House safety supervision intelligent early warning method based on multi-source data fusion

Abstract

The application discloses a kind of multi-source data fusion's house safety supervision intelligent early warning method, it is related to house safety monitoring technical field.The method includes: obtaining multiple sensor monitoring data, time alignment and standardization are formed multi-source time series data, and construct the mechanical topology graph between measuring point;Separate the synchronous component caused by environmental factors, obtain residual sequence;Based on multi-window and linear and nonlinear correlation measure, construct correlation matrix sample set;Emergent degree index is calculated using random matrix theory distribution upper bound, and local damage correlation emergent measuring point is identified by concentration degree test;Through topological graph connectivity verification and time delay analysis, the damage crack and propagation direction are inferred, and early warning information is generated.The application changes early warning from single measuring point threshold determination to the detection of correlation emergence between multiple measuring points, which helps to achieve early warning in the early stage of damage.

Description

Technical Field

[0001] This invention relates to the field of building safety monitoring technology, and more specifically, to a smart early warning method for building safety supervision that integrates multi-source data. Background Technology

[0002] Over long-term use, buildings are subject to factors such as material aging, foundation settlement, and environmental erosion, resulting in gradual internal structural damage. To monitor the safety status of buildings, various sensors, including crack gauges, inclinometers, settlement sensors, and accelerometers, are typically installed in key areas such as load-bearing walls, foundations, and floor slabs for continuous monitoring.

[0003] Chinese patent CN118780622A discloses a building safety monitoring system and method based on sensor data fusion. It collects data such as stress-strain, relative settlement, tilt, and cracks, preprocesses the data, determines the risk, and outputs alarm signals. A safety level prediction model is then established based on a test set. Chinese patent CN113587977A discloses a dynamic monitoring method for the collapse of old and dilapidated buildings based on multi-sensor data. It identifies weak points through ultimate load analysis and sets early warning thresholds. After analyzing mechanical properties and comparing thresholds using sensor data such as tilt angle, displacement, and strain, an early warning is triggered.

[0004] The aforementioned existing technologies all use the numerical value of each monitoring indicator as the early warning criterion, and the monitoring data of each measuring point is processed independently. Their common limitation is that the early warning judgment only focuses on whether the value of a single measuring point exceeds a set threshold or classification boundary, without considering the interrelationships between measuring points. However, the main characteristic of early structural damage is not a large change in the value of a single measuring point, but rather the redistribution of stress within the structure, which forces previously independently responding measuring points to exhibit a coordinated change trend. Even before cracks have significantly expanded or tilt has significantly increased, monitoring data from different locations may already show synchronous fluctuations that were not previously present. This change in cross-measuring point correlation cannot be perceived under the framework of independent judgment for a single indicator, leading to early warnings often being triggered only after the damage has progressed to a later stage and a single indicator has clearly exceeded its limit. Therefore, a smart early warning method for housing safety supervision based on multi-source data fusion is proposed to address the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a smart early warning method for building safety supervision based on multi-source data fusion. This method addresses the problem that the prior art, by processing monitoring data from each measuring point in isolation and using only single-index threshold comparison or feature classification as early warning criteria, cannot detect the emergent changes in the correlation between multiple measuring points from non-existent to existing and from weak to strong in the early stages of structural damage, resulting in insufficient early warning capabilities.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A smart early warning method for housing safety monitoring based on multi-source data fusion includes the following steps: S1: Acquire monitoring data collected by multiple types of sensors deployed at different stress points of the building structure, perform time alignment and standardization on the monitoring data to form multi-source building structure response time series data, and acquire a building mechanical topology map constructed with each monitoring location as a node and the node pairs connected by direct structural force transmission paths as edges.

[0007] Different sensors have different sampling times and data scales. Time alignment and standardization unify multi-source data to the same time reference and numerical magnitude, eliminating the interference of asynchronous sampling and dimensional differences on correlation analysis. The building mechanics topology map records the force transmission connectivity of each measuring point in the structural system, providing topological constraints for subsequent verification of the physical rationality of the detected set of correlated measuring points.

[0008] S2: Perform environmental common mode separation on the multi-source building structure response time series data, remove the synchronous change components between measuring points caused by environmental factors, and obtain the structural response residual sequence.

[0009] Environmental factors such as temperature changes and groundwater level fluctuations can cause synchronous seasonal or periodic drifts at all measuring points. These changes are not caused by structural damage. If correlation analysis is performed directly, the environment-driven synchronous drifts will form false strong correlations, leading to false alarms. By separating the environmental common modes, the aforementioned global synchronous components are removed from the measuring point sequences, retaining the residual components that mainly reflect the structural response characteristics, thus suppressing the interference of environmental covariance on subsequent analysis from the data source.

[0010] Furthermore, one way to implement the environmental common mode separation in step S2 is as follows: perform principal component analysis on the time series data of the multi-source building structure response, take the first principal component with the largest variance contribution as the environmental common mode component, subtract the projection of each time series data onto the environmental common mode component, and obtain the structural response residual sequence. The environmental factors include at least one of temperature and groundwater level. Principal component analysis can extract the most significant common variation direction driven by environmental factors in multi-point data. Projective subtraction removes synchronous drift in this direction from each sequence, retaining the residual components that reflect the structure's own response.

[0011] S3: At the time of the warning, based on the preset main time window and multiple sub-time windows obtained by shrinking the main time window at different ratios, the correlation matrix between the measurement points under each window is calculated using linear correlation measurement method and nonlinear correlation measurement method respectively, forming a correlation matrix sample set.

[0012] Sub-time windows of different lengths are used to observe changes in the correlation between measurement points from multiple time scales. Shorter windows are more sensitive to recent sudden changes in correlation, while longer windows are more robust to slowly accumulating correlation trends.

[0013] Linear correlation measurement captures the proportional coordination relationship between synchronous increases and decreases of measurement points, while nonlinear correlation measurement can detect non-monotonic and nonlinear coupling relationships between measurement points. The two methods complement each other, so that different types of damage correlation patterns have corresponding detection channels.

[0014] Furthermore, in step S3, the linear correlation measurement method is the Pearson correlation coefficient measurement, and the nonlinear correlation measurement method is the distance correlation coefficient measurement; the multiple sub-time windows are obtained by multiplying the main time window by a preset multiple scaling factors, and the scaling factors are all greater than 0 and less than 1.

[0015] Pearson correlation coefficient captures linear and consistent changes between measurement points, while distance correlation coefficient can detect statistical dependencies between measurement points that do not depend on monotonic trends. The two complement each other and can cover different types of damage association patterns. Multiple sub-time windows allow association analysis to be performed separately over shorter and longer historical periods, taking into account both sensitivity to sudden association changes and robustness to slow cumulative association trends.

[0016] S4: Perform eigenvalue decomposition on each correlation matrix in the correlation matrix sample set. Based on the upper bound of the distribution of eigenvalues ​​of a purely random correlation matrix in random matrix theory, calculate the emergence index that characterizes the degree of randomness of the correlation structure between measurement points. When the emergence index exceeds a preset threshold, perform a concentration test on the largest eigenvector to identify the set of measurement points where local damage correlations emerge.

[0017] Random matrix theory provides an unsupervised statistical benchmark for judging the significance of association structures, independent of historical damage labels: when multiple measurement points are independent of each other and contain only random noise, the eigenvalue distribution of their association matrix follows a definite theoretical upper bound; when organized and coordinated changes occur among measurement points due to damage, some eigenvalues ​​will exceed this upper bound. The emergence index measures the extent to which the largest eigenvalue exceeds the theoretical upper bound, thus measuring the magnitude of the transformation of the association structure among measurement points from a disordered random state to an ordered organized state.

[0018] The concentration test further analyzes the contribution distribution of each element in the largest eigenvector: If the contributions of almost all measurement points are relatively uniform, it indicates that the enhanced correlation originates from the common mode of the residual environment; If only a few measuring points contribute significantly, it indicates that the correlation is concentrated between specific measuring points, corresponding to the redistribution of force flow caused by local damage, thus distinguishing local damage synergy from diffuse environmental fluctuations.

[0019] Furthermore, the emergence index in step S4 is calculated as follows: The first element in the association matrix sample set... The largest eigenvalue of the correlation matrices is denoted as . Let the upper bound of the distribution be denoted as Take the corresponding values ​​from each correlation matrix. The median is used as an emergence index. The largest eigenvalue reflects the strength of the strongest association pattern in the association matrix. By comparing it with the upper bound of the random theoretical distribution and taking the median of the excess proportion, the degree of association deviation under multiple windows and multiple measures can be integrated. At the same time, the median can reduce the interference of random fluctuations in individual window data.

[0020] Furthermore, the upper bound of the distribution The dimension of the correlation matrix is ​​determined by the ratio of the number of data points in the time window used to calculate the correlation matrix. This ratio determines the theoretical upper bound of the distribution of eigenvalues ​​of the random matrix. This upper bound varies with the window length, providing an adaptive statistical benchmark for different windows.

[0021] Further, the concentration test in step S4 includes: calculating the sum of the fourth powers of each element of the maximum eigenvector as a concentration index; comparing the current concentration index with the baseline distribution of the concentration index calculated from the monitoring data during the healthy operation period of the building; if the current concentration index meets preset conditions, it is determined that a local damage association has emerged, and the measurement points whose absolute values ​​of the elements in the maximum eigenvector are greater than a set threshold are grouped into the measurement point set. The concentration index quantifies the degree of concentration of the measurement points participating in the association pattern. Comparison with the baseline during the healthy period can identify whether the current concentration has shown a statistically significant abnormal increase, thereby distinguishing between local damage synergy and diffuse environmental fluctuations.

[0022] Furthermore, the preset conditions are: the current concentration index exceeds three standard deviations of the baseline distribution during the healthy period, or the significance level obtained by the Man-Whitney U test is lower than the preset significance level. The three standard deviations method sets the anomaly judgment boundary based on the normal distribution assumption, while the Man-Whitney U test method does not depend on the distribution assumption and uses a non-parametric test to determine whether the current value and the baseline come from the same population. The two methods can be selected independently.

[0023] S5: Map the set of measuring points to the mechanical topology of the building, filter out the measuring points that are not connected in the mechanical topology by the connectivity test, obtain the effective damage association cluster, and perform time delay analysis on the residual sequence of each measuring point in the cluster to infer the crack initiation location and propagation direction of the damage.

[0024] Connectivity testing utilizes the prior topological relationships of the structural force transmission paths to verify whether the detected measuring points belong to the same mechanical force transmission system. Topologically dispersed measuring point combinations are eliminated, retaining clusters with consistent mechanical interpretations. Time-delay analysis determines the damage propagation sequence by analyzing the temporal relationship of responses between measuring points, thereby locating the crack initiation point and propagation path.

[0025] Further, the connectivity test in step S5 includes: calculating the proportion of the number of nodes contained in the largest connected subgraph of the measurement point set in the building's mechanical topology diagram to the total number of nodes in the measurement point set. If this proportion exceeds a preset proportion threshold, the measurement points corresponding to the largest connected subgraph are retained as valid damage association clusters. The largest connected subgraph reflects the degree of aggregation of the measurement point set on the force transmission path, and the proportion threshold is used to determine whether the association clusters are mechanically sufficiently cohesive.

[0026] Furthermore, the time delay analysis in step S5 includes: performing pairwise cross-correlation analysis on the residual sequences of each measuring point within the damage association cluster to determine the time delay corresponding to the maximum correlation coefficient; inferring the crack initiation location and propagation direction of the damage based on the time delay direction between each measuring point and its relative position in the building structure; and verifying the inferred propagation direction using the transfer entropy.

[0027] Time delay analysis determines the damage propagation sequence by the temporal relationship between the responses of the measurement points. Transfer entropy, based on information theory, measures the direction of information flow between the two sequences and provides independent verification for the propagation direction inferred by time delay analysis. When the results of the two are consistent, the direction is confirmed; when they are inconsistent, it is marked as pending confirmation.

[0028] S6: Generate early warning information based on the structural location and damage propagation direction corresponding to the damage association cluster.

[0029] The early warning information directly includes the inference results of the structural parts involved in the damage association cluster and the damage propagation path, providing a reference for on-site verification.

[0030] The technical effects and advantages of this invention are as follows: (1) This invention removes the synchronous drift between measuring points caused by factors such as temperature and groundwater level through environmental common mode separation, and obtains a residual sequence that reflects the structure's own response. Environmental covariance is the main source of false strong correlations across measuring points. Pre-separating it helps reduce false alarms in subsequent correlation analysis.

[0031] (2) This invention uses multiple time windows in combination with both linear and nonlinear correlation measurement methods to construct a correlation matrix sample set. Multiple windows cover different time scales, linear measurement captures the proportional coordination between measurement points, and nonlinear measurement detects non-monotonic coupling, so that there are corresponding detection channels for different rates and types of damage correlation.

[0032] (3) This invention utilizes the theoretical bound of the eigenvalue distribution of a purely random correlation matrix in random matrix theory. The degree to which the largest eigenvalue of the correlation matrix exceeds the upper bound is used as an emergence index to determine whether the correlation between measurement points has transitioned from a random state to an organized state. The entire process does not rely on historical damage annotation data, thus avoiding the difficulty of model training due to the scarcity of damage samples.

[0033] (4) This invention, in conjunction with concentration testing, distinguishes between global environmental residues and local damage synergy. When the emergence exceeds the limit, the energy concentration of the largest eigenvector is further examined, and only the case of highly coupled a few measuring points is judged as a damage signal, while diffuse environmental residue fluctuations are excluded, which helps to reduce the false alarm rate.

[0034] (5) This invention introduces a building mechanics topology map to verify the connectivity of the detected associated measurement point set, filters out mechanically unrelated measurement point combinations, and infers the crack initiation location and propagation direction of the damage through time delay analysis and propagation entropy verification, so that the early warning result is accompanied by information on the damage location and propagation path, providing a reference for subsequent on-site verification. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the overall execution process of the method of the present invention; Figure 2 This is a flowchart of the core process of correlation emergence detection in this invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1 As attached Figures 1 to 2 The multi-source data fusion-based intelligent early warning method for building safety supervision, as shown, shifts the focus of early warning from the numerical value of a single measuring point to the changes in the correlation between multiple measuring points. By detecting the emergence process of this correlation from non-existent to existing and from weak to strong, earlier damage warnings are achieved. The specific implementation of the present invention will be described in detail below with reference to steps S1 to S6.

[0038] S1. Obtain monitoring data and building mechanical topology diagram. Multiple types of sensors were deployed at key load-bearing locations in the building. Crack monitoring sensors were installed at the base of load-bearing walls where cracks were prone to occur, tilt monitoring sensors were installed in the middle of load-bearing walls, settlement monitoring sensors were installed at the four corners of the foundation, and vibration monitoring sensors were installed in the middle of the floor slabs. All sensors collected data at a uniform sampling frequency with a sampling interval of 1 hour, and the raw data were aggregated and sent to the analysis server.

[0039] In one implementation method, only crack monitoring sensors and tilt monitoring sensors are used to obtain the crack width change and wall tilt angle. The crack width change is the crack width value at the current sampling time minus the crack width value at the previous sampling time, and the wall tilt angle is the angle between the wall surface normal direction and the vertical direction.

[0040] As another implementation, settlement monitoring sensors are added to the aforementioned data to obtain data including local settlement difference. Local settlement difference is the difference between the readings of settlement monitoring sensors at different foundation locations.

[0041] In another implementation method, all four types of sensors are used to acquire changes in crack width, wall tilt angle, local settlement difference, and structural vibration dominant frequency shift. The structural vibration dominant frequency shift is obtained by acquiring the environmental vibration acceleration response of the floor slab and extracting the dominant frequency change. Any combination of at least two of the above four sensors can be implemented independently.

[0042] The raw monitoring data were time-aligned and standardized. Time alignment unified the data from all sensors to the same hourly time. If no corresponding data was collected at a certain hour, the linear interpolation of the two most recent data collected before and after that hour was used to complete the data. Time periods with continuous missing data exceeding 24 hours were marked as missing data segments and were not included in subsequent analysis.

[0043] Z-score standardization was used. For each time series, its mean was calculated. and standard deviation Then each value in the sequence Transform into This standardizes the series so that the mean becomes 0 and the standard deviation becomes 1. The standardized series constitute a multi-source building structure response time-series data matrix. The rows of the matrix correspond to Each monitoring indicator is listed in the corresponding column. Each observation time.

[0044] Simultaneously with sensor deployment, a building mechanics topology diagram is created. The installation location of each sensor is abstracted as a node in the diagram. If two measuring points are located at corresponding positions above and below the same load-bearing wall, at opposite ends of the same frame beam, or at a directly connected beam-column node, an undirected edge is added between the corresponding node pairs. These spatial relationships can be obtained from the axial positioning information in the building structural design drawings, or determined by obtaining the three-dimensional coordinates between the measuring points through on-site laser ranging scanning. The resulting topology diagram... It only records the mechanical connectivity between measuring points and does not include quantitative mechanical parameters.

[0045] S2. Environmental Common Mode Separation Environmental factors such as temperature and groundwater level can cause synchronous drift at all measuring points. Temperature changes lead to thermal expansion and contraction of building materials, resulting in seasonal fluctuations in crack width and wall tilt angle; changes in groundwater level alter the soil moisture content, causing settlement at each foundation measuring point to rise or fall synchronously. If these synchronous changes caused by environmental factors are not separated, they will form spurious strong correlations in subsequent correlation analysis.

[0046] As one implementation method, environmental common-mode separation is achieved through principal component analysis. This involves analyzing the standardized matrix. Calculate the covariance matrix and perform eigenvalue decomposition to obtain the principal components and their variance contribution rates, arranged in descending order of eigenvalues. The first principal component with the largest variance contribution rate is taken as the environmental common-mode component. This component reflects the overall synchronous drift trend of all measuring points caused by environmental factors.

[0047] For the Time series of monitoring indicators Calculate its in Projection coefficients on , for and inner product divided by Its own inner product. From Subtract The structural self-response residual sequence is obtained. For all This operation is performed on each monitoring indicator, and the total residual sequences form a residual matrix. .

[0048] As another implementation method, when the ratio of the variance contribution rates of the first principal component to the second principal component is less than 3, it indicates that the second principal component also has a non-negligible explanatory power. In this case, the temperature-related common mode component and the groundwater level-related common mode component are extracted from the first two principal components, respectively. The projections on these two components are subtracted from each monitoring index sequence, and the residuals are used to construct a matrix. .

[0049] As another implementation method, if groundwater level monitoring sensors are separately installed on-site, then linear regression is first performed on each monitoring index sequence with the groundwater level sequence as the independent variable. After subtracting the regression estimate, principal component analysis is then performed to obtain the residual matrix. .

[0050] S3. Construct the correlation matrix sample set After environmental common-mode separation, at the current time when an early warning is needed, from the residual matrix The system extracts corresponding residual sequence segments according to each time window, automatically skipping marked missing data segments during extraction, and constructs a sample set of the correlation matrix. The current time for triggering an early warning is set to 0:00 daily.

[0051] First, set the main time window length. The selection of the main time window length depends on the typical time scale of the damage evolution of the monitored building structure. For structural types with slower damage evolution, a longer window is used to cover a sufficient evolutionary process, while for structural types with faster damage evolution, a shorter window is used to maintain sensitivity to recent changes.

[0052] As one implementation method, for masonry structures, the cycles of crack propagation and wall tilting are typically measured in months. Take 60 days; As another implementation method, steel-structured houses exhibit a relatively rapid rate of damage evolution. Take 30 days; As another implementation method, for large-span spatial structures, the damage evolution is slower and more significantly affected by annual environmental cycles. Take 90d.

[0053] Based on the main time window, multiple sub-time windows are obtained by shrinking the window at different ratios. Changes in the correlation matrix within a short window reflect recent changes in the correlation structure, while changes in the correlation matrix within a long window reflect correlation trends over a longer time scale. As one implementation method, five scaling factors of 0.6, 0.7, 0.8, 0.9, and 1.0 are used to... Multiply by these 5 values ​​to obtain the lengths of the 5 sub-time windows, and so on. Taking d as an example, the lengths of the five sub-windows are 36d, 42d, 48d, 54d, and 60d, respectively. As another implementation, three scaling factors are used: 0.7, 0.8, and 1.0. All scaling factors are greater than 0 and less than 1.

[0054] Each sub-window should contain more than [number of data points]. The number of data points equals the number of days in the sub-window. If the number of data points in the sub-window corresponding to a certain scaling factor is not greater than... If the sample size of the sub-window is insufficient to support a stable association estimate, the scaling factor should be discarded.

[0055] For the selected Each proportional factor corresponds to Each sub-window uses two different association metrics to calculate the association matrix.

[0056] The first method is to use the Pearson correlation coefficient. For any two measurement points within the window... and The residual sequence segment is calculated. As a linear incidence matrix The Middle Line 1 The element values ​​of the column. The range of values ​​for this measure is... Positive values ​​indicate that the two measuring points change in the same direction, while negative values ​​indicate that they change in opposite directions, capturing the linear and consistent response between the measuring points.

[0057] The second method is to measure the distance correlation coefficient. For any two measurement points within the window... and Calculate the distance correlation coefficient from the residual sequence segments. As a nonlinear correlation matrix The Middle Line 1 The element values ​​of the column. The value range of this metric is [0,1]. A value of 0 indicates that the two sequences are statistically independent, and a value greater than 0 indicates that there is a statistical dependency. It can detect non-monotonic and nonlinear coupling between measurement points.

[0058] Both types of correlation matrices are A symmetric matrix has diagonal elements of 1. When calculating, only the upper or lower triangular part needs to be calculated due to the symmetry.

[0059] Will Each sub-window corresponds to and There are a total of 2m correlation matrices, ordered by the size of the sub-windows from smallest to largest, and within the same sub-window first... back The order of these elements forms the correlation matrix sample set.

[0060] S4. Emergence Calculation and Emergence Recognition of Local Damage Association For each association matrix in the association matrix sample set, perform eigenvalue decomposition and extract its maximum eigenvalue. ,in .

[0061] We use random matrix theory to determine whether these eigenvalues ​​significantly exceed the range of random fluctuations. When multiple time series are independent of each other and contain only random noise, the eigenvalue distribution of their sample correlation matrix follows... The distribution has a definite theoretical upper bound. Anything beyond The eigenvalues ​​indicate the presence of structural associations in the data that cannot be explained by randomness.

[0062] When each sequence has been Z-score standardized to a variance of 1 The dimension of the correlation matrix The number of data points in the window used to calculate this matrix. Sure, ,in Equal to the number of days in the corresponding sub-window .by , For example, As another implementation, if the variance was not normalized to 1 during standardization, then the mean of the variances of all residual sequences is taken. ,Will Multiply the result of the above formula as .

[0063] For the Calculate the number of correlation matrices. Take all 2m. The median is used as the emergence index at the current moment. ,Right now When 2m is an even number, the median is the arithmetic mean of the two middle values. Using the median reduces the interference from occasional fluctuations in individual window data.

[0064] When the structure is healthy, the correlation between the measurement points is close to a random state. The value is low; when damage causes coordinated changes at some measuring points, The value increases accordingly. Set an emergence threshold. As one implementation method, Set to 0.3, when Then proceed to the concentration test.

[0065] The concentration test is used to distinguish between two types of association reinforcement that differ in nature. When... Exceed Two scenarios may occur. The first is that the residual environmental common modes are not completely separated, with almost all measurement points participating in the associated mode, resulting in a relatively uniform contribution from each element in the maximum eigenvector. The second is force flow redistribution caused by local damage, where only a few measurement points located on the same force transmission path form tight coupling, and the energy of the eigenvector is highly concentrated on these measurement points. Only the second scenario represents true damage coordination.

[0066] The concentration test process is as follows: when Exceed When, select one of them The largest eigenvector corresponding to the window with the largest value Its elements are denoted as , Calculate the concentration index . The range of values ​​is When all measuring points contribute equally near When energy is highly concentrated at a few measuring points Approaching 1. For example, when uniformly distributed It is approximately 0.125.

[0067] Establish a baseline distribution for the building's healthy operating period. The healthy operating period refers to the initial operating phase after the building is put into monitoring and confirmed to have no structural damage. During this phase, the same window configuration and correlation measurement method as online monitoring are used, with daily sliding calculations. The values ​​are accumulated to form a baseline distribution sample. The baseline distribution is re-established after known repairs or reinforcements have been performed on the building structure to reflect the normal correlation level of the structure after the repairs.

[0068] As one implementation method, the mean of the baseline distribution sample is calculated. and standard deviation ,like If the concentration is significantly increased, it is determined that the concentration is in the baseline distribution. Taking a mean of 0.2 and a standard deviation of 0.05 as an example, the critical value for judgment is 0.35.

[0069] As another implementation method, for the current Perform the Mann-Whitney U test on the values ​​and baseline distribution samples. If the obtained significance level is... A value below 0.01 indicates a significantly increased concentration. The two determination methods can be used independently; one can be selected based on actual needs.

[0070] Once the judgment is established, Measurement points whose absolute value is greater than a set threshold are selected and formed into a measurement point set. In one implementation, the threshold is set to 0.3; in another implementation, the threshold is set to 0.25. Measurement point set. These are the core measurement points involved in this localized damage-related emergent event.

[0071] S5. Topology Verification and Time Delay Analysis Measurement point set There is anomaly in the synchronization of the measuring points at the data level, which is confirmed by the building mechanics topology diagram established with S1. A comparison was conducted to verify its mechanical rationality.

[0072] In the topology diagram Extract only those containing The derived subgraphs of the middle nodes are found using breadth-first search to identify the maximum connected subgraph. The number of nodes in the maximum connected subgraph is then calculated. occupy The proportion of the total number of nodes |S| .

[0073] As one implementation method, the ratio threshold is set to 0.6. When When, it indicates Most of the abnormal measurement points belong to the same force transmission system in terms of mechanical topology, and the measurement points contained in this maximum connected subgraph are considered as effective damage correlation clusters. For example, If a maximum connected subgraph covers 4 of the 5 measurement points, If the maximum connected subgraph only covers two of them, then these four measurement points will be retained for subsequent analysis; If the abnormal measurement points are relatively dispersed mechanically, the event will be marked as a warning signal instead of triggering an early warning.

[0074] As another implementation method, the ratio threshold is set to 0.5, which is suitable for situations where the sensors are sparsely distributed and the number of measuring points on the same force transmission path is small.

[0075] After obtaining effective damage correlation clusters, the initiation location and propagation direction of the damage are inferred through time delay analysis. For the residual sequences of each measurement point within the cluster, the residual sequence segments corresponding to the main time window used for emergence detection are extracted, and pairwise cross-correlation analysis is performed.

[0076] For any two measuring points and Fixed measuring points The residual sequence of the measurement points The residual sequence is slid forward and backward along the time axis, with each sliding step being one sampling interval. The cross-correlation coefficient is calculated at each step. The step size that maximizes the cross-correlation coefficient is searched, and this step size is the measurement point. Relative to the measuring point latency ,Right now A positive time delay indicates the measurement point is... Leading the measurement points A negative time delay indicates that the measurement point... Lagging behind the measuring point The number of sampling intervals for the delay. h yields the actual lead time in units of h.

[0077] By combining the lead-lag relationship between each measuring point with their spatial location within the building, the direction of damage propagation can be inferred: When a measuring point located at a lower part of the structure precedes a measuring point located at a higher part, it is inferred that the damage develops from the bottom upwards. When there is a sequential relationship between measuring points at the same horizontal level, it is inferred that the damage propagates in the horizontal direction; When time delay conflicts occur, the direction inference is based on the pair of relations with the largest mutual relation number.

[0078] The above inferences were verified using transfer entropy, which was calculated using the same residual sequence segment as in time-delay analysis. For each pair of measurement points within the cluster... and ,calculate and Compare the sizes of the two. If Then the net direction of the information flow is from arrive .

[0079] Compare the asymmetric direction of the transfer entropy with the propagation direction inferred from time delay analysis: If the two are consistent, then the direction of propagation is confirmed; If the two are inconsistent, the direction of propagation will be marked as pending confirmation, and the warning level of the event will be lowered accordingly. If the difference in propagation entropy between the two directions is less than 10% of the smaller value, the direction is considered uncertain and is also marked as pending confirmation.

[0080] S6. Generate early warning information Based on the effective damage association clusters and their analysis results, early warning information containing early warning level and associated measurement point information is generated.

[0081] The warning level is determined by the emergence index. and connectivity ratio Determined comprehensively. Establish hierarchical classification rules in advance, for example... and It was determined to be a red alert at that time. and A situation that meets the criteria but not the above standards is designated as a yellow alert; other situations that meet the criteria but not the above standards are designated as a warning signal. The above thresholds can be adjusted according to the importance level of the building.

[0082] The associated measurement point information is obtained through a pre-established mapping table between sensor numbers and structural locations. This table records the correspondence between each sensor number and the structural location where it is installed. The output includes a textual description of the structural location of each measurement point in the effective damage association cluster, the lead-hysteresis relationship between measurement points, the damage initiation location, and the propagation direction. Based on the verification results of the propagation entropy, the propagation direction description is marked as either verified or pending confirmation.

[0083] The warning information is sent to building safety management personnel through a display interface or push notification. The information includes the warning level identifier, a list of associated monitoring points, the corresponding structural location, the inference of the direction of damage propagation, and the verification status.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent early warning of housing safety supervision through multi-source data fusion, characterized in that, Includes the following steps: S1. Acquire monitoring data collected by multiple types of sensors deployed at different stress locations on the building structure, perform time alignment and standardization on the monitoring data, and form multi-source building structure response time-series data. And obtain the building mechanics topology map constructed with each monitoring location as a node and the node pairs connected by direct structural force transmission paths as edges; S2. Perform environmental common mode separation on the multi-source building structure response time series data, remove synchronous change components between measuring points caused by environmental factors, and obtain the structural response residual sequence. S3. At the time of warning, based on the preset main time window and multiple sub-time windows obtained by shrinking the main time window at different ratios, the correlation matrix between the measurement points under each window is calculated using linear correlation measurement method and nonlinear correlation measurement method respectively, forming a correlation matrix sample set. S4. Perform feature decomposition on each association matrix in the association matrix sample set. Based on the upper bound of the distribution of eigenvalues ​​of pure random association matrices in random matrix theory, calculate the emergence index that characterizes the degree of deviation of the association structure between measurement points from randomness. When the emergence index exceeds a preset threshold, perform a concentration test on the largest eigenvector to identify the set of measurement points where local damage associations emerge. S5. Map the set of measuring points to the mechanical topology of the building, filter out the measuring points that are not connected in the mechanical topology by the connectivity test, obtain the effective damage association cluster, and perform time delay analysis on the residual sequence of each measuring point in the cluster to infer the crack initiation location and propagation direction of the damage. S6. Generate early warning information based on the structural location and damage propagation direction corresponding to the damage association cluster.

2. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 1, characterized in that, The multiple types of sensors include at least two of the following: crack monitoring sensors, tilt monitoring sensors, settlement monitoring sensors, and vibration monitoring sensors; The multi-source building structure response time series data includes at least two of the following: crack width change, wall tilt angle, local settlement difference, and structural vibration dominant frequency offset.

3. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 1, characterized in that, The environmental common-mode separation in step S2 includes: Principal component analysis was performed on the time series data of the multi-source building structure response, and the first principal component with the largest variance contribution was taken as the environmental common mode component. Subtract the projection of each time series data onto the common mode component of the environment to obtain the structural response residual sequence; The environmental factors include at least one of temperature and groundwater level.

4. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 1, characterized in that, The linear correlation metric in step S3 is the Pearson correlation coefficient, and the nonlinear correlation metric is the distance correlation coefficient. The multiple sub-time windows are obtained by multiplying the main time window by a set of preset scaling factors, where each scaling factor is greater than 0 and less than 1.

5. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 1, characterized in that, The emergence index mentioned in step S4 is calculated as follows: The first in the correlation matrix sample set The largest eigenvalue of the correlation matrices is denoted as . Let the upper bound of the distribution be denoted as , Take the corresponding values ​​of each correlation matrix The median is used as an emergence indicator.

6. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 5, characterized in that, The upper bound of the distribution It is determined by the ratio of the dimension of the correlation matrix to the number of data points in the time window used to calculate the correlation matrix.

7. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 1, characterized in that, The concentration test in step S4 includes: The sum of the fourth powers of each element in the largest eigenvector is used as a concentration index. The current concentration index is compared with the baseline distribution of the concentration index calculated from the monitoring data during the healthy operation period of the building. If the current concentration index meets the preset conditions, then it is determined that a local damage-related emergence has occurred. The measurement points that have an absolute value greater than a set threshold in the largest eigenvector are then grouped into the measurement point set.

8. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 7, characterized in that, The preset conditions are: The current concentration index exceeds three standard deviations of the baseline distribution during the healthy period, or the significance level obtained by the Mann-Whitney U test is lower than the preset significance level.

9. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 1, characterized in that, The connectivity test in step S5 includes: Calculate the proportion of the number of nodes in the largest connected subgraph of the measurement point set in the building mechanics topology graph to the total number of nodes in the measurement point set. If the ratio exceeds the preset ratio threshold, the measurement point corresponding to the largest connected subgraph is retained as an effective damage association cluster.

10. The intelligent early warning method for housing safety supervision based on multi-source data fusion according to claim 1, characterized in that, The time delay analysis described in step S5 includes: Perform pairwise cross-correlation analysis on the residual sequences of each measurement point within the damage correlation cluster to determine the time delay corresponding to the maximum correlation coefficient; Based on the time delay direction between each measuring point and its relative position in the building structure, the initiation location and propagation direction of the damage can be inferred. The propagation direction was then verified using the propagation entropy.

Images