A gouta multi-source disease digital twin detection early warning system

Abstract

This invention relates to the fields of computer data processing, digital twins, and ancient building detection and early warning technologies, specifically a digital twin detection and early warning system for multi-source defects in ancient pagodas. The system establishes twin semantic coordinates based on the pagoda's hierarchy, facade orientation, components, and adjacency relationships. It converts inspection images, 3D measurements, deformation monitoring, and historical repair information into defect events. Based on topological continuity, defect morphology direction, temporal correlation, and source credibility, it forms defect entities and evidence chains. The residual between the predicted twin state and the observed state is decomposed according to source, component location, and time window. Source mapping parameters are corrected based on single-source concentrated residuals, and the digital twin state is updated based on multi-source continuous residuals. The system also adjusts the association weights of subsequent defect events to generate early warning information. This invention can distinguish between data mapping deviations and actual defect evolution, reduce duplicate defect registration, erroneous merging, and twin state drift, and maintain a consistent correspondence between defect entities, evidence chains, and component states.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing, digital twins, and ancient building detection and early warning technology, specifically a digital twin detection and early warning system for multi-source defects in ancient pagodas. Background Technology

[0002] Digital twin detection of ancient pagoda defects typically uses 3D models or component information models as spatial carriers, importing inspection images, 3D measurement results, tilt and settlement data, vibration data, environmental records, and historical repair information into a unified platform. Existing systems generally attach data from different sources to corresponding tower layers, facades, or component locations through coordinate registration, timestamp alignment, and component number matching. For defects such as cracks, erosion, detachment, and deformation, defect records are created based on image recognition results, measurement anomalies, or manual annotations, and the component status is updated according to the acquisition time. Some solutions also use the data difference between adjacent time points to correct model parameters, or merge duplicate records according to preset fixed weights of data sources, and output alarms when defect values ​​exceed preset ranges. This type of implementation mainly relies on multi-source data aggregation and model status display. Cross-source correlation usually depends on spatial distance, component number, or temporal proximity. Model correction is usually based on the overall deviation between observed and predicted values, lacking joint processing of defect semantics, component topological relationships, data source errors, and the continuous evolution of defects.

[0003] The multi-source disease data of Guta exhibit differences in spatial reference, acquisition cycle, disease description granularity, and error characteristics. This makes it difficult for the system to reliably determine in the digital twin model whether the observational differences stem from data mapping bias or from actual changes in the disease itself. When the same crack is described separately by images, 3D measurements, and manual records, merging them solely by coordinates or time can easily lead to duplicate registration of the same disease or the merging of adjacent but discontinuous anomalous areas into a single disease entity. When a source experiences coordinate drift, scale deviation, or temporal misalignment, the overall residual may be directly interpreted as a change in component state and continued to be written into the twin model. Errors in correlation and updates propagate continuously in subsequent cycles, causing a loss of stable correspondence between disease entities, evidence chains, and component states, ultimately resulting in early warnings based on a shifted model state. Therefore, existing technologies lack a closed-loop processing mechanism that can utilize component topology and cross-source continuity to constrain disease correlations and distinguish between data bias and actual disease evolution through residual source decomposition. Summary of the Invention

[0004] The purpose of this invention is to provide a digital twin detection and early warning system for multi-source diseases in ancient pagodas, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a digital twin detection and early warning system for multi-source diseases of ancient pagodas, comprising a processor and a memory, wherein the memory stores program instructions, and the processor executes the program instructions to establish twin semantic coordinates based on the pagoda's hierarchy, facade orientation, component identification, and adjacency relationship.

[0006] Multi-source disease observation data are converted into disease events with spatiotemporal envelopes, disease categories, source identifiers, and confidence intervals.

[0007] Based on topological continuity, morphological direction of the disease, temporal correlation, and source credibility, the disease events are fused to form disease entities and evidence chains.

[0008] The residuals between the predicted and observed states of the twin are decomposed according to the data source, component location, and time window. The source mapping parameters are corrected or the digital twin states are updated based on the single-source concentrated residuals or the multi-source continuous residuals.

[0009] The association weights of subsequent disease events are adjusted based on the updated digital twin status, and early warning information is generated based on the corrected disease entity status.

[0010] Preferably, the twin semantic coordinates establish a hierarchical index according to the tower body, level, facade, component and local area of ​​the component, and record the geometric boundary, parent-child containment relationship and adjacency direction for each level of index.

[0011] The disease event is generated by projecting the image disease area, the three-dimensional measurement abnormal area, the deformation monitoring sequence and historical repair information onto the local area of ​​the corresponding component. Each disease event includes the observation time interval, spatial coverage, disease characterization vector, source error description and semantic index path. The processor retains the bidirectional mapping relationship between the original coordinates and semantic coordinates for the cross-source projection results.

[0012] Preferably, the processor constructs a candidate event association graph based on the component adjacency relationship. The nodes in the candidate event association graph correspond to disease events, and the edges correspond to cross-source event combinations that satisfy the component reachability relationship.

[0013] The spatial envelope overlap, morphological main direction consistency, temporal evolution continuity, and source complementarity are jointly calculated for each edge. The associated path is determined according to the constraints of continuity within the same component, directional propagation between adjacent components, and prohibition of direct merging between non-adjacent components.

[0014] Disease events aggregated along the associated path are written into the same disease entity and form a traceable chain of evidence according to the data source and observation time sequence.

[0015] Preferably, the processor maps the twin prediction state and the disease observation state to the same component semantic index to generate residual samples, and then forms a residual tensor according to the data source dimension, component topology dimension and sliding time window.

[0016] Source cohesion, cross-source consistency, and spatial continuity along the component adjacency path are extracted from the residual tensor. When the residual continuously aggregates in a single source dimension and lacks cross-source correspondence, a source deviation label is generated. When the residual appears synchronously in multiple source dimensions and is continuously distributed along the adjacency path, a disease evolution label is generated. The source deviation label and the disease evolution label are respectively associated with the source mapping parameter set and the digital twin state variable set.

[0017] Preferably, when establishing a bidirectional mapping relationship, the processor represents the spatial coverage of each data source as a local envelope with uncertain boundaries, and clips the cross-level projection based on the component geometric boundaries and parent-child inclusion relationships.

[0018] For observation results that simultaneously cover multiple components, they are divided into multiple sub-events according to the overlap ratio between the local envelope and the component boundary, and the original event identifier is written into each sub-event.

[0019] For observations from the same source obtained at different times, the semantic coordinates are corrected using the mapping offset of the stable component region. At the same time, the correction amount is written into the source error description for use in the disease event association and residual decomposition.

[0020] Preferably, the processor performs topologically constrained path expansion on the candidate event association graph. When multiple candidate paths exist between a new disease event and an existing disease entity, the processor compares the evidence source coverage, morphological extension continuity, and temporal sequence consistency of each candidate path, and retains the path that satisfies the reachability direction of the component and has no evidence conflict.

[0021] When the same diseased entity has mutually separated spatial envelopes and discontinuous evolutionary sequences, the corresponding evidence chain is split into multiple diseased entities.

[0022] When the associated paths of multiple diseased entities converge within the same component region and the source evidence is complementary, the diseased entities are merged and the entity identifiers before merging are retained.

[0023] Preferably, the processor extracts the residual baseline of the same data source within the stable component region for the source deviation mark, and decomposes the residual baseline into spatial translation component, scale deviation component and time drift component, and updates the coordinate mapping and time alignment parameters of the corresponding source according to each component.

[0024] For the disease evolution marker, the component semantic index, disease representation vector and observation time series corresponding to the continuous residuals are written into the digital twin state variable, and the state update range is restricted to not cross the component region that lacks adjacency relationship or lacks evidence chain connection.

[0025] The mapping parameters and state variables before and after the update are all saved as versioned records.

[0026] Preferably, the processor generates a correlation feedback quantity based on the source deviation marker and the disease evolution marker, the correlation feedback quantity including source credibility correction quantity, component path continuity correction quantity and disease morphology consistency correction quantity.

[0027] When subsequent disease events enter the candidate event association graph, the source credibility correction amount is applied to the candidate edge of the corresponding source, the component path continuity correction amount is applied to the corresponding adjacent path, and the disease morphology consistency correction amount is applied to the matching item of the corresponding disease category.

[0028] When the versioned digital twin state reverts, the associated feedback quantity corresponding to the reverted version is cancelled, and the historical weights of the corresponding candidate edges are restored.

[0029] Preferably, the processor performs local reassociation on the affected evidence chain after the diseased entity is split, merged, or the source mapping parameters are updated.

[0030] The local reassociation starts with the semantic index of the changed component, retrieves associated edges along the adjacent paths of the component that share the original event identifier, the observation period, or the disease characterization, and only recalculates candidate paths that have a topological reachability relationship with the changed region.

[0031] For disease events whose attribution changes before and after reassociation, record the original disease entity identifier, the target disease entity identifier, and the residual marker that triggered the change, and bind the record to the corresponding digital twin state version.

[0032] Preferably, the processor stores the correspondence between the associated feedback quantity, the attribution of the disease entity, and the residual label in multiple consecutive update cycles, and performs a consistency check on the attribution changes of the same disease event in adjacent update cycles.

[0033] When the attribution change is triggered by a single source confidence correction, the digital twin state variable update is frozen and the source bias decomposition is re-executed.

[0034] When the attribution change has both cross-source consistent residuals and continuous component paths, the digital twin state update is retained and the corresponding associated path is solidified.

[0035] The results of freezing, re-decomposing, retaining, and solidifying are written into the versioned evidence chain.

[0036] Compared with existing technologies, the beneficial effects of this invention are as follows: 1. Based on the pagoda's hierarchy, facade orientation, component identification, and adjacency relationships, twin semantic coordinates are established. Disease observations from different sources are uniformly converted into disease events with spatiotemporal envelopes, disease categories, source identification, and confidence intervals. Within the reachable range of components, spatial overlap, morphological orientation, temporal continuity, and source complementarity are combined to form disease entities and evidence chains. Therefore, cross-source disease association no longer relies solely on coordinate proximity or temporal proximity. Multiple observations of the same disease can be merged along the component topology and evolutionary sequence, isolating adjacent but non-continuous anomalous areas. The residual between the twin predicted state and the observed state is decomposed according to data source, component location, and time window. Single-source concentrated residuals are used to correct source mapping parameters, while multi-source continuous residuals are used to update disease states, preventing measurement drift and mapping errors from being transmitted to the digital twin state, and maintaining a consistent correspondence between disease entities, evidence chains, and component states.

[0037] 2. The system performs quantitative corrections for spatial translation, scale bias, and temporal drift of data sources, and saves version records of mapping parameters, twin states, and evidence chains, enabling subsequent disease events to be re-associated in the corrected semantic coordinates. When disease entities are split, merged, or source mapping parameters change, local re-association is performed only along the adjacent paths of the affected components, avoiding repeated calculations of irrelevant areas. The association feedback quantity is further adjusted according to source credibility, component path continuity, and disease morphology consistency to reduce the recurrence of similar mismatches within continuous update cycles. For cases where the attribution of disease events changes, the system freezes state updates, re-decomposes source bias, or solidifies association paths based on single-source residuals or cross-source continuous residuals, and writes the processing results into a versioned evidence chain, facilitating the restoration of historical states and verification of the data sources, association processes, and state change records upon which the early warning is based. Attached Figure Description

[0038] Figure 1 A flowchart for semantic mapping and disease entity construction of multi-source disease data.

[0039] Figure 2 This is a candidate decomposition graph for semantic mapping credibility.

[0040] Figure 3 This is a flowchart of the disease association correction process based on topological constraints and residual decomposition.

[0041] Figure 4 This is a graph showing the correlation degree of candidate edges.

[0042] Figure 5 This is a time slice plot of the residual tensor.

[0043] Figure 6Convergence plot of source correction residuals.

[0044] Figure 7 This is a trend chart showing the feedback of candidate association weights.

[0045] Figure 8 This is a decomposition diagram of the early warning state variables.

[0046] Figure 9 A flowchart for the generation of disease correlation and early warning driven by versioned evidence chain. Detailed Implementation

[0047] In one embodiment, a digital twin detection and early warning system for multi-source defects in ancient pagodas includes a processor and a memory. The memory stores program instructions, and the processor executes the program instructions to establish twin semantic coordinates based on the pagoda's hierarchy, facade orientation, component identification, and adjacency relationships. It converts multi-source defect observation data into defect events with spatiotemporal envelopes, defect categories, source identifiers, and confidence intervals. Based on topological continuity, defect morphology direction, temporal correlation, and source credibility, the defect events are fused to form defect entities and evidence chains. The residual between the twin prediction state and the observation state is decomposed according to data source, component location, and time window. Source mapping parameters are corrected or the digital twin state is updated based on single-source concentrated residuals or multi-source continuous residuals. The association weights of subsequent defect events are adjusted according to the updated digital twin state, and early warning information is generated based on the corrected defect entity state.

[0048] In this embodiment, the processor organizes program instructions according to a data access layer, a semantic mapping layer, a disease association layer, a residual discrimination layer, a twin update layer, and an early warning output layer. The processing results are exchanged between the layers through a unified disease event data structure. The data access layer reads inspection images, three-dimensional measurement data, tilt settlement sequences, vibration responses, environmental records, and historical repair information. The semantic mapping layer projects various observation results onto the semantic space of the ancient pagoda components. The disease association layer merges cross-source observations belonging to the same disease process into disease entities. The residual discrimination layer distinguishes between source mapping deviations and actual disease changes. The twin update layer saves the corrected component status and historical versions. The early warning output layer generates an early warning record containing component location, disease category, status level, and evidence index based on the current status, direction of change, and completeness of evidence of the disease entity.

[0049] Specifically, the digital twin model of the ancient pagoda adopts a hierarchical component data organization method. The pagoda body is the root node, each pagoda layer is the first-level child node, each facade is the second-level child node, brackets, columns, beams, walls, eaves or other modeled components are the third-level child nodes, and the local areas divided on the surface or inside of the components are the fourth-level child nodes. Each node is configured with a unique semantic identifier, geometric boundary, parent node identifier, child node set, adjacent node set and orientation attribute. The adjacent node set records the relationships such as vertical connection, left and right adjacency, internal and external correspondence and cross-component continuity, so that image coordinates, three-dimensional spatial coordinates, measurement point numbers and manual inspection descriptions can be converted into the same semantic index path.

[0050] When establishing twin semantic coordinates, the processor reads the geometric range of components in the existing digital model of the ancient pagoda and converts the geometric range of components into a semantic region jointly described by bounding boxes, surface meshes, and local parameter coordinates. The bounding boxes are used for fast candidate component retrieval, the surface meshes are used to determine the projection position of the observation point on the surface of the component, and the local parameter coordinates are used to record the lateral, longitudinal, and depth positions of the defects relative to the component's own reference plane. When the observation data lacks complete three-dimensional coordinates, the processor parses the semantic description based on the tower layer, facade, component name, and adjacent reference objects, and establishes candidate positions with uncertain boundaries in the corresponding semantic region.

[0051] Please refer to Figure 1 After the multi-source disease observation data enters the system, the original data identifier and source identifier are retained respectively. Crack areas, erosion areas or peeling areas in the inspection images are converted into spatial envelopes through pixel masks and image shooting poses. Local depressions, protrusions, misalignments or missing areas in the three-dimensional measurement data are converted into spatial envelopes through anomaly point sets. The tilt settlement sequence is converted into temporal anomaly envelopes according to the component to which the measuring point belongs and the observation period. Historical repair information is converted into historical event envelopes according to the repair location, repair object and recording date. All types of envelopes are associated with disease category, observation value, time interval, source error description and semantic index path.

[0052] The processor uses semantic mapping confidence to determine the correspondence between each piece of observation data and the candidate component region. The semantic mapping confidence is weighted according to the degree of coordinate inclusion, geometric overlap, hierarchical semantic consistency, and source location reliability. The calculation formula is as follows: , ;in, This represents the semantic mapping confidence level of the i-th observation data. This indicates the degree to which the observation center falls within the geometric boundary of the candidate component. This indicates the degree of overlap between the observation space envelope and the local region of the candidate component. This indicates the degree of consistency between the names of tower floors, facades, and components in the observation description and the hierarchical level of the candidate semantic index. This indicates the reliability of the location of the corresponding data source. , , and These represent the weights of the four indicators, and their sum is 1. When a certain inspection image... Take 0.9, Take 0.8, Take 1, When the weight is 0.7 and all four weights are 0.25, the result is... The processor maps the observation data to the corresponding local region of the component and saves 0.85 as the mapping confidence information. When multiple candidate regions obtain similar confidence, multiple candidate mappings are retained instead of immediately determining a unique attribution.

[0053] Please refer to Figure 2 , Figure 2 The horizontal axis represents the semantic mapping candidate sampling stage, consisting of the initial segment, the first segment, the progressive segment, the middle-early segment, the middle segment, the middle-late segment, the late segment, and the final segment. The vertical axis represents the mapping confidence level. The solid black line represents the change in mapping confidence level between each candidate observation data and the candidate component region of the ancient pagoda, the dashed gray line represents the candidate retention threshold, and the dotted gray line represents the main candidate threshold.

[0054] from Figure 2 As can be seen, the mapping confidence curve is generally located near the candidate retention threshold and shows a slow downward trend. This indicates that when the same batch of inspection images, 3D measurement anomalies, and deformation monitoring records enter the semantic coordinate system, differences will arise due to variations in the degree of spatial envelope overlap, the consistency of component hierarchical descriptions, and the reliability of source location. The curve from the initial segment to the previous segment is close to the retention threshold, indicating that multiple component local areas have obtained similar scores at this stage. The processor should not immediately determine a unique attribution but should retain multiple candidate mappings as described in the original text.

[0055] The curve gradually decreases after the middle section, indicating that although some candidate components have a spatial proximity relationship with the observation envelope, their hierarchical semantic path or source location reliability is insufficient, and they cannot be used as the main candidate to directly enter the solidification process of the disease entity.

[0056] The disease events are stored using a unified data structure, which includes at least an event identifier, original data identifier, source identifier, observation start time, observation end time, semantic index path, spatial envelope, disease category, disease characterization vector, observation confidence interval, and source error description. The disease characterization vector stores crack length, crack direction, regional boundary morphology, local deformation direction, temporal change trend, or repair status according to the data type. The observation confidence interval stores the range in which the observed value may fall. The source error description stores coordinate deviation, scale deviation, time deviation, and candidate ranges for undetermined semantic mapping.

[0057] In a preferred embodiment, the processor uses different disease characterization extraction methods for different data sources, but converts them to a unified characterization dimension after the event is generated. The image disease area is represented by the main direction of the region, the skeleton direction, the boundary curvature and the area distribution. The three-dimensional measurement anomaly is represented by the normal offset, the local elevation change, the spatial extension direction and the point set density. The deformation monitoring sequence is represented by the trend slope, the periodic component, the abrupt change location and the duration. The historical repair information is represented by the disease type before repair, the repair area, the repair time and the subsequent review status. Missing characterization components are recorded with an unavailable identifier to avoid directly assigning missing data to a disease-free state.

[0058] Please refer to Figure 3 After the processor completes the conversion of disease events, it establishes a candidate event association graph according to the adjacency relationship of components. Each node in the candidate event association graph corresponds to a disease event. Any two disease events are only considered as having the same semantic index, being adjacent to each other's components, or having a traceable propagation path. No candidate edge is established for event combinations whose components are not adjacent and whose disease spatial envelopes do not share a boundary. For event combinations whose collection times are far apart but are connected by historical repair information, candidate edges with historical association tags are established, thereby limiting the retrieval scope of cross-source disease merging.

[0059] The degree of association for each candidate edge is determined based on the spatial envelope overlap, the consistency of the main direction of the lesion morphology, the continuity of temporal evolution, the complementarity of sources, and the amount of topological conflict. The degree of association is calculated according to the following formula: ;in, This indicates the degree of correlation between the i-th disease event and the j-th disease event. The component topology reachability flag is set to 1 if it is reachable and 0 if it is not reachable. Indicates the overlap of spatial envelopes. This indicates the degree of consistency in the main direction of the disease morphology. Indicates the continuity of time evolution. Indicates the complementarity of sources. Indicates the amount of conflict of evidence. to This represents the non-negative weight of each item, when two events are in the same component and their principal directions are continuous. Set to 1 when the two events are located on non-adjacent facades and do not share a common boundary. If the value is set to 0, no associated path will be established even if the time is close. When the amount of evidence conflict increases, the degree of association will decrease accordingly and the path will be retained as a candidate edge to be verified.

[0060] Please refer to Figure 4 , Figure 4 The horizontal axis represents the candidate edge sampling stage, and the vertical axis represents the degree of correlation between candidate edges. The solid black line is the curve showing the change in the degree of correlation between candidate events, the dashed gray line is the threshold to be verified, and the dotted gray line is the merging threshold. The legend indicates the judgment meaning of the curves and threshold lines.

[0061] from Figure 4 As can be seen, the candidate edges in the initial segment have a high degree of correlation, indicating that the corresponding disease events have strong continuity in terms of component semantic index, spatial envelope position, morphological main direction, and temporal evolution relationship. This meets the constraints of establishing candidate edges when the semantic indexes are the same, the components are adjacent, or there is a traceable propagation path. The curve shows a significant drop after entering the first segment, indicating that although some events are close in time, their spatial boundaries, component adjacency relationships, or complementary evidence sources are insufficient, and they need to enter the verification state.

[0062] The curve from the middle to the end continues to be below the merging threshold, indicating that the amount of conflicting evidence has an inhibitory effect on the degree of correlation, which can prevent disease events that are not adjacent to each other or lack shared boundaries from being directly merged.

[0063] The spatial envelope overlap is calculated based on the intersection and union of the two disease events projected onto the local coordinates of the same component. The consistency of the main direction of the disease morphology is calculated based on the angle between the direction of the crack skeleton, the extension direction of the anomalous point cloud, or the direction of the deformation gradient. The continuity of temporal evolution is calculated based on the time interval between two observations and whether the changes in disease characteristics conform to a continuous change relationship. The complementarity of sources is calculated based on whether the two events come from mutually independent data sources and the characterization dimensions that each source can provide. The amount of evidence conflict is calculated based on the degree of inconsistency in disease categories, separation of spatial boundaries, opposite change directions, or inconsistency in historical repair status.

[0064] Disease events are aggregated into disease entities along reachable paths in the candidate event association graph. Each disease entity stores an entity identifier, its current component, coverage area, disease category, first discovery time, most recent observation time, state variables, and evidence chain. The evidence chain stores each disease event and its source in chronological order, and records the association edges, degree of association, semantic mapping version, and data processing version between events. When the same disease entity contains multiple data sources, the original observations of each source are retained without using fused values ​​to overwrite the original observations, so that subsequent residual decomposition can locate the source of the anomaly.

[0065] The digital twin model generates component observation states according to the disease entities. The component observation states consist of the spatial location, disease category, morphological parameters, change trend, and evidence completeness of each disease entity within the component. The twin prediction states are derived from the component states of the previous update cycle, the evolution records of confirmed disease entities, and the component adjacency states. Before entering the residual calculation, the observation states and prediction states are converted to the same component semantic index, the same disease representation dimension, and the same time base. Representation components that cannot be directly corresponding are not included in this residual calculation and are reserved as supplementary markers.

[0066] In this embodiment, the processor forms a residual tensor according to the data source, component location, and time window. The basic calculation relationship of the residual tensor is as follows: ;in, This represents the state residual of data source s within component location c and time window t. This indicates the observation status of the disease originating from the corresponding source. The digital twin model predicts the state of component position c within a time window t. When both the image source and the 3D measurement source obtain positive residuals in the same component region, the multiple source dimensions in the residual tensor have a unidirectional distribution. When only one image source continuously produces residuals in multiple stable regions with approximate directions, the residual tensor exhibits a single source dimension concentration and insufficient cross-source correspondence.

[0067] Please refer to Figure 5 , Figure 5 The horizontal axis represents the residual sampling stage, and the vertical axis represents the state residuals. The solid black line represents the state residual sequence formed by each source within the corresponding component and time window, the gray dashed line represents the attention threshold, the gray dotted line represents the review threshold, and the other gray dashed line represents the reverse attention threshold.

[0068] from Figure 5As can be seen, the state residual curve exhibits periodic peaks and low-value declines, indicating differences between the observed state and the twin predicted state in different component windows, based on sources such as inspection images, 3D measurements, and tilt settlement sequences. Some peaks exceed the verification threshold, indicating a significant positive deviation between the observed state and the predicted state. If this deviation occurs simultaneously from multiple independent sources and is continuously distributed along adjacent components, it corresponds to the disease evolution marker generation conditions described in the text.

[0069] The curve also shows rapid declines or isolated peaks within a single stage, indicating that the residuals may be caused by source mapping bias, spatial projection error, or time base drift.

[0070] The processor extracts source cohesion, cross-source consistency, and spatial continuity from the residual tensor. Source cohesion describes whether residuals from the same source maintain similar directions and magnitudes across multiple stable components or multiple time windows. Cross-source consistency describes whether residuals from different sources exhibit the same direction within the same component region. Spatial continuity describes whether residuals form a continuous distribution along adjacent paths of components. When high source cohesion and low cross-source consistency occur simultaneously, a source deviation marker is generated. When multiple independent sources form residuals in the same direction in adjacent component regions and spatial continuity satisfies the component reachability relationship, a disease evolution marker is generated.

[0071] The determination of single-source concentrated residuals is not based on a single instantaneous observation value. The processor summarizes the residual distribution corresponding to the same source within a sliding time window, and selects the component area with no active disease entities, stable historical state and consistent observations from other sources as stable component area. The residuals with the same spatial direction or time offset characteristics within the stable component area are used as evidence of source deviation, so as to avoid misjudging a single local disease change as a mapping deviation of the entire data source.

[0072] The determination of multi-source continuous residuals is based on the mutual support between independent sources and the adjacency relationship of components. The processor checks whether the spatial envelopes corresponding to the residuals of each source overlap in the same component region or extend continuously along the adjacent components, and checks whether the residuals correspond to the disease characteristics have compatible morphological directions and temporal order. When the direction of image crack extension is consistent with the direction of extension of three-dimensional anomaly point set and subsequent component deformation observations appear along the same adjacent path, the corresponding residuals are written into the evidence set associated with the disease evolution marker.

[0073] The source deviation marker is used to trigger source mapping parameter correction. The processor extracts the residual baseline from the same source from the stable component region and decomposes the residual baseline into spatial translation component, scale deviation component and time drift component. The spatial translation component is determined according to the uniform offset direction between the observation envelope and the boundary of the stable component. The scale deviation component is determined according to the ratio of the observation range to the model range in multiple stable component regions. The time drift component is determined according to the time difference of periodic observation characteristics or synchronous events in different sources. After each component obtains a compatible result in multiple stable regions, it is written into the source mapping parameter set.

[0074] The correction of spatial location and time reference is expressed by the following formula: , ;in, , and Represents the spatial coordinates of the i-th observation data before correction. , and Represents the corrected spatial coordinates. This represents the scale correction amount for source s. , and This represents the translation correction amount of source s in three spatial directions. Indicates the observation time before correction. Indicates the corrected observation time. Indicate source The time drift correction when there is no scale bias at a certain source. Set to 1; when the observed coordinates are uniformly offset laterally, only adjust... When its recording time lags behind the unified time benchmark, it will be processed through the corresponding... Move the observed event to the correct time window.

[0075] Please refer to Figure 6 , Figure 6 The horizontal axis represents the source correction sampling stage, and the vertical axis represents the residual baseline. The solid black line represents the residual baseline change in the stable component region, the dashed gray line represents the upper limit of the stable band, and the dotted gray line represents the correction trigger line.

[0076] from Figure 6 As can be seen, the initial residual baseline shows a significant rise, followed by a rapid decline and gradual approach to the stable zone. This indicates that after the processor detects a consistent offset at the boundary of a stable component, the location of a stable texture, or the surface of a stable 3D surface, it performs spatial translation, scale, and temporal drift corrections on the corresponding sources. This change is consistent with the technical constraint in the paper that displacements occurring only in a single disease area are not considered as global mapping corrections, emphasizing that the correction parameters originate from the common offset of the stable region, rather than local changes within the disease area itself.

[0077] The curve from the middle to the end approaches the low residual region, indicating that after the corrected observation coordinates and observation time are reprojected onto twin semantic coordinates, the error in the stable region converges significantly.

[0078] After the processor completes the source mapping parameter correction, it does not directly delete the data records before correction. Instead, it saves the coordinate mapping relationship, time mapping relationship, correction basis, residual mark and parameter version before and after correction, and reprojects the affected disease events to twin semantic coordinates. For events that still fall into the local area of ​​the original component after reprojection, its spatial envelope is updated. For events that cross the component boundary after reprojection, candidate components are re-determined and enter the local re-association process.

[0079] Disease evolution markers are used to trigger digital twin state updates. The processor writes the component semantic index, disease representation vector, observation time series and cross-source evidence corresponding to the continuous residuals into the corresponding disease entity. The current spatial range of the disease entity is adjusted to cover the region that can cover the newly added observations and meet the component boundary constraints. The time series of disease representations is updated according to continuous observations. The state update range is limited to the component region that the evidence chain can connect and does not cross the component boundary that lacks adjacency relationship, lacks disease morphological continuity or lacks temporal continuity.

[0080] When updating the digital twin state, the processor creates a new version for each state change. The new version saves the previous version identifier, the diseased entity that changed, the state before the change, the state after the change, the disease event that triggered the change, the corresponding residual tensor region, and the discrimination flag. The state of the unchanged components remains unchanged through version reference. When subsequent evidence shows that an update originated from a data mapping deviation, the state before the change can be restored along the version relationship, and the attribution of the disease event and the associated feedback quantity associated with that version can be revoked.

[0081] Furthermore, the processor generates associated feedback quantities based on the updated digital twin state. The associated feedback quantities include source confidence correction quantity, component path continuity correction quantity, and disease morphology consistency correction quantity. The source confidence correction quantity is determined by the frequency of occurrence of source deviation markers, the convergence of corrected residuals, and the consistency of stable component regions. The component path continuity correction quantity is determined by the extension of confirmed disease entities along adjacent paths. The disease morphology consistency correction quantity is determined by the stability of the main direction and boundary morphology of similar diseases in continuous observation.

[0082] The candidate event association weights are updated and fed back according to the following formula: ;in, This represents the candidate association weight between event i and event j in the nth update cycle. Indicates the candidate association weights for the next update cycle. This indicates the amount of source credibility correction. This represents the component path continuity correction amount. This indicates the correction amount for the consistency of disease morphology. , and This represents the update coefficient corresponding to the correction amount, when a source still produces single-source concentrated residuals in the stable region after correction. Take a negative value when multiple sources confirm the extension of the defect along the same component path. Take a positive value when the shape and direction of the newly added disease event are opposite to those of the confirmed entity. Take the negative value.

[0083] Please refer to Figure 7 , Figure 7 The horizontal axis represents the feedback update sampling stage, and the vertical axis represents the updated weights. The solid black line represents the update trend of candidate event association weights, the dashed gray line represents the stable merging line, and the dotted gray line represents the strong evidence line.

[0084] from Figure 7 As can be seen, the weight curves rise and fall periodically, indicating that the weights of candidate edges are not fixed all at once, but are continuously adjusted according to the source correction results, the continuity of the adjacent paths of components, and the consistency of the disease morphology. The curves rise to near the strong evidence line in multiple stages, indicating that the corresponding candidate edges have obtained cross-source evidence support, their morphological direction is stable and continuous along the component path, and they are suitable for being solidified into the disease entity evidence chain.

[0085] Meanwhile, some stage curves fell rapidly, indicating that the corresponding candidate edges were affected by negative feedback from source bias, path discontinuity, or morphological conflict.

[0086] The correlation feedback quantity is applied only to the candidate edges corresponding to its evidence range, the source credibility correction quantity is applied to the candidate edges containing the corresponding source, the component path continuity correction quantity is applied to the confirmed adjacent paths and their direct extension paths, and the disease morphology consistency correction quantity is applied to the morphology matching items corresponding to the disease category. The processor does not use the feedback quantity of a certain local disease to change the correlation relationship of all ancient pagoda components, so as to avoid the diffusion effect of local evidence on irrelevant components.

[0087] When generating early warning information based on the corrected state of the diseased entity, the processor reads the current component location, spatial coverage, disease category, continuous observation sequence, number of cross-source evidence, latest update time and associated stable state of the diseased entity, and compares it with the preset state boundary of the component. The early warning information includes the diseased entity identifier, component semantic index, disease state, change direction, evidence chain entry point and digital twin state version. When the diseased entity is still in the source deviation re-decomposition state or the attribution freeze state, the early warning information is marked as pending verification instead of directly treating the single-source anomaly as the confirmed disease state.

[0088] The early warning status of the diseased entity can be calculated using the following formula: ;in, This represents the warning status quantity for the diseased entity e. This indicates the degree to which the observed defect value is close to the corresponding component's state boundary. Indicates the degree of change within a continuous time window. Indicates the extent to which the damaged entity extends along the adjacent path of the component. Indicates the consistency of the chain of evidence. , , and This represents the weight of the four state components. When the observed damage values ​​are not yet close to the state boundary but extend continuously along adjacent components and the cross-source evidence is consistent, and As the corresponding value increases, when the observed disease value is close to the state boundary but originates from a single source with a bias marker, Keep the value low and retain the warning information in a pending verification state.

[0089] Please refer to Figure 8 , Figure 8 The horizontal axis represents the sampling stage of the early warning status, and the vertical axis represents the status quantity. The solid black line represents the change in the early warning status quantity corresponding to the disease entity, the gray dashed line represents the attention level, and the gray dotted line represents the high level.

[0090] from Figure 8 As can be seen, the initial stage state variables are below the attention level, indicating that although the disease entity has formed an evidence chain, the degree of proximity of the component state boundary, the degree of change of the continuous window, or the degree of extension along the adjacent path has not yet reached a high level. The curve of the advancement stage rises rapidly and approaches the high level, indicating that the disease observation value at this stage is close to the component state boundary, and multiple sources or adjacent component paths support the disease evolution. A higher-level warning record should be generated and the evidence chain entry should be bound.

[0091] The curve drops significantly after the middle and early stages, indicating that when the consistency of the evidence chain is insufficient, the source deviation has not been resolved, or the disease entity is still in a state of re-decomposition, the system will retain the pending verification mark instead of directly solidifying the single-source anomaly as a confirmed disease state.

[0092] In a preferred embodiment, the twin semantic coordinates establish a hierarchical index according to the tower body, level, facade, components, and local areas of components, and record the geometric boundaries, parent-child containment relationships, and adjacency directions for each level of index. The tower body index is used to distinguish different ancient tower objects, the level index is used to determine the vertical position, the facade index is used to determine the orientation area, the component index is used to determine the object bearing the disease, and the local area index of the component is used to record the specific distribution range of the disease on the surface or inside the component. Each level of index uses the combination of the parent identifier and the current identifier to generate a complete semantic path.

[0093] The geometric boundary includes the three-dimensional enclosure range of the component, the observable surface range, and the local partition boundary. The parent-child containment relationship is used to limit the observation event to be traced back from the local area to its own component, facade, and tower layer. The adjacency direction is used to record the spatial direction that the disease may continuously cross. When two components are only close in coordinates in the digital model but have no connection relationship, they are not registered as adjacent components. When the same component is divided into multiple local areas, each local area registers a local adjacency relationship according to the shared boundary.

[0094] The aforementioned defects are generated by projecting the image defect area, the 3D measurement anomaly area, the deformation monitoring sequence, and historical repair information onto the corresponding local area of ​​the component. The image defect area is projected based on the shooting pose, the correspondence between image feature points and component surface control points. The 3D measurement anomaly area is directly retrieved based on the coordinates of the point set and projected onto the surface grid. The deformation monitoring sequence is bound to the local area of ​​the component based on the measurement point identifier. The historical repair information is parsed using semantic indexes based on the tower layer, facade, component name, and location description recorded in the records.

[0095] Each disease event includes the observation time interval, spatial coverage, disease characterization vector, source error description, and semantic index path. The processor retains the bidirectional mapping relationship between the original coordinates and semantic coordinates for the cross-source projection results. The mapping from the original coordinates to the semantic coordinates is used for disease event attribution and cross-source association, while the mapping from the semantic coordinates to the original coordinates is used to look up the image pixel region, three-dimensional measurement point set, deformation monitoring record, or historical document location. The bidirectional mapping relationship is stored together with the source mapping parameter version.

[0096] In another preferred embodiment, when establishing a bidirectional mapping relationship, the processor represents the spatial coverage of each data source as a local envelope with uncertain boundaries. The uncertain boundaries extend outward from the spatial coverage according to the source error description. The image projection error corresponds to forming a two-dimensional extension range along the surface of the component, the three-dimensional measurement error corresponds to forming a three-dimensional extension range in the spatial direction, and the artificial position description corresponds to forming a discrete extension range composed of candidate local regions. The local envelope stores the distinguishing markers between the observation core region and the uncertain boundary region.

[0097] The processor clips cross-level projections based on the geometric boundaries of components and the parent-child containment relationship. When the spatial envelope of the defect exceeds the current component boundary but is still within the parent facade, the excess part is checked for overlap with adjacent components. When the spatial envelope of the defect exceeds the parent facade, the facade affiliation is recalculated based on the original projection relationship. The original envelope is not deleted during the clipping process. Instead, the original envelope, the clipped envelope, and the clipped area are saved separately so that they can be restored and reprojected when the source mapping parameters change.

[0098] For observation results that simultaneously cover multiple components, the processor splits them into multiple sub-events according to the overlap ratio between the local envelope and the boundaries of each component. Each sub-event inherits the disease category, time interval, and source identifier of the original event, and saves the spatial envelope and semantic index path within the corresponding component. The original event identifier is written into the parent event field of each sub-event. When candidate events are associated, multiple sub-events are identified as originating from the same observation result by sharing the original event identifier, but sub-events within different components are not directly merged into the same disease entity based on this.

[0099] For observations from the same source obtained at different times, the processor selects a stable component region as a registration reference, compares the mapping offsets of adjacent time observations on the stable component boundary, stable texture position, or stable 3D surface, and uses the common offset components of multiple stable component regions to correct semantic coordinates. Displacements that only appear in a single disease area are not used as global mapping corrections. The corrections are written into the source error description and bound to the observation time interval for subsequent disease event association and residual decomposition.

[0100] In a further embodiment, the processor constructs a candidate event association graph based on the component adjacency relationship. The nodes in the candidate event association graph correspond to the disease events, and the edges correspond to cross-source event combinations that satisfy the component reachability relationship. The component reachability relationship is determined by the same local area, shared component boundaries, explicit vertical connection relationship, or continuous paths in historical disease propagation records. The candidate edges store the starting event identifier, the target event identifier, the component path, the set of association indicators, and the set of evidence conflicts.

[0101] The processor determines the associated path according to the constraints of continuity within the same component, directional propagation between adjacent components, and prohibition of direct merging between non-adjacent components. Continuity within the same component requires that the spatial envelope of the disease overlaps, connects, or has a connectable region along a common main direction. Directional propagation between adjacent components requires that the main direction of the disease can be continuously extended through a shared boundary. Even if non-adjacent components have similar disease categories and similar observation times, no direct associated edge is established, unless there are continuous events in the chain of evidence that pass through intermediate components.

[0102] Disease events aggregated along the associated path are written into the same disease entity and form a traceable evidence chain according to the data source and observation time sequence. Each event node in the evidence chain stores the predecessor event, successor event, candidate edge association degree and processing version. When a new event has multiple candidate paths with an existing disease entity, the processor constructs candidate evidence chains respectively, without covering other candidate paths in a single comparison. The retained path is determined after the source complementarity, morphological continuity and time consistency checks are completed.

[0103] In a preferred embodiment, the processor performs topologically constrained path expansion on the candidate event association graph. When multiple candidate paths exist simultaneously between a new disease event and an existing disease entity, the processor compares the evidence source coverage, morphological extension continuity, and temporal sequence consistency of each candidate path. The evidence source coverage is determined based on the number of independent sources in the path and the disease representation dimensions that each source can provide. The morphological extension continuity is determined based on the connection relationship between the main direction of adjacent events and the envelope boundary. The temporal sequence consistency is determined based on whether the event time in the path matches the disease expansion order.

[0104] The processor retains only paths that satisfy the reachability of components and have no evidence conflict. Evidence conflict includes incompatible disease categories appearing in the same component area at the same time, unconnectable intervals between disease spatial envelopes, subsequent observations showing that the disease range is smaller than the previously confirmed range that has not been repaired, and the same source simultaneously mapping an original observation to mutually exclusive component areas. Paths with evidence conflict are marked as paths to be verified and the source of the conflict is retained.

[0105] When the same diseased entity has mutually separated spatial envelopes and discontinuous evolutionary sequences, the processor uses the break point in the evidence chain as the dividing point to divide the corresponding event into multiple sub-chains and generate new diseased entity identifiers for each. Each new diseased entity inherits the event records in the original entity that correspond to its own spatial envelope and time sequence. The original diseased entity enters the splitting history record. The entity identifiers before and after the splitting, as well as the spatial separation evidence and time discontinuity evidence that triggered the splitting, are stored together.

[0106] When the associated paths of multiple diseased entities converge within the same component area and the source evidence is complementary, the processor checks the disease category, morphological direction, time series and historical repair status of each entity. If there is no mutually exclusive evidence, the multiple diseased entities are merged. The merged diseased entity retains all entity identifiers and evidence chains before the merger, and records the merger source through the entity relationship field. If the subsequent residual judgment indicates that there is a mapping deviation in one of the sources, the entity affiliation before the merger can be restored based on the merger history.

[0107] In a preferred embodiment, the processor maps the twin prediction state and the disease observation state to the semantic index of the same component to generate residual samples, and then forms a residual tensor according to the data source dimension, the component topology dimension and the sliding time window. The data source dimension distinguishes images, three-dimensional measurements, deformation monitoring, vibration response, environmental records and historical repair information. The component topology dimension is arranged according to the semantic index and the adjacent path. The sliding time window moves according to the comparable time range of the observation data.

[0108] Source cohesion, cross-source consistency, and spatial continuity along the adjacent path of the component are extracted from the residual tensor. When the residual continuously aggregates in a single source dimension and lacks cross-source correspondence, a source deviation label is generated. When the residual appears synchronously in multiple source dimensions and is continuously distributed along the adjacent path, a disease evolution label is generated. The source deviation label is associated with the source identifier, the time range of influence, the stable component region, and the candidate mapping parameters. The disease evolution label is associated with the disease entity, the component path, the cross-source event set, and the state update range.

[0109] In a further embodiment, the processor extracts the residual baseline of the same data source within the stable component region for the source deviation marker, and decomposes the residual baseline into spatial translation component, scale deviation component and time drift component. The decomposition process uses the common residual relationship of multiple stable component regions to constrain each component to prevent the local disease state of a single region from entering the source mapping parameters. After the decomposition is completed, the coordinate mapping and time alignment parameters of the corresponding source are updated according to each component.

[0110] For disease evolution markers, the processor writes the component semantic index, disease representation vector and observation time sequence corresponding to the continuous residuals into the digital twin state variable, and restricts the state update range to not cross the component region that lacks adjacency relationship or evidence chain connection. The updated state variable and the original state variable are saved separately according to version. The version record includes state update time, data source set, disease event set, association path and residual discrimination result.

[0111] In a preferred embodiment, please refer to Figure 9The processor generates associated feedback quantities based on source deviation markers and disease evolution markers. When a subsequent disease event enters the candidate event association graph, it applies source credibility correction quantities to the candidate edges of the corresponding source, applies component path continuity correction quantities to the corresponding adjacent paths, and applies disease morphology consistency correction quantities to the matching items of the corresponding disease categories. The associated feedback quantities are bound to the state version that generates the feedback to avoid the correction relationships formed by different versions from overlapping each other.

[0112] When the versioned digital twin state is rolled back, the processor reads the state change record after the rolled-back version, cancels the associated feedback quantity corresponding to the rolled-back version, restores the historical weight of the corresponding candidate edge, and resets the evidence events of the newly added disease entities in the rolled-back version to the pending association state. The original data, original disease events and historical evidence chains are all retained. The rollback operation only changes the current activity state and the current candidate association relationship.

[0113] In another preferred embodiment, after the disease entity is split, merged, or the source mapping parameters are updated, the processor performs local reassociation on the affected evidence chain. The local reassociation starts from the semantic index of the changed component, searches along the component adjacency path for associated edges that share the original event identifier, the observation period, or the disease characterization, and limits the search depth to no more than the range of components that the existing evidence chain can continuously reach.

[0114] The processor only recalculates candidate paths that have topological reachability to the changed region. It does not maintain the original association results for disease entities that do not share adjacent paths, original events, observation periods, or disease characteristics with the changed components. For the affected candidate paths, it recalculates the spatial envelope overlap, morphological main direction consistency, temporal evolution continuity, source complementarity, and evidence conflict, and redetermines the path affiliation based on the updated source mapping parameters.

[0115] For disease events whose attribution changes before and after reassociation, the processor records the original disease entity identifier, the target disease entity identifier, the original semantic index, the new semantic index, the residual marker that triggered the change, and the reassociation path, and binds the records to the corresponding digital twin state version. When a disease event is transferred from one disease entity to another disease entity, the transfer record is retained in the original evidence chain instead of directly deleting the event history.

[0116] In a further embodiment, the processor stores the correspondence between the associated feedback quantity, the attribution of the disease entity, and the residual label in multiple consecutive update cycles, and performs a consistency check on the attribution change of the same disease event in adjacent update cycles. The consistency check reads the semantic mapping version of the event, the candidate association weight, the source credibility correction amount, the component path continuity correction amount, and the disease morphology consistency correction amount to determine the direct triggering source of the attribution change.

[0117] When the attribution change is triggered by a single source confidence correction, the processor freezes the update of the digital twin state variable associated with the corresponding disease event, restores the event to the candidate attribution set before the state update, and re-extracts the residual baseline of the source in the stable component region. It then re-executes the decomposition of spatial translation components, scale deviation components, and time drift components. The state change of the disease entity triggered by the event alone is not fixed until the source deviation is corrected.

[0118] When the attribution change has both cross-source consistent residuals and continuous component paths, the processor retains the digital twin state update, solidifies mutually supporting disease events and their associated edges to the corresponding evidence chain, and marks the associated path as a confirmed path. The solidification operation saves the data source, component adjacent path, disease morphology direction and observation time sequence involved in the confirmation. Subsequent new events still need to satisfy the corresponding component reachability relationship and evidence consistency relationship before they can be connected to the confirmed path.

[0119] The results of freezing, re-decomposition, state preservation, and path solidification are all written into a versioned evidence chain. The versioned evidence chain can trace back from the current disease entity to the original observation data, semantic coordinate mapping, candidate event association graph, residual tensor, source deviation marker, disease evolution marker, mapping parameter correction record, and state update record, so that every change in disease attribution and component state change in the digital twin detection and early warning process has a corresponding data basis and processing record.

Claims

1. A digital twin detection and early warning system for multi-source diseases in ancient pagodas, characterized in that, Includes a processor and a memory, wherein the memory stores program instructions, and the processor executes the program instructions to: establish twin semantic coordinates based on the pagoda's hierarchy, facade orientation, component identification, and adjacency relationships; Multi-source disease observation data are converted into disease events with spatiotemporal envelopes, disease categories, source identifiers, and confidence intervals; Based on topological continuity, morphological direction of the disease, temporal correlation, and source credibility, the disease events are fused to form disease entities and evidence chains; The residuals between the predicted and observed states of the twin are decomposed according to the data source, component location and time window, and the source mapping parameters are corrected or the digital twin states are updated based on single-source concentrated residuals or multi-source continuous residuals, respectively. The association weights of subsequent disease events are adjusted based on the updated digital twin status, and early warning information is generated based on the corrected disease entity status.

2. The digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 1, characterized in that, The twin semantic coordinates are indexed hierarchically according to the tower body, level, facade, components and local areas of components, and the geometric boundaries, parent-child containment relationships and adjacency directions are recorded for each level of index; The disease event is generated by projecting the image disease area, the three-dimensional measurement abnormal area, the deformation monitoring sequence and historical repair information onto the local area of ​​the corresponding component. Each disease event includes the observation time interval, spatial coverage, disease characterization vector, source error description and semantic index path. The processor retains the bidirectional mapping relationship between the original coordinates and semantic coordinates for the cross-source projection results.

3. The digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 1, characterized in that, The processor constructs a candidate event association graph based on the adjacency relationship of components. The nodes in the candidate event association graph correspond to disease events, and the edges correspond to cross-source event combinations that satisfy the reachability relationship of components. The spatial envelope overlap, morphological main direction consistency, temporal evolution continuity and source complementarity are jointly calculated for each edge, and the associated path is determined according to the constraints of continuity within the same component, directional propagation between adjacent components and prohibition of direct merging between non-adjacent components. Disease events aggregated along the associated path are written into the same disease entity and form a traceable chain of evidence according to the data source and observation time sequence.

4. The digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 1, characterized in that, The processor maps the twin prediction state and the disease observation state to the semantic index of the same component to generate residual samples, and then forms a residual tensor according to the data source dimension, component topology dimension and sliding time window. Source cohesion, cross-source consistency, and spatial continuity along the component adjacency path are extracted from the residual tensor. When the residual continuously aggregates in a single source dimension and lacks cross-source correspondence, a source deviation label is generated. When the residual appears synchronously in multiple source dimensions and is continuously distributed along the adjacency path, a disease evolution label is generated. The source deviation label and the disease evolution label are respectively associated with the source mapping parameter set and the digital twin state variable set.

5. The digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 2, characterized in that, When establishing a bidirectional mapping relationship, the processor represents the spatial coverage of each data source as a local envelope with uncertain boundaries, and clips the cross-level projection based on the geometric boundaries of the components and the parent-child inclusion relationship. For observation results that simultaneously cover multiple components, they are divided into multiple sub-events according to the overlap ratio between the local envelope and the component boundary, and the original event identifier is written into each sub-event. For observations from the same source obtained at different times, the semantic coordinates are corrected using the mapping offset of the stable component region. At the same time, the correction amount is written into the source error description for use in the disease event association and residual decomposition.

6. The digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 3, characterized in that, The processor performs topologically constrained path expansion on the candidate event association graph. When multiple candidate paths exist between a new disease event and an existing disease entity, it compares the evidence source coverage, morphological extension continuity, and temporal sequence consistency of each candidate path, and retains the path that satisfies the reachability direction of the component and has no evidence conflict. When mutually separated spatial envelopes and discontinuous evolutionary sequences appear within the same diseased entity, the corresponding evidence chain is split into multiple diseased entities; When the associated paths of multiple diseased entities converge within the same component region and the source evidence is complementary, the diseased entities are merged and the entity identifiers before merging are retained.

7. The digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 4, characterized in that, The processor extracts the residual baseline of the same data source within the stable component region for the source deviation mark, and decomposes the residual baseline into spatial translation component, scale deviation component and time drift component, and updates the coordinate mapping and time alignment parameters of the corresponding source according to each component. For the disease evolution marker, the component semantic index, disease representation vector and observation time series corresponding to the continuous residuals are written into the digital twin state variable, and the state update range is restricted to not cross the component region that lacks adjacency relationship or lacks evidence chain connection. The mapping parameters and state variables before and after the update are all saved as versioned records.

8. The digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 4, characterized in that, The processor generates a correlation feedback quantity based on the source deviation marker and the disease evolution marker. The correlation feedback quantity includes a source credibility correction quantity, a component path continuity correction quantity, and a disease morphology consistency correction quantity. When subsequent disease events enter the candidate event association graph, the source credibility correction amount is applied to the candidate edge of the corresponding source, the component path continuity correction amount is applied to the corresponding adjacent path, and the disease morphology consistency correction amount is applied to the matching item of the corresponding disease category. When the versioned digital twin state reverts, the associated feedback quantity corresponding to the reverted version is cancelled, and the historical weights of the corresponding candidate edges are restored.

9. The digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 4, characterized in that, After the diseased entity is split, merged, or the source mapping parameters are updated, the processor performs local reassociation on the affected evidence chain; The local reassociation starts with the semantic index of the changed component, searches along the adjacent path of the component for associated edges that share the original event identifier, the observation period, or the disease characterization, and only recalculates candidate paths that have a topological reachability relationship with the changed region. For disease events whose attribution changes before and after reassociation, record the original disease entity identifier, the target disease entity identifier, and the residual marker that triggered the change, and bind the record to the corresponding digital twin state version.

10. A digital twin detection and early warning system for multi-source diseases of ancient pagodas according to claim 8, characterized in that, The processor stores the correspondence between the associated feedback quantity, the attribution of the disease entity, and the residual label in multiple consecutive update cycles, and performs a consistency check on the attribution changes of the same disease event in adjacent update cycles. When the attribution change is triggered by a single source confidence correction, freeze the digital twin state variable update and re-execute the source bias decomposition; When the attribution change has both cross-source consistent residuals and continuous component paths, the digital twin state update is retained and the corresponding associated path is solidified. The results of freezing, re-decomposing, retaining, and solidifying are written into the versioned evidence chain.

Images