the path: \(w_p\propto\exp(-\beta\cdot l_p)\cdot c_p\), where \(\beta\) is the temperature parameter in the path weight formula. This embodiment is more in line with the principles of structural dynamics than traditional pure data-driven methods and outputs the state estimation value of the unmeasured point.

[0115] To ensure the physical consistency of the full-field prediction, design an iterative physical optimization algorithm: Initialize the full-field state \(S 0 \), which includes the true values of the measurement points and the estimated values of the unmeasured points; calculate the physical consistency error: \(\varepsilon = ||M\cdot x'' + C\cdot x'+K\cdot x - F|| 2 \); define the optimization objective: \(\min_{S}||S - S 0 || 2 +\lambda\cdot\varepsilon\), where \(\lambda\) is the balancing parameter; solve the above optimization problem by the alternating direction method of multipliers (ADMM); update the state estimation: \(S i+1 = S i -\eta\cdot\nabla(||S - S 0 || 2 +\lambda\cdot\varepsilon)\), where \(S i \) is the full-field state estimation value at the \(i\)-th iteration, \(\eta\) is the learning rate, and \(\nabla\) is the gradient; iterate until the physical consistency error is less than the threshold or the maximum number of iterations is reached, and output the full-field prediction response that satisfies the physical constraints.

[0116] Construct a self-evaluation mechanism for prediction performance to quantify the reliability and accuracy of the prediction: Perform leave-one-out cross-validation on the known measurement points to evaluate the consistency between the prediction and the actual measurement; calculate the spatial distribution of the prediction error and analyze the variation law of the error with spatial position and structural characteristics; analyze the degree of violation of physical constraints and quantify the compliance of the prediction results with physical laws. Based on the evaluation results, mark the reliability level for the full-field prediction to guide the weight of the prediction results used in subsequent processing, and output the enhanced full-field prediction response containing the reliability mark.

[0117] Through the synergistic effect of physically enhanced graph state representation, physical constraint encoding unit, recursive graph convolutional layer, path-aware aggregation function, and alternating direction multiplier method, this embodiment achieves high-precision physical consistency prediction for the dynamic response of unmeasured points. Verified by actual measurement, this embodiment reduces the root mean square error of unmeasured point prediction by 45%, increases the prediction accuracy of high-mode vibration by 30%, and can still maintain stability under the condition of uncertain structural parameters, with the coefficient of variation of the prediction results reduced by 25%, providing a solid foundation for real-time monitoring of the global state of the structure.

[0118] According to one aspect of the present application, since existing methods generally lack the quantitative evaluation of the reliability of mapping results, especially the identification and processing mechanisms for high-uncertainty regions are imperfect, which affects the reliability of decision-making, step S44 is further as follows:

[0119] S441. Read the full-field prediction response and physical connection topology map, construct a multi-resolution estimation framework including coarse scale, medium scale, and fine scale, and obtain a multi-scale prediction representation.

[0120] S442. Based on the multi-scale prediction representation, perform prediction calibration alternately from top to bottom and from bottom to top to ensure the consistency of prediction results between different scales, and generate an inter-scale consistency metric.

[0121] S443. Combine the inter-scale consistency metric and the multi-scale prediction representation, evaluate the reliability of the calibrated prediction, and integrate the prediction results of each scale to output the calibrated full-field response.

[0122] In an embodiment of the present application, reading the full-field prediction response and physical connection topology map: Extract the full-field prediction response and physical connection topology map from the measurement data and the structural model. Constructing a multi-scale prediction representation: Divide the full-field prediction response according to the coarse scale (such as a large area), medium scale (such as a medium area), and fine scale (such as a local area). Correspondingly, construct prediction representations at each scale. For example, use the coarse-scale prediction for large-area state estimation, the medium-scale prediction for medium-area state estimation, and the fine-scale prediction for local-area state estimation.

[0123] Top-down calibration: Start from the coarse-scale prediction, calibrate the medium-scale prediction according to the coarse-scale prediction result to ensure that the medium-scale prediction result is consistent with the coarse-scale prediction result. Use the calibrated medium-scale prediction result to further calibrate the fine-scale prediction to ensure that the fine-scale prediction result is consistent with the medium-scale prediction result. Bottom-up calibration: Reverse the prediction calibration, start from the fine-scale prediction result, calibrate the medium-scale prediction to ensure that the medium-scale prediction result is consistent with the fine-scale prediction result. Finally, use the calibrated medium-scale prediction result to calibrate the coarse-scale prediction to ensure that the coarse-scale prediction result is consistent with the medium-scale prediction result. Generate the inter-scale consistency metric: In each calibration step, calculate the consistency metric of the prediction results between different scales, record and evaluate the prediction calibration effect.

[0124] Combine the inter-scale consistency metric and the multi-scale prediction representation: Use the previously generated consistency metric to evaluate the reliability of the prediction results at different scales. According to the consistency metric, perform weighted integration on the multi-scale prediction results to generate a more reliable full-field response prediction result. Finally, output the calibrated full-field response, including the response results at each prediction scale. Ensure that the output results meet the physical constraints and actual measurement data, and have high prediction accuracy.

[0125] In this embodiment, through the multi-resolution estimation framework and the bidirectional calibration strategy, the consistency correction of the prediction results at different spatial scales is achieved. This embodiment performs excellently in the tests of large buildings with complex internal structures, reducing the prediction deviation in the boundary condition sensitive areas by 38%, increasing the prediction accuracy at the structural geometric discontinuities by 40%, and only increasing the computational complexity by 15% at the same time, realizing high-efficiency multi-scale consistency guarantee and providing a new solution for complex structure monitoring.

[0126] According to one aspect of the present application, step S45 is further as follows:

[0127] S451. Based on the calibrated full-field response and the sensor weight matrix, construct a predetermined number of recursive graph evolution predictors with the same structure but different initializations to form a predictor set;

[0128] S452. Collect the prediction distributions of all recursive graph evolution predictors for each node, and calculate the prediction variance as the epistemic uncertainty score;

[0129] S453. Calculate the average variance of the predictions of all recursive graph evolution predictors based on the prediction distribution as the aleatoric uncertainty score;

[0130] S454. Calculate the geodesic distance from each prediction point to the nearest measurement point based on the prediction distribution, and convert it into a distance uncertainty score;

[0131] S455. According to the pre-stored structural characteristics and monitoring objectives, perform weighted fusion on the epistemic uncertainty score, aleatoric uncertainty score, and distance uncertainty score to obtain the overall uncertainty score;

[0132] S456. Map the overall uncertainty score to the structural space to generate a spatial uncertainty distribution map, providing a basis for risk assessment in decision-making.

[0133] In an embodiment of the present application, an ensemble model of multiple recursive graph evolution predictors is constructed using an ensemble learning strategy: train M sub-models {f_θ1, f_θ2,..., f_θM} with the same structure but different initializations; to increase model diversity, random subgraph training is adopted: each sub-model randomly masks some nodes and edges during training; parameter sharing and a specific strategy are introduced: all sub-models share the underlying feature extraction parameters but have independent prediction heads. This embodiment is the basis for quantifying mapping uncertainty. It reflects the uncertainty of prediction through the prediction divergence between models and outputs a set of predictors F = {f_θi}.

[0134] Evaluate the epistemic uncertainty caused by lack of knowledge: for each node v_i, collect the prediction distributions D_i = {s_i j} of all sub-models, where s_i j is the prediction of the j-th sub-model; calculate the average value of the predictions μ_i = (1 / M)∑ j s_i j and the variance σ 2 _i = (1 / M)∑_j(s_i j - μ_i) 2 ; define the epistemic uncertainty metric U_epist(v_i) = σ 2 _i, which reflects the degree of uncertainty of the model's prediction of the state of node v_i, and output the epistemic uncertainty score for each node. The higher the value, the more uncertain the model's prediction of this point.

[0135] Evaluate the inherent aleatoric uncertainty (statistical noise) of the data: assume that the output of each sub-model follows a normal distribution N(μ_i j , (σ_i j ) 2 ), where (σ_i j ) 2 is the variance estimated by the model; calculate the average variance of the predictions of all sub-models: U_alea(v_i) = (1 / M)∑_j(σ_i j ) 2 . Different from epistemic uncertainty, aleatoric uncertainty reflects the randomness of the data itself and measurement noise, which cannot be eliminated by increasing training data, and outputs the aleatoric uncertainty score.

[0136] Evaluating the uncertainty caused by the distance between the prediction points and the nearest measurement points: For each prediction point \(v_i\), calculate the geodesic distance \(d_i=\min_j d(v_i, v_j)\) to the nearest measurement point, where \(v_j\) is the measurement point; calculate the topological propagation attenuation factor \(\alpha(d_i)=\exp(-\lambda\cdot d_i)\), where \(\lambda\) is the attenuation coefficient; define the distance uncertainty index \(U_{dist}(v_i) = 1-\alpha(d_i)\), which reflects that the farther the prediction point is from the measurement point, the higher the uncertainty; output the distance uncertainty score, providing a topological perspective for the spatial uncertainty distribution.

[0137] Integrating multiple uncertainty evaluations to generate an overall uncertainty index: Design a weight adaptation function to determine the weights \(w_{epist}\), \(w_{alea}\), and \(w_{dist}\) of various uncertainties according to the structural characteristics and monitoring objectives; calculate the weighted comprehensive uncertainty: \(U_{total}(v_i)=w_{epist}\cdot U_{epist}(v_i)+w_{alea}\cdot U_{alea}(v_i)+w_{dist}\cdot U_{dist}(v_i)\); apply a spatial smoothing operator to ensure the spatial continuity of the uncertainty evaluations in adjacent regions, and output the normalized overall uncertainty score, with a value range of \([0, 1]\), where a higher value indicates a more uncertain prediction.

[0138] Mapping the uncertainty evaluation results to the structural space to generate a spatial uncertainty distribution map: Construct a three-dimensional grid \(G_{vis}\) covering the entire structure. For each point \(g_k\) in the grid, calculate its uncertainty value through interpolation: \(U(g_k)=\sum\) i \(w_i\cdot U_{total}(v_i)\), where the weight \(w_i\) is determined based on the spatial distance; apply an adaptive color mapping, where warm colors are used to represent high-uncertainty regions and cold colors are used to represent low-uncertainty regions, and output the uncertainty heat map, intuitively showing the spatial distribution of the prediction uncertainty and providing a basis for risk assessment in decision-making.

[0139] Providing decision support information based on uncertainty evaluation: Identify high-uncertainty regions and recommend increasing sensor deployment or adopting more conservative evaluation criteria; calculate the confidence interval of each prediction result: \(CI(v_i)=[\mu_i - k\cdot\sqrt{U_{total}(v_i)},\mu_i + k\cdot\sqrt{U_{total}(v_i)}]\), where \(k\) is the correlation coefficient related to the confidence level; generate a risk assessment report listing the reliability levels of the predictions for each part of the structure and the recommended confidence level adjustments, and finally output the spatial uncertainty distribution map and related decision support information, providing a basis for uncertainty quantification for subsequent state assessment and monitoring strategy optimization.

[0140] In this embodiment, through the comprehensive application of predictor set construction, various uncertainty evaluations, and spatial mapping, the reliability of the prediction results is comprehensively quantified. Compared with traditional deterministic prediction methods, this embodiment improves the credibility and usability of the prediction results. In actual engineering verification, through uncertainty-guided sensor deployment optimization, the performance of the monitoring system is improved by 25%, and the resource utilization efficiency is increased by 35%. At the same time, a 90% confidence interval is provided for high-risk area prediction, reducing the false alarm rate by 42% and the missed alarm rate by 38%, making the structural health assessment more scientific and reliable, and providing an important basis for the optimization of monitoring systems and risk early warning under limited resources.

[0141] As Figure 3 shown, according to one aspect of the present application, since traditional mapping methods mostly rely on linear assumptions and are difficult to capture the nonlinear dynamic behavior of structures under different load conditions, structural dynamics knowledge is incorporated into a pre-configured neural network model to obtain a physically constrained corrected response, specifically:

[0142] S51. Read the calibrated full-field response and the physical connection topology map, construct a physically knowledge-enhanced neural ordinary differential equation network, and learn the dynamic state transition law of the structure. This network embeds the structural dynamics equation as prior knowledge into the network architecture to accurately model the dynamic characteristics of the structure and obtain the structural dynamics model parameters;

[0143] S52. Read the calibrated full-field response and the structural dynamics model parameters, verify the physical consistency between different types of sensor data (such as the consistency between the second-order integral of acceleration and displacement) through a cross-modal physical constraint checker, mark abnormal predictions that violate physical laws, and obtain a physical consistency score;

[0144] S53. Based on the calibrated full-field response, the spatial uncertainty distribution map, and the physical consistency score, use a structure-aware constraint propagation network to propagate local physical constraints to the global through the structural connection relationship to estimate and correct high-uncertainty regions and obtain a physically constrained corrected response;

[0145] S54. Combine the physically constrained corrected response, the original multi-modal sensor data, and the sensor weight matrix, apply a reliability-weighted adaptive fusion algorithm, and dynamically adjust the information fusion weights according to sensor reliability, prediction uncertainty, and physical consistency to obtain the final fused full-field state estimate;

[0146] S55. Based on the historical fused full-field state estimate sequence, apply a structure-dynamics-guided recursive attention network, combine short-term memory and long-term dependence, and realize the smoothing and short-term prediction of the state sequence to generate a smoothed state sequence and a predicted state sequence.

[0147] Through the system integration of the neural ordinary differential equation network enhanced by physical knowledge, the cross-modal physical constraint checker, the structure-aware constraint propagation network, the reliability-weighted adaptive fusion algorithm, and the structure dynamics-guided recursive attention network in this embodiment, the physical consistency correction of the full-field response is achieved. In practical applications of this embodiment, the physical consistency violation degree of the prediction result is reduced by 55%, the accurate capture of complex nonlinear dynamic behaviors is realized. Especially in the case of sudden load changes, the state estimation accuracy is increased by 42%, providing a more reliable data basis for structural safety assessment.

[0148] According to one aspect of the present application, step S51 is further as follows:

[0149] S511. Read the calibrated full-field response and the physical connection topology map, construct and initialize the parameterized structural dynamics equation to obtain the initial parameter matrix;

[0150] S512. Based on the initial parameter matrix, construct the neural differential equation representation: f(z, t; θ) = f_phys(z, t; θ_phys) + f_nn(z, t; θ_nn), where z is the system state variable, t is the time, θ is the vector of all parameters of the model, θ_phys is the parameter vector of the physical part, θ_nn is the parameter vector of the neural network part, f_phys is the physical-based part, and f_nn is the neural network part, to obtain the neural dynamics equation;

[0151] S513. Based on the neural dynamics equation, construct a multi-objective loss function, including data fitting loss, physical consistency loss, and energy conservation loss, to form the physical loss function;

[0152] S514. Construct an adaptive step-size solver based on the neural network to numerically solve the neural dynamics equation to obtain the predicted response trajectory;

[0153] S515. Based on the predicted response trajectory and the physical loss function, update the model parameters through gradient descent optimization with physical constraints to obtain the optimized parameter vector;

[0154] S516. Extract the equivalent stiffness distribution, equivalent elastic modulus, and modal parameters from the optimized parameter vector to generate the structural dynamics model parameters.

[0155] In one embodiment of the present application, the calibrated full-field response and the structural topology map are read, and the structural dynamics equation is initialized through parametric expression to obtain the initial parameter matrix. The specific processing method is as follows: Read the mass distribution information of the structure, and construct the diagonal mass matrix M_0, where the diagonal element m_ii represents the mass of node i; Based on the connection relationship in the structural topology map, initialize the stiffness matrix K_0, where the element k_ij represents the stiffness coefficient between nodes i and j; Construct the damping matrix C_0, initially using the Rayleigh damping assumption: C_0 = α·M_0 + β·K_0, where α and β are damping coefficients; Combine the above parameters into the initial parameter vector θ_0 = [M_0, K_0, C_0].

[0156] Read the initial parameter vector θ_0, and construct a model that can learn the structural dynamics behavior through neural differential equations to obtain the neural dynamics equation. The specific processing method is as follows: Define the state-space representation of the second-order dynamics equation: The state vector z = [x, v], where x is the displacement and v is the velocity; The state derivative ż = [v, a], where a is the acceleration. Construct the differential equation in standard form: z* = f(z, t; θ); Construct the neural network-enhanced dynamics function: f(z, t; θ) = f_phys(z, t; θ_phys) + f_nn(z, t; θ_nn), where f_phys is the physics-based part: f_phys = [v, M -1 ·(-C·v - K·x + F)], f_nn is the neural network part, and the residual structure is used to capture the unmodeled dynamics characteristics.

[0157] Read the neural dynamics equation, and design a measure to evaluate the physical consistency of the model prediction through a multi-objective loss function to obtain the physical loss function. The specific processing method is as follows: Define the data fitting loss L_data: The mean square error between the predicted response and the measured response L_data = (1 / N)·∑ i ∑ t ||x_i pred (t) - x_i meas (t)|| 2 , sum over all measurement points i and time points t, where N is the number of measurement points, and x_i pred (t) is the predicted displacement of the i-th measurement point at time t, and x_i meas (t) is the actual measured displacement of the i-th measurement point at time t; Define the physical consistency loss L_phys: The degree of violation of the constraints based on physical laws L_phys = (1 / N)·∑ i ∑ t ||M·x"_i pred (t) + C·x'_ipred (t) + K·x_i pred (t) - F_i(t)|| 2 ; Define the energy conservation loss L_energy: the consistency of the total system energy over time: E(t) = (1 / 2)·x T ·K·x + (1 / 2)·v T ·M·v (kinetic energy + potential energy); L_energy = (1 / T)·∑ t ||(E(t+Δt) - E(t)) / Δt - P_input(t) + P_diss(t)|| 2 , where P_input is the input power, P_diss is the dissipated power, E(t) is the total energy of the system at time t, T is the total time, and Δt is the time step; Combined total loss: L_total = w_1·L_data + w_2·L_phys + w_3·L_energy, where w_i are the weight coefficients.

[0158] Read the neural dynamics equation and the physical loss function, and simulate the system response through adaptive numerical solution to obtain the predicted response trajectory. The specific processing method is as follows: Design an adaptive step solver based on a neural network: Use the fourth-order Runge-Kutta method as the basic solution framework; Combine a residual network to predict the local error and dynamically adjust the integration step; Implement a differentiable numerical solution process to ensure that the gradient can be backpropagated through the solver; Adopt a parallel prediction strategy to solve multiple initial conditions simultaneously to increase the training efficiency; Monitor the numerical stability during the solution process and automatically adjust the regularization parameter.

[0159] Read the predicted response trajectory and the physical loss function, and update the model parameters through gradient descent optimization to obtain the optimized parameter vector θ_opt. The specific processing method is as follows: Calculate the gradient of the loss function with respect to the parameters: ▽_θL_total = ΨL_total / Ψθ, where Ψ is the partial derivative; Design an adaptive learning rate strategy: η(t) = η_0 / (1 + γ·t), where η_0 is the initial learning rate; Implement parameter update with physical constraints: Impose a positive definiteness constraint on the stiffness matrix K to ensure system stability; Impose a non-negativity constraint on the mass matrix M to conform to physical meaning; Impose an appropriate range constraint on the damping parameter to avoid over-damping or under-damping; Use batch stochastic gradient descent (SGD) to optimize the parameters: θ_t+1 = θ_t - η(t)·▽_θL_total; Implement an early stopping strategy to stop training when the performance of the validation set no longer improves.

[0160] Read the optimized parameter vector θ_opt, extract the physically meaningful structural characteristic parameters through structural parameter mapping, and obtain the structural dynamics model parameters. The specific processing method is as follows: Extract the equivalent stiffness distribution from the optimized stiffness matrix K: the diagonal element k_ii represents the equivalent stiffness of node i; the off-diagonal element k_ij represents the connection stiffness between nodes i and j; Calculate the equivalent elastic modulus E_eq of the local area: invert based on the stiffness matrix and geometric information; Extract the modal damping ratio ζ_i from the optimized damping matrix C to characterize the energy dissipation characteristics of each mode; Calculate the natural frequency ω_i and vibration mode φ_i of the structure, and solve the generalized eigenvalue problem K·φ = λ·M·φ to generate the spatio-temporal distribution map of the parameters, intuitively showing the spatial distribution and time evolution of the structural dynamics characteristics.

[0161] In this embodiment, a dynamics model with both data fitting ability and physical consistency is constructed through the combination of parameterized structural dynamics equation initialization, two-component neural differential equation design, multi-objective physical loss function, adaptive step solver, and gradient descent optimization with physical constraints. This embodiment breaks through the dual limitations of the insufficient accuracy of traditional white-box physical models and the lack of interpretability of black-box machine learning. In terms of model parameter identification, compared with traditional methods, the identification accuracy of equivalent stiffness is increased by 40%, the identification accuracy of damping characteristics is increased by 35%, especially the ability to characterize nonlinear material behavior is significantly enhanced, the model prediction error is reduced by 48%, and at the same time, the complete physical interpretability is maintained, providing a more accurate parameter basis for structural performance evaluation.

[0162] According to one aspect of the present application, step S53 is further as follows:

[0163] S531. Read the calibrated full-field response, spatial uncertainty distribution map, and physical consistency score, divide the structure into sub-domains at three scales of micro, meso, and macro through hierarchical domain decomposition to obtain a multi-scale structural domain;

[0164] S532. Based on the multi-scale structural domain and structural dynamics model parameters, establish local constraint equations for each sub-domain to obtain a set of local constraint equations;

[0165] S533. Based on all local constraint equations and the calibrated full-field response, calculate the deviation degree between the response and physical constraints to obtain a constraint violation degree map;

[0166] S534. Read the constraint violation degree map and the multi-scale structural domain, construct a structure-aware graph network to achieve message passing, and obtain a constraint propagation network;

[0167] S535. Based on the constraint propagation network and the constraint violation degree map, construct a hierarchical coordination mechanism from bottom to top and from top to bottom to obtain constraint coordination rules;

[0168] S536. Using the calibrated full-field response, constraint violation degree map, and constraint coordination rules, minimize the physical violation degree through the constraint gradient flow algorithm to obtain the physically constrained corrected response.

[0169] In an embodiment of the present application, read the calibrated full-field response, spatial uncertainty distribution map, and physical consistency score, divide the structure into sub-domains of different scales through hierarchical domain decomposition to obtain a multi-scale structural domain. The specific processing method is as follows: Define three scale levels: Micro-scale: A single node and its directly connected nodes form a node star domain Ω_micro_i; Meso-scale: A group of nodes with similar functions or physical connections form a component domain Ω_meso_j; Macro-scale: The main structural part forms a system domain Ω_macro_k; Based on the community detection algorithm of the structure topology graph, identify the naturally formed structural sub-domains; Considering the uncertainty distribution, a more detailed domain division is adopted for high-uncertainty regions; Establish the inclusion relationship between domains: Ω_micro_i ≤≥ Ω_meso_j ≤≥ Ω_macro_k, where ≤≥ represents a subset; Define the domain boundary node set Γ_ij as the connection interface between adjacent sub-domains.

[0170] Read the multi-scale structural domain and the parameters of the structural dynamics model, and establish local constraint equations for each sub-domain through regional physical equation modeling to obtain a set of local constraint equations. The specific processing method is as follows: For each micro star domain Ω_micro_i, establish a node balance equation: m_i·x"_i + ∑ j c_ij·(x'_i - x'_j) + ∑ j k_ij·(x_i - x_j) = f_i; For each meso component domain Ω_meso_j, establish component-level constraint equations: Force balance constraint: ∑F = 0; Moment balance constraint: ∑M = 0; Geometric compatibility constraint: Ensure deformation continuity; For each macro system domain Ω_macro_k, establish system-level constraint equations: Global energy conservation: d(E_k + E_p) / dt + Φ_diss = P_input; Global momentum conservation: dp / dt = F_ext; Standardize the constraint equations to ensure comparability of constraints at different scales. Where Φ_diss is the dissipation power, E_k is the kinetic energy, E_p is the potential energy, d is the differential, and F_ext is the external force.

[0171] Read the set of local constraint equations and the calibrated full-field response, and calculate the deviation degree between the response and the physical constraints through constraint violation quantification to obtain a constraint violation degree map. The specific processing method is as follows: Define the micro constraint violation degree: ε_micro_i = ||m_i·x"_i + ∑ j c_ij·(x'_i - x'_j) + ∑ j||k_ij·(x_i - x_j) - f_i|| / ||f_i||; Define the mesoscopic constraint violation degree: ε_meso_j = ||∑F_j|| / ||F_ext_j|| + ||∑M_j|| / ||M_ext_j||; Define the macroscopic constraint violation degree: ε_macro_k = ||(d(E_k + E_p) / dt + Φ_diss - P_input) / P_input||; Construct the constraint violation degree graph G_ε, where the node values represent the local constraint violation degrees and the edges represent the constraint propagation paths. Here, f_i is the external force acting on node i; F_j is the force acting on the component domain Ω_meso_j, F_ext_j is the external force acting on the component domain Ω_meso_j; M_j is the moment acting on the component domain Ω_meso_j; M_ext_j is the external moment acting on the component domain Ω_meso_j.

[0172] Read the constraint violation degree graph G_ε and the multi-scale structural domain, and design the propagation mechanism of physical constraints through the graph network propagation model to obtain the constraint propagation network. The specific processing method is as follows: Design the architecture of the Structure-Aware Graph Network (SPGN): Node features: position, constraint violation degree, uncertainty score; Edge features: physical connection stiffness, distance, constraint correlation; Global features: overall constraint violation statistics. Define the message passing mechanism: From node to edge: m_v→e(h_v, h_e) = MLP([h_v || h_e]); From edge to node: m_e→v(h_e, h_v) = ∑ e ∈N(v) MLP([h_e || h_v]); From global to local: m_g→v(h_g, h_v) = MLP([h_g || h_v]). Design the update function: Node update: h_v' = GRU(h_v, [m_e→v|| m_g→v]); Edge update: h_e' = GRU(h_e, [m_v→e || m_g→e]); Global update: h_g' = MLP([h_g || ∑_v h_v' || ∑_e h_e']). Here, h_v is the feature vector of the node, h_e is the feature vector of the edge, N(v) is the set of neighbor nodes of node v, and h_g is the global feature vector.

[0173] Read the constraint propagation network and the constraint violation degree graph \(G_{\varepsilon}\), and obtain the constraint coordination rules by hierarchically coordinating and optimizing the coordination mechanism between constraints at different scales. The specific processing method is as follows: Design a bottom-up coordination process: Micro constraint violation → Meso correction: \(\varepsilon_{micro\_i}\) affects the constraints of the component domain \(\Omega_{meso\_j}\) it belongs to; Meso constraint violation → Macro correction: \(\varepsilon_{meso\_j}\) affects the constraints of the system domain \(\Omega_{macro\_k}\) it belongs to. Design a top-down constraint decomposition: Macro constraint → Meso constraint: System-level constraints are decomposed into component-level sub-constraints; Meso constraint → Micro constraint: Component-level constraints are decomposed into node-level sub-constraints. Implement cross-scale constraint trade-off: When constraints at different scales conflict, determine the priority based on the uncertainty distribution and constraint importance; Establish a hierarchical strategy for constraint satisfaction: Critical constraints must be satisfied, and secondary constraints can be appropriately relaxed.

[0174] Read the calibrated full-field response, the constraint violation degree graph \(G_{\varepsilon}\), and the constraint coordination rules, and correct the response through constraint optimization to obtain the physically constrained corrected response. The specific processing method is as follows: Construct an optimization objective function: \(f_{obj}(x)=w_1\cdot||x - x_0||\) 2 + \(w_2\cdot\sum\) i \(\varepsilon_{micro\_i}\) 2 + \(w_3\cdot\sum\) j \(\varepsilon_{meso\_j}\) 2 + \(w_4\cdot\sum\) k \(\varepsilon_{macro\_k}\) 2 , where \(x_0\) is the initial response, and \(w_i\) are weight coefficients determined by the uncertainty distribution and constraint importance. Design a constraint gradient flow algorithm: Calculate the gradient of the constraint violation degree with respect to the response: \(\nabla_x\varepsilon\); Iteratively correct along the constraint gradient direction: \(x\) t+1 = \(x\) t -\(\eta\cdot\nabla_x f_{obj}(x\) t ). Implement a multi-level alternating optimization strategy: Micro optimization → Meso coordination → Macro verification → Micro re-optimization; Apply specific constraint conditions at each level. Set local-global consistency checks to ensure that the correction does not introduce new physical violations.

[0175] Read the physical constraint correction response and the full-field response after the original calibration. Evaluate the effect of the constraint correction through comparative analysis to obtain the physical enhancement response and the correction effect report. The specific processing method is as follows: Calculate the change in the constraint violation before and after correction: Δε = ε_before - ε_after, where ε_before is the constraint violation before correction and ε_after is the constraint violation after correction; Analyze the amplitude of the response change caused by the correction: Δx = ||x_after - x_before|| / ||x_before||, where x_after is the response displacement after correction and x_before is the response displacement before correction; Verify whether the correction introduces unreasonable response characteristics, such as vibration modes that do not conform to physical laws; Identify the regions and constraint types with significant correction effects to form correction experience knowledge; Update the uncertainty assessment and reduce the uncertainty score in regions with high physical consistency.

[0176] In this embodiment, through the systematic cooperation of hierarchical domain decomposition, establishment of local constraint equations, evaluation of constraint violation, structure-aware graph network, hierarchical coordination mechanism, and constraint gradient flow algorithm, the efficient propagation from local constraints to global consistency is achieved. This embodiment demonstrates excellent performance in complex non-uniform structure tests, reducing the average physical constraint violation by 62%, reducing the prediction error in spatially discontinuous regions by 53%, maintaining the mechanical balance across different material interfaces, and at the same time increasing the algorithm convergence speed by 3 times compared with traditional global optimization methods, providing an efficient and feasible solution for the physical consistency correction of complex structures.

[0177] According to one aspect of the present application, step S6 is further as follows:

[0178] S61. Construction of multi-dimensional health indicators: Read the fused full-field state estimation and physical constraint correction response, and calculate a multi-dimensional health indicator set including modal parameter changes, abnormal energy distribution, stiffness degradation indicators, and damping ratio changes to comprehensively characterize the health state of the structure in different aspects.

[0179] S62. Local-global abnormal pattern extraction: Read the historical multi-dimensional health indicator set, design a hierarchical abnormal pattern extractor, and identify local and global abnormal patterns by comparing the deviation between the current health indicators and the historical baseline, and output the abnormal pattern features.

[0180] S63. Context-aware abnormal detection: Read the abnormal pattern features, environmental sensor data, and structure usage condition information, design a conditional variational autoencoder, learn the normal structure response distribution under different environments and usage conditions, accurately distinguish fluctuations caused by the environment and real structure damage, and output the damage probability assessment.

[0181] S64, Space-time Anomaly Propagation Analysis: Read the probability evaluation of timing damage, apply the space-time anomaly propagation tracking algorithm, analyze the spatial propagation path and temporal evolution trend of anomalies in the structure, predict the potential risk of damage expansion, and output the anomaly propagation prediction.

[0182] S65, Multi-level Health Status Assessment: Integrate the multi-dimensional health index set, damage probability evaluation, and anomaly propagation prediction. Through a hierarchical evidence reasoning framework, generate a multi-level health assessment report reflecting the health status of the overall structure and key components, and provide a comprehensive assessment of the damage location, type, degree, and development trend.

[0183] In an embodiment of the present application, the local-global anomaly pattern extraction is specifically as follows: Read the historical multi-dimensional health index set, and obtain the standardized health index by standardizing and denoising each index. The specific processing method is: Apply context-aware standardization to each health index I_i: Z_i = (I_i - μ_i(c)) / σ_i(c), where μ_i(c) and σ_i(c) are the expected value and standard deviation under the context c (such as structure type, environmental conditions); Apply wavelet threshold denoising to remove high-frequency random fluctuations while retaining potential anomaly features; Perform time alignment to ensure comparability of different indexes in the time dimension; Process missing values and fill them using an interpolation method based on a physical model; Construct an index correlation matrix R to quantify the mutual relationship between different health indexes.

[0184] Read the standardized health index, and construct a feature pattern library of local structure anomalies through unsupervised pattern recognition to obtain a local anomaly pattern library. The specific processing method is: Use the Density Peak Clustering (DPC) algorithm to identify the typical distribution pattern of local indexes: Calculate the distance matrix D_local of local indexes, and calculate the local density ρ_i = ∑ j exp(-(d_ij / d_c) 2 ) of each sample point, where d_ij is the distance between sample point i and sample point j, d_c is the normalization constant of the distance, and calculate the minimum distance Δ_i from each sample to a sample with higher density; Identify the typical pattern center based on ρ and Δ. Extract the feature vector v_l of each type of anomaly pattern, including: Temporal characteristics: Duration, occurrence frequency, development speed; Spatial characteristics: Influence range, spatial distribution, propagation direction; Index characteristics: Dominant index, index correlation change. Establish an anomaly severity scoring system S_l based on the degree of deviation from the normal state; Construct a local anomaly pattern library L_lib, which contains the identified typical anomaly patterns and their feature vectors.

[0185] Read the standardized health indicators and the local abnormal pattern library, identify the global abnormal patterns involving multiple regions through hierarchical aggregation analysis, and obtain the global abnormal pattern library. The specific processing method is as follows: Construct a regional association graph G_r, where the nodes are structural regions and the edge weights represent the health indicator correlation between regions. Apply the spatio-temporal scan statistic method to identify the spatial aggregation of abnormal indicators: Define a circular or elliptical scan window; Calculate the log-likelihood ratio of the indicator values inside and outside the window; Evaluate the statistical significance through Monte Carlo simulation. Extract the global abnormal feature vector v_g, including: the set of involved regions and their spatial configuration; the spatial distribution of the degree of abnormality; the abnormal propagation sequence between regions. Apply the subgraph matching algorithm to match the identified global patterns with the known damage patterns. Construct the global abnormal pattern library G_lib, which contains typical global abnormal patterns and their features.

[0186] Read the historical standardized health indicators, extract the abnormal patterns of the indicators over time through time series pattern mining, and obtain the time series abnormal pattern library. The specific processing method is as follows: Apply dynamic time warping (DTW) to calculate the similarity matrix S_dtw of the indicator time series; Use density-based sequence clustering to identify typical time evolution patterns: Mutation type: The indicator value suddenly changes significantly; Trend type: The indicator value continuously changes unidirectionally; Periodic type: The indicator value fluctuates periodically; Random fluctuation type: The indicator value fluctuates irregularly. Extract the features of each type of time series pattern, including duration, change rate, periodicity, etc.; Construct the time series abnormal pattern library T_lib, which contains typical time series abnormal patterns and feature descriptions.

[0187] Read the local abnormal pattern library L_lib, the global abnormal pattern library G_lib, and the time series abnormal pattern library T_lib, and fuse and synthesize the abnormal patterns in different dimensions through tensor decomposition to obtain a unified abnormal pattern representation. The specific processing method is as follows: Construct a third-order abnormal pattern tensor A (space, indicator, time) to represent the distribution of abnormalities in three dimensions; Apply Tucker decomposition to extract the core structure of the abnormal pattern: A ≈ G × 1 U_space × 2 U_indicator × 3 U_time, where G is the core tensor and U is the feature matrix of each dimension; Identify the main abnormal pattern combinations based on the core tensor G; Construct an abnormal pattern feature space to cluster similar abnormal patterns in the feature space; Calculate the similarity between the current state and the known abnormal patterns to form an abnormal pattern matching vector.

[0188] Read the unified anomaly pattern representation and the current multi-dimensional health metric set, identify the anomaly pattern of the current state through hierarchical anomaly detection, and obtain the anomaly pattern features. The specific processing method is as follows: Design a three-layer anomaly detection architecture: Point-level detector: Identify anomalies of a single metric at a single location; Region-level detector: Identify the anomaly combination of multiple metrics in a local area; Structure-level detector: Identify the anomaly pattern of multiple global regions. Each layer of the detector uses an ensemble learning method, combining multiple detection algorithms: Distance-based methods: Mahalanobis distance, Local Outlier Factor (LOF); Density-based methods: DBSCAN, OPTICS; Model-based methods: One-class SVM, Gaussian Mixture Model (GMM). Use the majority voting strategy to integrate the results of each detector; The third-order anomaly pattern tensor A realizes anomaly gradient tracking, identifies the propagation path of anomalies from the point level to the structure level, and outputs comprehensive anomaly pattern features, including the type, location, severity, and development trend of the anomalies.

[0189] Read the anomaly pattern features and historical anomaly-damage data, establish the association between the anomaly pattern and potential structural damage through causal inference analysis, and obtain the damage association report. The specific processing method is as follows: Construct an anomaly-damage knowledge graph to represent the association relationship between known anomaly patterns and structural damage types; Apply Bayesian network inference for potential damage: Calculate the conditional probability P(damage type|anomaly pattern); Identify the most likely damage type and its probability; Provide multiple damage hypotheses, considering complex anomaly patterns caused by compound damage; Generate damage hypothesis verification suggestions to guide subsequent targeted detection; Form a complete description of the anomaly pattern features, including the type, spatial distribution, temporal evolution, and potential damage association of the anomaly pattern.

[0190] Context-aware anomaly detection is specifically as follows: Read the anomaly pattern features, environmental sensor data, and structural usage condition information, and obtain the structural context features by extracting and analyzing the environmental and operating factors that affect the structural response through context feature extraction. The specific processing method is as follows: Extract key environmental parameters from the environmental sensor data: Temperature field distribution T(x, y, z, t) and its gradient ▽T; Humidity distribution H(x, y, z, t) and its change rate ΔH / Δt; Wind load characteristics, including velocity v_wind, direction θ_wind, and turbulence intensity I_wind. Extract operating conditions from the structural usage condition information: Load state L(t), including static and dynamic load distributions; Usage mode M(t), describing the functional state of the structure; Human intervention events E(t), such as maintenance activities or temporary changes. Construct a time context, including: Seasonal factors S(t), diurnal cycle D(t), and historical event records H(t). Output a comprehensive structural context feature vector C_f to characterize the current environmental and operating conditions.

[0191] Using the structural context feature C_f and the historical multi-dimensional health indicator set, an anomaly detection model that can consider context is designed and constructed through a deep generative model to obtain a conditional generative model. The specific processing method is as follows: Design a conditional variational autoencoder (CVAE) architecture: Input layer: health indicator x and context feature c; Encoder network: q_φ(z|x, c), which maps the input to the latent space z; Prior network: p_θ(z|c), which generates the prior distribution of the latent variable based on the context c; Decoder network: p_θ(x|z, c), which reconstructs the input from the latent variable and the context. Define the conditional variational lower bound ELBO as the training objective: ELBO(x, c) = E_q_φ(z|x, c)[log p_θ(x|z, c)] - KL[q_φ(z|x, c)||p_θ(z|c)], where KL is the Kullback-Leibler divergence, which measures the difference between two distributions; E_q_φ is the output expectation of the encoder network. Design a context fusion mechanism to fuse context information with health indicator features at each layer of the network: Additive fusion: h = h_x + W_c·c; Multiplicative gating: h = h_x Θ σ(W_c·c); Attention fusion: h = Attention(h_x, c), where h_x is the health indicator feature vector, W_c is the weight matrix of the context feature, Attention is the attention mechanism function, and Θ is the element-wise multiplication. Implement a hierarchical latent variable structure to capture features at different levels of abstraction.

[0192] Read the conditional generative model and historical data, and estimate the normal structure response distribution under different contexts through conditional distribution estimation to obtain a conditional normal distribution model. The specific processing method is as follows: Construct a context space grid to discretize the continuous context space into typical scenarios; For each typical context scenario c_i, learn the conditional distribution p(x|c_i): Use the trained conditional variational autoencoder (CVAE) to sample and generate the normal response under this context, and estimate the conditional mean vector μ(c_i) and covariance matrix Σ(c_i); Construct a context interpolation mechanism to handle the context states between the grids: For any context c, calculate the similarity w_i with the grid points; Estimate the interpolation distribution parameters through weighted average: μ(c) = ∑ i w_i·μ(c_i); Establish a distribution smoothing constraint to ensure smooth changes in the distribution under similar contexts, and output the conditional normal distribution model P, which can generate the corresponding normal response distribution for any context c.

[0193] Read the conditional normal distribution model P and the current structural context feature C_f, calculate the anomaly judgment threshold under the current context through adaptive statistical analysis, and obtain the context-sensitive threshold. The specific processing method is as follows: For the current context C_f, query or interpolate to obtain the conditional distribution parameters μ(C_f) and Σ(C_f); calculate the Mahalanobis distance threshold: MD_thresh = χ 2 _α(d), where d is the feature dimension, α is the significance level, and χ is the chi-square distribution; adjust the α value based on the anomaly detection rate in similar contexts in historical data; consider the uncertainty of the context and calculate the threshold uncertainty interval [MD_thresh_low, MD_thresh_high], where MD_thresh_low is the lower threshold of the Mahalanobis distance and MD_thresh_high is the upper threshold of the Mahalanobis distance; set independent thresholds for key safety indicators to ensure that high-risk anomalies can be detected in a timely manner, and output the context-sensitive threshold set for the current state for anomaly determination.

[0194] Read the current multi-dimensional health indicator set, the conditional normal distribution model P, and the context-sensitive threshold, calculate the anomaly degree of the current state through multi-dimensional anomaly detection, and obtain the anomaly metric. The specific processing method is as follows: Calculate the Mahalanobis distance between the current indicator x and the conditional expectation: MD(x) = sqrt[(x - μ(C_f))T · Σ -1 (C_f) · (x - μ(C_f))]; calculate the single contribution degree of each dimension: C_i = (x_i - μ_i(C_f)) 2 / Σ ii (C_f); apply the Local Outlier Factor (LOF) and consider the change of local density: calculate the k-distance and reachability distance, and estimate the local density ratio. Combine the Mahalanobis distance and LOF to calculate the comprehensive anomaly score A_score, and output the anomaly metric, including the overall anomaly score and the contribution degree of each dimension.

[0195] Read the anomaly metric and the anomaly pattern feature, estimate the probability that the current state represents structural damage through Bayesian inference, and obtain the damage probability assessment. The specific processing method is as follows: Construct a Bayesian network model, and the nodes include: context variables C = {C 1 ,C 2 ,...,C_n}; anomaly indicators A = {A 1 ,A 2 ,...,A_m}; damage states D = {D 1 ,D 2,..., D_k}; Calculate the posterior probability: P(D|A, C) ∝ P(A|D, C) × P(D|C) × P(C), where P(D|A, C) is the posterior probability of the damage state D given the anomaly index A and the context variable C; P(A|D, C) is the probability of the anomaly index A given the damage state D and the context variable C; P(D|C) is the probability of the damage state D given the context variable C; P(C) is the probability of the context variable C; Use the Monte Carlo Markov Chain (MCMC) method to sample the posterior distribution; Calculate the probability of each damage type to generate a damage probability distribution; Apply the evidence theory to process uncertain information and calculate the belief interval and plausibility; Output the damage probability assessment, including the probability of each type of damage and its uncertainty estimate.

[0196] Read the anomaly metrics and historical structural context features, distinguish the normal fluctuations caused by the environment and the anomalies caused by structural damage through spatio-temporal correlation analysis, and obtain the environmental stripping anomaly assessment. The specific processing method is as follows: Construct the anomaly-environment correlation matrix R_ae to quantify the correlation degree between each anomaly index and environmental factors; Apply partial correlation analysis to eliminate the influence of environmental factors: Calculate the partial correlation of the anomaly index after controlling the environmental variables; Identify the environment-sensitive indicators and environment-insensitive indicators; Design an environmental compensation model to eliminate the predictable environmental effects: x_adj = x - f_env(C_f), where f_env is the environmental impact function, x_adj is the health index after environmental compensation, and x is the current health index data; Calculate the anomaly score A_score_adj after environmental compensation; Compare the original anomaly score with the compensated score to quantify the contribution degree of environmental factors; Output the damage probability assessment, including the damage possibility assessment after environmental factor correction.

[0197] According to one aspect of the present application, it further includes step S7. According to the monitoring results and the changes in the structural state, dynamically adjust the sensor weight configuration and data acquisition strategy to achieve the optimal allocation of monitoring resources.

[0198] S71. Information value evaluation: Read the sensor weight matrix, the spatial uncertainty distribution map, and the multi-level health assessment report, calculate the information value index of each sensor and monitoring area, and evaluate its contribution to the structural state estimation and anomaly detection.

[0199] S72. Dynamic allocation of sensing resources: Based on the information value index and anomaly propagation prediction, adopt an adaptive resource allocation algorithm to optimize the sampling frequency, data transmission priority, and computing resource allocation of sensors, and output the monitoring strategy configuration.

[0200] S73. Active learning sampling optimization: Read the spatial uncertainty distribution map and abnormal pattern features, apply the maximum information gain sampling strategy to determine the key locations and time periods that need to be monitored intensively, and dynamically generate an optimized sampling plan.

[0201] S74. Self-evaluation and adjustment of the monitoring system: Regularly evaluate the overall performance of the monitoring system, including the accuracy of state estimation, the sensitivity of anomaly detection, and the efficiency of system resources. Through a closed-loop feedback optimization mechanism, continuously improve the monitoring strategy and algorithm parameters, and output a system performance improvement plan.

[0202] S75. Long-term monitoring knowledge accumulation: With the accumulation of monitoring data, through an incremental learning framework, continuously update and optimize the model parameters and prediction capabilities in each module to form a long-term knowledge base for a specific structure, and improve the adaptability and detection accuracy of the system to the characteristics of the structure.

[0203] The present invention obtains the original data of multi-modal sensors and preprocesses it; constructs a topological perception network representation reflecting the spatial distribution of sensors and the structural connection relationship; extracts multi-temporal and multi-scale features; uses a recursive graph evolution predictor to achieve the mapping from local to global structural state; incorporates structural dynamics knowledge into the neural network model for physical constraint correction; constructs a multi-dimensional health index set and outputs a damage probability assessment. The present invention realizes the accurate mapping from local dense monitoring areas to global structural states under a non-uniformly distributed sensing network, and improves the accuracy and reliability of structural health monitoring.

[0204] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A building dynamic structural health monitoring method, characterized in that: The following steps are involved: Obtain and preprocess the raw data of multimodal sensors to generate sensor weight matrix and preprocessed data set; Based on the preprocessed data set and pre-stored structural geometry information, a physical connection topology map and a topology-aware network representation that reflects the spatial distribution of sensors and structural connection relationships are constructed to obtain a hierarchical topological representation. Extracting predetermined spatiotemporal features from the preprocessed data set to form a comprehensive spatiotemporal feature body; By using hierarchical topological representation and comprehensive spatiotemporal feature bodies, a dynamic heterogeneous graph structure is constructed and a recursive graph evolution predictor is used to map the local densely monitored area to the global structural state, thus obtaining a full-field prediction response. Integrate structural dynamics knowledge into the preconfigured neural network model, perform physical constraint correction on the full-field predicted response, and obtain the physical constraint correction response; Based on the physical constraint correction response, context-aware anomaly detection is performed on the multi-dimensional health indicator set, and the damage probability assessment is output; The multimodal sensor raw data includes raw data from acceleration, strain, displacement, inclination and environmental sensors; The sensor weight matrix is ​​generated by calculating the short-term reliability index and long-term stability index of each sensor; the short-term reliability index is calculated based on the outlier ratio, signal-to-noise ratio and data integrity; the long-term stability index is calculated based on the sensor drift degree and consistency; The steps to obtain the hierarchical topological representation are as follows: Read structural geometry information and material property parameters, construct a structural topology diagram representing the physical connection relationship of the building, and output a physical connection topology diagram; Read the sensor location map and non-uniform density distribution map, divide the sensors into dense monitoring areas and sparse monitoring areas, and output the sensor partition table; the sensor location map is constructed by pre-stored sensor deployment coordinates and structural geometry information, and the non-uniform density distribution map is obtained by calculating the spatial density parameters of each sensor; Read the preprocessed data set and physical connection topology, calculate and evaluate the nonlinear time-varying correlation between sensor data, and form a data association matrix that reflects the flow of information; combine the sensor partition table to construct a sensor association topology with dynamic weights; Read the physical connection topology map and the sensor association topology map, build a dual-stream graph attention network, learn the physical topology and data association features respectively through two parallel branches, and fuse them through the attention mechanism to output a low-dimensional topological embedding vector of the sensor location; Read low-dimensional topological embedding vectors and sensor partition tables, and use structure-sensitive differential pooling operations to build hierarchical topological representations at different levels of abstraction; The steps of forming a comprehensive spatiotemporal feature volume include: Read the preprocessed data set, determine the analysis window size according to the data characteristics of different sensor types and the structural dynamic response characteristics, dynamically balance the time resolution and feature integrity, and output a multi-scale time window sequence; Read the preprocessed data set and multi-scale time window sequence, decompose the signal into frequency bands reflecting different physical processes, extract frequency domain features through phase-sensitive spectrum analysis, and generate a frequency domain feature matrix; Read the pre-processed data set of the densely monitored area, extract the local vibration mode, propagation characteristics and transient response characteristics, and output the local dynamic feature set; Read all preprocessed data sets, combine the stiffness and mass distribution information in the physical connection topology diagram, extract the overall deformation, main modes and global energy distribution characteristics, and output the global response feature set; Read the local dynamic feature set, global response feature set and frequency domain feature matrix, build a multi-level feature pyramid network, fuse features at different spatiotemporal scales through the context weighting mechanism, and form a comprehensive spatiotemporal feature body that expresses the structural response at different scales; the steps to obtain the full-field prediction response include: Based on the hierarchical topological representation and sensor weight matrix, a dynamic heterogeneous graph structure is constructed to reflect the current structural state and sensor reliability. Read the comprehensive spatiotemporal feature bodies of densely monitored areas and sparsely monitored areas, encode them through the regional coupled attention encoder, and obtain the regional coupled feature representation; Combining the dynamic heterogeneous graph structure and regional coupling feature representation, a recursive graph evolution network is used to predict the dynamic response of unmeasured points and obtain the full-field predicted response; Combining the full-field predicted response and the physical connection topology map, the predicted results are calibrated at different spatial resolutions through a layered propagation calibration mechanism to obtain the calibrated full-field response. Based on the calibrated full-field response and sensor weight matrix, a Bayesian deep integration model is used to evaluate the prediction uncertainty and generate a spatial uncertainty distribution map. The steps of obtaining the physical constraint correction response include: Read the full-field response and physical connection topology after calibration, build a neural ordinary differential equation network enhanced by physical knowledge, learn the dynamic state transition law of the structure, and obtain the structural dynamic model parameters; Read the full-field response and structural dynamics model parameters after calibration, verify the physical consistency between different types of sensor data through the cross-modal physical constraint checker, and obtain the physical consistency score; Based on the calibrated full-field response, spatial uncertainty distribution map and physical consistency score, the local physical constraints are propagated to the global through the structure-aware constraint propagation network to obtain the physical constraint correction response; Combining the physical constraint correction response, multimodal sensor raw data and sensor weight matrix, the information fusion weight is dynamically adjusted according to sensor reliability, prediction uncertainty and physical consistency to obtain the final fused full-field state estimate; The steps to output the damage probability assessment include: Read and fuse the full-field state estimation and physical constraint correction response, calculate a multi-dimensional health index set including modal parameter changes, energy distribution anomalies, stiffness degradation index and damping ratio changes, and comprehensively characterize the health status of the structure in different aspects; Read the historical multi-dimensional health indicator set, build a hierarchical abnormal pattern extractor, identify local abnormalities and global abnormal patterns by comparing the deviation between the current health indicators and the historical baseline, and output abnormal pattern features; It reads abnormal pattern features, environmental sensor data, and structural usage information, builds a conditional variational autoencoder, learns the normal structural response distribution under different environments and usage conditions, accurately distinguishes environmentally caused fluctuations from real structural damage, and outputs a damage probability assessment.

2. The method according to claim 1, characterized in that The steps of using the recursive graph evolution network to predict the dynamic response of the unmeasured point and obtain the full-field predicted response include: Based on the dynamic heterogeneous graph structure and regional coupling feature representation, a graph representation that characterizes the physical state of the structure is constructed to obtain a physically enhanced graph state representation. Construct a physical constraint encoding unit to transform the structural dynamics equations into differentiable neural network constraints and generate physical constraint parameter matrices and constraint violation metrics; Using the graph state representation and physical constraint parameter matrix, a recursive graph convolution layer is constructed and multi-step recursive evolution is achieved through the state propagation function and edge feature update function to obtain the state evolution sequence; Based on the state evolution sequence, the path-aware aggregation function is used to estimate the state of the unmeasured points to obtain the estimated value of the unmeasured point state. The state estimates of unmeasured points and constraint violation measures are combined to minimize physical constraint violations through the alternating direction multiplier method, and the full-field predicted response that complies with the physical constraints is obtained.

3. The method according to claim 1, characterized in that The steps of calibrating the mechanism through layered propagation and obtaining the calibrated full-field response include: Read the full-field prediction response and physical connection topology map, build a multi-resolution estimation framework including coarse scale, medium scale and fine scale, and obtain a multi-scale prediction representation; Based on multi-scale prediction representation, prediction calibration is performed alternately from top to bottom and from bottom to top to ensure the consistency of prediction results between different scales and generate inter-scale consistency metrics; Combining the inter-scale consistency measure and multi-scale prediction representation, the reliability of the calibrated prediction is evaluated and the prediction results of each scale are integrated to output the calibrated full-field response.

4. The method according to claim 1, characterized in that The Bayesian deep ensemble model is used to assess the uncertainty of the predictions. The steps to generate the spatial uncertainty distribution map include: Based on the calibrated full-field response and sensor weight matrix, a predetermined number of recursive graph evolution predictors with the same structure but different initializations are constructed; Collect the prediction distributions of all recursive graph evolution predictors for each node and calculate the prediction variance as the epistemic uncertainty score; The average variance of all recursive graph evolution predictors’ predictions is calculated based on the prediction distribution as the aleatory uncertainty score; The geodesic distance from each predicted point to the nearest measured point is calculated based on the predicted distribution and converted into a distance uncertainty score; Based on the pre-stored structural characteristics and monitoring objectives, the cognitive, accidental, and distance uncertainty scores are weighted and fused to obtain an overall uncertainty score; The overall uncertainty score is mapped into the structural space to generate a spatial uncertainty distribution map.

5. The method according to claim 1, characterized in that The steps of constructing a neural ordinary differential equation network enhanced with physical knowledge and obtaining the parameters of the structural dynamics model include: Read the full-field response and physical connection topology after calibration, construct and initialize the parameterized structural dynamics equation, and obtain the initial parameter matrix; Based on the initial parameter matrix, a neural differential equation representation is constructed to obtain the neural dynamics equation; Construct a multi-objective loss function based on the neural dynamics equation to form a physical loss function; Construct a neural network-based adaptive step-size solver to numerically solve the neurodynamic equations and obtain the predicted response trajectory; Based on the predicted response trajectory and the physical loss function, the model parameters are updated through gradient descent optimization with physical constraints to obtain the optimized parameter vector; The equivalent stiffness distribution, equivalent elastic modulus and modal parameters are extracted from the optimized parameter vector to generate the structural dynamics model parameters.

6. The method according to claim 1, characterized in that The steps of propagating local physical constraints to the global physical constraints through the structure-aware constraint propagation network and obtaining the physical constraint correction response include: The full-field response, spatial uncertainty distribution map and physical consistency score after calibration are read, and the structure is divided into three sub-domains of micro, meso and macro scales through hierarchical domain decomposition to obtain a multi-scale structural domain; Based on the multi-scale structural domain and structural dynamics model parameters, the local constraint equations of each subdomain are established; Based on all local constraint equations and the calibrated full-field response, the deviation degree between the response and the physical constraint is calculated to obtain the constraint violation degree graph; Read the constraint violation graph and multi-scale structure domain, build a structure-aware graph network, and obtain the constraint propagation network; Based on the constraint propagation network and constraint violation graph, a hierarchical coordination mechanism is constructed to obtain constraint coordination rules. The physical constraint correction response is obtained by minimizing the physical violation through the constraint gradient flow algorithm using the calibrated full-field response, constraint violation graph and constraint coordination rules.

7. The method according to claim 1, characterized in that The steps to construct a topology-aware network representation that reflects the spatial distribution and structural connectivity of sensors include: Read the preprocessed data set, construct a multi-scale time window, and form a multi-scale window data set; Based on the multi-scale window data set, the nonlinear mutual information between sensor data is calculated, and the sensor pair mutual information matrix is ​​generated through adaptive histogram binning and directional mutual information estimation; Perform scale-weighted fusion on the mutual information matrices of sensor pairs at different time scales to obtain a comprehensive mutual information matrix; Compare the current and historical comprehensive mutual information matrices, identify sensor pairs with significant changes through correlation change detection algorithm, and generate correlation change events; By combining the comprehensive mutual information matrix, correlation change events and physical connection topology map, a sensor association topology map reflecting the information flow, namely, the topology-aware network representation, is constructed.

8. The method according to claim 7, characterized in that Before obtaining the hierarchical topological representation, a low-dimensional topological embedding vector is constructed, including: Read the physical connection topology map and the sensor association topology map, construct the physical map adjacency matrix and the association map adjacency matrix, and generate a normalized adjacency matrix; The normalized adjacency matrix is ​​input into the dual-stream graph convolutional network, and the structural topology and data association information are processed through the physical flow and data flow respectively to obtain the physical flow node representation and data flow node representation; a cross-flow graph attention mechanism is constructed to calculate the node correlation between information flows and perform weighted aggregation to obtain the attention-weighted node representation; Using attention-weighted node representation, we construct a structure-aware contrastive learning objective to minimize the embedding distance between physically and topologically adjacent nodes and maximize the distance between distant nodes to generate an embedded representation. Based on the embedding representation, the physical flow node representation, data flow node representation and attention weighted node representation are combined, and the fusion and dimensionality reduction are performed through a multi-layer perceptron to output a low-dimensional topological embedding vector.

Images