Building health state detection system

Abstract

The invention discloses a building health state detection system, and relates to the technical field of photogrammetry and image processing, the system firstly obtains four-dimensional space-time coordinate data of a plurality of monitoring points on a target building based on satellite image analysis, constructs a four-dimensional space-time network model containing spatial topology and time sequence evolution information according to the four-dimensional space-time coordinate data, and then obtains a four-dimensional space-time network model containing space topology and time sequence evolution information; the system maps network model nodes to a predefined standard reference network associated with a local gravity direction, then, the system analyzes dynamic relevance in the four-dimensional network model by adopting a graph self-attention network, and meanwhile, calculates relative relation changes of monitoring points relative to the predefined standard reference network through geometric analysis; the deformation interpretation module distinguishes and determines internal relative deformation, overall displacement and absolute posture change (including settlement and inclination) of the building according to the two types of analysis results and by referring to surrounding environment information, and the state evaluation module integrates the information to generate a health state report containing risk levels.

Description

Technical Field

[0001] The present invention relates to the technical fields of photogrammetry and image processing, and in particular to a building health status detection system. Background Art

[0002] Buildings, especially large-scale infrastructures such as bridges, dams, high-rise buildings, and tunnels, will be affected by various factors such as loads, environmental erosion, material aging, and geological activities during their service life, and may suffer from diseases such as deformation, displacement, and cracking. Therefore, it is crucial to regularly and accurately detect the health status of buildings, which is a key link to ensure their safe operation, formulate scientific maintenance strategies, and extend their service life.

[0003] Traditional building health monitoring methods include manual visual inspection, deployment of contact sensors (such as strain gauges, displacement gauges, inclinometers), geodetic surveying (such as leveling, GPS / GNSS, total station), and three-dimensional laser scanning developed in recent years. However, these methods often have some limitations: manual inspection is subjective, inefficient, and difficult to detect early minor damages; contact sensors require physical deployment, are costly, easily damaged, and can only monitor limited points; although geodetic surveying has high accuracy, it is usually also point-based monitoring, with high labor intensity and high costs and low efficiency for monitoring large areas or inaccessible areas; although three-dimensional laser scanning can obtain high-precision geometric models, it usually requires ground operations, is affected by line-of-sight occlusion, and has high costs and complexity for periodic monitoring of large structures or areas. Therefore, it is of great practical significance and application value to develop automated, non-contact, large-scale, and high-efficiency building health status detection technologies.

[0004] In order to achieve more efficient structural deformation monitoring, some explorations have been carried out in the prior art, for example: Chinese invention patent CN113819902B discloses a three-dimensional measurement method of a traveling trajectory line based on a single-axis gyroscope with IPC classification number G01C. This method measures the three-dimensional coordinates of each point on the traveling route in a local geographic coordinate system and compensates for gyroscope errors (including the influence of the earth's rotation) to improve the Z coordinate accuracy, and is used for deformation detection of structures such as bridges and dams. This method has its advantages in improving the measurement accuracy of specific sensors and is convenient for local high-precision trajectory measurement.

[0005] Chinese invention patent CN115908424B discloses a building health detection method based on three-dimensional laser scanning with IPC classification number G06T. This method identifies building anomalies, conducts health assessment and early warning by planning scanning sites, obtaining point cloud data, generating three-dimensional models, and combining image comparison. This method utilizes the advantages of three-dimensional laser scanning technology to obtain high-precision geometric models and is committed to the informatization management of the entire building life cycle.

[0006] Although the above prior art has made certain progress in specific aspects, there are still certain limitations. In particular, there are still deficiencies in realizing large-scale, long-term, all-round, and automated monitoring. Specifically, it is manifested in that: the sensor physically moves along a specific trajectory, making it difficult to achieve rapid and non-contact monitoring of buildings in wide areas or areas inconvenient for personnel to reach. For terrestrial three-dimensional laser scanning, on-site operations are required, which are easily blocked by the ground line of sight. Moreover, for the periodic monitoring of large areas or numerous buildings, the cost is high and the efficiency is low. There is limited ability to reveal complex relative deformation patterns inside the structure (such as torsion and shear strain distribution), and there is a lack of a mechanism for directly analyzing the attitude changes of the building relative to the absolute gravity reference (such as pitch angle and roll angle changes). Summary of the Invention

[0007] The object of the present invention is to overcome the deficiencies of the prior art and propose a building health status detection system to solve the above existing problems.

[0008] The object of the present invention is achieved by the following technical solutions: A building health status detection system, comprising a four-dimensional spatio-temporal network construction module, configured to construct a four-dimensional spatio-temporal network model of a target building based on four-dimensional spatio-temporal coordinate data, the model including nodes representing the spatio-temporal positions of monitoring points, spatial edges connecting the nodes within the same timestamp, and temporal edges connecting the nodes of the same monitoring point at different timestamps; a standard network mapping module, configured to map the nodes in the four-dimensional spatio-temporal network model to the reference points of a predefined standard reference network associated with the local gravity direction, and establish a mapping relationship; a first analysis module, configured to process the four-dimensional spatio-temporal network model using a graph self-attention network, for outputting attention weights representing the internal spatio-temporal dynamic correlation between nodes and dynamic embedding vectors encoding the node states and neighborhood information; a second analysis module, configured to process the mapping relationship, and output, through geometric analysis, a time-varying index of the spatio-temporal relative geometric relationship of the monitoring points relative to the reference points of the standard reference network; a deformation interpretation module, configured to perform the following steps: Step S1, based on the time-varying pattern of the dynamic embedding vectors or attention weights, filter out the common mode effect by comparing with the patterns of surrounding reference buildings, and determine the internal relative deformation information of the target building; Step S2, fit a rigid body transformation or track the centroid based on the four-dimensional spatio-temporal coordinate data, and determine the overall displacement information of the target building by comparing with the movements of surrounding reference buildings; Step S3, based on the time-varying index, determine the absolute attitude change information of the target building relative to the standard reference network, including the absolute vertical settlement rate, pitch angle, and roll angle, through a conversion model; a status evaluation module, configured to integrate the internal relative deformation information, overall displacement information, and absolute attitude change information, and generate a health status evaluation report including a risk level assessment according to a preset threshold or trend analysis; a verification module, configured to compare the health status evaluation report with the ground measured reference data to generate an accuracy evaluation index; and a calibration module, configured to generate a calibration feedback instruction according to the accuracy evaluation index, so as to adjust the parameters or models of at least one other module of the system through iterative feedback.

[0009] When constructing the spatial edges, the four-dimensional spatio-temporal network construction module connects the adjacent monitoring point nodes in terms of structure according to the known roof geometric topology of the target building or the roof geometric topology extracted from images.

[0010] The standard network mapping module is configured to map each monitoring point node in the four-dimensional spatio-temporal network model to three reference points with predetermined relative geometric positions in the standard reference network.

[0011] The attention weights and dynamic embedding vectors output by the first analysis module include the attention weight values between nodes varying with time and / or the coordinates of the node embedding vectors varying with time.

[0012] The time-varying index of the space-time relative geometric relationship output by the second analysis module includes the time-varying rate or the accumulated variation of the relative distance or relative angle defined by the monitoring point node and the reference point set of the mapped standard reference network.

[0013] When determining the internal relative deformation information, the deformation interpretation module compares the time evolution pattern of the dynamic embedding vector or attention weight of the target building with the corresponding time evolution pattern obtained from the surrounding stable reference buildings to identify and quantify specific internal deformations indicating stress concentration or uneven settlement.

[0014] When determining the absolute attitude change information, the deformation interpretation module uses the time change index of the relative geometric relationship in space and time through a preset conversion model to calculate the absolute vertical settlement rate expressed in millimeters / year or similar units and the tilt angle change value including the pitch angle and roll angle expressed in degrees or radians.

[0015] The ground measured reference data used by the verification module is selected from at least one of: the three-dimensional coordinate time series of the global navigation satellite system (GNSS) receiver deployed on the target building, the elevation change data of the precise leveling measurement, the angle change data recorded by the inclinometer, or the multi-period point cloud data obtained by the ground three-dimensional laser scanning (TLS).

[0016] The calibration feedback instructions generated by the calibration module are used to adjust at least one of the following: the model structure or hyperparameters of the graph self-attention network in the first analysis module, the definition parameters of the standard reference network, the mapping rules in the standard network mapping module, or the classification threshold used to distinguish different deformation levels in the deformation interpretation module.

[0017] The system also includes a data interface and preprocessing module, which is used to receive the initial monitoring point detection results generated by analyzing a series of high-resolution optical satellite images (wherein the satellite images contain stereoscopic observation information) acquired over time and covering the target building, and perform coordinate solution, time series matching and noise filtering, and output four-dimensional space-time coordinate data to the four-dimensional space-time network construction module.

[0018] The beneficial effects of the present invention are: 1. By constructing a four-dimensional spatiotemporal network model and combining it with analytical methods such as graph self-attention networks, this system can deeply explore the complex spatiotemporal dynamic correlations between building monitoring points, revealing internal relative deformation patterns and early minor changes that are difficult to discover by displacement alone. At the same time, by introducing a standard reference network associated with gravity and specialized geometric analysis, the absolute posture (settlement, tilt, including pitch angle and roll angle) of the building is solved. Combined with the analysis of the overall displacement, the system can provide multi-dimensional and comprehensive deformation information including internal relative deformation, overall displacement and absolute posture changes.

[0019] 2. The system adopts a multi-strategy deformation interpretation method. By comparing and analyzing the dynamic patterns or overall movements of the target building and the surrounding stable reference buildings, it can effectively filter out regional common-mode noises (such as residual atmospheric delays, seasonal influences, etc.), and clearly distinguish the specific internal deformations of the building and the overall displacements relative to the environment. Combining with the absolute pose analysis based on the standard reference network, it realizes the effective decoupling and identification of deformation information with different physical sources and manifestations.

[0020] 3. The system is built-in with a verification module and a calibration module, forming a closed-loop feedback mechanism. By comparing the system output with the ground-measured reference data (such as high-precision measurement results of GNSS, leveling, inclinometers, etc.), it can objectively evaluate the accuracy of the system monitoring results. More importantly, the calibration module can automatically or semi-automatically generate feedback instructions according to the evaluation results, and perform iterative adjustment and optimization on the key parameters, model structures or processing rules inside the system, so as to continuously improve the monitoring accuracy of the system and the reliability of long-term operation.

[0021] 4. This system not only provides the original deformation data, but also integrates and analyzes the interpreted multi-dimensional deformation information through the state evaluation module. Combining with the preset threshold and change trend judgment, it automatically generates a health status report including risk level assessment, which provides an intuitive and quantitative decision-making basis for managers, helps to realize state-based predictive maintenance, and improves the efficiency and safety of building asset management.

[0022] 5. This invention mainly relies on high-resolution satellite images as the data source to achieve non-contact remote monitoring of building health, especially suitable for scenarios where it is difficult for ground personnel to approach or the cost of deploying traditional sensors is high (such as high-rise buildings, large bridges, buildings in dangerous areas, etc.). The automated processing flow designed by the system (from data preprocessing, analysis to evaluation) reduces manual intervention and has the potential to be applied to large-scale, periodic building censuses and long-term health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is the system architecture diagram of the present invention; Figure 2 is the timing diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0024] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0025] It should be noted that in the following solutions, the orientation concepts of "left", "right", "up", "down", "front", "back", "inside", and "outside" are all relative directions, and will not be listed one by one here.

[0026] Embodiment 1: As Figure 1 and Figure 2 shown, this embodiment details the front-end part of the basic construction and core analysis of a building health status detection system. This embodiment aims to illustrate how the system constructs a structured four-dimensional spatio-temporal network model starting from the input of raw data, establishes a mapping relationship with the standard reference network, and uses advanced analysis modules to generate key intermediate indicators required for subsequent deformation interpretation. The implementation method of this embodiment mainly corresponds to the functions of the four-dimensional spatio-temporal network construction module, the standard network mapping module, the first analysis module, the second analysis module, and the data interface and preprocessing module.

[0027] In this embodiment, the building health status detection system specifically implements its basic construction and core analysis functions in the following manner: The data interface and preprocessing module is configured with a data interface for receiving a series of high-resolution optical satellite images covering the target building and its surrounding area obtained over time provided by commercial satellite suppliers (such as Maxar, Airbus, etc.). The spatial resolution of these images is better than 0.5 meters and contains stereoscopic observation information (for example, stereoscopic image pairs in the same orbit or different orbits) to support the solution of three-dimensional coordinates. The received images are first subjected to strict preprocessing, including radiometric calibration, atmospheric correction, orthorectification, and high-precision temporal image registration. Computer vision techniques are used, such as an automatic building contour recognition algorithm (such as U-Net or MaskR-CNN) combined with a convolutional neural network, to determine the scope of the target building. On this basis, deep learning corner detection algorithms such as SuperPoint, or a combination of traditional methods such as Harris and FAST and sub-pixel precise positioning, are used to identify clear and stable corners such as the building roof, and an initial monitoring point detection result is generated. For temporal images, a robust feature matching and tracking algorithm (such as photometric matching based on a Siamese network combined with graph optimization tracking with geometric constraints, or combined with Kalman filter prediction) is used to continuously track the corresponding positions of these monitoring points in different images.

[0028] Using the stereoscopic observation information of satellite images, the three-dimensional geodetic coordinates (longitude, latitude, elevation) or projected coordinates (x, y, z) of the successfully tracked monitoring points at each time point are calculated through photogrammetric principles (such as space resection, forward intersection, bundle adjustment). Necessary noise filtering and outlier rejection are performed on the calculation results. After the above processing, the data interface and preprocessing module finally output reliable four-dimensional spatio-temporal coordinate data containing the three-dimensional coordinates of each monitoring point at each time stamp and its time information.

[0029] The four-dimensional spatio-temporal network construction module receives the four-dimensional spatio-temporal coordinate data from the data interface and preprocessing module. For each target building, at each time stamp tk, the corresponding three-dimensional coordinates (xi, yi, zi) of the monitoring points and the time stamp tk are encapsulated into a network node vi(tk).

[0030] Within the same time stamp tk, based on the known roof geometric topology information of the target building (which can be imported from design drawings, BIM models, or inferred by analyzing the spatial distribution pattern of monitoring points), an undirected spatial edge eij(tk) is established between the monitoring point nodes that are physically adjacent on the representative structure (such as along the same ridge line, eaves line). For the same physical monitoring point i, a directed time edge vi(tk) → vi(tk+1) is established between consecutive time stamps tk and tk+1, representing the temporal evolution relationship of this point.

[0031] The constructed nodes, spatial edges, time edges and their attributes (such as coordinates, time stamps, edge types) are organized into a graph data structure (for example, represented using an adjacency list), and can be stored in a graph database (such as Neo4j) or in-memory data structure.

[0032] Generate a structured four-dimensional spatio-temporal network model for each target building, which encodes the spatial form of the building and its dynamic changes over time.

[0033] The standard network mapping module receives the four-dimensional spatio-temporal network model, the geographical location of the target building, and the predefined standard reference network construction rules and mapping rules.

[0034] According to the geographical location of the building, use a high-precision geoid model (such as EGM2008) or a local gravity field model to determine the local gravity vertical direction. On this basis, instantiate a standard reference network associated with the building. The standard reference network can be designed as a virtual coordinate system or a set of points, whose Z-axis is strictly aligned with the local gravity reverse direction (vertically upward), and the X and Y axes are defined according to the geographical directions (such as due north, due east). The reference point has predefined and stable relative coordinates in this standard reference network.

[0035] Using specific geometric mapping rules, each monitoring point node vi(tk) in the four-dimensional spatio-temporal network model is mapped to three reference points with predetermined relative geometric positions in its associated standard reference network. For example, these three points can be a projection point at the origin of the standard reference network, and two other reference points located on the horizontal and vertical planes of the standard reference network respectively, which can reflect the relative attitude information of the node. The mapping relationship between the monitoring point node and its corresponding standard reference network reference point is established and stored. This relationship can be attached to the model as a node attribute or stored as independent associated data.

[0036] The first analysis module receives the four-dimensional spatio-temporal network model and uses a spatio-temporal graph neural network model, specifically a Graph Attention Network (GAT) or its spatio-temporal extended form (such as ST-GAT). This network takes the four-dimensional spatio-temporal network model (including node coordinate features, spatial edges, and temporal edges) as input. Through network-level calculations, the model can learn the complex spatio-temporal dependencies between nodes. The self-attention mechanism enables the model to dynamically assign different importance (attention weights) to different neighbor nodes (spatial or temporal), capturing the collaborative / differential patterns of the transfer or deformation of internal forces in the structure. At the same time, the model generates a dynamic embedding vector for each node vi(tk), which is a highly condensed representation of the node's own information and its spatio-temporal neighborhood structure and dynamics.

[0037] Output internal spatio-temporal dynamic correlation indicators, specifically including the numerical values of the attention weights between nodes varying with time and the coordinate values of the dynamic embedding vector varying with time.

[0038] The second analysis module receives the mapping relationship and the four-dimensional spatio-temporal coordinate data of each monitoring point node. For each monitoring point node vi(tk) and its set of standard reference network reference points associated through the mapping relationship, this module performs geometric calculations.

[0039] Calculate and track the relative distances (e.g., the distances to each reference point) and relative angles (e.g., the angles formed by the node and two reference points, or the angles of the node relative to the coordinate axes of the standard reference network) of the monitoring point node vi(tk) with respect to the three standard reference network reference points it corresponds to. Calculate the rates of change or cumulative changes of these geometric parameters over time, and output spatio-temporal relative geometric relationship time change indicators, specifically in the form of time series data, rates of change, or cumulative changes of relative distances / angles.

[0040] Working process The workflow starts with the data interface and preprocessing module. First, it receives a series of high-resolution optical satellite images of the target building obtained in chronological order. These images preferably include stereoscopic observation capabilities. Through the image processing and computer vision algorithms integrated within the data interface and preprocessing module (including object recognition, feature point detection, cross-temporal tracking, three-dimensional coordinate calculation based on stereo image pairs, and necessary noise filtering and optimization), the original image data is transformed into a set of reliable and spatio-temporally aligned four-dimensional spatio-temporal coordinate data. This data is the basis for all subsequent analyses, and it assigns three-dimensional spatial coordinates (x, y, z) to each monitored point of the target building that is stably tracked over a continuous time series (t1, t2,..., tn).

[0041] Subsequently, the four-dimensional spatio-temporal network construction module receives this time-series coordinate data. It not only stores these points but also gives structural meaning to these discrete spatio-temporal coordinate points. For each timestamp tk, the module constructs a spatial graph based on the geometric topology of the building (such as the connection relationship of the roof edge lines), uses nodes to represent the monitored points, and uses spatial edges to represent their structural connections. More crucially, it connects the nodes of the same physical monitored point at different timestamps tk and tk+1 by establishing time edges. In this way, the four-dimensional spatio-temporal network construction module transforms (constructs) the input point cloud time-series data into a dynamic four-dimensional spatio-temporal network model that can reflect the spatial form of the building and its evolution over time.

[0042] On this basis, the standard network mapping module starts to work. It receives the constructed four-dimensional spatio-temporal network model and instantiates a standard reference network aligned with the local gravity field according to the geographical location information of the target building and predefined rules. The core task is to perform the mapping: for each dynamic node in the four-dimensional spatio-temporal network model, according to specific geometric rules (for example, the 1-to-3 mapping adopted in this embodiment), establish a clear correspondence relationship, that is, a mapping relationship, with a set of static reference points in the associated standard reference network. This step provides a key reference benchmark for subsequent analysis of absolute attitude changes.

[0043] Next, the system starts two core analysis engines in parallel to deeply mine the data generated previously: The first analysis module focuses on the four-dimensional spatio-temporal network model itself. It utilizes the powerful capabilities of the graph self-attention network to deeply analyze the internal structure encoded in the four-dimensional spatio-temporal network model and its dynamic evolution. By learning the spatio-temporal dependencies (attention weights) between nodes and generating dynamic embedding vectors reflecting the node states, it aims to capture the complex patterns and correlations of the relative movements between different parts within the building. Its output mainly reflects the internal relative dynamic characteristics of the building.

[0044] The second analysis module utilizes the mapping relationship and the coordinate data of the monitoring points. It does not focus on the relationships between points inside the building, but rather on the relative geometric relationships (such as distance and angle) between each monitoring point and its corresponding standard reference network reference point. By calculating and tracking the changes of these geometric parameters over time, the module aims to quantify the movement of the monitoring points relative to the external absolute reference system (the direction of gravity), and the output time-varying indicators mainly reflect the absolute attitude change trend of the building.

[0045] The attention weights and dynamic embedding vectors output by the first analysis module, together with the time-varying indicators output by the second analysis module, jointly constitute the core output of this embodiment (the system foundation construction and core analysis stage). The dynamic embedding vector and the time-varying indicator respectively reveal the behavioral characteristics of the building from two different dimensions of "internal relative dynamics" and "external absolute reference". They are structuredly stored and transmitted to the subsequent deformation interpretation module as the key input basis for interpreting the three specific deformation types of internal relative deformation, overall displacement, and absolute attitude change.

[0046] Through the above detailed workflow, this embodiment gradually transforms the original satellite remote sensing observation data into structured analysis-ready data containing deep features and multi-dimensional reference information, laying a solid foundation for the subsequent comprehensive building health status assessment.

[0047] Application Case Consider long-term health monitoring of a super high-rise building located in the central area of a city. Due to the complex surrounding environment, it is difficult and costly to deploy ground monitoring means. The technical solution of this embodiment can play an important role: the data interface and preprocessing module regularly (such as every quarter) obtains high-resolution stereo satellite images of the building's coverage area, and through an automated processing process, identifies and tracks dozens of corner points on the roof and its attached structures, generating four-dimensional spatio-temporal coordinate data of these corner points with a time span of several years.

[0048] The four-dimensional spatio-temporal network construction module constructs a four-dimensional spatio-temporal network model of the super high-rise building based on the four-dimensional spatio-temporal coordinate data and the basic structural information of the building. The standard network mapping module then establishes an associated standard reference network for the building and maps the network nodes to reference points to establish a mapping relationship.

[0049] The first analysis module analyzes the four-dimensional spatio-temporal network model, and the output dynamic embedding vector reveals the subtle differences in the response patterns of different parts of the building (such as the core tube and the outer frame) under wind load or temperature changes. The second analysis module analyzes the mapping relationship, and the output time-varying indicators directly reflect the small swing or overall settlement trend of the building top relative to the direction of gravity.

[0050] These structured four-dimensional spatio-temporal network models, dynamic embedding vectors, and time-varying indicators generated by this embodiment provide high-quality inputs and bases for subsequent interpretation of possible differential settlements, overall tilts, torsional deformations, etc. of the super high-rise building.

[0051] Through the four-dimensional spatio-temporal network construction module, discrete monitoring point data is transformed into a four-dimensional spatio-temporal network model with topological structure and temporal relationship, greatly enhancing the analyzability of the data and facilitating subsequent application of models such as graph neural networks.

[0052] The standard reference network associated with gravity and its mapping relationship established through the standard network mapping module introduce a stable absolute reference system into the system, making it possible to distinguish relative deformations and absolute attitude changes (such as settlements and tilts) subsequently.

[0053] The first analysis module uses the graph self-attention mechanism to automatically learn and extract deep features (attention weights and dynamic embedding vectors) that reflect the complex dynamic responses inside the building, and has the potential to detect early, small, or non-linear deformation patterns.

[0054] The second analysis module directly quantifies the change in the geometric relationship (time-varying indicator) of the monitoring points relative to the stable reference system, providing a direct basis for subsequent calculation of absolute settlements and tilt angles.

[0055] The structured model and two types of core analysis indicators generated by this embodiment, dynamic embedding vectors and time-varying indicators, lay a solid data and feature foundation for the realization of refined and multi-strategy deformation interpretation and state evaluation in subsequent embodiments.

[0056] Combined with the data interface and preprocessing module, the entire process is based on satellite remote sensing data, has the potential for non-contact, large-scale, and periodic monitoring, and has a high degree of automation, making it suitable for monitoring scenarios where it is difficult or inconvenient for humans to intervene.

[0057] Embodiment 2: As Figure 1 and Figure 2 shown, this embodiment describes how the building health status detection system performs core deformation interpretation and state evaluation on the basis of the foregoing Embodiment 1 (system construction and core analysis foundation). This embodiment focuses on transforming the intermediate analysis results (internal correlation indicators and attitude change-related indicators) generated by Embodiment 1 into building deformation information with clear physical meanings, and finally forming a health status report including risk assessment. The implementation method of this embodiment mainly corresponds to the functions of the deformation interpretation module and the state evaluation module, and also covers the logic of determining the overall displacement information.

[0058] In this embodiment, the deformation interpretation and state evaluation functions of the system are implemented through the following modules and their internal logics: The input of the deformation interpretation module is: the attention weights and dynamic embedding vectors from the first analysis module, the time variation index of the spatio-temporal relative geometric relationship from the second analysis module, and the original four-dimensional spatio-temporal coordinate data from the data interface and preprocessing module (or its cache).

[0059] The surrounding environment reference information obtained from the system or externally input (for example, the analysis results obtained by applying the process of Example 1 to adjacent buildings identified as stable).

[0060] The algorithm within the deformation interpretation module analyzes the time series of the input dynamic embedding vectors or attention weights. For example, time series clustering or anomaly detection algorithms can be used to identify monitoring points or groups of points where the trajectory of the embedding vectors deviates significantly (relative to its own historical trajectory or relative to other nodes), or to analyze the anomalies and continuous changes in the connection strength in the attention weight map.

[0061] To distinguish real internal deformations from regional common-mode effects (such as large-scale ground settlement, seasonal thermal expansion and contraction, satellite orbit or atmospheric error residuals), the system compares the above change patterns of the target building with the corresponding patterns extracted from the surrounding environment reference information (i.e., stable reference buildings). For example, it calculates the correlation or difference degree of the change patterns between the target building and the reference building.

[0062] Only those relative change patterns that are significantly different from the reference buildings and have statistical significance are identified as specific internal deformations indicating internal stress concentration, uneven settlement, loosening of component connections or local damage. The internal relative deformation information is quantified as the relative displacement rate between point pairs, strain indices or marked as abnormal regions.

[0063] The deformation interpretation module uses the input four-dimensional spatio-temporal coordinate data. For each time stamp tk, it calculates the geometric centroid of the point cloud formed by all monitoring points of the target building, or fits a rigid body transformation model (including translation and rotation) by the least squares method.

[0064] Extract the translation vector of the target building at time tk relative to time tk−1.

[0065] Meanwhile, using the surrounding environment reference information (specifically, the four-dimensional spatio-temporal coordinate data of the reference building), calculate the average translation vector of the reference building.

[0066] Subtract the average translation vector of the reference building from the translation vector of the target building to obtain the specific overall displacement of the target building relative to its locally stable environment, forming overall displacement information, which usually appears as velocity vectors (such as velocities in the horizontal directions X and Y) and cumulative displacement trajectories.

[0067] The deformation interpretation module receives the time-varying index of the spatio-temporal relative geometric relationship from the second analysis module, which is usually a time series of parameters such as the distance and angle of the monitoring point relative to the reference points of its standard reference network.

[0068] Apply a conversion model (e.g., based on trigonometric functions, coordinate transformation, or small calibration models) established in advance according to the standard reference network structure and mapping rules, which converts the input relative geometric change index into absolute attitude parameters with clear physical units and directions.

[0069] The output absolute attitude change information specifically includes: the absolute vertical settlement rate of the whole building or specific parts relative to the gravity direction (e.g., in millimeters per year), and the change value of the tilt angle describing its tilt state over time, which is further decomposed into the pitch angle and roll angle commonly used in engineering (e.g., in degrees or arcseconds).

[0070] The deformation interpretation module finally outputs three independent but possibly related deformation information sets: internal relative deformation information, overall displacement information, and absolute attitude change information.

[0071] The state evaluation module receives three types of deformation information from the deformation interpretation module: internal relative deformation information, overall displacement information, and absolute attitude change information, and may also include auxiliary information such as the type of building, design specifications, historical monitoring data, and preset deformation control thresholds.

[0072] First, cross-validate the three types of input information to check for logical contradictions. For example, whether significant tilting is accompanied by a corresponding internal differential settlement pattern.

[0073] Compare the calculated deformation amounts (such as settlement rate, tilt angle, maximum relative displacement) with the allowable thresholds in the design specifications or industry standards corresponding to this type of building.

[0074] Analyze the time series of deformation parameters (such as settlement rate, tilt angle change rate) to determine whether the deformation tends to converge, remains stable, or shows a trend of accelerating development.

[0075] Based on the results of the above integrated analysis (deformation magnitude, rate, trend, distribution, and relationship with thresholds), the state evaluation module uses preset evaluation rules (possibly based on expert systems or machine learning models) to classify the current health state of the building, for example, into risk level evaluations such as "normal", "concerned", "alert", and "dangerous".

[0076] Integrate the quantified internal relative deformation information, overall displacement information, and absolute attitude change information of the deformation index, time series charts, possible deformation visualization diagrams (such as nephograms and vector diagrams), threshold comparison results, and the final risk level assessment to generate a comprehensive health status assessment report.

[0077] Working process As the core processing unit at this stage, the deformation interpretation module is first activated. It receives the key outputs from the two parallel analysis modules in the first embodiment: namely, the attention weights and dynamic embedding vectors generated by the first analysis module, which reflect the internal dynamic relevance, and the time-varying index of the spatio-temporal relative geometric relationship generated by the second analysis module, which reflects the geometric changes relative to the absolute reference system. At the same time, in order to perform a comprehensive interpretation, the deformation interpretation module also calls the original four-dimensional spatio-temporal coordinate data (for overall motion analysis) and the surrounding environment reference information obtained through system processing (for filtering common-mode effects).

[0078] After receiving the required inputs, the deformation interpretation module internally starts its three unique interpretation strategies either in parallel or in a predetermined logical order, independently analyzing data in different dimensions: Through time series analysis, pattern recognition, or anomaly detection algorithms, deeply explore the time evolution characteristics of the attention weights and dynamic embedding vectors. After identifying significant changes in the correlation between internal nodes of the target building or abnormal drifts in node states, through strict comparison with the corresponding patterns of stable buildings in the surrounding environment reference information, effectively filter out regional common-mode effects. Finally, this strategy outputs quantified internal relative deformation information, indicating possible stress concentration or uneven deformation regions and degrees inside the building.

[0079] Calculate the centroid position of the monitored point cloud of the target building at each time step or obtain the overall translation vector by fitting a rigid body transformation model. Then, compare this motion vector with the average motion vector calculated from the surrounding environment reference information (coordinate data of the reference building) in the same way (vector subtraction). Thus, separate and output the overall displacement information of the target building relative to its local stable environment (such as horizontal velocity and cumulative displacement trajectory).

[0080] Apply a preset conversion model based on the geometric and mapping relationships of the standard reference network to convert the time-varying index of the geometric relationship between the monitored points and the reference points of the standard reference network (such as relative distance and angular rate of change) into absolute attitude parameters used in engineering practice. Finally, output the absolute attitude change information including the absolute vertical settlement rate and the time series of the decomposed pitch angle and roll angle.

[0081] After the deformation interpretation module completes the calculations of the three strategies, it integrates their respective output results, namely the internal relative deformation information, the overall displacement information, and the absolute attitude change information. This integrated structured data set containing multi-dimensional deformation information is passed to the state evaluation module.

[0082] The state evaluation module receives the complete set of deformation information from the module, including the internal relative deformation information, the overall displacement information, and the absolute attitude change information, and it performs the following operations: conducts consistency checks and comprehensive analyses of the information. For example, it verifies whether the observed inclination conforms to the internal deformation pattern, quantitatively compares key deformation parameters (such as the maximum settlement rate, the maximum inclination angle, the maximum relative displacement, etc.) with the safety thresholds or performance criteria preset in the system for this type of building, analyzes the time series data of each deformation parameter to determine whether the deformation tends to be stable, continues to develop, or shows a trend of accelerating deterioration. Based on the results of the above quantitative comparison and trend analysis, it applies the built-in evaluation rules or models to objectively assess the risk level of the current health state of the building (for example, "normal", "concerned", "alert", "dangerous").

[0083] Automatically compile all key quantitative deformation indicators, time series graphs, risk level assessment results, and possible suggestions into a structured and information-rich health state assessment report.

[0084] The state evaluation module outputs the final health state assessment report. This report is not only the final deliverable of this embodiment (the deformation interpretation and state evaluation stage), which can be consulted and used for decision-making by users, but also serves as a key input and will be passed to the verification module in Embodiment 3 for external verification of the system performance.

[0085] This embodiment effectively converts the intermediate analysis data generated in Embodiment 1 into a comprehensive, in-depth, and actionable understanding and assessment of the health state of the building.

[0086] By implementing three different interpretation strategies, the system can comprehensively diagnose the health status of the building from multiple dimensions such as internal relative motion, overall motion relative to the environment, and attitude change relative to the absolute reference system, avoiding the one-sidedness that may be brought by a single indicator.

[0087] The state evaluation module not only presents the data, but also gives a clear risk level assessment through integrated analysis, threshold comparison, and trend judgment, providing direct and intelligent support for the maintenance decisions of managers (such as whether inspection is needed, when to inspect, and the key points of inspection).

[0088] Common-mode noise is effectively suppressed by comparing with the surrounding environment reference information, improving the relative accuracy and reliability of internal deformation and overall displacement monitoring. By introducing a standard reference network and transformation, the ability to quantitatively monitor absolute settlement and inclination is provided.

[0089] It can automatically generate a health status assessment report containing rich information, unify the assessment criteria and output format, and improve the efficiency and standardization of monitoring work.

[0090] Embodiment 3: As Figure 1 and Figure 2 As shown, this embodiment details the crucial subsequent steps after the building health status detection system completes deformation interpretation and status assessment (as described in Embodiment 2): verification and adaptive calibration of the system output results. This embodiment focuses on ensuring the accuracy and reliability of the system monitoring results and establishing a feedback mechanism to achieve continuous optimization and performance improvement of the system. The implementation method of this embodiment mainly corresponds to the functions of the verification module and the calibration module.

[0091] In this embodiment, the verification and calibration functions of the system are implemented through the following modules and their internal logics: The verification module receives the health status assessment report from the status assessment module, which contains quantified deformation information (such as key indicators of internal relative deformation, overall displacement rate, settlement rate and inclination angle of absolute attitude change, etc.), and independently obtained, high-precision ground measured reference data.

[0092] The verification module is configured with interfaces for accessing or importing ground measured reference data. These data have diverse sources, specifically selected from: three-dimensional coordinate time series obtained by continuous or periodic observation of global navigation satellite system (GNSS) receivers deployed at key positions on the target building; elevation change data obtained by traditional precise leveling; angle change time series recorded by inclinometers installed on the structure; or homologous point displacement information extracted from multi-period high-density point cloud data obtained by terrestrial three-dimensional laser scanning (TLS).

[0093] The spatio-temporal alignment algorithm executed by the verification module matches the deformation amount of the monitoring points (or specific positions deduced therefrom) calculated by this system in the health status assessment report with the measurement results in the ground measured reference data at the same geographical location and the same time period (or the closest time point), which may involve operations such as coordinate system transformation, time interpolation, and point position association (for example, associating the roof corner points monitored by satellites with the GNSS points deployed on the ground).

[0094] For the successfully matched data pairs, calculate the difference (i.e., error) between the output value of this system and the true value measured on the ground. For example, calculate the difference in settlement rate, the difference in tilt angle, the difference in displacement vectors at specific points, etc.

[0095] Based on the calculated errors, statistically generate a series of accuracy evaluation indicators. These indicators may include: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Bias, standard deviation of the error, and the correlation coefficient between the two, etc., to comprehensively and quantitatively evaluate the performance of the system in monitoring various deformation parameters.

[0096] Output a set of quantified accuracy evaluation indicators, reflecting the precision level and error characteristics of the current monitoring results of the system. The calibration module receives the accuracy evaluation indicators from the verification module and also refers to the current configuration parameters of the system and historical calibration records.

[0097] The calibration module receives and analyzes the accuracy evaluation indicators. For example, if the RMSE is large, analyze whether the random error or systematic bias is dominant, whether the bias is positive or negative, and whether the error is related to a specific time period, specific area, or specific deformation type. Through this analysis, try to infer which upstream processing link the main source of the error may lie in (for example, satellite data preprocessing, three-dimensional coordinate solution accuracy, standard reference network definition, applicability of the analysis model [3,4], bias of the interpretation algorithm, etc.).

[0098] Based on the results of the error analysis and diagnosis, the calibration module applies a preset calibration strategy (which can be a rule-based expert system or a machine learning-based adaptive optimization algorithm) to generate specific calibration feedback instructions. The goal of these instructions is to adjust the internal parameters or models of the system in order to reduce the identified errors in subsequent processing. The specific objects that can be adjusted include (but are not limited to): The model structure (such as the number of layers, node feature dimension) or hyperparameters (such as learning rate, regularization coefficient) of the graph self-attention network in the first analysis module; the definition parameters of the standard reference network (such as fine-tuning of the origin position, axis orientation); the mapping rules in the standard network mapping module (such as adjusting the mapping algorithm or parameters); the classification threshold used in the deformation interpretation module to distinguish different deformation levels (risk levels); it can be recommended to adjust certain parameters in the data interface and preprocessing module (such as filtering intensity).

[0099] The generated calibration feedback instructions are passed back to the corresponding modules or parameter configuration files in the system. After receiving the instructions, the system can automatically update the relevant parameters or prompt the operator for manual adjustment. These adjustments will take effect when the next monitoring task is executed or the model is retrained, thus realizing the iterative feedback optimization of the system and outputting calibration feedback instructions pointing to specific modules or parameters.

[0100] Working process The working process of this embodiment constitutes a closed-loop for quality assurance and performance improvement of the entire monitoring system. After the health status assessment report is generated in Embodiment 2, the verification module is activated. It actively obtains or receives relevant ground measured reference data. The core task of the module is to perform strict spatio-temporal alignment, and compare the key deformation quantification results in the system report with the reference data one by one, calculating a series of accuracy evaluation indicators reflecting the current system performance. These objective indicators are then sent to the calibration module. The calibration module conducts in-depth analysis on these indicators, aiming to diagnose the error sources and determine the optimization direction. Based on the diagnostic results, it generates specific and operable calibration feedback instructions, clearly indicating which parameters or models of which modules (such as to etc.) in the system need to be adjusted. These instructions are fed back to the system through an internal mechanism, enabling the system to apply these adjustments in subsequent operations and achieve self-correction and continuous improvement of performance. This process of "evaluation - diagnosis - feedback - adjustment" forms a complete iterative optimization loop.

[0101] By comparing with recognized high-precision ground measurement methods, the verification module provides objective and quantitative accuracy evaluation indicators for the system output results, enabling users to understand the reliability level of the monitoring results. The iterative feedback mechanism established by the calibration module enables the system to self-adjust and optimize according to the actual verification results, and is expected to gradually improve the monitoring accuracy and adapt to changes in different monitoring objects and environmental conditions. Through continuous verification and calibration cycles, systematic biases or model drifts that may exist in the system can be detected and corrected in a timely manner, thus significantly enhancing the stability and reliability of the system during long-term operation. The clear verification link and transparent (or traceable) calibration process enhance users' trust in the results provided by this automated and intelligent monitoring system. The diagnostic analysis of the error sources by the calibration module can also provide valuable feedback information for the continuous research and development of the system and algorithm improvement, guiding the focus of subsequent optimization.

[0102] The above is only the preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in related fields. As long as the changes and variations made by those skilled in the art do not depart from the spirit and scope of the present invention, they should all be within the protection scope of the appended claims of the present invention.

Claims

1. A building health status detection system, characterized in that, Including: A four-dimensional spacetime network construction module, configured to construct a four-dimensional spacetime network model of a target building based on four-dimensional spacetime coordinate data. The model includes nodes representing the spacetime positions of monitoring points, spatial edges connecting the nodes within the same timestamp, and temporal edges connecting the nodes of the same monitoring point at different timestamps; A standard network mapping module, configured to map the nodes in the four-dimensional spacetime network model to the reference points of a predefined standard reference network associated with the local gravity direction, and establish a mapping relationship; A first analysis module, configured to process the four-dimensional spacetime network model using a graph self-attention network, and output attention weights representing the internal spacetime dynamic correlation between nodes and dynamic embedding vectors encoding the node states and neighborhood information; A second analysis module, configured to process the mapping relationship, and output, through geometric analysis, a time variation index of the spacetime relative geometric relationship of the monitoring points with respect to the reference points of the standard reference network; A deformation interpretation module, configured to perform the following steps: Step S1, based on the time variation pattern of the dynamic embedding vectors or attention weights, filter the common mode effect by comparing with the patterns of surrounding reference buildings, and determine the internal relative deformation information of the target building; Step S2, based on the four-dimensional spacetime coordinate data, fit a rigid body transformation or track the centroid, and determine the overall displacement information of the target building by comparing with the movements of surrounding reference buildings; Step S3, based on the time variation index, determine, through a conversion model, the absolute attitude change information of the target building with respect to the standard reference network, including the absolute vertical settlement rate, pitch angle, and roll angle; A state evaluation module, configured to integrate the internal relative deformation information, overall displacement information, and absolute attitude change information, and generate a health state evaluation report including a risk level assessment according to a preset threshold or trend analysis; A verification module, configured to compare the health state evaluation report with the ground measured reference data, and generate an accuracy evaluation index; And a calibration module, configured to generate a calibration feedback instruction according to the accuracy evaluation index, so as to iteratively feedback and adjust the parameters or models of at least one other module of the system.

2. The building health status detection system according to claim 1, characterized in that: When constructing the spatial edges, the four-dimensional spacetime network construction module connects the adjacent monitoring point nodes in terms of structure according to the known roof geometric topology of the target building or the roof geometric topology extracted from images.

3. The building health status detection system according to claim 1, characterized in that: The standard network mapping module is configured to map each monitoring point node in the four-dimensional spacetime network model to three reference points with predetermined relative geometric positions in the standard reference network.

4. The building health status detection system according to claim 1, characterized in that: The attention weights and dynamic embedding vectors output by the first analysis module include the attention weight values between nodes varying with time and / or the coordinates of the node embedding vectors varying with time.

5. The building health status detection system according to claim 1, wherein: The time variation index of the spacetime relative geometric relationship output by the second analysis module includes the time variation rate or cumulative change amount of the relative distance or relative angle defined by the monitoring point nodes and the set of reference points of the standard reference network they are mapped to.

6. The building health status detection system according to claim 1, wherein: When determining the internal relative deformation information, the deformation interpretation module compares the temporal evolution patterns of the dynamic embedding vectors or attention weights of the target building with the corresponding temporal evolution patterns obtained from the surrounding stable reference buildings to identify and quantify specific internal deformations indicating stress concentration or uneven settlement.

7. The building health status detection system according to claim 1, wherein: When determining the absolute attitude change information, the deformation interpretation module calculates the absolute vertical settlement rate expressed in millimeters per year or a similar unit and the change value of the tilt angle including the pitch angle and the roll angle expressed in degrees or radians by passing the temporal change index of the spatio-temporal relative geometric relationship through a preset conversion model.

8. A building health status detection system according to claim 1, characterized in that: The ground-measured reference data used by the verification module is selected from at least one of the following: the three-dimensional coordinate time series of the global navigation satellite system (GNSS) receivers deployed on the target building, the elevation change data obtained by precise leveling, the angle change data recorded by inclinometers, or the multi-period point cloud data obtained by terrestrial three-dimensional laser scanning (TLS).

9. The building health status detection system according to claim 1, wherein: The calibration feedback instruction generated by the calibration module is used to adjust at least one of the following: the model structure or hyperparameters of the graph self-attention network in the first analysis module, the defined parameters of the standard reference network, the mapping rules in the standard network mapping module, or the classification threshold used in the deformation interpretation module to distinguish different deformation levels.

10. A building health status detection system according to claim 1, characterized in that: The system further includes a data interface and preprocessing module, which is configured to receive the initial monitoring point detection results generated by analyzing a series of high-resolution optical satellite images covering the target building obtained over time (wherein the satellite images contain stereo observation information), and perform coordinate solution, time series matching, and noise filtering, and output the four-dimensional spatio-temporal coordinate data to the four-dimensional spatio-temporal network construction module.

Images