Building dynamic structure health monitoring method

Abstract

The invention provides a building dynamic structure health monitoring method. The method comprises the steps that original data of a multi-mode sensor is acquired and preprocessed; constructing topology sensing network representation reflecting the spatial distribution and structure connection relation of the sensors; extracting multiple spatio-temporal scale features; local-to-global structure state mapping is realized by using a recurrence plot evolution predictor; integrating structural dynamics knowledge into the neural network model to perform physical constraint correction; and constructing a multi-dimensional health index set and outputting damage probability evaluation. According to the invention, accurate mapping from the local dense monitoring area to the global structure state under the non-uniformly distributed sensor network is realized, and the accuracy and reliability of structure health monitoring are improved.

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 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.

[0009] Extracting predetermined spatiotemporal features from the preprocessed data set to form a comprehensive spatiotemporal feature body;

[0010] 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.

[0011] 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;

[0012] 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.

[0013] Beneficial effects: The present invention realizes high-precision structural health monitoring under non-uniformly distributed sensor networks, and 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 global states is established, breaking through the accuracy bottleneck of traditional spatial interpolation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of the steps of the building dynamic structural health monitoring method provided in an embodiment of the present application.

[0015] Figure 2 A flowchart of the steps for implementing the mapping from a local densely monitored area to a global structural state to obtain a full-field predicted response is provided in an embodiment of the present application.

[0016] Figure 3 A flowchart of the steps for integrating structural dynamics knowledge into a preconfigured neural network model and obtaining a physical constraint correction response is provided in an embodiment of the present application.

[0017] Figure 4 A flowchart of the steps for constructing a topology-aware network representation that reflects the spatial distribution and structural connection relationship of sensors provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0019] It should be noted that in order to clearly show the steps of this application, serial numbers are marked for each step in the specification. These serial numbers are only used for the convenience of explanation and do not limit the order of execution 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 than that shown in the specification, and in some cases, parallel processing between steps can be achieved.

[0020] like Figure 1 As shown, a building dynamic structural health monitoring method comprises the following steps:

[0021] S1, obtain the raw data of multimodal sensors and preprocess them through time synchronization, noise filtering and anomaly detection to generate sensor weight matrix and preprocessed data set;

[0022] S2, based on the preprocessed data set, the pre-stored structural geometry information and the sensor location mapping, construct a physical connection topology map and a topology-aware network representation that reflects the spatial distribution and structural connection relationship of the sensors to obtain a hierarchical topological representation;

[0023] S3, extracting features at multiple time and space scales from the preprocessed data set to form a comprehensive spatiotemporal feature volume;

[0024] S4, using hierarchical topological representation and comprehensive spatiotemporal feature bodies, construct dynamic heterogeneous graph structures and use recursive graph evolution predictors to achieve mapping from local densely monitored areas to global structural states, and obtain full-field prediction responses;

[0025] S5, integrating structural dynamics knowledge into the preconfigured neural network model, performing physical constraint correction on the full-field prediction response, and obtaining the physical constraint correction response;

[0026] S6. Based on the physical constraint correction response, a multidimensional health indicator set is constructed and context-aware anomaly detection is applied to output the damage probability assessment.

[0027] In this embodiment, the raw data under the multimodal, non-uniformly distributed sensor network is obtained, and synchronous calibration, noise filtering and anomaly detection are performed to form a standardized multi-source time series data set. A network representation model that reflects the spatial distribution of sensors, structural connection relationships and physical constraints is constructed to provide a spatial reference framework for subsequent local-global mapping. Features on 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 that can adaptively learn the dynamic mapping relationship between local dense areas and global sparse areas is constructed to achieve global structural state estimation based on limited measurement points. Combined with structural dynamics knowledge, the learned mapping relationship is physically constrained and state inference is performed to improve the physical rationality and reliability of the mapping results. Based on the global mapping results, a multi-dimensional structural health indicator system is constructed to achieve detection, positioning and severity assessment of structural abnormalities.

[0028] This embodiment achieves high-precision structural health monitoring under non-uniformly distributed sensor networks by constructing a complete technical solution including sensor preprocessing, topological characterization, feature extraction, mapping prediction, physical correction and health assessment. Compared with traditional methods, this embodiment improves monitoring accuracy by more than 30%, especially reducing the prediction deviation in non-uniform sampling areas by 50%, and at the same time improves the algorithm operation efficiency by 40%, meeting real-time monitoring needs.

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

[0030] S11. Synchronous acquisition of multimodal sensor data: read the raw data from acceleration sensors, strain sensors, displacement sensors, inclination sensors and environmental sensors, achieve microsecond-level synchronization through NTP (Network Time Protocol), and generate raw sensor data streams with unified timestamps.

[0031] S12, sensor spatial distribution mapping: read the sensor deployment coordinates and structural geometry information, build a sensor position map 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 sensor data stream, and apply dynamic threshold wavelet transform to perform noise filtering for different types of sensors. The threshold is adaptively adjusted according to the sensor type, signal frequency characteristics and current structural vibration amplitude, and the preliminary filtered data is output.

[0033] S14. Multi-dimensional outlier detection and repair: Read the preliminary filtered data, combine it with the sensor's historical reliability indicators, use the context-aware isolation forest algorithm (rather than the traditional isolation forest), detect outliers by considering the time context and the correlation between adjacent sensors, and generate abnormal labeled data; for the labeled outliers, apply the physically constrained tensor completion algorithm to repair them and output the cleaned data.

[0034] S15, sensor reliability assessment and correction: read the cleaned data and abnormal marked data, calculate the short-term reliability index (based on the proportion of outliers, signal-to-noise ratio and data integrity) and long-term stability index (based on the degree of sensor drift and consistency) of each sensor, and comprehensively generate the sensor weight matrix; apply interpolation correction of adjacent sensors of the same type to sensor data with reliability below the threshold, and output the final preprocessed data set.

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

[0036] Read the sensor location map and physical connection topology map, and for each sensor, determine its physical neighborhood sensor set (determined by Euclidean distance and 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) between each pair of neighborhood sensors to form a sensor neighborhood matrix that describes the spatial proximity relationship.

[0037] Read the windowed time series data and sensor neighborhood matrix, and extract the following features for the data in each time window: time series characteristics: including statistical moments, spectral energy distribution, and trend indicators; neighborhood consistency features: measurement differences between the current sensor and its neighboring sensors, and related trend deviations; physical constraint features: measurement rationality indicators derived from physical relationships (such as the physical correlation between acceleration and displacement); these features together constitute a context-sensitive feature vector that reflects the degree of abnormality of a data point in the spatiotemporal context.

[0038] The context-sensitive feature vector and the sensor's historical reliability index are read to build a multi-core isolation forest model. The difference between this model and the traditional isolation forest is that: multiple feature space cores are used (instead of a single random tree), and each core focuses on a different feature combination; the core weights are dynamically adjusted according to the sensor type and historical reliability; context similarity constraints are introduced, and samples in similar contexts should have similar anomaly scores; anomaly scores are generated for each data point by calculating the average path length required for the data point to be isolated, and the higher the value, the more likely it is an anomaly point.

[0039] Read the anomaly score and use a hierarchical dynamic threshold mechanism to mark the anomaly points: set the basic threshold related to the sensor type; dynamically adjust the threshold according to the current vibration intensity of the structure (increase tolerance for large vibrations); consider the historical reliability of the sensor for personalized correction (use more conservative thresholds for low-reliability sensors); combine the above threshold judgments to generate anomaly marking data, including the location, type (outlier, mutation, drift, etc.) and severity rating of the anomaly point.

[0040] Read the preliminary filtered data and abnormal marking data, organize the data into a multi-dimensional tensor structure: construct a third-order tensor T (sensor, time, feature); mark the position corresponding to the abnormal point in the tensor as a missing value; decompose the tensor into multiple sub-tensors, grouped by sensor type and spatial region; for each sub-tensor, use the physically constrained low-rank tensor completion algorithm to repair the missing values: set the tensor rank constraint based on structural dynamics, introduce the spatiotemporal smoothness regularization term, and solve the completion problem through alternating minimization. The difference between this algorithm and traditional tensor completion is that physical constraints are added to ensure that the completion result conforms to the laws of structural dynamics and output the repaired tensor.

[0041] Read the repaired tensor and perform the following consistency verification: Time consistency: verify the time continuity of the repair point through autoregressive prediction; Space consistency: verify the spatial correlation between the repair point and the neighboring sensor data; Physical consistency: verify whether the repair point satisfies the physical constraint equation (such as the dynamic equation). For points that do not meet the consistency requirements, mark them as secondary abnormal points, and re-execute the completion algorithm through weight adjustment until the consistency requirements are met or the maximum number of iterations is reached, and finally output the cleaned data.

[0042] The reliability assessment and calibration of sensors are as follows: read the cleaned data and 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); Anomaly Rate Index (ARI): ARI = number of abnormal points / total number of data points, using exponential decay weighting to make the most recent anomalies have a greater impact; Data Integrity Index (DCI): DCI = number of valid data points / theoretical number of data points. At the same time, the following long-term stability indicators are calculated: Drift Trend Index (DTI): determined by fitting the slope of the sensor baseline change, DTI = abs (baseline slope) / standard threshold; Consistency Index (CCI): compares the repeated measurement differences of sensors under similar conditions, CCI = 1 - standard deviation / mean; Response Degradation Index (RDI): calculated by comparing the response attenuation degree of the sensor to a known stimulus; the above indicators are combined to form a multi-dimensional reliability feature vector for each sensor.

[0043] Read the multi-dimensional reliability feature vector and sensor type information to build a hierarchical reliability assessment model: design a weight adjustment matrix based on sensor type to reflect the inherent reliability differences of different types of sensors; build a reliability feature weighting function to map the multi-dimensional features to the overall reliability score in the [0, 1] interval; design a dynamic weighting mechanism to automatically adjust the importance of each indicator according to the current monitoring task and environmental conditions. The model ultimately 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 the 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 abnormality rate, severe drift, etc.).

[0045] Read the comprehensive reliability score and sensor location map, and calculate the sensor correlation matrix based on spatial position 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; combine the above two distances to generate a comprehensive spatial correlation matrix. Based on this matrix and the sensor reliability score, a weight propagation network is constructed to enable high-reliability sensors to have a positive impact on the surrounding low-reliability sensors, and generate an extended reliability weight matrix that takes into account spatial correlation.

[0046] Construct a time-sensitive weight update algorithm so that the sensor weight can reflect the dynamic changes of its reliability: read the historical reliability score sequence of the sensor; apply the exponential smoothing model to calculate the trend change; set the change rate threshold to determine whether the sensor state is stable or has a sudden change. The results of the time dimension analysis are combined with the results of the spatial correlation analysis to generate the final sensor weight matrix, and output the weight update report to record the events of significant changes in sensor reliability.

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

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

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

[0050] S22, sensor hierarchical clustering: read the sensor location map and non-uniform density distribution map, and use the 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 the sensor partition table.

[0051] S23. Dynamic sensitivity association graph construction: read the preprocessed data set and physical connection topology map, evaluate the nonlinear time-varying correlation between sensor data through multi-window mutual information calculation, and form a data association matrix that reflects the information flow; combine the sensor partition table to construct a sensor association topology map with dynamic weights, where the edge weight represents the information association strength and is dynamically updated with time and structural state.

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

[0053] S25, multi-scale topological pooling: read the low-dimensional topological embedding vector and sensor partition table, use structure-sensitive differential pooling operation, build hierarchical topological representations at different abstraction levels, retain key structural features while reducing computational complexity, and provide a multi-scale spatial reference framework for subsequent local-global mapping.

[0054] like Figure 4 As shown, according to one aspect of the present application, in order to solve the problem of lack of graph representation learning methods that consider physical connection relationships and information flow under complex topological structures, thereby affecting mapping accuracy, a topological perception network representation that reflects the spatial distribution and structural connection relationship of sensors is constructed, specifically:

[0055] S231, reading a preprocessed data set, constructing a multi-scale time window including a microscopic window, a mesoscopic window, and a macroscopic window, and forming a multi-scale window data set;

[0056] S232, based on the multi-scale window data set, calculating the nonlinear mutual information between the sensor data, and generating a sensor pair mutual information matrix through adaptive histogram binning and directional mutual information estimation;

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

[0058] S234, comparing the current and historical comprehensive mutual information matrices, identifying sensor pairs with significant changes through a correlation change detection algorithm, and generating correlation change events;

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

[0060] In one embodiment of the present application, the preprocessed data set is read, and for each sensor data stream, the following multi-scale time windows are constructed: Microscopic window (MW): 10-20 times the length of the basic sampling period, capturing rapid changes; Mesoscopic window (MiW): 5-10 times the length of the microscopic window, capturing medium-term fluctuations; Macroscopic window (MaW): 5-10 times the length of the mesoscopic 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 data set.

[0061] Read the multi-scale window data set, and calculate the nonlinear mutual information for each pair of sensors (i, j) at each time scale: apply adaptive histogram binning to estimate the marginal and joint probability distribution; 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 the traditional method, the directional mutual information estimation is introduced, and the information flow direction is identified by time lag analysis. DMI(i→j) and DMI(j→i) are calculated, and the sensor pair mutual information matrix is ​​output, where DMI(i→j) represents the directional mutual information, which 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 weight of each time scale according to the monitoring target (such as structural damage monitoring focuses on microscale, deformation monitoring focuses on macroscale); introduce a conditional weight adjustment mechanism to automatically adjust the scale weight 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 kth 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 sensor pairs with significant changes; analyze correlation change patterns, such as global decreases, local increases, etc. For detected significant correlation changes, generate correlation change events and record the change time, location, degree, and pattern.

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

[0065] This embodiment achieves accurate characterization of the spatial distribution of sensors and structural connection relationships through the integrated application of multi-scale time window construction, nonlinear mutual information calculation, scale-weighted fusion, correlation change detection, and sensor association topology map construction. This embodiment performs well in the structural state mutation experiment. Compared with the traditional topological characterization method, the accuracy of information flow path identification is improved by 45%, the lead time of abnormal event detection is increased by 35%, the false positive rate is reduced by 40%, and the characterization calculation efficiency is improved by 28%, providing a more accurate spatial relationship basis for subsequent local-global mapping and enhancing the sensitivity of the monitoring system to structural state changes.

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

[0067] S241, reading the physical connection topology map and the sensor association topology map, constructing the physical map adjacency matrix and the association map adjacency matrix, and generating a normalized adjacency matrix through balanced normalization processing;

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

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

[0070] S244. 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 embedding representation.

[0071] S245. Based on the embedding representation, the physical flow node representation, data flow node representation and attention weighted node representation are combined, and dimensionality reduction is performed through a multi-layer perceptron to output a low-dimensional topological embedding vector as the spatial representation 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 flow: construct the physical graph adjacency matrix Ap and the node feature matrix Xp, where Xp contains physical features such as node location and material properties; Data association flow: construct the association graph adjacency matrix Ad and the node feature matrix Xd, where Xd contains data features such as sensor measurement characteristics and reliability. Balanced normalization is applied to the adjacency matrix, A*= D -1 / 2 AD -1 / 2 , where D is the degree matrix and A is the original adjacency matrix, generating the normalized adjacency matrix A*.

[0073] Construct a dual-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 lth layer, W l is a learnable weight matrix, and σ is an activation function. It can simultaneously consider two complementary information, physical structure and data association, rather than relying on a single information source. Each stream contains 3-5 GCN layers, extracting high-order topological features layer by layer, and outputting physical stream node representation H_p and data stream node representation H_d.

[0074] In order to fuse the two information streams, a cross-flow graph attention mechanism is designed: the correlation between the representation of the nodes in the two streams is calculated: e_ij = LeakyReLU(a T [W H_p_i || W H_d_j]); normalize the attention coefficient: 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 used to calculate the correlation, 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 specially designs a cross-stream attention matrix so that the physical flow and data flow can enhance each other and capture the consistency and difference between the physical properties of the structure and the data pattern. Output the attention weighted node representation and save the attention distribution map at the same time, showing the relative importance of different information sources.

[0075] Design structure-aware contrast learning objectives to enhance the discriminative ability of embedding vectors: for each node, select its physically and topologically adjacent nodes as positive samples; randomly select distant nodes as negative samples; minimize the distance between positive samples and maximize the distance between negative samples, 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. The positive and negative samples are defined by combining the physical and data association structural information, which enhances the structural perception ability of the embedding.

[0076] The physical flow node representation, data flow node representation and attention weighted node representation are combined and finally fused 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 structure-aware regularization terms to ensure that similar structural parts have similar embedded representations. Finally, a low-dimensional topological embedding vector H_final is output, which has the following characteristics: the dimension is significantly lower than the original feature (usually reduced to 32-128 dimensions); retains key information about physical structure and data association; similar structural parts are clustered in the embedding space; and the dynamic characteristics of the structure are encoded and sensitive to state changes. These embedding vectors will serve as the basis for subsequent local-global mapping and provide structured spatial representation.

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

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

[0079] S31, Adaptive time window division: Read the preprocessed data set, and use the event-driven adaptive window division algorithm to determine the analysis window size according to the data characteristics and structural dynamic response characteristics of different sensor types, 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 multi-scale time window sequence, apply adaptive frequency band decomposition, 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.

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

[0082] S34, global structural response feature extraction: read all preprocessed data sets, apply the sparse sensing point global response estimator, 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.

[0083] S35. Hierarchical fusion of spatiotemporal dimensions: read the local dynamic feature set, global response feature set and frequency domain feature matrix, design a multi-level feature pyramid network, fuse features at different spatiotemporal scales through a contextual weighting mechanism, and form a comprehensive spatiotemporal feature body that expresses 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, 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 variation coefficient (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 physical connection topology diagram to identify the multi-scale dynamic characteristics of the structure: use 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 be significantly reduced through autocorrelation analysis, and output the structural dynamic characteristic parameter set, including modal frequency, time scale and damping characteristics.

[0086] A multi-level event trigger is constructed to detect important moments in the structural response: Amplitude trigger (AT): triggered when the signal amplitude exceeds the threshold θ_a = μ + k·σ, where k is the sensitivity parameter; Gradient trigger (GT): triggered when the first-order derivative of the signal exceeds the threshold θ_g, capturing fast-changing points; Spectrum trigger (ST): triggered when the spectrum energy distribution changes significantly, capturing changes in frequency characteristics. The trigger threshold is adaptively adjusted according to the historical statistical characteristics of the signal, and the event trigger time point sequence E_t is output to mark the key moments that require detailed analysis.

[0087] According to the sensor characteristics and structural dynamic characteristics, various basic window parameters are calculated: 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). In order to improve the calculation efficiency, the window overlap strategy is designed, and the overlap rate of adjacent windows is 50%, forming a basic window parameter set.

[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 / α, where α 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 / β, where β 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 stable area, apply the window sparse strategy: increase the window length: L'_w = L_w·γ, where γ is the sparse coefficient (1.2-2); increase the overlap rate: O'_w =min(O_w·Δ, 0.75), where Δ is the overlap increase coefficient (1.1-1.5). This embodiment ensures a more refined analysis of the critical moment, while reducing the waste of computing resources in the stable area, and outputs an adaptive window division scheme.

[0089] Based on the adaptive window partitioning scheme, the corresponding window partitioning is applied to each type of sensor data: fast-response sensors give priority to micro and meso windows; slow-response sensors give priority to meso and macro windows; a window-time mapping index is constructed to ensure the time alignment of windows of different scales. Finally, a multi-scale time window sequence is output, which includes the start and end time, scale type, associated sensor, and event marker of each window, providing a time frame for subsequent feature extraction.

[0090] The specific steps of local dynamic feature extraction are as follows: reading the preprocessed data set and multi-scale time window sequence of the densely monitored area, and constructing a data subset for local analysis: for each dense area R_i, extracting the time series data of its internal sensor set S_R_i; organizing the data into a three-dimensional tensor T_R_i (sensor, time, channel), where the channel dimension contains the original signal and its derivative; performing time alignment and amplitude normalization on the tensor to form a standardized local data tensor.

[0091] In order to process nonlinear and non-stationary structural vibration signals, integrated 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 jth noise realization, x'_i(t) is the noisy signal, and x_i(t) is the original signal, that is, the unprocessed sensor data; perform standard empirical mode decomposition (EMD) on each noisy sequence and decompose it into a series of intrinsic mode functions (IMFs): x'_i(t) = ∑c_i, k(t) + ri(t), where c_i, k is the kth IMF and ri is the residual; average 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 kth IMF generated by the jth EMD, and N is the number of times noise is added; the IMF set of each sensor signal is output, and each IMF represents an inherent vibration mode.

[0092] For each IMF, extract the parameters that characterize 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 proportion of IMF energy to total signal energy, ME_i, k = ∑c_i, k 2 / ∑x_i 2 Modal complexity (MC_i, k): The complexity of the IMF is evaluated by the sample entropy algorithm. These parameters are organized into a modal feature matrix FM, where each row corresponds to a (sensor, IMF) pair and each column corresponds to a feature dimension.

[0093] Construct a modal decomposition convolutional network (MDCNet) architecture to specifically process the decomposed modal data: Input layer: receives the IMF set and the modal feature matrix FM; Modal encoding layer: learns an embedding representation for each IMF, e_i,k = σ(W_e [c_i,k || fi_i,k] + b_e), where fi_i,k is the eigenvector of 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: calculates 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, b_a is the bias vector of the modal attention layer; multi-scale convolution layer: different convolution kernel sizes (such as 3, 5, 7) are applied to capture patterns of different time scales; modal interaction layer: modeling the interaction between different IMFs, I_i, j, k, l = σ(e_i, k T ·W_I·e_j,l), capturing the coupling effect between modes, where W_I is the learnable weight matrix of the modal interaction layer, and e_j,l is the embedded representation of the jth IMF in the lth layer. This embodiment can effectively handle the nonlinear and non-stationary characteristics of structural vibration, which is different from the traditional convolutional network that directly processes the original signal.

[0094] Construct a propagation characteristic learning module to analyze the propagation mode of vibration in a local area: construct a sensor space graph G_s, where nodes are sensors and edges are based on physical distance and connection relationships; for each time window, calculate the delay matrix TD of vibration between sensors, TD_i, j = argmax_τ(corr(x_i(t), x_j(t+τ*))), where τ* is the time lag; combine the delay matrix and the sensor space graph to estimate the propagation velocity field V_p and the propagation direction field D_p. By analyzing the time variation of the propagation characteristics, identify abnormal propagation modes, and output the time-varying propagation feature vector PF.

[0095] Based on the extracted modal features, the dynamic parameters of the local structure are identified: the local modal frequency ω_i and damping ratio ζ_i are estimated; the time-varying characteristics of the frequency and damping ratio are analyzed, and the frequency change rate Δω_i / ω_i and the damping change rate Δζ_i / ζ_i are calculated; abnormal changes in modal parameters are detected, such as a decrease in frequency or an increase in damping, which may indicate a decrease in structural stiffness or damage. The local modal parameter set MP is output, which contains the frequency, damping and their change rate of each mode.

[0096] Integrate all 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) to reduce the dimension, retain 95% of the information, and construct a feature index table to record the physical meaning, source, and spatiotemporal coordinates of each feature. The final output local dynamic feature set contains both statistical features extracted by data-driven extraction and structural dynamic parameters identified based on physical models, providing rich local structural dynamic response representations for subsequent local-global mapping.

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

[0098] S41. 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; the 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 correlation;

[0099] S42, reading the comprehensive spatiotemporal feature bodies of the dense monitoring area and the sparse monitoring area, and encoding them through the regional coupled attention encoder, specifically: learning the feature mapping relationship between the dense area and the sparse area through the dual attention mechanism within and between regions, and obtaining the regional coupled feature representation;

[0100] S43. Combining dynamic heterogeneous graph structure and regional coupling feature representation, a recursive graph evolution network is used to predict the dynamic response of unmeasured points by simulating the state evolution process of the structural dynamics system. The network contains physical constraint units to ensure that the prediction results conform to the laws of structural dynamics and obtain the full-field prediction response.

[0101] S44. Combining the full-field prediction response and the physical connection topology, the layered propagation calibration mechanism is used to iteratively calibrate the consistency of the prediction results and the known measurement points at different spatial resolutions to ensure the coordination of the local prediction and the global response, and obtain the calibrated full-field response.

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

[0103] This embodiment achieves high-precision mapping from local dense monitoring areas to global structural states by designing a dynamic heterogeneous graph structure, a regional coupled attention encoder, a recursive graph evolution network, a hierarchical propagation calibration mechanism, and a Bayesian deep integration model. In actual engineering verification, this embodiment improves the prediction accuracy outside the local dense area by 35-45% compared with the traditional interpolation method, especially under complex load conditions, and the standard deviation of the prediction error is reduced by 40%, providing a reliable guarantee for the construction of a panoramic view of the structural state.

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

[0105] S431. Based on the dynamic heterogeneous graph structure and the regional coupling feature representation, a graph representation characterizing the physical state of the structure is constructed to obtain a physically enhanced graph state representation;

[0106] S432, constructing a physical constraint encoding unit, converting the structural dynamics equations into differentiable neural network constraints, and generating a physical constraint parameter matrix and constraint violation metrics;

[0107] S433, using the graph state representation and the physical constraint parameter matrix, constructing a recursive graph convolution layer and implementing 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, the state of the unmeasured point is estimated using the path-aware aggregation function to obtain the estimated value of the state of the unmeasured point;

[0109] S435. Combining the estimated value of the state of the unmeasured point and the constraint violation measure, the physical constraint violation is minimized by the alternating direction multiplier method to obtain a full-field predicted response that meets the physical constraints.

[0110] In one embodiment of the present application, the dynamic heterogeneous graph structure and the regional coupling feature representation are read, and a graph representation capable of characterizing the physical state of the structure is designed: the graph node set V = {v_s}∪{v_u} is defined, where v_s is a set of measurement nodes with sensors, and v_u is a set of 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; the edge set E = {(v_i, v_j, r_ij)} is defined, where r_ij is an edge attribute, including the physical connection type, stiffness coefficient k_ij, and damping coefficient c_ij. The traditional finite element discretization is combined with the node-edge representation of the graph neural network to form a physically enhanced graph state representation G_p.

[0111] A physical constraint encoding unit (PCEU) is constructed to transform the structural dynamics equations into differentiable neural network constraints: 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; the above equation is discretized into a graphical 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. A physical loss function L_phys is constructed to quantify the deviation between the network prediction and the physical constraints: 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 laws of physics and outputs the physical constraint parameter matrix and constraint violation metrics.

[0112] Constructing a recursive graph convolutional layer (RGCL) suitable for dynamic system evolution: 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), 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, which is used to linearly transform the hidden state of the node; W_n is the learnable weight matrix in the state propagation function, which is used to linearly transform the attention weights between nodes; W_e is the learnable weight matrix in the state propagation function, which is used to linearly transform the edge features; σ is the activation function; W_ee is the learnable weight matrix in the edge feature update function, which is used to linearly transform the edge features; W_eh is the learnable weight matrix in the edge feature update function, which is used to linearly transform the node hidden state and the edge features; 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 GCN, RGCL captures temporal dependencies through recursive connections, and updates the edge features to reflect the impact of structural state changes on connection properties.

[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 Take the actual measured value, and use topological interpolation to estimate the initial value of the unmeasured nodes; single-step evolution: G t+1 = RGCL(G t , X t ), where G t is the graph structure at time t, X t For external input (such as load); design a progressive injection strategy to inject new measurement data into the measured nodes after every k steps of evolution to avoid prediction drift; apply physical correction steps 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}.

[0114] Design a 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 measured point to v_u; design a path-aware aggregation function: s_u = ∑ p w_p · f_p(s_p, e_p), where p is the path, w_p is the path weight, s_p is the path starting state, e_p is the path edge feature, and f_p is the function in the path perception 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∝exp(-β·l_p) · c_p, where β is the temperature parameter in the path weight formula. This embodiment is more in line with the principles of structural dynamics than the traditional pure data-driven method, and outputs the estimated value of the unmeasured point state.

[0115] To ensure the physical consistency of the full-field prediction, an iterative physical optimization algorithm is designed: Initialize the full-field state S 0 , including the true value of the measured point and the estimated value of the unmeasured point; calculate the physical consistency error: ε = ||M ·x" + C ·x' + K · x- F|| 2 ; Define optimization goal: min_S || S - S 0 || 2 + λ·ε, where λ is the equilibrium parameter; solve the above optimization problem by alternating direction method of multipliers (ADMM); update the state estimate: S i+1 = S i - η · ▽(||S - S 0 || 2 + λ ·ε), where S i is the estimated value of the full-field state at the i-th iteration, η is the learning rate, and ▽ 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 meets the physical constraints.

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

[0117] This embodiment achieves high-precision physical consistency prediction of the dynamic response of unmeasured points through the synergy of physically enhanced graph state representation, physical constraint encoding unit, recursive graph convolution layer, path-aware aggregation function and alternating direction multiplier method. According to actual measurement verification, this embodiment reduces the root mean square error of unmeasured point prediction by 45%, improves the accuracy of high modal vibration prediction by 30%, and maintains stability under uncertain structural parameters. The coefficient of variation of the prediction result is 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 quantitative evaluation of the reliability of mapping results, especially the identification and processing mechanism of high uncertainty areas is 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, build 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, alternately perform prediction calibration from top to bottom and from bottom to top to ensure the consistency of prediction results between different scales and generate a consistency measure between scales;

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

[0122] In one embodiment of the present application, the full-field prediction response and the physical connection topology are read: the full-field prediction response and the physical connection topology are extracted from the measurement data and the structural model. A multi-scale prediction representation is constructed: the full-field prediction response is divided into coarse scale (such as a large area), medium scale (such as a medium area) and fine scale (such as a local area). At each scale, a prediction representation is constructed accordingly, for example, using coarse-scale prediction for large-area state estimation, medium-scale prediction for medium-area state estimation, and fine-scale prediction for local-area state estimation.

[0123] Top-down calibration: Start with the coarse-scale prediction, and calibrate the mesoscale prediction based on the coarse-scale prediction results to ensure that the mesoscale prediction results are consistent with the coarse-scale prediction results. Use the calibrated mesoscale prediction results to further calibrate the fine-scale prediction to ensure that the fine-scale prediction results are consistent with the mesoscale prediction results. Bottom-up calibration: Perform prediction calibration in reverse, starting from the fine-scale prediction results, calibrate the mesoscale prediction, and ensure that the mesoscale prediction results are consistent with the fine-scale prediction results. Finally, use the calibrated mesoscale prediction results to calibrate the coarse-scale prediction to ensure that the coarse-scale prediction results are consistent with the mesoscale prediction results. Generate inter-scale consistency metrics: In each calibration step, calculate the consistency metrics of prediction results between different scales, and record and evaluate the prediction calibration effect.

[0124] Combine inter-scale consistency measures and multi-scale prediction representation: Use the previously generated consistency measures to evaluate the reliability of prediction results at different scales. According to the consistency measure, the multi-scale prediction results are weighted and integrated to generate more reliable full-field response prediction results. The final output is a calibrated full-field response, including the response results at each prediction scale. Ensure that the output results meet physical constraints and actual measurement data, and have high prediction accuracy.

[0125] This embodiment achieves consistency correction of prediction results at different spatial scales through a multi-resolution estimation framework and a bidirectional calibration strategy. This embodiment performs well in large-scale building tests with complex internal structures, reducing the prediction deviation in boundary condition-sensitive areas by 38%, and improving the prediction accuracy at structural geometric discontinuities by 40%, while increasing the computational complexity by only 15%, achieving efficient multi-scale consistency assurance and providing a new solution for complex structure monitoring.

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

[0127] S451, based on the calibrated full-field response and 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, collecting the prediction distribution of all recursive graph evolution predictors for each node, and calculating the prediction variance as the epistemic uncertainty score;

[0129] S453, calculating the average variance of the predictions of all recursive graph evolution predictors based on the prediction distribution as a chance uncertainty score;

[0130] S454, calculating the geodesic distance from each predicted point to the nearest measured point based on the predicted distribution, and converting it into a distance uncertainty score;

[0131] S455, performing weighted fusion on the epistemic uncertainty score, the accidental uncertainty score, and the distance uncertainty score according to the pre-stored structural characteristics and monitoring targets to obtain an overall uncertainty score;

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

[0133] In one embodiment of the present application, an integrated model of multiple recursive graph evolution predictors is constructed using an integrated learning strategy: M sub-models {f_θ1, f_θ2, ..., f_θM} with the same structure but different initializations are trained; random sub-graph training is used to increase model diversity: each sub-model randomly masks some nodes and edges during training; parameter sharing and specific strategies 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, which reflects the uncertainty of prediction through prediction divergence between models, and outputs a predictor set F = {f_θi}.

[0134] Evaluate epistemic uncertainty caused by missing knowledge: For each node v_i, collect the prediction distributions D_i = {s_i j}, where s_i j is the prediction of the jth sub-model; calculate the average prediction μ_i = (1 / M)∑ j s_i j and variance σ 2 _i = (1 / M)∑_j(s_i j - μ_i) 2 ; Define epistemic uncertainty index U_epist(v_i) = σ 2 _i, reflects the uncertainty of the model's prediction of the node v_i state, and outputs the epistemic uncertainty score of each node. The higher the value, the more uncertain the model's prediction of the point.

[0135] Assess the aleatoric uncertainty (statistical noise) inherent in 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 of the model estimate; calculate the average variance of all sub-model predictions: U_alea(v_i) = (1 / M)∑_j(σ_i j ) 2 ,Unlike epistemic uncertainty, aleatoric uncertainty reflects the randomness and measurement noise of the data itself, which cannot be eliminated by increasing training data and outputting aleatoric uncertainty scores.

[0136] Evaluate the uncertainty caused by the distance between the predicted point and the nearest measured point: for each predicted point v_i, calculate its geodesic distance d_i = min_j d(v_i, v_j) to the nearest measured point, where v_j is the measured point; calculate the topological propagation attenuation factor α(d_i) = exp(-λ·d_i), where λ is the attenuation coefficient; define the distance uncertainty index U_dist(v_i) = 1 - α(d_i), reflecting that the farther the predicted point is from the measured point, the higher the uncertainty; output the distance uncertainty score to provide a topological structural perspective for the spatial uncertainty distribution.

[0137] Comprehensive multiple uncertainty assessments to generate an overall uncertainty index: design a weight adaptive function to determine the weights of various types of uncertainty according to structural characteristics and monitoring objectives: w_epist, w_alea and w_dist; calculate the weighted comprehensive uncertainty: U_total(v_i) = w_epist·U_epist(v_i) + w_alea·U_alea(v_i) + w_dist·U_dist(v_i); apply a spatial smoothing operator to ensure that the uncertainty assessment of adjacent areas has spatial continuity, and output a normalized overall uncertainty score with a value range of [0, 1]. The higher the value, the more uncertain the prediction.

[0138] Map the uncertainty assessment results to the structural space and generate a spatial uncertainty distribution map: Construct a three-dimensional grid G_vis covering the entire structure, and for each point g_k in the grid, calculate its uncertainty value by interpolation: U(g_k) =∑ i w_i·U_total(v_i), where the weight w_i is determined based on the spatial distance; adaptive color mapping is applied, high uncertainty areas are represented by warm colors, low uncertainty areas are represented by cold colors, and uncertainty heat maps are output to intuitively display the spatial distribution of prediction uncertainty and provide a risk assessment basis for decision-making.

[0139] Based on uncertainty assessment, decision support information is provided: high uncertainty areas are identified, and it is recommended to increase sensor deployment or adopt more conservative evaluation criteria; the confidence interval of each prediction result is calculated: CI(v_i) = [μ_i - k·sqrt(U_total(v_i)), μ_i + k·sqrt(U_total(v_i))], where k is the confidence level correlation coefficient; a risk assessment report is generated, listing the predicted reliability levels of each part of the structure and the recommended confidence adjustments, and finally outputting a spatial uncertainty distribution map and related decision support information to provide an uncertainty quantification basis for subsequent status assessment and monitoring strategy optimization.

[0140] This embodiment achieves comprehensive quantification of the reliability of prediction results through the construction of predictor sets, the comprehensive application of multiple uncertainty assessments and spatial mapping. Compared with traditional deterministic prediction methods, this embodiment improves the credibility and availability of prediction results. In actual engineering verification, through uncertainty-guided sensor deployment optimization, the monitoring system performance is improved by 25%, and the resource utilization efficiency is improved 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 structural health assessment more scientific and reliable, and providing an important basis for monitoring system optimization and risk warning under limited resource conditions.

[0141] like Figure 3 As 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 the structure under different load conditions, the structural dynamics knowledge is integrated into the preconfigured neural network model, and the physical constraint correction response is obtained, specifically:

[0142] S51, 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 embed the structural dynamic equation into the network architecture as prior knowledge to achieve accurate modeling of the dynamic characteristics of the structure and obtain the structural dynamic model parameters;

[0143] S52, read the full-field response and structural dynamics model parameters after calibration, verify the physical consistency between different types of sensor data (such as the consistency between the second-order integral of acceleration and displacement) through the 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, spatial uncertainty distribution map and physical consistency score, the local physical constraints are propagated to the global through the structural connection relationship through the structure-aware constraint propagation network, so as to realize the estimation correction of the high uncertainty area and obtain the physical constraint correction response;

[0145] S54, combining the physical constraint correction response, the multimodal sensor raw data and the sensor weight matrix, applying the reliability weighted adaptive fusion algorithm, dynamically adjusting the information fusion weight according to the sensor reliability, prediction uncertainty and physical consistency, and obtaining the final fused full-field state estimate;

[0146] S55. Based on the historical fusion of the full-field state estimation sequence, a recursive attention network guided by structural dynamics is applied, combining short-term memory and long-term dependency to achieve smoothing and short-term prediction of the state sequence, generating a smooth state sequence and a predicted state sequence.

[0147] This embodiment achieves physical consistency correction of the full-field response 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 recursive attention network guided by structural dynamics. In practical applications, this embodiment reduces the physical consistency violation of the prediction results by 55%, and accurately captures complex nonlinear dynamic behaviors. In particular, under load mutation conditions, the state estimation accuracy is improved 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:

[0149] S511, reading the full-field response and physical connection topology after calibration, constructing and initializing the parameterized structural dynamics equation, and obtaining an initial parameter matrix;

[0150] S512. Based on the initial parameter matrix, construct a 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 time, θ is all parameter vectors 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 physics-based part, and f_nn is the neural network part, and obtain the neural dynamics equation;

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

[0152] S514, constructing an adaptive step-size solver based on a neural network, numerically solving the neural dynamics equations, and obtaining a predicted response trajectory;

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

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

[0155] In one embodiment of the present application, the full-field response and structural topology after calibration are read, and the structural dynamics equation is initialized through parameterized expression to obtain an initial parameter matrix. The specific processing method is: read the mass distribution information of the structure, construct a 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, 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, and initially use the Rayleigh damping assumption: C_0 = α·M_0 + β·K_0, where α and β are damping coefficients; combine the above parameters into an initial parameter vector θ_0 = [M_0, K_0, C_0].

[0156] Read the initial parameter vector θ_0, and build a model that can learn the structural dynamic behavior through the neural differential equation to obtain the neural dynamic equation. The specific processing method is: define the state space representation of the second-order dynamic equation: state vector z = [x, v], where x is displacement and v is velocity; state derivative ż = [v, a], where a is acceleration. Construct the standard form of differential equations: z* = f(z, t; θ); construct the neural network enhanced dynamic 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, which uses the residual structure to capture the unmodeled dynamic characteristics.

[0157] Read the neural dynamics equation, establish a metric for evaluating the physical consistency of the model prediction through multi-objective loss function design, and obtain the physical loss function. The specific processing method is: 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 all measurement points i and time points t, where N is the number of measurement points, x_i pred (t) is the predicted displacement of the ith measurement point at time t, x_i meas (t) is the actual measured displacement of the i-th measurement point at time t; physical consistency loss L_phys is defined as the degree of constraint violation 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 energy conservation loss L_energy: the consistency of the total energy of the system 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 is the weight coefficient.

[0158] Read the neural dynamics equation and physical loss function, simulate the system response through adaptive numerical solution, and obtain the predicted response trajectory. The specific processing method is: design an adaptive step-size solver based on neural network: use the 4th-order Runge-Kutta method as the basic solution framework; combine the residual network to predict the local error and dynamically adjust the integral step size; implement a differentiable numerical solution process to ensure that the gradient can be back-propagated through the solver; adopt a parallel prediction strategy to solve multiple initial conditions at the same time to increase training efficiency; monitor the numerical stability during the solution process and automatically adjust the regularization parameter.

[0159] Read the predicted response trajectory and physical loss function, update the model parameters through gradient descent optimization, and obtain the optimized parameter vector θ_opt. The specific processing method is: 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 updates with physical constraints: impose positive definiteness constraints on the stiffness matrix K to ensure system stability; impose non-negative constraints on the mass matrix M to conform to physical meaning; impose appropriate range constraints on the damping parameters to avoid over-damping or under-damping; use batch stochastic gradient descent (SGD) to optimize parameters: θ_t+1 = θ_t - η(t)·▽_θL_total; implement an early stopping strategy and 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 dynamic model parameters. The specific processing method is: extract the equivalent stiffness distribution from the optimized stiffness matrix K: the diagonal element k_ii represents the equivalent stiffness of node i; the non-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 structural natural frequency ω_i and vibration mode φ_i, and generate the space-time distribution diagram of the parameters by solving the generalized eigenvalue problem K·φ = λ·M·φ, which intuitively displays the spatial distribution and time evolution of the structural dynamic characteristics.

[0161] This embodiment constructs a dynamic model with both data fitting capability and physical consistency by combining 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 insufficient accuracy of traditional white-box physical models and lack of interpretability of black-box machine learning. In terms of model parameter identification, compared with traditional methods, the accuracy of equivalent stiffness identification is improved by 40%, and the accuracy of damping characteristic identification is improved by 35%. In particular, the ability to characterize nonlinear material behavior is significantly enhanced, and the model prediction error is reduced by 48%. 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:

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

[0164] S532, based on the multi-scale structural domain and the structural dynamics model parameters, establishing the local constraint equations of each subdomain to obtain a set of local constraint equations;

[0165] S533, based on all local constraint equations and the calibrated full-field response, calculating the degree of deviation between the response and the physical constraint, and obtaining a constraint violation degree graph;

[0166] S534, read the constraint violation graph and the multi-scale structure domain, build a structure-aware graph network to realize message passing, and obtain a constraint propagation network;

[0167] S535. Based on the constraint propagation network and the constraint violation graph, a bottom-up and top-down hierarchical coordination mechanism is constructed to obtain the constraint coordination rules;

[0168] S536. Using the calibrated full-field response, the constraint violation graph and the constraint coordination rule, the physical violation is minimized through the constraint gradient flow algorithm to obtain the physical constraint correction response.

[0169] In one embodiment of the present application, the full-field response, spatial uncertainty distribution map and physical consistency score after calibration are read, and the structure is divided into sub-domains of different scales through hierarchical domain decomposition to obtain a multi-scale structural domain. The specific processing method is: define three scale levels: microscale: a single node and its directly connected nodes, forming a node star domain Ω_micro_i; mesoscale: a group of nodes with similar functions or physical connections, forming a component domain Ω_meso_j; macroscale: the main structural part, forming a system domain Ω_macro_k; a community detection algorithm based on a structural topology map, identifying naturally formed structural subdomains; considering the uncertainty distribution, a more detailed domain division is used for high uncertainty areas; establishing an inclusion relationship between domains: Ω_micro_i ≤≥ Ω_meso_j ≤≥ Ω_macro_k, where ≤≥ represents a subset; defining a domain boundary node set Γ_ij as a connection interface between adjacent subdomains.

[0170] Read the parameters of the multi-scale structural domain and structural dynamics model, establish the local constraint equations of each subdomain through regional physical equation modeling, and obtain the local constraint equation set. The specific processing method is: for each microscopic star domain Ω_micro_i, establish the 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 mesoscopic 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 macroscopic 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 dissipated power, E_k is kinetic energy, E_p is potential energy, d is differential, and F_ext is external force.

[0171] Read the local constraint equation set and the calibrated full-field response, and obtain the constraint violation degree graph by quantifying the deviation between the response and the physical constraint through constraint violation. The specific processing method is: define the micro constraint violation degree: ε_micro_i = ||m_i·x"_i + ∑ j c_ij·(x'_i - x'_j) + ∑ jk_ij·(x_i - x_j) - f_i|| / ||f_i||;Define meso-constraint violation: ε_meso_j = ||∑F_j|| / ||F_ext_j|| + ||∑M_j|| / ||M_ext_j||;Define macro-constraint violation: ε_macro_k = ||(d(E_k + E_p) / dt + Φ_diss - P_input) / P_input||;Construct constraint violation graph G_ε, node values ​​represent local constraint violations, and edges represent constraint propagation paths. Where f_i is the external force acting on node i; F_j is the force acting on Ω_meso_j in the component domain, F_ext_j is the external force acting on Ω_meso_j in the component domain; M_j is the moment acting on Ω_meso_j in the component domain; M_ext_j is the external moment acting on Ω_meso_j in the component domain.

[0172] Read the constraint violation graph G_ε and the multi-scale structure domain, design the propagation mechanism of physical constraints through the graph network propagation model, and obtain the constraint propagation network. The specific processing method is: design the structure-aware graph network (SPGN) architecture: node features: location, constraint violation, uncertainty score; edge features: physical connection stiffness, distance, constraint correlation; global features: overall constraint violation statistics. Define the message passing mechanism: node to edge: m_v→e(h_v, h_e) = MLP([h_v || h_e]); edge to node: m_e→v(h_e, h_v) = ∑ e ∈N(v) MLP([h_e || h_v]); global to local: m_g→v(h_g, h_v) = MLP([h_g || h_v]). Design 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']). Where 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 constraint violation graph G_ε, and design the coordination mechanism between constraints of different scales through hierarchical coordination optimization to obtain the constraint coordination rules. The specific processing method is: design a bottom-up coordination process: micro constraint violation → meso correction: ε_micro_i affects the constraints of the component domain Ω_meso_j to which it belongs; meso constraint violation → macro correction: ε_meso_j affects the constraints of the system domain Ω_macro_k to which it belongs. Design constraint decomposition from top to bottom: 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. Realize cross-scale constraint trade-offs: when constraints of different scales conflict, determine the priority based on the uncertainty distribution and constraint importance; establish a hierarchical strategy for constraint satisfaction: key constraints must be satisfied, and secondary constraints can be appropriately relaxed.

[0174] Read the full-field response after calibration, the constraint violation graph G_ε and the constraint coordination rule, correct the response through constraint optimization, and obtain the physical constraint correction response. The specific processing method is: Construct the optimization objective function: f_obj(x) = w_1·||x - x_0|| 2 + w_2·∑ i ε_micro_i 2 + w_3·∑ j ε_meso_j 2 + w_4·∑ k ε_macro_k 2 , where x_0 is the initial response and w_i is the weight coefficient, which is determined by the uncertainty distribution and the importance of the constraint. Design the constraint gradient flow algorithm: Calculate the gradient of the constraint violation with respect to the response: ▽_x ε; Iterate and correct along the constraint gradient direction: x t+1 = x t -η·▽_x f_obj(x t ). Implement a multi-level alternating optimization strategy: micro optimization → meso coordination → macro verification → micro re-optimization; apply specific constraints at each level. Set up local-global consistency checks to ensure that corrections do not introduce new physical violations.

[0175] Read the physical constraint correction response and the original full-field response after calibration, evaluate the effect of constraint correction by comparative analysis, and obtain the physical enhancement response and correction effect report. The specific processing method is: calculate the change in 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 response change caused by 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 areas and constraint types with significant correction effects, and form correction experience knowledge; update uncertainty assessment and reduce uncertainty scores in areas with high physical consistency.

[0176] This embodiment achieves efficient propagation from local constraints to global consistency through the systematic coordination of hierarchical domain decomposition, establishment of local constraint equations, constraint violation evaluation, structure-aware graph network, hierarchical coordination mechanism and constraint gradient flow algorithm. This embodiment shows excellent performance in complex non-uniform structure testing, reducing physical constraint violations by an average of 62%, reducing prediction errors in spatial discontinuous regions by 53%, maintaining mechanical balance across interfaces of different materials, and the algorithm convergence speed is 3 times faster than traditional global optimization methods, providing an efficient and feasible solution for physical consistency correction of complex structures.

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

[0178] S61. Construction of multi-dimensional health indicators: Read and fuse the full-field state estimation and physical constraint correction response, calculate a set of multi-dimensional health indicators including modal parameter changes, energy distribution anomalies, stiffness degradation indicators and damping ratio changes, and comprehensively characterize the health status 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, 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.

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

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

[0182] S65, Multi-level health status assessment: Integrate multi-dimensional health indicator sets, damage probability assessment and abnormal propagation prediction, and generate a multi-level health assessment report that reflects the health status of the overall structure and key components through a hierarchical evidence reasoning framework, providing a comprehensive assessment of the damage location, type, degree and development trend.

[0183] In one embodiment of the present application, local-global abnormal pattern extraction is specifically as follows: read the historical multi-dimensional health indicator set, process each indicator by standardization and noise reduction, and obtain standardized health indicators. The specific processing method is: apply context-aware standardization to each health indicator 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 context c (such as structural type, environmental conditions); apply wavelet threshold denoising to remove high-frequency random fluctuations while retaining potential abnormal features; perform time alignment to ensure that different indicators are comparable in the time dimension; handle missing values ​​and fill them with interpolation methods based on physical models; construct an indicator correlation matrix R to quantify the relationship between different health indicators.

[0184] Read the standardized health indicators, build the characteristic pattern library of local structural abnormalities through unsupervised pattern recognition, and obtain the local abnormal pattern library. The specific processing method is: use the density peak clustering (DPC) algorithm to identify the typical distribution pattern of local indicators: calculate the distance matrix D_local of the local indicators, calculate the local density ρ_i = ∑ j exp(-(d_ij / d_c) 2 ), where d_ij is the distance between sample point i and sample point j, d_c is the normalization constant of the distance, and the minimum distance Δ_i from each sample to the sample with higher density is calculated; based on ρ and Δ, the typical pattern center is identified. The characteristic vector v_l of each type of abnormal pattern is extracted, including: time characteristics: duration, frequency of occurrence, development speed; spatial characteristics: impact range, spatial distribution, propagation direction; indicator characteristics: dominant indicators, indicator correlation changes. Based on the degree of deviation from the normal state, an abnormal severity scoring system S_l is established; a local abnormal pattern library L_lib is constructed, which contains the identified typical abnormal patterns and their characteristic vectors.

[0185] Read the standardized health indicators and 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: construct a regional association graph G_r, the nodes are structural regions, and the edge weights represent the correlation of health indicators between regions. Apply the spatiotemporal scanning statistical method to identify the spatial aggregation of abnormal indicators: define a circular or elliptical scanning window; calculate the log-likelihood ratio of the indicator values ​​inside and outside the window; and 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 pattern with the known damage pattern. Construct a global abnormal pattern library G_lib, which contains typical global abnormal patterns and their characteristics.

[0186] Read historical standardized health indicators, extract abnormal patterns of indicators changing over time through time series pattern mining, and obtain a time series abnormal pattern library. The specific processing method is: 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: indicator value suddenly changes significantly; trend type: indicator value continues to change in one direction; periodic type: indicator value fluctuates periodically; random fluctuation type: indicator value fluctuates irregularly. Extract the characteristics of each type of time series pattern, including duration, change rate, periodicity, etc.; build a 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 integrate the abnormal patterns of different dimensions through tensor decomposition and fusion to obtain a unified abnormal pattern representation. The specific processing method is: construct a third-order abnormal pattern tensor A (space, index, 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_index×3 U_time, where G is the core tensor and U is the feature matrix of each dimension; identify the main abnormal pattern combination based on the core tensor G; construct an abnormal pattern feature space so that similar abnormal patterns are gathered in the feature space; calculate the similarity between the current state and the known abnormal pattern to form an abnormal pattern matching vector.

[0188] Read the unified abnormal pattern representation and the current set of multi-dimensional health indicators, identify the abnormal pattern of the current state through hierarchical anomaly detection, and obtain the abnormal pattern characteristics. The specific processing method is: design a three-layer anomaly detection architecture: point-level detector: identify the anomaly of a single indicator at a single location; regional-level detector: identify the abnormal combination of multiple indicators in a local area; structural-level detector: identify the abnormal pattern of global multiple regions. Each layer of detector uses an integrated 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 abnormal pattern tensor A realizes abnormal gradient tracking, identifies the propagation path of anomalies from the point level to the structural level, and outputs comprehensive abnormal pattern characteristics, including the type, location, severity and development trend of the anomaly.

[0189] Read the abnormal pattern characteristics and historical abnormal-damage data, establish the association between abnormal patterns and potential structural damage through causal reasoning analysis, and obtain a damage association report. The specific processing method is: construct an abnormal-damage knowledge graph to represent the association between known abnormal patterns and structural damage types; apply Bayesian network to infer potential damage: calculate the conditional probability P (damage type | abnormal pattern); identify the most likely damage type and its probability; provide multiple damage hypotheses, considering the complex abnormal patterns caused by compound damage; generate damage hypothesis verification suggestions to guide subsequent targeted detection; form a complete description of abnormal pattern characteristics, including the type, spatial distribution, time evolution and potential damage association of abnormal patterns.

[0190] Context-aware anomaly detection is specifically as follows: read the anomaly pattern features, environmental sensor data and structural usage information, analyze the environmental and operational factors that affect the structural response through context feature extraction, and obtain the structural context features. The specific processing method is as follows: extract key environmental parameters from environmental sensor data: temperature field distribution T(x, y, z, t) and its gradient ▽T; humidity distribution H(x, y, z, t) and its rate of change ΔH / Δt; wind load characteristics, including speed v_wind, direction θ_wind and turbulence intensity I_wind. Extract operating conditions from structural usage information: load state L(t), including static load and dynamic load distribution; 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 contextual features C_f and the historical multi-dimensional health indicator set, a deep generative model is designed to build an anomaly detection model that can consider the context, and a conditional generative model is obtained. The specific processing method is: design a conditional variational autoencoder (CVAE) architecture: input layer: health indicator x and contextual features c; encoder network: q_φ(z|x, c), mapping the input to the latent space z; prior network: p_θ(z|c), generating a prior distribution of latent variables based on the context c; decoder network: p_θ(x|z, c), reconstructing the input from the latent variables and context. Define the conditional variational lower bound ELBO as the training target: 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 the 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_xΘσ(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-by-element multiplication. A hierarchical latent variable structure is implemented to capture features at different levels of abstraction.

[0192] Read the conditional generation model and historical data, learn the normal structural response distribution under different contexts through conditional distribution estimation, and obtain the conditional normal distribution model. The specific processing method is: construct a context space grid, discretize the continuous context space into typical scenes; for each typical context scene 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 the context, estimate the conditional mean vector μ(c_i) and covariance matrix Σ(c_i); construct a context interpolation mechanism to process the context state between grids: for any context c, calculate the similarity w_i with the grid point; estimate the interpolation distribution parameters by weighted average: μ(c) = ∑ i w_i·μ(c_i); establish distribution smoothing constraints to ensure smooth distribution changes 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 abnormal judgment threshold under the current context through adaptive statistical analysis, and obtain the context-sensitive threshold. The specific processing method is: 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 under 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 low threshold of the Mahalanobis distance and MD_thresh_high is the high threshold of the Mahalanobis distance; set independent thresholds for key security indicators to ensure that high-risk anomalies can be detected in a timely manner, and output a context-sensitive threshold set for the current state for anomaly determination.

[0194] Read the current multidimensional health indicator set, conditional normal distribution model P and context-sensitive threshold, calculate the abnormality of the current state through multidimensional anomaly detection, and obtain the abnormality measurement index. The specific processing method is: 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 individual contribution of each dimension: C_i = (x_i - μ_i(C_f)) 2 / Σ ii (C_f); Apply the local anomaly factor (LOF) to consider the change of local density: calculate the k-distance and reachable 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 measurement indicators, including the overall anomaly score and the contribution of each dimension.

[0195] Read the abnormal metric indicators and abnormal pattern features, estimate the probability that the current state indicates structural damage through Bayesian reasoning, and obtain the damage probability assessment. The specific processing method is: construct a Bayesian network model, the nodes include: context variable C = {C1, C2, ..., C_n}; abnormal indicator A = {A1, A2, ..., A_m}; damage state D = {D1, D2, ..., 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 damage state D given the abnormal indicator A and context variable C; P(A|D, C) is the probability of abnormal indicator A given the damage state D and context variable C; P(D|C) is the probability of damage state D given the context variable C; P(C) is the probability of the context variable C; the Monte Carlo Markov chain (MCMC) method is used to sample the posterior distribution; the probability of each damage type is calculated to generate a damage probability distribution; the evidence theory is used to process uncertain information, and the confidence interval and plausibility are calculated; the damage probability assessment is output, including the probability of each type of damage and its uncertainty estimation.

[0196] Read the abnormal measurement indicators and historical structural context features, distinguish the normal fluctuations caused by the environment and the abnormalities caused by structural damage through spatiotemporal correlation analysis, and obtain the environmental stripping abnormal assessment. The specific processing method is: construct the abnormal-environment correlation matrix R_ae to quantify the correlation between each abnormal indicator and the environmental factors; apply partial correlation analysis to eliminate the influence of environmental factors: calculate the partial correlation of abnormal indicators after controlling environmental variables; identify environmental sensitive indicators and environmental insensitive indicators; design an environmental compensation model to eliminate predictable environmental effects: x_adj = x - f_env(C_f), where f_env is the environmental impact function, x_adj is the health indicator after environmental compensation, and x is the current health indicator data; calculate the abnormal score A_score_adj after environmental compensation; compare the original abnormal score with the compensated score to quantify the contribution of environmental factors; output damage probability assessment, including damage possibility assessment after environmental factor correction.

[0197] According to one aspect of the present application, step S7 is also included, dynamically adjusting the sensor weight configuration and data acquisition strategy according to the monitoring results and changes in the structural state to achieve optimal allocation of monitoring resources.

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

[0199] S72. Dynamic allocation of sensor resources: Based on information value indicators and anomaly propagation prediction, an adaptive resource allocation algorithm is used to optimize the sensor's sampling frequency, data transmission priority, and computing resource allocation, and output monitoring strategy configuration.

[0200] S73, Active learning sampling optimization: read the spatial uncertainty distribution map and abnormal pattern characteristics, apply the maximum information gain sampling strategy, determine the key locations and time periods that require enhanced monitoring, and dynamically generate an optimized sampling plan.

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

[0202] S75. Long-term monitoring knowledge accumulation: With the accumulation of monitoring data, the model parameters and prediction capabilities in each module are continuously updated and optimized through an incremental learning framework to form a long-term knowledge base for specific structures, thereby improving the system's adaptability to the structural characteristics and detection accuracy.

[0203] The present invention obtains and preprocesses the raw data of multimodal sensors; constructs a topological perception network representation that reflects the spatial distribution of sensors and the structural connection relationship; extracts multi-temporal and spatial scale features; uses a recursive graph evolution predictor to achieve local to global structural state mapping; incorporates structural dynamics knowledge into a neural network model for physical constraint correction; constructs a multidimensional health indicator set and outputs a damage probability assessment. The present invention achieves accurate mapping from local dense monitoring areas to global structural states under a non-uniformly distributed sensor network, thereby improving the accuracy and reliability of structural health monitoring.

[0204] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within 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.

2. The method according to claim 1, characterized in that The steps to achieve mapping from the local dense monitoring area to the global structural state and 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.

3. The method according to claim 2, 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.

4. The method according to claim 2, 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.

5. The method according to claim 2, 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.

6. The method according to claim 4, characterized in that The steps to incorporate structural dynamics knowledge into the preconfigured neural network model and obtain 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.

7. The method according to claim 6, 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.

8. The method according to claim 6, 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 a constraint violation diagram; 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.

9. 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, a sensor association topology map reflecting the information flow, namely, the topology-aware network representation, is constructed.

10. The method according to claim 9, 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 embedding 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