Construction background stratum concealed disease early warning system based on multi-source fusion

Abstract

The present application relates to computer data processing, space-time information fusion and concealed disease early warning technology, specifically to a structure background stratum concealed disease early warning system based on multi-source fusion. The system obtains structure monitoring, geological survey, deformation detection and geophysical data, constructs an observation support domain containing spatial range, detection depth and time lag, and establishes a layered space-time calculation domain according to the structure and background stratum interface; according to the stratum continuity relationship and interface compatibility relationship, a physical consistency constraint is formed, combined with residual stability, cross-source mutual verification degree, data timeliness and historical drift state to generate a dynamic credibility tensor, through cross-source evidence map and leave-one-source reconstruction bias iteration to obtain disease hidden state field and uncertainty field, and thus determine the early warning area. The present application can correct the space-time difference of multi-source data, suppress the deviation of low credibility data and the overlap of false evidence, maintain the continuity of disease position, range and horizon, and realize early warning basis tracing.

Description

Technical Field

[0001] This invention relates to computer data processing, spatiotemporal information fusion, and hidden disease early warning technology, specifically a hidden disease early warning system for the background strata of buildings based on multi-source fusion. Background Technology

[0002] Early warning systems for hidden geological defects in structures typically rely on geological survey data, structure deformation monitoring data, geophysical testing results, historical construction records, and maintenance records. Existing systems generally convert borehole-revealed stratigraphic information, settlement or displacement time series, detected profile anomalies, and historical data into data points under a unified coordinate system, and then establish a regular grid based on the structure's foundation location and stratigraphic distribution. For data with different spatial sampling locations, neighbor interpolation, distance weighting, or surface fitting methods are typically used to complete the grid values; for data with different sampling periods, time resampling, moving averages, or window statistics are used to form a unified time series. After completing spatiotemporal alignment, the system normalizes various data types, composing feature vectors from indicators such as anomaly amplitude, rate of change, cumulative amount, and spatial gradient. Then, weighted summation, feature stitching, cluster analysis, or classification models are used to obtain an anomaly score for each grid. When the anomaly score continuously exceeds a set condition, the system marks the corresponding grid as a suspected loosening, voiding, seepage anomaly, or local settlement area, and generates an early warning result based on the structure's impact range.

[0003] In relatively similar multi-source fusion schemes, data from different sources are typically assigned fixed weights according to a preset confidence level, or the fusion coefficient is determined based on the recognition accuracy in the training samples. Geological exploration data serves as the basic constraint for stratigraphic boundaries and media categories, deformation monitoring data reflects the response changes of structures and the surface, geophysical detection results are used to supplement the distribution of underground anomalies, and historical construction and maintenance records are used to correct the prior probabilities of local areas. After mapping various types of data to the same grid, the system performs summation, voting, or probability synthesis on the multi-source features at the same location, and uses spatial smoothing constraints to reduce isolated anomalies. Some schemes also include a data quality evaluation process, adjusting weights based on the proportion of missing data, noise levels, and acquisition time, and using other sources to complete the calculation when one source is missing. The calculation object of this type of method is still mainly a unified grid value, assuming that different data have the same spatial orientation and temporal meaning within the same grid. Physical constraints are often added to the objective function as smoothing terms, neighborhood consistency terms, or boundary penalty terms. The fusion result is generally represented as a static anomaly distribution map or a graded risk map.

[0004] When multi-source data is mapped to a unified grid, its actual observation range, effective detection depth, sampling interval, response delay, and detection mechanism are not considered as components of the fusion calculation. Consequently, values ​​from different sources within the same grid may correspond to different volumes, stratigraphic levels, or time periods of geological conditions. Fixed weights or weights determined based on overall data quality are insufficient to reflect drift, distortion, or conflicts with other sources occurring in local areas or specific time windows. The resulting fusion process may treat non-overlapping observational evidence as mutually corroborating, or propagate biases from low-confidence data to adjacent grids, spreading during spatial smoothing. When resolution differences, time lags, or local conflicts exist between data points, the system struggles to distinguish between the actual disease state and data mapping errors within the unified computational domain. This leads to reconstructed locations, extents, and continuity of hidden diseases deviating from the actual background geological conditions, resulting in a lack of stable data support for the warning area. Summary of the Invention

[0005] The purpose of this invention is to provide an early warning system for hidden geological defects in buildings based on multi-source fusion, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A multi-source fusion-based early warning system for hidden geological defects in building background strata includes a data access processor, a spatiotemporal modeling processor, and a defect reconstruction processor.

[0008] The data access processor acquires data from structure monitoring, geological exploration, deformation detection, and geophysical detection, and generates an observation support domain that includes spatial coverage, detection depth, sampling interval, and time lag.

[0009] The spatiotemporal modeling processor establishes a layered spatiotemporal computing domain based on the interface between the structure and the background strata, sets hidden state variables of the disease in it, and generates physical consistency constraints based on the stratum continuity relationship and interface compatibility relationship.

[0010] The disease reconstruction processor forms a dynamic credibility tensor based on local residual stability, cross-source mutual verification degree, data timeliness, and historical drift state. It establishes a cross-source evidence map based on the overlapping relationship of the observation support domain, identifies conflicting data by combining physical consistency residuals and single-source support domain deviations, and iteratively updates the disease latent state field and uncertainty field under the constraints of the dynamic credibility tensor and physical consistency. The warning area is determined based on the update results.

[0011] Preferably, the spatiotemporal modeling processor establishes an observation support domain object for each type of data, consisting of a central location, spatial influence boundary, effective detection depth, time action interval, and observation response kernel, and projects the observation support domain object onto the layered spatiotemporal computing domain according to the background stratigraphic coordinates and the structure interface coordinates.

[0012] For any two observed support domain objects after projection, calculate the spatial overlap ratio, depth overlap ratio, temporal overlap ratio, and response kernel similarity, and combine each ratio with the response kernel similarity to form a support domain correlation matrix;

[0013] The disease reconstruction processor establishes cross-source evidence edges in overlapping observation ranges based on the support domain correlation matrix, retains local evidence nodes in non-overlapping observation ranges, and constrains the update of the disease latent state variables through the cross-source evidence edges and the local evidence nodes.

[0014] Preferably, the spatiotemporal modeling processor generates a layered topology map based on the stratigraphic boundaries, the influence zone of the foundation of the structure, and the existing anomaly area. The nodes of the layered topology map correspond to the computational subdomains in the layered spatiotemporal computational domain. The node attributes include the stratigraphic category, the medium continuity identifier, and the boundary status. The edge attributes include the adjacency type, the interface direction, and the cross-layer transfer permission.

[0015] Based on the layered topology diagram, construct continuous constraints within the layer, cross-layer interface compatibility constraints, and structure contact boundary constraints, and convert each constraint into a sparse constraint operator that acts on the latent state variables of the disease.

[0016] In each round of graph constraint iteration, the disease reconstruction processor first updates the hidden state of nodes based on observation evidence, and then uses the sparse constraint operator to decompose and write back the update components that do not satisfy the intra-layer continuity relationship, cross-layer interface relationship or contact boundary relationship.

[0017] Preferably, the disease reconstruction processor maintains and calculates the credibility status corresponding to each data source subdomain and time window, and forms a residual stability component based on the residual sign consistency and residual volatility within the continuous window, forms a cross-source mutual verification component based on the consistency of evidence direction with other sources within the overlapping support domain, forms a timeliness component based on the interval between the acquisition time and the current reconstruction time, and forms a drift component based on the distribution migration between the historical benchmark segment and the current data segment.

[0018] The residual stable components, cross-source mutual verification components, time-dependent components, and drift components are arranged according to their source, spatial dimension, and time dimension to form the dynamic credibility tensor. The contribution of the observation residuals from each source to the update of the latent state variable of the disease is modulated domain by domain through the dynamic credibility tensor.

[0019] Preferably, the spatiotemporal modeling processor performs adaptive subdivision of the hierarchical spatiotemporal computational domain based on the support domain correlation matrix, dividing regions with dense cross-source connections and inconsistent main directions or attenuation patterns of observation response kernels into multiple sub-computational domains, and maintaining regions with sparse cross-source connections and continuous observations from the same source as joint computational domains.

[0020] A fine-grained projection index is established for each sub-computation domain, and an aggregated projection index is established for the joint computation domain. The conservation mapping relationship between the fine-grained projection index and the aggregated projection index is recorded.

[0021] The disease reconstruction processor uses the conservation mapping relationship to transfer observation residuals and hidden state increments between different spatial resolutions, and normalizes and allocates observation residuals and hidden state increments according to the effective overlap volume between each observation support domain and the corresponding computation domain during the transfer process.

[0022] Preferably, the spatiotemporal modeling processor expands the hierarchical topology graph into a two-layer graph consisting of an observation node layer and a constraint node layer. The nodes in the observation node layer carry multi-source evidence for each computational subdomain, and the nodes in the constraint node layer carry the intra-layer continuous constraints, cross-layer interface compatibility constraints, and structure contact boundary constraints, respectively. The two node layers are connected by constraint action edges.

[0023] Constraint activation sequences are generated based on the number of observation nodes associated with each constraint node, the degree of conflict in evidence directions, and the historical satisfaction status. The corresponding sparse constraint operators are then selectively invoked according to the constraint activation sequences.

[0024] The disease reconstruction processor maps the hidden state update of the observation node layer to the constraint node layer. After processing by the activated sparse constraint operator, it writes back to the associated observation node according to the constraint action edge.

[0025] Preferably, the disease reconstruction processor establishes drift latent variables for each data source, and combines the drift latent variables with the credibility status of the corresponding source and the disease latent state variables to form a joint state vector;

[0026] Within each time window, the drift latent variable is first updated based on the quantile mapping deviation between the historical benchmark segment and the current data segment. Then, the residual stable component and cross-source mutual verification component are corrected based on the drift latent variable. Subsequently, the dynamic credibility tensor is updated with the corrected residual stable component and cross-source mutual verification component.

[0027] The disease reconstruction processor solves for the disease latent state increment based on the updated dynamic credibility tensor, and feeds back the prediction residual generated by the disease latent state increment to the drift latent variable, so that the drift latent variable, dynamic credibility tensor and disease latent state variable are updated alternately according to the same time window.

[0028] Preferably, the disease reconstruction processor constructs a set of source subsets for computational subdomains with inconsistent cross-source evidence directions, and performs disease hidden state reconstruction on each source subset to obtain the corresponding subset hidden state field and reconstruction residual field.

[0029] A conflict topology graph is formed based on the spatial offset, morphological differences, and physical constraint violation degree of each reconstructed residual field between the hidden state fields of each subset. The nodes of the conflict topology graph correspond to the source subsets, and the edges correspond to the conflict associations between two source subsets.

[0030] Extract common conflict components that cannot be eliminated by a single source and source-specific components that disappear with the elimination of a specific source from the conflict topology graph. Write the common conflict components into an independent anomaly evidence channel and write the source-specific components into the credibility status update channel of the corresponding source.

[0031] Preferably, the disease reconstruction processor organizes the joint state vector into a sliding state chain in the time dimension, and establishes drift continuation edges, credibility inheritance edges, and hidden state evolution edges between adjacent time windows;

[0032] The drift continuation edge constrains the direction of change of latent drift variables from the same source in adjacent windows;

[0033] Based on the aforementioned credibility inheritance edge, the credibility state after source mutual verification is passed; based on the aforementioned hidden state evolution edge, the disease hidden state increment that satisfies the hierarchical topological relationship is passed.

[0034] When the prediction residual is interpreted by both the drift latent variable and the disease latent state variable in a continuous window, the drift latent variable and the disease latent state variable are frozen respectively, and two local reconstructions are performed. The corresponding drift continuation edge and latent state evolution edge are updated according to the difference in the residual structure of the two local reconstructions, and then the sliding state chain is jointly smoothed and solved.

[0035] Preferably, the disease reconstruction processor organizes the common conflicting components in the independent anomaly evidence channels into a conflict evidence chain by interleaving and continuing them according to the time progression relationship, and generates a combination of source subsets of common conflict classes, corresponding observation support domains, and physical constraint violation locations for each conflict evidence chain;

[0036] In subsequent time windows, directional backtracking is performed on new observation evidence that overlaps with the supporting domain of the conflicting evidence chain, and the directional consistency relationship between the new observation evidence and the source subset combination, the topological adjacency relationship with the physical constraint violation position, and the temporal continuity relationship with the conflicting evidence chain are calculated respectively.

[0037] The conflict evidence chain is updated based on the direction consistency relationship, topological adjacency relationship and time continuity relationship, and the updated common conflict component is used as an additional observation term to participate in the joint reconstruction of the disease hidden state field and uncertainty field.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. Using the observation support domain as the basic object for multi-source data fusion computation, the spatial coverage boundaries, effective detection depths, sampling intervals, temporal action intervals, and observation response relationships of various observations are jointly mapped to the hierarchical spatiotemporal computation domain. This ensures that cross-source evidence connections are established only when there is a correspondence between the observation range and temporal meaning of different sources. The hierarchical topology map and its intra-layer continuity constraints, cross-layer interface compatibility constraints, and structure contact boundary constraints limit the update range of the latent state of the disease. The dynamic credibility tensor, composed of the residual stable state, cross-source mutual verification state, data timeliness state, and historical drift state, modulates the contribution of different sources within specific regions and time windows. Source subset reconstruction, conflict topology analysis, and leave-one-out source comparison are used to retain common conflicts and weaken source-specific biases. The above processing forms a reconstruction process that integrates observation range correction, local credibility control, and physical relationship constraints. This enables multi-resolution, multi-temporal, and locally conflicting data to participate in the calculation under consistent spatial and temporal semantics, reducing false overlap, error diffusion, and abnormal over-smoothing. It yields a disease latent state field and early warning area consistent with the background stratigraphic relationship and the interface relationship of structures.

[0040] 2. Source contribution information and uncertainty information are generated simultaneously outside the latent state field of the disease. The correspondence between each observation evidence and the computational subdomain is recorded through the support domain correlation matrix, enabling the early warning area to be traced back to the specific data source, observation range, and time window. The computational domain is adaptively subdivided based on the cross-source connectivity density and observation response differences. A finer spatial representation is retained in complex overlapping areas, while a joint computational domain is used in areas with continuous co-source connections and sparse cross-source connections. The observation residuals and latent state increments are transferred using the conservation mapping relationship, which avoids the loss of local details and unnecessary computational expansion caused by a unified grid. Observation nodes and constraint nodes in the two-layer graph participate in iteration according to the constraint activation sequence, reducing the interference of irrelevant constraints on the local state. Drift latent variables, dynamic confidence tensors, and disease latent state variables are updated alternately according to the time window, which can incorporate continuous data source shifts and stratigraphic state changes into the joint state estimation. Conflict evidence chains and subsequent directional back-evidence continuously retain and update data differences that cannot be immediately attributed, maintaining the continuity and verifiability of the early warning results under conditions of data loss, source drift, and new observations. Attached Figure Description

[0041] Figure 1The flowchart shows the overall processing flow of the early warning system for hidden defects in the background strata of buildings based on multi-source fusion.

[0042] Figure 2 Flowchart for the construction, correlation calculation, and adaptive partitioning of the hierarchical spatiotemporal computation domain for multi-source observation support domain;

[0043] Figure 3 This is a residual decomposition diagram for physical consistency constraints.

[0044] Figure 4 A flowchart for the joint reconstruction of dynamic credibility modulation, physical consistency constraints, source conflict identification, and hidden state of disease;

[0045] Figure 5 Time series plot of deviation index contributing to dynamic credibility;

[0046] Figure 6 A horizontal bar chart showing the source conflict identification quantity;

[0047] Figure 7 A target decomposition diagram is reconstructed by combining the latent state of the disease and its uncertainty.

[0048] Figure 8 The target decomposition map is updated alternately for source drift. Detailed Implementation

[0049] In one embodiment, please refer to Figure 1The multi-source fusion-based early warning system for hidden defects in the background strata of structures includes a data access processor, a spatiotemporal modeling processor, and a defect reconstruction processor. The data access processor receives monitoring data of the structure, geological survey data, deformation detection data, geophysical detection data, and historical construction and maintenance data corresponding to the current calculation area. For each input record, it retains the data source identifier, acquisition time, spatial coordinates, stratum affiliation, sampling interval, effective detection depth, observation direction, and original quality identifier. Then, it transforms the data from different coordinate references to a reference coordinate system jointly determined by the structure foundation outline and the background stratum layering interface. The spatiotemporal modeling processor does not directly write the transformed data into a unified regular grid. Instead, it generates an observation support domain based on the spatial coverage, detection depth, sampling interval, and time lag of various observations, and establishes a layered spatiotemporal computation domain based on the interface between the structure and the background stratum. In the empty computational domain, latent state variables representing the degree of loosening, voiding, seepage disturbance, local settlement trend, and stress anomaly are set. At the same time, physical consistency constraints are generated based on the stratigraphic continuity and interface compatibility. The disease reconstruction processor forms a dynamic credibility tensor based on local residual stability, cross-source mutual verification degree, data timeliness, and historical drift state. It establishes a cross-source evidence map based on the overlap relationship between observation support domains and identifies conflicting data by combining physical consistency residuals and one-source reconstruction deviation. Then, under the joint constraints of the dynamic credibility tensor and physical consistency constraints, the latent state field and uncertainty field of the disease are iteratively updated. Finally, the warning area is determined based on the spatial connectivity range, temporal duration, and distribution of the uncertainty field of the latent state field of the disease. This makes the entire calculation process a closed-loop processing chain from original observation, support domain expression, layered state reconstruction, conflict identification, uncertainty quantification to warning area output.

[0050] In this embodiment, the data access processor performs timestamp correction on the settlement, horizontal displacement, tilt, crack changes, and strain time series in the structure monitoring data; performs spatial entityization on the stratigraphic interfaces, medium types, borehole depths, and local adverse geological descriptions in the geological exploration data; establishes a unified geometric index for discrete measuring points, profile lines, and planar deformation results in the deformation detection data; retains the original detection direction and effective influence range for profile anomalies, velocity differences, dielectric response differences, or wave impedance differences in the geophysical detection data; and extracts the excavation disturbance range, backfill area, reinforcement area, and effective time interval for historical construction and maintenance data. All input data are written into an observation record table with source, spatial, and temporal dimensions. The numerical fields in the record table only undergo dimensional unification and outlier marking, without direct averaging of cross-source values. The access processor also adds observation type codes to the observation records according to the data acquisition method. Point observations correspond to finite neighborhood influences, line observations correspond to banded influences extending along the profile, area observations correspond to planar coverage, and volume observations correspond to three-dimensional influences with depth boundaries. After reading the observation type codes, the spatiotemporal modeling processor selects the corresponding support domain generation method and maps the sampling time to the state action time through time lag correction. For example, when the settlement time series of a structure reflects the cumulative response after changes in the background stratum state, the action interval of the settlement record is extended forward to a time window that can be jointly explained with the adjacent geophysical detection results. When geological exploration data is used as relatively stable prior information, it is given an action interval spanning multiple time windows, but its direct update of the disease state increment is restricted, so that each type of data enters the subsequent calculation in a form consistent with its own observation mechanism.

[0051] Specifically, please refer to Figure 2The spatiotemporal modeling processor establishes an observation support domain object for each type of data, consisting of a central location, spatial influence boundary, effective detection depth, temporal action interval, and observation response kernel. The central location is represented by three-dimensional coordinates in a reference coordinate system. The spatial influence boundary is represented by polygons, buffer lines, ellipsoids, or layered cylinders. The effective detection depth is determined by the original detection record or existing exploration depth. The temporal action interval is jointly determined by the sampling time, sampling period, and response lag. The observation response kernel is used to describe the sensitivity of the observation to different locations within the support domain. Point deformation measurements can use a radial kernel that decays with distance, while profile detection can use a kernel that decays slowly along the profile direction and is perpendicular to the profile. Anisotropic kernels that decay rapidly along the profile direction are used. Drilling data can be revealed using kernels that are segmented and constant along the borehole axis and decay rapidly outside the borehole. When the spatiotemporal modeling processor projects the observed support domain objects onto the layered spatiotemporal computational domain, it retains the geometric segments that intersect with the computational subdomains and their response kernel integral values, and writes the projection results into the support domain index table. The support domain index table contains at least the observation record number, computational subdomain number, spatial overlap ratio, depth overlap ratio, temporal overlap ratio, and response kernel similarity. The disease reconstruction processor only creates data items for observation records and computational subdomains that are associated in the support domain index table. Observation records that do not overlap with the support domain are not forcibly associated due to coordinate proximity.

[0052] The overall correlation between observation support domains is calculated according to the following formula:

[0053] ;

[0054] in, This represents the comprehensive correlation between the i-th observation support domain and the j-th observation support domain. This indicates the proportion of the spatial overlap volume to the smaller support domain volume of both. This indicates the overlap ratio of the effective detection depth ranges. Indicates the overlap ratio of time intervals. This represents the kernel-normalized inner product of two observed responses. , , and They are respectively represented as , , and The corresponding non-negative weights sum to 1, and the weights are determined by the spatial, depth, time, and response direction sensitivity of the corresponding data type; for example, when two observations completely overlap on a plane but their depth intervals do not overlap. , ,but The two do not establish a cross-source evidence boundary; when the two observations have high overlap in space, depth, and time and their response directions are consistent, all four ratios are close to 1. Maintaining a larger value allows the corresponding evidence to participate in mutual verification within the same computational subdomain.

[0055] Furthermore, the spatiotemporal modeling processor establishes a layered spatiotemporal computational domain based on the stratigraphic interfaces, the foundation bottom and side boundaries of structures, the boundaries of existing anomaly areas, and the density of observed support domains in the geological exploration data. The computational domain employs irregular subdomains that can conform to the interfaces near the stratigraphic interfaces, larger joint subdomains in areas with continuous media and low observation density, and smaller, more subdivided subdomains in areas with complex overlapping support domains or concentrated historical anomalies. Each computational subdomain records its respective stratigraphic layer, its relative position to the structure foundation, the set of adjacent subdomains, and its boundary type. Latent state variables of the defects are stored in the computational subdomain in vector form, with each component of the vector corresponding to the states of loosening, voiding, seepage disturbance, local settlement, and stress anomaly, respectively. The initial value of intensity and state intensity is jointly given by geological exploration priors, historical anomaly records, and current observation residuals. Among them, geological exploration priors only limit the range of states that the medium can produce, historical anomaly records are used to provide spatial continuity priors, and current observation residuals are used to form the update direction within the time window. For state components that cannot be directly observed in certain types of data, the disease reconstruction processor maps the state vector to observables through the corresponding observation response model before calculating the residuals. For example, settlement monitoring data mainly constrains the local settlement component and indirectly constrains the loose and void components, while geophysical detection data mainly constrains the medium anomaly component and passes it to adjacent state components through hierarchical topological relationships, avoiding the equivalent interpretation of all data sources as the same disease index.

[0056] In a preferred embodiment, the physical consistency constraint consists of intra-layer continuity constraints, inter-layer interface compatibility constraints, and structure contact boundary constraints. Intra-layer continuity constraints calculate the state differences between adjacent computational subdomains within the same layer, but allow for reduced constraint strength at known fault zones, backfill boundaries, or maintenance interfaces. Inter-layer interface compatibility constraints limit the inter-layer propagation of state increments based on the types of media between upper and lower layers, interface inclination, and allowed state types. For example, seepage disturbances are allowed to propagate along connected interfaces, while voided states cannot unconditionally traverse intact and stable layers. Structure contact boundary constraints are based on the contact area between the foundation bottom, sidewalls, and strata. The relationship defines the mapping relationship between local settlement and stress anomaly states on both sides of the boundary. The spatiotemporal modeling processor writes the above constraints into the sparse constraint operator. Each row of the sparse constraint operator corresponds to a local relation, and each column corresponds to a state component in a computational subdomain. Non-zero coefficients only connect topologically adjacent variables that allow state transfer. In each iteration, the disease reconstruction processor first generates candidate state increments based on multi-source observation residuals, then uses the sparse constraint operator to calculate the degree of violation of the physical relation by the candidate state increments, and decomposes the violation components so that the retained state increments satisfy the layer continuity, interface compatibility and contact boundary constraints.

[0057] The constraint amount corresponding to the physical consistency constraint is calculated according to the following formula:

[0058] ;

[0059] in, Represents the physical consistency constraint quantity. Represents the set of adjacent subdomains at the same level. Represents the set of adjacent subdomains at a cross-layer interface. Represents the set of subdomains representing the contact boundaries of structures. This represents the latent disease state vector of the m-th computational subdomain. , and These represent the constraint coefficients for intra-layer, inter-layer, and contact boundaries, respectively. and The matrix represents the selection of state components that can be passed across layers. Represents the contact boundary state mapping matrix. This represents the boundary reference vector determined by the contact state of the structure; for example, taking the larger value between two adjacent subdomains within a complete continuous layer. When both occur, mutations without observational support increase the likelihood of occurrence. Reduce at the known backfill boundary At that time, the state differences supported by data are allowed to be preserved on both sides of the boundary.

[0060] Please refer to Figure 3 , Figure 3 The horizontal axis represents the iteration number of the latent state field reconstruction of the disease, ranging from 1 to 12; the vertical axis represents the physical consistency constraint residual, with larger values ​​indicating a more significant violation of stratigraphic continuity, interface transmission, or contact boundary relationships of structures by the candidate state increment. Thick solid lines correspond to continuity constraints of adjacent subdomains within the same layer, dotted dashed lines correspond to cross-layer interface allowable transmission constraints, square dotted lines correspond to contact boundary constraints of structures, and triangular dotted lines represent the total of the three types of residuals. Figure 3 The line types and markings of each curve in the text correspond to the intra-layer continuity relationship, inter-layer compatibility relationship, and structure contact boundary limitation relationship described in the text.

[0061] from Figure 3As can be seen, the total residual gradually decreased from approximately 2.77 in the first round to approximately 1.72 in the 12th round, indicating that the candidate disease states gradually satisfied the local physical relationships in the layered spatiotemporal computational domain after multiple rounds of constraint projection. The continuous residual within the same layer always accounted for the majority, and decreased from approximately 1.86 to approximately 1.28, indicating that the state abrupt changes between adjacent subdomains within the strata were the focus of constraint correction; the cross-layer transfer residual decreased from approximately 0.49 to approximately 0.25, reflecting that the cross-layer interface gradually became more compatible in terms of the state components allowed for transfer; the contact boundary residual was small and continued to decrease, indicating that the state mapping near the foundation bottom and sidewalls was stably corrected during iteration.

[0062] In this embodiment, please refer to Figure 4 The reconstruction processor maintains and calculates the credibility state corresponding to each data source subdomain and time window. The local residual stability is determined by the residual sign consistency, the median absolute value of the residual, and the residual dispersion within the continuous window. The cross-source mutual verification degree is determined by the degree of support of other data sources with supporting domain correlation within the same calculation subdomain for the same state update direction. The data timeliness is determined by the degree of overlap between the observation period and the current reconstruction time window. The historical drift state is determined by the quantile offset, mean offset, and local trend change between the baseline time period and the current time period.

[0063] The specific calculation methods for each component are as follows:

[0064] 1. Residual stable components Given a source s, a computational subdomain m, and a time window t, take the observation residual sequence corresponding to N consecutive historical time windows, including the current time window. The sign consistency rate is calculated by determining the proportion of windows in the statistical sequence whose residual signs match the residual signs of the current time window out of the total number of windows. ; Calculate the median of the absolute values ​​of the sequence residuals with the standard deviation of the sequence residuals The ratio of the two values ​​yields the volatility stability rate; the residual stability component is calculated using the following formula:

[0065]

[0066] In the formula, , The weight coefficients are non-negative and satisfy the following conditions: , To prevent extremely small positive numbers from being divided by zero.

[0067] 2. Cross-source mutual verification component : Obtain all other data sources within the current computational subdomain m that overlap with the supporting domain of source s, and form a mutual verification source set. The observation residuals from each source are mapped and converted into constraint direction vectors for the latent state increments of the disease through their respective observation responses. , Calculate the cosine similarity of the direction vector between source s and each source s' in the set, and take the average of all similarities as the cross-source mutual verification component, calculated using the following formula:

[0068]

[0069] In the formula, This represents the number of elements in the mutual verification source set.

[0070] 3. Time-sensitive quantity Let the center time of the current reconstruction time window t be... The observation data acquisition time of source s within the computational subdomain m is The effective time window width for this type of data is The aging component is calculated using an exponential decay method.

[0071]

[0072] 4. Drift component : Select historical baseline data from source s within the computational subdomain m, and calculate the mean of the observations in the baseline period. Standard deviation Upper quartiles and lower quartiles ; Calculate the mean of the data segment corresponding to the current time window. Upper quartiles and lower quartiles Calculate the normalized mean offset and the normalized interquartile range offset separately, and then sum them by weight to obtain the drift component:

[0073]

[0074] In the formula, , The weight coefficients are non-negative and satisfy the following conditions: , To prevent extremely small positive numbers from being divided by zero.

[0075] The disease reconstruction processor arranges the above components into a dynamic confidence tensor according to the source dimension, computational subdomain dimension, and time window dimension. Before each state update, it reads the corresponding tensor elements to modulate the observation residuals. Data with continuous local residual directions and support from other sources receive higher contributions, while data with drastic residual reversals, time intervals far from the current time window, or continuous drift receive lower contributions. The confidence state is not stored as a fixed data source level, but is updated independently within each computational subdomain and time window. The same data source can maintain a high contribution in one region, while its contribution decreases in another region due to the edge of the supporting domain, local noise, or drift, so that the confidence expression is consistent with the actual range of data influence.

[0076] The elements in the dynamic credibility tensor are generated according to the following formula:

[0077] ;

[0078] in, This indicates the confidence level of source s in the computational subdomain m and time window t. Represents the stable component of the residual. Indicates cross-source mutual verification components. Indicates time-sensitive components. Indicates the historical drift component, to This represents the non-negative scaling factor for each component. This indicates the operation of the natural exponent; for example, when the residual direction of a certain source is consistent within a continuous window, the update direction is consistent with that of adjacent sources, and the data is within the current scope. , and Take the larger positive value Take the smaller value. Approaching 1; when a source exhibits persistent baseline drift and conflicts with the direction of other sources, Increase and make As the number of observations decreases, the weight of the observed residuals in the state update also decreases.

[0079] Please refer to Figure 5 , Figure 5 The horizontal axis represents the sampling window number, ranging from 0 to 500; the vertical axis represents the dynamic reliability contribution deviation index, ranging from approximately 0.15 to 0.55. The thick solid line is the deviation index time series curve after comprehensive conversion based on residual stable state, cross-source mutual verification state, data timeliness state, and historical drift state. The dashed line and dotted line represent the first warning threshold and the second warning threshold, respectively. Figure 5 This is used to demonstrate the dynamic changes in the credibility contribution of the same data source across different computational subdomains and time windows, rather than fixing the data source to a single level.

[0080] from Figure 5 As can be seen, the exponents from sequence 0 to approximately 100 mainly fluctuate between 0.16 and 0.25, near the first threshold, indicating that the residual direction and cross-source mutual verification relationship still have slight instability in the early window. From approximately sequence 100 to approximately 230, the exponents continuously rise and repeatedly approach the second threshold, indicating that historical drift components or insufficient mutual verification begin to affect the credibility modulation. After approximately sequence 230, the curve repeatedly exceeds the second threshold and rises above 0.50 in the latter part, indicating that the contribution deviation of this source in the local region has formed a continuous trend, and it is necessary to reduce the influence of its corresponding residual in the state update and trigger a review. The curve still retains obvious fluctuations, indicating that the system does not simply remove sources, but continuously corrects the credibility according to the actual mutual verification relationship within each window.

[0081] Specifically, the cross-source evidence graph uses observation records or local evidence aggregated by source as evidence nodes. Edge construction is based on the condition that the comprehensive correlation of the observation support domain reaches the connection condition and the state update direction is comparable. Evidence edges record the correlation of the support domain, the mapping relationship of observation directions, the correspondence of state components, and the temporal sequence. The disease reconstruction processor extracts interconnected evidence subgraphs from the evidence graph within each computational subdomain and forms mutually corroborating components based on the directional consistency relationship in the evidence subgraphs. If the original observations from two sources have different physical meanings, they are first converted into constraint directions on the latent state increments of the disease using their respective observation response models, and then consistency is achieved. In comparison, evidence that cannot be mapped to a common state component does not participate in direct mutual verification. For evidence subgraphs with inconsistent directions, the disease reconstruction processor performs a one-source reconstruction for each source, that is, temporarily removes the observations of that source in the current local area, recalculates the disease latent state field and physical consistency residuals while keeping other conditions unchanged, and compares the reconstruction position, shape and constraint violation degree before and after removal. If the conflict disappears and the physical consistency residual decreases after removing a source, the corresponding conflict is assigned to the source-specific component. If removing any single source cannot eliminate the conflict, the part that cannot be eliminated is retained as a common conflict component and written into the independent anomalous evidence channel.

[0082] Source conflict identification score is calculated using the following formula:

[0083] ;

[0084] in, This represents the conflict identification result after removing the q-th source. This represents the normalized difference in spatial location and morphology of the latent disease state field before and after removal. This represents the normalized change in the physical consistency constraint quantity before and after removal. This represents the normalized change in the uncertainty field before and after removal. , and They are respectively represented as , and The corresponding non-negative coefficients sum to 1; for example, when the location of the diseased area remains basically unchanged after removing a certain source, the physical consistency constraint changes little, and the uncertainty does not increase significantly. If the source remains small, it is not a primary source of conflict; if removing a source causes existing local anomalies to disappear and physical consistency constraints to decrease, while consistency among other sources improves, then... and The larger value is selected, and the corresponding source is written into the credibility status update channel.

[0085] Please refer to Figure 6 , Figure 6 The horizontal axis represents the source conflict score, ranging from approximately 0 to 0.62; the vertical axis represents the observation source category, including horizontal displacement monitoring, geophysical profiles, structure settlement monitoring, crack tilt observation, construction and maintenance records, and borehole exploration. Each horizontal column represents the combined results of changes in the latent state morphology of the defect, changes in physical constraint residuals, and changes in uncertainty after temporarily removing the source under the same stratified spatiotemporal computational domain, the same initial state, and the same constraints; the vertical dashed line represents the key review threshold value, and the number at the end of the column represents the corresponding conflict score.

[0086] from Figure 6 As can be seen, the conflict scores for each source are significantly lower than the key review threshold, with the highest being approximately 0.14 for borehole exploration, approximately 0.12 for construction and maintenance records, and approximately 0.10 for crack tilt observation, structure settlement monitoring, geophysical profiles, and horizontal displacement monitoring. This distribution indicates that the current local conflicts do not exhibit source-specific issues that would disappear significantly after the removal of a single source, but are more likely due to differences in the edges of the observation support domain, time windows, or inconsistencies in the interpretation of state components. Because different textures are used to fill the columns and numerical values ​​are labeled, the differences in contribution from each source can be clearly distinguished. In engineering practice, all sources should be retained for reconstruction, but the review priority should be increased for areas containing borehole exploration and construction and maintenance records, and local weight modulation should be implemented in dynamic credibility updates.

[0087] In one embodiment, the disease reconstruction processor organizes observations, dynamic confidence tensors, physical consistency constraints, and independent anomaly evidence into an iterative reconstruction target. A single iteration includes observation residual calculation, confidence modulation, candidate state increment solving, physical constraint projection, conflict evidence correction, and uncertainty updating. The observation residual is formed by the difference between the actual observed value and the predicted value obtained from the corresponding observation response model of the current disease latent state. The confidence-modulated residual is used to solve for the candidate state increment of each computational subdomain. The candidate state increment is decomposed by a sparse constraint operator and written back into the disease latent state field. Common conflict components are retained as additional observations but with higher initial uncertainties. Source-specific components are not directly deleted but their corresponding source uncertainty is reduced. Based on the credibility within the region and time window, the disease reconstruction processor determines whether to end the reconstruction of the current time window according to the changes in state before and after iteration, the changes in physical consistency constraints, and the changes in observation residuals. After the reconstruction ends, an uncertainty field is generated based on the coverage of each computational subdomain, the credibility distribution, the degree of conflict residue, and the sensitivity of state solution for multi-source observations. The warning region consists of computational subdomains where the hidden state of the disease exceeds the corresponding state reference range and is spatially continuous and temporally persistent. When the local state is high but the uncertainty is also high, it is marked as a region to be verified. When the local state has continuous multi-source support and the uncertainty remains low, it is written into the warning region output table. The output table simultaneously saves the source contribution, the range of the support domain, the state components, and the time window.

[0088] The joint update of the latent state of the disease and its uncertainty is expressed by the following formula:

[0089] ;

[0090] in, This represents the latent state of the disease in the (r+1)th round. This represents the uncertainty field in the (r+1)th round. This represents the observations of source s in subdomain m and time window t. This represents the corresponding observation response mapping. This represents a robust loss function that uses a slowly increasing form for large residuals. This represents the conflict-preserving term corresponding to the shared conflicting components. This represents the uncertainty constraint term formed by the coverage quantity, confidence dispersion, and state sensitivity. , and They are respectively represented as , and The corresponding non-negative coefficients; for example, when a computational subdomain is covered by only a single source and the source has low credibility. Limit the uncertainty to an excessively small value; when a computational subdomain has multiple high-confidence sources and the residual directions are consistent, the observation terms jointly constrain X, and the uncertainty can be reduced in the joint solution.

[0091] Please refer to Figure 7 , Figure 7 The horizontal axis represents the iteration round of joint reconstruction, ranging from 1 to 12; the vertical axis represents the target contribution. The thick solid line represents the contribution of the observation item, the dotted dashed line represents the contribution of the physical constraint item, the square dotted line represents the contribution of the conflict constraint item, the triangular dotted line represents the contribution of the uncertainty item, and the diamond dotted line represents the total of the targets. Figure 7 The corresponding approach in the paper, which organizes observation residuals, dynamic reliability, physical consistency, common conflicts, and uncertainty constraints into an iterative objective, is used to illustrate how the system can obtain a stable latent state field of the disease while maintaining multi-source observation information.

[0092] from Figure 7 As can be seen, the total target value rapidly decreased from approximately 7.05 in the first round to approximately 3.37 in the fourth round, then slowly decreased and stabilized around approximately 2.15 in the twelfth round, indicating that the reconstruction process underwent a convergence phase from large-scale corrections to local fine-tuning. The observation term decreased from approximately 6.13 to approximately 1.35, which is the main source of the decrease in the total target value, indicating that the observation residuals after confidence modulation were gradually absorbed into the latent state field of the disease. The physical constraint term decreased from approximately 2.76 to approximately 1.71, reflecting the suppression of state abrupt changes without physical support by constraint projection. The conflict constraint term fluctuated slightly around 0.50, indicating that common conflicts were retained as supplementary evidence rather than simply deleted. The uncertainty term gradually increased to approximately 0.70, indicating that in areas with insufficient evidence or residual conflicts, the system actively maintained a high degree of uncertainty to avoid outputting overly certain warning conclusions.

[0093] Preferably, the spatiotemporal modeling processor performs adaptive subdivision of the hierarchical spatiotemporal computational domain based on the support domain correlation matrix. Regions with dense cross-source connections and inconsistent principal directions or attenuation patterns of observation response kernels are divided into multiple sub-computational domains, ensuring that different observation directions fall into their respective sub-domains. Regions with sparse cross-source connections and continuous observations from the same source remain as joint computational domains, avoiding the formation of unnecessary fragmented grids in regions lacking multi-source resolution. After subdivision, a fine-grained projection index is established for each sub-computational domain, while an aggregated projection index is established for the joint computational domain. The two types of indexes are connected through a conservation mapping relationship, which records the volume of the parent and child computational domains. The correspondence between the product ratio, the response kernel integral ratio, and the state components is defined. When the observation residual is transferred from the joint computation domain to the subdivided subdomain, it is allocated according to the effective overlap volume and the response kernel integral ratio. When the hidden state increment is aggregated back from the subdivided subdomain to the joint computation domain, it is normalized and synthesized according to the subdomain volume and the reciprocal of the uncertainty. For example, if a certain profile detection overlaps with multiple point deformation observations in a local area and the response directions are different, after the local area is subdivided, the anomaly in the profile direction can be continuously expressed along the corresponding subdomain. The point observation only corrects the subdomains covered by its supporting domain. The uncovered areas are not directly affected by the point observation. The total state between the parent and child computation domains is kept consistent through conservation mapping.

[0094] In a preferred embodiment, the spatiotemporal modeling processor expands the hierarchical topology graph into a two-layer graph consisting of an observation node layer and a constraint node layer. Each node in the observation node layer corresponds to a computational subdomain and carries observation residuals from various sources, latent state vectors of defects, confidence states, and uncertainty states. The constraint node layer is configured with intra-layer continuous constraint nodes, cross-layer interface compatibility constraint nodes, and structure contact boundary constraint nodes. The two node layers are connected by constraint action edges, which record the state component selection relationship, transmission direction, and current activation state. The spatiotemporal modeling processor determines the constraint node based on the number of observation nodes associated with it, the degree of conflict in evidence directions, and historical fullness. The system generates a constraint activation sequence based on the state. Local areas with sufficient observation and consistent evidence direction are prioritized for activation of constraint nodes directly corresponding to their topological relationship. Areas with high evidence conflict are activated later for strong continuous constraints while retaining shared conflict components. The disease reconstruction processor maps the candidate state increments from the observation node layer to the constraint node layer. After processing by the activated sparse constraint operators, the data is written back to the associated observation nodes along the constraint action edges. Unactivated constraint nodes do not participate in the state correction of the current round. In the next round, they are reordered according to the updated conflict degree and historical satisfied states. This allows physical constraints to participate in reconstruction gradually according to local evidence states, rather than acting on the entire region with a fixed intensity at the same time.

[0095] Furthermore, the disease reconstruction processor establishes drift latent variables for each data source and combines the drift latent variables, the corresponding source's confidence status, and the disease latent status variables into a joint state vector. The drift latent variables are used to record source baseline offset, local scale changes, and slow trend changes. Within each time window, the disease reconstruction processor extracts multiple quantiles from the historical benchmark segment and establishes a mapping with the corresponding quantiles of the current data segment, updating the drift latent variables according to the quantile mapping deviation.

[0096] The drift latent variables are presented as a two-dimensional vector containing an additive baseline term and a multiplicative proportional term, denoted as... ,in, For transpose, This is an additive baseline drift, used to characterize the overall baseline shift of the observed data. This is the multiplicative proportional drift coefficient, used to characterize the amplitude proportional shift of the observed data.

[0097] The specific update method for the drift latent variable is as follows: p quantiles are uniformly selected from the historical baseline segment to obtain the baseline observation value corresponding to each quantile. Extract the actual observed values ​​of the corresponding quantiles in the current data segment. The objective is to minimize the squared error of the quantile mapping, solved using least squares.

[0098]

[0099] The obtained b and k are the values ​​corresponding to the current time window. and This completes the update of the drift hidden variables.

[0100] The specific process of correcting the stable residual component and the cross-source mutual verification component by drifting latent variables is as follows: First, the drifting latent variables are used to modify the original observation values. Perform correction to obtain the corrected observations. Then, the drift-corrected observation residuals are calculated based on the corrected observation values. ,in The predicted value is the output of the observation response mapping; the original residual is replaced with the corrected observation residual, and the residual stable component and cross-source mutual verification component are recalculated according to the calculation rules of the aforementioned residual stable component and cross-source mutual verification component to obtain the residual stable component and cross-source mutual verification component after drift correction, thereby eliminating residual fluctuations and direction differences that can be explained by data source drift.

[0101] The drift latent variables are then used to correct the current observation value or observation response mapping, so that the residual stable component and cross-source mutual verification component do not directly absorb changes that can be explained by source drift. The corrected residual stable component and cross-source mutual verification component are used to update the dynamic confidence tensor. The disease reconstruction processor then solves the disease latent state increment based on the updated dynamic confidence tensor and feeds back the prediction residual generated by the disease latent state increment to the drift latent variables. Drift update, confidence update and disease state update are executed alternately within the same time window until the joint state vector change meets the stopping condition. If a certain type of change can be explained by the drift latent variables and does not satisfy the spatial continuity relationship on the hierarchical topology, it is retained in the source drift channel. If a certain type of change is still supported by multiple sources after drift correction and satisfies the physical consistency constraint, it is written into the disease latent state field.

[0102] The objective quantity corresponding to the joint state alternation update is calculated according to the following formula:

[0103] ;

[0104] Where J represents the joint state objective quantity, This represents the latent drift variable of source s in subdomain m and time window t. This represents the drift vector of source s within time window t. Represents the sum of absolute values. This represents the drift time continuity constraint coefficient. This represents the physical consistency constraint coefficient. This represents the uncertainty constraint coefficient; for example, when a source exhibits a consistent and slowly changing offset across the entire coverage area, the coefficients of uncertainty constraint coefficients in adjacent time windows. It can continuously change and absorb the corresponding shift; when an anomaly only appears in a local adjacent subdomain and is observed by other sources, the drift vector from a single source is difficult to explain the spatial structure at the same time, and the anomaly is retained in X.

[0105] Please refer to Figure 8 , Figure 8 The horizontal axis represents the number of alternating update iterations, ranging from 1 to 12; the vertical axis represents the contribution of the drift target. The thick solid line represents the observation residual term after drift correction, the dotted dashed line represents the drift continuity term, the square dotted line represents the physical constraint term, the triangular dotted line represents the uncertainty term, and the diamond dotted line represents the target total. Figure 8 It is used to describe the process of drift hidden variables, dynamic confidence status, disease hidden status and uncertainty status being updated alternately within the same time window, and is particularly suitable for distinguishing between slow shift of the source baseline and changes in the actual local disease status.

[0106] from Figure 8As can be seen, the observed residual term decreased from approximately 16.6 to approximately 14.8, indicating that the drift correction can absorb some of the source baseline shift, but it does not interpret all residuals as drift. The drift continuity term remained stable around 20, indicating that the drift vector from the same source mainly continues continuously along the time dimension without any sudden jumps. The physical constraint term decreased from approximately 2.7 to approximately 1.7, indicating that the state increment after drift correction is still constrained by the hierarchical topological relationship. The uncertainty term gradually increased from approximately 0.6 to approximately 6.5, reflecting that continuous drift will increase the interpretation uncertainty in some areas. The target total quickly approached the stable range after the first round, indicating that the system can retain the drift channel while avoiding misreporting slow global shifts as local hidden defects, and retaining necessary uncertainty information for subsequent verification.

[0107] In one embodiment, for computational subdomains with inconsistent cross-source evidence directions, the disease reconstruction processor constructs a set of source subsets. These source subsets include at least a baseline subset retaining all sources, a leave-one subset formed by removing individual sources, and complementary subsets divided according to observation type. For each source subset, disease latent state reconstruction is performed under the same hierarchical spatiotemporal computational domain, the same initial state, and the same physical consistency constraints, resulting in a subset latent state field and a reconstruction residual field. Subsequently, a conflict topology graph is constructed based on the spatial offset, connectivity differences, state component direction differences, and the degree of physical constraint violation among the latent state fields of each subset. The nodes in the conflict topology graph... Each point corresponds to a source subset. Node attributes record the local morphology and constraint residuals of the latent state field of the subset. Edge attributes record the conflict location, conflict state components, and conflict duration between two source subsets. The disease reconstruction processor extracts source-specific components that disappear with the removal of a specific source from the conflict topology graph and writes the source-specific components into the confidence state update channel of the corresponding source. At the same time, it extracts common conflict components that cannot disappear by removal of any single source and retains them in the independent anomaly evidence channel. The data in the independent anomaly evidence channel is not directly equivalent to the disease state, but participates in the reconstruction as an additional observation item that needs to be continuously verified in subsequent time windows.

[0108] Preferably, the disease reconstruction processor organizes the joint state vector into a sliding state chain in the time dimension. Each time node of the sliding state chain stores the drift latent variables, dynamic confidence tensor, disease latent state field, and uncertainty field for the corresponding time window. Drift continuation edges, confidence inheritance edges, and latent state evolution edges are established between adjacent time nodes. Drift continuation edges limit the direction and sparsity of change of drift latent variables from the same source in adjacent time windows. Confidence inheritance edges pass the confidence state corrected by cross-source mutual verification to the next time window and allow new observations to readjust. Latent state evolution edges pass the disease latent state increment that satisfies the spatial continuity condition according to the hierarchical topological relationship. When a certain prediction residual can be explained by both the drift latent variable and the latent state variable of the disease within a continuous time window, the disease reconstruction processor freezes the drift latent variable and the latent state variable of the disease and performs two local reconstructions respectively. When freezing the drift latent variable, only the latent state of the disease is allowed to absorb the residual, and when freezing the latent state variable of the disease, only the drift latent variable is allowed to absorb the residual. Then, the spatial structure, cross-source consistency and physical constraint violation position of the residuals obtained from the two local reconstructions are compared. Based on this, the drift continuation edge and the latent state evolution edge are updated. After the update is completed, joint smoothing is performed on multiple time nodes covered by the current sliding window so that the state in the early window can be corrected based on subsequent observations.

[0109] The consistency of transitions in a sliding state chain is measured by the following formula:

[0110] ;

[0111] in, This represents the consistency quantity of transitions in a sliding state chain. This represents all the latent drift variables within the time window t. This represents all the confidence states within the time window t. This represents all the hidden states of diseases within the time window t. This indicates a drift continuation mapping. This represents a credibility inheritance mapping. This represents the hidden state evolution mapping generated based on the hierarchical topology. , and These are non-negative constraint coefficients; for example, when new observations prove that early anomalies originated from persistent drift, the updated... Allowing drift to propagate continuously along the time chain Reduce the propagation of the corresponding state; when new observations form continuous anomalies in adjacent subdomains, This propagation relationship is preserved and the latent state of the disease after joint smoothing remains continuous across multiple time windows.

[0112] In a preferred embodiment, the disease reconstruction processor organizes the common conflicting components in the independent anomaly evidence channels into conflict evidence chains according to spatial connectivity and temporal continuity. Each conflict evidence chain stores the source subset combination that generated the common conflicting component, the corresponding observation support domain, the conflicting state component, the physical constraint violation location, and the duration of the time window. When new observational evidence overlaps with the support domain of the conflict evidence chain in a subsequent time window, the disease reconstruction processor performs directional evidence return, converting the new observational evidence into the hidden state constraint direction of the disease through the corresponding observation response model, calculating its directional consistency with the original source subset combination, and reading the topological adjacency relationship between the computational subdomain where the new observation is located and the original physical constraint violation location. The system calculates the continuity between the new observation's effective interval and the time interval of the conflict evidence chain. New observations that are consistent in direction, topologically adjacent, and temporally continuous are written into the conflict evidence chain, and the uncertainty of the corresponding shared conflict component is reduced. Observations that are inconsistent in direction or lack topological continuity are retained as new independent evidence branches. The updated shared conflict component participates again in the joint reconstruction of the disease latent state field and uncertainty field as an additional observation item. If the shared conflict component gradually gains cross-source support in multiple consecutive time windows and the physical constraint violation position moves along the allowed intra-layer or cross-layer path, the corresponding state is transferred to the disease latent state field. If subsequent observations cannot support its continuation, its uncertainty is gradually increased and it stops being transmitted to the warning area.

Claims

1. A multi-source fusion-based early warning system for hidden geological defects in building structures, characterized in that, This includes a data access processor, a spatiotemporal modeling processor, and a disease reconstruction processor; The data access processor acquires data from structure monitoring, geological exploration, deformation detection, and geophysical detection, and generates an observation support domain that includes spatial coverage, detection depth, sampling interval, and time lag. The spatiotemporal modeling processor establishes a layered spatiotemporal computing domain based on the interface between the structure and the background strata, sets hidden state variables of the disease in it, and generates physical consistency constraints based on the stratum continuity relationship and interface compatibility relationship. The disease reconstruction processor forms a dynamic credibility tensor based on local residual stability, cross-source mutual verification degree, data timeliness, and historical drift state. It establishes a cross-source evidence map based on the overlapping relationship of the observation support domain, identifies conflicting data by combining physical consistency residuals and single-source support domain deviations, and iteratively updates the disease latent state field and uncertainty field under the constraints of the dynamic credibility tensor and physical consistency. The warning area is determined based on the update results.

2. The early warning system for hidden geological defects in structures based on multi-source fusion as described in claim 1, characterized in that, The spatiotemporal modeling processor establishes an observation support domain object for each type of data, consisting of a central location, spatial influence boundary, effective detection depth, time action interval, and observation response kernel. The observation support domain object is then projected onto the layered spatiotemporal computing domain according to the background stratigraphic coordinates and the structure interface coordinates. For any two observed support domain objects after projection, calculate the spatial overlap ratio, depth overlap ratio, temporal overlap ratio, and response kernel similarity, and combine each ratio with the response kernel similarity to form a support domain correlation matrix; The disease reconstruction processor establishes cross-source evidence edges in overlapping observation ranges based on the support domain correlation matrix, retains local evidence nodes in non-overlapping observation ranges, and constrains the update of the disease latent state variables through the cross-source evidence edges and the local evidence nodes.

3. The early warning system for hidden geological defects in structures based on multi-source fusion as described in claim 1, characterized in that, The spatiotemporal modeling processor generates a layered topology map based on the stratigraphic boundaries, the influence zone of the foundation of the structure, and the existing anomaly area. The nodes of the layered topology map correspond to the computational subdomains in the layered spatiotemporal computational domain. The node attributes include the stratigraphic category, the medium continuity identifier, and the boundary status. The edge attributes include the adjacency type, the interface direction, and the cross-layer transfer permission. Based on the layered topology diagram, construct continuous constraints within the layer, cross-layer interface compatibility constraints, and structure contact boundary constraints, and convert each constraint into a sparse constraint operator that acts on the latent state variables of the disease. In each round of graph constraint iteration, the disease reconstruction processor first updates the hidden state of nodes based on observation evidence, and then uses the sparse constraint operator to decompose and write back the update components that do not satisfy the intra-layer continuity relationship, cross-layer interface relationship or contact boundary relationship.

4. The early warning system for hidden geological defects in structures based on multi-source fusion as described in claim 1, characterized in that, The disease reconstruction processor maintains and calculates the credibility status corresponding to each data source subdomain and time window. It forms a residual stability component based on the residual sign consistency and residual volatility within the continuous window, a cross-source mutual verification component based on the consistency of evidence direction with other sources within the overlapping support domain, a timeliness component based on the interval between the acquisition time and the current reconstruction time, and a drift component based on the distribution migration between the historical benchmark segment and the current data segment. The residual stable components, cross-source mutual verification components, time-dependent components, and drift components are arranged according to their source, spatial dimension, and time dimension to form the dynamic credibility tensor. The contribution of the observation residuals from each source to the update of the latent state variable of the disease is modulated domain by domain through the dynamic credibility tensor.

5. The early warning system for hidden geological defects in structures based on multi-source fusion as described in claim 2, characterized in that, The spatiotemporal modeling processor performs adaptive subdivision of the hierarchical spatiotemporal computational domain based on the support domain correlation matrix, dividing regions with dense cross-source connections and inconsistent main directions or attenuation patterns of observation response kernels into multiple sub-computational domains, and maintaining regions with sparse cross-source connections and continuous observations from the same source as joint computational domains. A fine-grained projection index is established for each sub-computation domain, and an aggregated projection index is established for the joint computation domain. The conservation mapping relationship between the fine-grained projection index and the aggregated projection index is recorded. The disease reconstruction processor uses the conservation mapping relationship to transfer observation residuals and hidden state increments between different spatial resolutions, and normalizes and allocates observation residuals and hidden state increments according to the effective overlap volume between each observation support domain and the corresponding computation domain during the transfer process.

6. The early warning system for hidden geological defects in structures based on multi-source fusion according to claim 3, characterized in that, The spatiotemporal modeling processor expands the hierarchical topology graph into a two-layer graph consisting of an observation node layer and a constraint node layer. The nodes in the observation node layer carry multi-source evidence for each computational subdomain, and the nodes in the constraint node layer carry the intra-layer continuous constraints, cross-layer interface compatibility constraints, and structure contact boundary constraints, respectively. The two node layers are connected by constraint action edges. Constraint activation sequences are generated based on the number of observation nodes associated with each constraint node, the degree of conflict in evidence directions, and the historical satisfaction status. The corresponding sparse constraint operators are then selectively invoked according to the constraint activation sequences. The disease reconstruction processor maps the hidden state update of the observation node layer to the constraint node layer. After processing by the activated sparse constraint operator, it writes back to the associated observation node according to the constraint action edge.

7. The early warning system for hidden geological defects in structures based on multi-source fusion according to claim 4, characterized in that, The disease reconstruction processor establishes drift latent variables for each data source and combines the drift latent variables with the credibility status of the corresponding source and the disease latent state variables to form a joint state vector; Within each time window, the drift latent variable is first updated based on the quantile mapping deviation between the historical benchmark segment and the current data segment. Then, the residual stable component and cross-source mutual verification component are corrected based on the drift latent variable. Subsequently, the dynamic credibility tensor is updated with the corrected residual stable component and cross-source mutual verification component. The disease reconstruction processor solves for the disease latent state increment based on the updated dynamic credibility tensor, and feeds back the prediction residual generated by the disease latent state increment to the drift latent variable, so that the drift latent variable, dynamic credibility tensor and disease latent state variable are updated alternately according to the same time window.

8. The early warning system for hidden geological defects in structures based on multi-source fusion according to claim 6, characterized in that, The disease reconstruction processor constructs a set of source subsets for computational subdomains with inconsistent cross-source evidence directions, and performs disease hidden state reconstruction on each source subset to obtain the corresponding subset hidden state field and reconstruction residual field. A conflict topology graph is formed based on the spatial offset, morphological differences, and physical constraint violation degree of each reconstructed residual field between the hidden state fields of each subset. The nodes of the conflict topology graph correspond to the source subsets, and the edges correspond to the conflict associations between two source subsets. Extract common conflict components that cannot be eliminated by a single source and source-specific components that disappear with the elimination of a specific source from the conflict topology graph. Write the common conflict components into an independent anomaly evidence channel and write the source-specific components into the credibility status update channel of the corresponding source.

9. The early warning system for hidden geological defects in structures based on multi-source fusion according to claim 7, characterized in that, The disease reconstruction processor organizes the joint state vector into a sliding state chain in the time dimension, and establishes drift continuation edges, credibility inheritance edges, and hidden state evolution edges between adjacent time windows; The drift continuation edge constrains the direction of change of latent drift variables from the same source in adjacent windows; Based on the aforementioned credibility inheritance edge, the credibility state after source mutual verification is passed; based on the aforementioned hidden state evolution edge, the disease hidden state increment that satisfies the hierarchical topological relationship is passed. When the prediction residual is interpreted by both the drift latent variable and the disease latent state variable in a continuous window, the drift latent variable and the disease latent state variable are frozen respectively, and two local reconstructions are performed. The corresponding drift continuation edge and latent state evolution edge are updated according to the difference in the residual structure of the two local reconstructions, and then the sliding state chain is jointly smoothed and solved.

10. The early warning system for hidden geological defects in building background strata based on multi-source fusion according to claim 8, characterized in that, The disease reconstruction processor organizes the common conflict components in the independent abnormal evidence channels into a conflict evidence chain by interleaving and continuing them according to the time progression relationship, and generates a combination of source subsets of common conflict classes, corresponding observation support domains, and physical constraint violation locations for each conflict evidence chain. In subsequent time windows, directional backtracking is performed on new observation evidence that overlaps with the supporting domain of the conflicting evidence chain, and the directional consistency relationship between the new observation evidence and the source subset combination, the topological adjacency relationship with the physical constraint violation position, and the temporal continuity relationship with the conflicting evidence chain are calculated respectively. The conflict evidence chain is updated based on the direction consistency relationship, topological adjacency relationship and time continuity relationship, and the updated common conflict component is used as an additional observation term to participate in the joint reconstruction of the disease latent state field and uncertainty field.

Images