BIM twin dynamic update method

By combining multimodal perception data with BIM constraint diagrams, component field change events are generated and verified, solving the problems of manual updates and insufficient evidence in BIM models, and achieving the accuracy and traceability of dynamic updates.

CN123069503APending Publication Date: 2026-11-10CHINA RAILWAY CONSTRUCTION ENGINEERING GROUP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611243285.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

The current BIM model updates rely on manual processing and lack a continuous data conversion chain, which leads to incorrect component association and erroneous writing when there is insufficient evidence, and there is a lack of closed-loop processing mechanism.

Method used

Multimodal sensing data is encapsulated as observation events, BIM is parsed into constraint diagrams, the corresponding results of components and candidate values ​​of fields are jointly solved, field change events are generated, and evidence sufficiency gating is implemented. The main model is updated through shadow subgraph verification and local incremental patching.

Benefits of technology

It enables dynamic updating of BIM models while maintaining model consistency and traceability of evidence, reducing erroneous writes and the involvement of irrelevant components, and improving the accuracy and recoverability of updates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069503A_ABST
    Figure CN123069503A_ABST
Patent Text Reader

Abstract

This invention discloses a dynamic update method for BIM twins, belonging to the field of building information model data processing. To solve the problem that traditional BIM relies on manual updates and is difficult to continuously evolve with the construction site, this invention encapsulates multimodal perception data into observation events, jointly obtains candidate values ​​of component fields based on BIM constraint diagrams, forms field change events and implements evidence gating, generates targeted supplementary collection tasks for the fields to be verified and associated with the original events, and atomically writes only the local patches, modification history and evidence index that meet the evidence conditions and pass the shadow subgraph verification into the main version, thereby realizing continuous, local and traceable updates of the BIM model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of building information modeling data processing, and more particularly to a method for dynamic updating of BIM twins. Background Technology

[0002] During the construction process, the Building Information Model (BIM) developed in the design phase typically serves as the basis for schedule management, quality inspection, and information delivery. As the construction entity is installed, adjusted, and its status changes, the component positions, dimensions, construction status, and attributes in the model need to correspond to the actual site conditions. Current practices largely rely on personnel periodically reviewing and then centrally modifying the model. There is a lack of a continuously executable conversion chain between perceived data and model fields, thus the update cycle is constrained by manual processing.

[0003] Point cloud data, construction images, positioning measurements, and component identification data can reflect changes on-site, but different sensor sources have differences in coordinates, time, coverage, and measurement quality. When there is occlusion, coexistence of similar components, or discontinuous observations, directly writing the model based on a single matching result can easily lead to incorrect component associations. At the same time, location, size, construction status, and typified attributes are constrained by connection, support, assembly, construction sequence, and field value rules, and modifying a single field in isolation may disrupt the consistency of the local model.

[0004] The existing update process lacks a closed-loop handling mechanism for situations where field evidence is insufficient. Repeatedly scanning the entire construction site increases the amount of perceived data and processing burden; writing data to the main model when evidence is insufficient may lead to erroneous results in subsequent versions. Model changes also need to synchronously record the modification history and original evidence to support subsequent tracing, isolation, and recovery.

[0005] Therefore, a BIM twin dynamic update method is needed to address the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a dynamic update method for BIM twins. Addressing the problems of existing technologies relying on manual updates, lack of continuous transformation chains between on-site observations and component fields, and susceptibility to erroneous writes when evidence is insufficient, this invention proposes a technical solution that encapsulates multimodal sensing data into observation events, parses the initial BIM into constraint diagrams, jointly solves the corresponding results of components and candidate values ​​of fields, generates field change events, and implements evidence sufficiency gating. For fields to be verified, a targeted supplementary collection task associated with the original change event is generated based on the evidence gap, and the original event is backfilled by new observations before resolving. For fields that can be committed, verification is performed in the shadow subgraph corresponding to the affected dependency closure, and local incremental patches, modification history, and evidence indexes are written in the same transaction, thereby enabling the main model to continuously evolve while maintaining constraint consistency and evidence traceability.

[0007] This invention provides a BIM twin dynamic update method, comprising: S1, acquiring multimodal sensing data from the construction site and generating observation events carrying spatiotemporal references, sensor sources, and measurement quality; S2, parsing the initial BIM and generating a BIM constraint graph with component identifiers and component fields as node data and component engineering relationships as edges; S3, establishing candidate relationships between observations and components based on the observation events and the BIM constraint graph, jointly solving the observation matching residuals and constraint violation quantities to obtain the corresponding results for components and candidate values ​​for component fields; S4, converting the differences between the candidate values ​​of the component fields and the current field values ​​into component field change events that are associated with observation evidence, and implementing field evidence sufficiency gating on the component field change events. S5. Based on the evidence gap of the field to be verified, generate a supplementary observation task for the original component field change event, backfill the new observation obtained from the supplementary observation to the original component field change event, re-execute the joint solution and field evidence sufficiency gating, and obtain the updated gating result; S6. Based on the initial gating result, the updated gating result, and the BIM constraint diagram, extract the affected dependency closure of the submittable field, write the field candidate value in the shadow subgraph and perform constraint verification. When the verification passes, write the local incremental patch, modification history and evidence index to the BIM main version in the same transaction. When the verification fails or the field belongs to the field to be verified or the conflict field, keep the corresponding field value unchanged.

[0008] Optionally, S1 includes: acquiring at least two types of sensing data from point cloud, construction images, positioning measurements, component identification, construction equipment operating conditions, and environmental or quality sensing sequences;

[0009] Unify all sensing data into the on-site coordinate system and time axis;

[0010] The observed events are encapsulated according to the data acquisition time, on-site coordinates, sensor source identifier, measurement covariance or sensor quality marker, and raw data index.

[0011] Optionally, S2 includes: creating component nodes for each component in the initial BIM, which include a globally unique identifier for the component, pose parameters, dimension parameters, finite construction status, typified attributes, and version number;

[0012] The edges are established based on the connection relationships, support relationships, inclusion relationships, assembly relationships, and construction sequence relationships in the initial BIM;

[0013] Establish spatial collision association edges based on the intersection relationship of the component's geometric bounding body in the on-site coordinate system;

[0014] Establish attribute derivation relationships based on source field references in the typed attribute calculation rules;

[0015] The BIM constraint diagram is obtained by associating the field data type, allowed value set, unit, allowed range, and constraint level flag indicating whether deviation from the corresponding constraint is allowed to the corresponding component node.

[0016] Optionally, S3 includes: establishing a deterministic candidate relationship between the observation event carrying the component identifier and the component node with the same component identifier;

[0017] For observation events for which no deterministic candidate relationship has been established, a set of candidate components is generated based on spatial neighborhood, geometry, component type, adjacency relationship, and construction time window;

[0018] Within the candidate component set, the observation matching residuals are weighted according to the measurement quality of each sensing mode, and then jointly solved by combining the topological residuals of connection or support relationships, the construction sequence violation amount of finite construction states, and the value violation amount of typified attributes. The corresponding component results are output, along with candidate values ​​for component fields such as pose parameters, size parameters, finite construction states, and typified attributes.

[0019] Optionally, S4 includes: setting field update thresholds for pose parameters, size parameters, finite construction states, and typified attributes respectively, wherein the field update thresholds are determined by at least one of the modeling tolerance of the corresponding field in the initial BIM, sensor calibration error, and project acceptance allowable deviation;

[0020] When the difference between a candidate value and the current value of a component field exceeds the corresponding field update threshold, a component field change event is generated, which includes the component's globally unique identifier, field path, previous value, candidate value, evidence quality information, timestamp, original data index, and association constraints.

[0021] Furthermore, the field evidence sufficiency gating includes: for each component field change event, obtaining the measurement interval width of the candidate value, the effective observation coverage, the occlusion rate, the cross-modal candidate value discreteness, the continuous observation holding time, and the constraint residual;

[0022] The acceptance boundaries for the aforementioned items are determined based on the sensor calibration samples, the visible surface range of the components, the corresponding field update thresholds, the construction plan time window, and the BIM constraint diagram.

[0023] Define field value constraints whose constraint level flag indicates that deviation is not allowed as field value constraints that cannot be relaxed.

[0024] When all items meet the corresponding acceptance boundaries, there are no mutually exclusive candidates, and the non-relaxable field value constraints are not violated, the corresponding field is marked as a submittable field.

[0025] When mutually exclusive candidates exist and their measurement intervals do not overlap, or when the non-relaxable field value constraint is violated, the corresponding field will be marked as a conflicting field.

[0026] The remaining fields are marked as fields to be verified.

[0027] Optionally, S5 includes: for fields to be verified that have mutually exclusive candidates, extracting features that are different from the predicted observations and can be observed by available sensors from the field differences of the mutually exclusive candidates and the adjacent nodes of the BIM constraint diagram to form a supplementary feature set; calculating the discriminative power of each feature in the supplementary feature set against the mutually exclusive candidates; determining the feature with the highest discriminative power as the supplementary observation feature; and when the supplementary feature set is empty, relabeling the corresponding field as a conflicting field and stopping the generation of supplementary observation tasks.

[0028] For fields to be verified that do not have mutually exclusive candidates, the absolute value of the difference between each observation feature that does not meet the acceptance boundary and the corresponding acceptance boundary is calculated. The absolute value of the difference is divided by the non-zero normalized scale of the corresponding observation feature, and the observation feature with the largest ratio is determined as the observation feature to be supplemented. The non-zero normalized scale is the difference between the upper and lower bound values ​​of the corresponding observation feature in the sensor calibration sample. When the difference is zero, the data acquisition system takes the record increment that is pre-configured for the observation feature and is greater than zero.

[0029] For fields to be verified that have been identified as having additional observation features, the additional observation features are back-projected onto the field coordinate system to determine the three-dimensional sensing area. Based on the observation object type, spatial resolution, field of view, reachable pose, and available measurement time period of the available sensors, a supplementary acquisition task is generated, which includes the globally unique identifier of the component, the field path, the three-dimensional sensing area, the required sensing mode, the spatial resolution requirement, the target viewpoint or measurement time window, and the original component field change event identifier.

[0030] Furthermore, the new observations obtained from the supplementary sampling task are added to the original data index according to the original component field change event identifier, the previous value and event identifier of the component field change event are retained, and the corresponding results of the component, the candidate values ​​of the component field, and the items required for the field evidence sufficiency gating are updated using the new observations;

[0031] When the updated corresponding field is marked as a committable field, it is sent to the affected dependency closure. When it is still marked as a field to be verified, the current field value is maintained and the supplementary collection task is updated. When it is marked as a conflict field, the automatic writing of the field is stopped and the conflict record is saved.

[0032] Optionally, S6 includes: for submittable fields of pose or size parameters, extending the affected dependency closure along the edges of connectivity, support, and spatial collision.

[0033] For committable fields in a finite construction state, extend the affected dependency closures along the assembly relationship and construction sequence relationship;

[0034] For committable fields of typed properties, extend the affected dependency closures along containment and property derivation relationships;

[0035] Perform geometric collision and connectivity continuity checks, construction sequence checks, data type and value range checks, and information delivery requirement checks on the shadow subgraph, and output the check results and recovery patches that are opposite to the field writing direction.

[0036] Furthermore, when the verification results show that all verifications have passed, the difference between the shadow subgraph and the BIM master version is converted into a local incremental patch that only contains the components and relationships within the affected dependency closure. The local incremental patch, component version number, modification subject, and modification time constitute the modification history and the evidence index of component field change events to the original data are written in a database transaction.

[0037] If any validation fails or the database transaction commit fails, the database transaction is rolled back or the recovery patch is applied, and the component field change event that failed the validation is written as an exception record to the conflict isolation zone.

[0038] The beneficial effects of this invention are:

[0039] 1. By unifying multimodal sensing data into observation events with spatiotemporal reference, source, and measurement quality, and combining them with BIM constraint diagrams to jointly solve component correspondence and field candidate values, different field observations can participate in matching and constraint determination in the same component field processing chain.

[0040] 2. By implementing evidence sufficiency gating on component field change events and generating supplementary collection tasks related to the original event for evidence gaps in the fields to be proven, it is possible to prevent fields under occlusion or candidate ambiguity from being directly written, and to concentrate supplementary perception on the areas and time periods that affect the determination of specific fields.

[0041] 3. By performing geometry, construction sequence, field type, and information delivery requirement validations on the shadow subgraphs of affected dependency closures, and atomically committing local patches, modification history, and evidence indexes, the involvement of irrelevant components in updates can be reduced, while maintaining consistency, recoverability, and change traceability of the main version. Attached Figure Description

[0042] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0043] Figure 1This is a flowchart of the BIM twin dynamic update method of the present invention.

[0044] Figure 2 This is a flowchart of the targeted supplementary sampling and backfilling mechanism for fields to be verified. Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0046] refer to Figures 1-2 The BIM twin dynamic update method includes: S1, acquiring multimodal sensing data from the construction site and generating observation events carrying spatiotemporal references, sensor sources, and measurement quality; S2, parsing the initial BIM and generating a BIM constraint diagram with component identifiers and component fields as node data and component engineering relationships as edges; S3, establishing candidate relationships between observations and components based on the observation events and the BIM constraint diagram, jointly solving the observation matching residuals and constraint violation quantities to obtain the corresponding results for components and candidate values ​​for component fields; S4, converting the differences between the candidate values ​​of component fields and the current field values ​​into component field change events associated with observation evidence, implementing field evidence sufficiency gating on the component field change events, and obtaining a package... The initial gating results include submittable fields, fields to be verified, and conflicting fields; S5, generate supplementary observation tasks for related original component field change events based on the evidence gaps of fields to be verified, backfill the new observations obtained from the supplementary observations to the original component field change events, re-execute joint solution and field evidence sufficiency gating, and obtain updated gating results; S6, extract the affected dependency closures of submittable fields based on the initial gating results, updated gating results, and BIM constraint diagram, write field candidate values ​​in the shadow subgraph and perform constraint verification. If the verification passes, write local incremental patches, modification history, and evidence indexes to the BIM main version in the same transaction. If the verification fails or the field belongs to a field to be verified or a conflicting field, keep the corresponding field value unchanged.

[0047] In this specific embodiment, S1 includes:

[0048] In this specific embodiment, the on-site data processing system is deployed on the edge server of the prefabricated building construction area. The construction area adopts an on-site coordinate system with the intersection of the first-floor axis grid as the origin and millimeters as the length unit, and uses Coordinated Universal Time millisecond timestamps as the unified time axis. In each two-second acquisition cycle, the system receives point cloud frames output by the 3D laser scanner, construction images output by the fixed camera, prism positioning measurements output by the total station, and QR code component labels attached to the surface of the prefabricated components. At the same time, it receives the working condition records of the tower crane hook height and slewing angle, so that each acquisition cycle contains at least two kinds of sensing data: point cloud and construction images.

[0049] Point cloud frames use ordered point records with intensity fields. Each point record includes the sensor's three-dimensional coordinates, reflection intensity, echo number, and equipment clock. Construction image records include image file index, camera number, exposure time, pixel size, and lens calibration version. Total station records include station number, prism number, angle measurement value, slope distance, and measurement standard deviation. Component identification records include decoded string, pixel four-corner coordinates, decoded confidence level, and associated image index. Equipment condition records include equipment number, sampling time, hook height, rotation angle, and validity mark. The system generates the original record key by combining the sensor source identifier and the equipment's serial number and saves it to the append-only original data warehouse.

[0050] Before deployment, the system uses field control points and calibration boards to determine the rigid transformation of each sensor to the field coordinate system, and then... Transform point cloud points, prism points, and spatial points obtained from image backprojection, where, This represents a four-dimensional homogeneous coordinate column vector in the field coordinate system, where the first three components are in millimeters and the fourth component is fixed as a dimensionless number one. Indicate source To the on-site coordinate system The 4x4 homogeneous transformation matrix has a dimensionless top-left 3x3 rotation component, translation columns in millimeters for the first three rows, and a dimensionless row of zero, zero, zero, and one for the last row. Indicate source A four-dimensional homogeneous coordinate column vector in its own coordinate system, with the first three components measured in millimeters and the fourth component fixed as a dimensionless number one. This indicates the sensor source index obtained from the sensor source table of the original data warehouse in this round. As an index, it is not applicable to engineering units. This indicates the site coordinate system identifier obtained from the project coordinate system registration form. As an identifier, it is not applicable to engineering units. The transformation matrix is ​​written into the calibration parameter library by the least squares registration results of no less than six non-coplanar control points and carries the effective version.

[0051] Device clocks and unified timelines are aligned according to To perform the conversion, among which, This represents the acquisition time on a unified timeline, in milliseconds. This indicates a unified reference timeline identifier obtained from the system timeline configuration table. As an identifier, it is not applicable to engineering units. Indicate source The dimensionless clock scale coefficient must be within the positive range of 0.999 to 1.001 given by the device clock drift specification. Indicate source Recorded device time, in milliseconds. Indicate source Offset relative to a uniform time axis, in milliseconds. This indicates the sensor source index obtained from the sensor source table of the original data warehouse in this round. As an index, it is not applicable to engineering units. The scale coefficient and offset are obtained by continuously collecting 60 seconds of synchronous pulses at the start of each shift and performing linear fitting. If the value is not within the positive range or the 95th percentile of the synchronization residual exceeds 20 milliseconds, the corresponding source will be marked as having insufficient time quality and will not participate in cross-modal same-period fusion.

[0052] The system removes outliers from the point cloud with a radius of 30 mm and a lower limit of four neighboring points. It performs camera distortion correction on construction images and limits the reprojection error of the four corners of component markers to three pixels. It retains valid points in total station records according to the condition that the measurement standard deviation is no more than three mm. It performs physical range and continuity verification on equipment condition records. Each cleaned record still retains the original record key, the cleansing rule version, and the reason for removal, so that any subsequent observation result can be traced back to the original data that has not been rewritten.

[0053] For point clouds, images, and positioning records within the same unified time window, the system uses the intersection of the bounding boxes of the field coordinates and the acquisition time difference not exceeding one hundred milliseconds as the cross-modal grouping condition. When there is a grouping conflict, records are first selected in ascending order of time difference, then in ascending order of calibration residual. If they are still the same, the unique grouping result is determined by the dictionary order of the original record key. Equipment condition records that lack spatial coordinates are mapped to the corresponding field area according to the equipment number and the work area in the construction plan. The area mapping record uses the equipment number and the planned task identifier as the joint key. The value fields include the field grid code, the effective time period, and the plan version. When the joint key is not matched, the record is retained and marked as spatial pending without forging coordinates.

[0054] The system saves data mode versions and content summaries for each source in the raw data warehouse. When the point cloud frame missing ratio exceeds 20%, the image exposure time exceeds the calibration range, the total station prism number is inconsistent with the component identification table, or the equipment operating condition sampling interval exceeds four seconds, the system does not discard the original records. Instead, it writes the anomaly type, discovery time, and resampling suggestion in the grouped records. Only fields that pass the spatial, temporal, and source integrity checks are included in the observation event payload. Anomaly fields still participate in subsequent tracing through the raw data index.

[0055] The system calculates measurement quality information for each grouping result. Point cloud records are written with triaxial measurement covariance and effective point ratio. Image records are written with reprojection error, sharpness and decoding confidence. Total station records are written with measurement standard deviation. Equipment condition records are written with continuous sampling integrity rate. Covariance is obtained from the sample covariance of thirty consecutive measurements of the same static control point and is measured in square millimeters. Quality fields all carry calibration batch, calculation window and effective range. Fields that exceed the effective range are not deleted but are written with quality invalidation flag.

[0056] After completing the alignment and quality calculation, the system encapsulates each grouping result as an observation event. The event record sequentially includes the globally unique identifier of the event, the unified acquisition time, the field coordinates or field area, the sensor source identifier set, the measurement covariance or sensor quality mark, the raw data index set, the calibration version, the cleaning rule version, and the processing status. The globally unique identifier of the event is generated by the project number, the unified acquisition time, and the group sequence number. The raw data index stores the raw record key and its content summary, thereby preventing the same data from being entered into the database repeatedly.

[0057] Observed events are written into the observation event table and a composite index established by field grid coding and time window. Events containing component identifiers and with a decoding confidence of not less than 0.95 are also written into the component identifier inverted index. Events that do not meet the time quality or spatial quality conditions are retained in the observation event table and marked as pending calibration. The observation event set, calibration version, and original data index output by S1 serve as the unified input for establishing candidate relationships in S3 and for subsequent evidence tracing.

[0058] In this specific embodiment, S2 includes:

[0059] The map maintenance module reads the IFC components, attribute sets, spatial hierarchy and engineering relationships of the initial BIM, and uses the IFC global identifier as the globally unique identifier of the component to create a component node record for each component. The node fields include the globally unique identifier of the component, component category, pose parameters in the site coordinate system, length, width and height dimension parameters, five limited construction states: not installed, hoisting, in place, fixed and accepted, typed attribute key value set, current version number and geometric bounding volume index.

[0060] Pose parameters are stored using translation vectors and unit quaternions, with translation amounts in millimeters and dimensional parameters in millimeters. Limited construction states are composed of status codes and the time when the status takes effect. Typed attributes are composed of field paths, data types, units, and values. When importing nodes, a unique identifier is checked for global uniqueness. Unit fields are uniformly converted to the project unit dictionary. If duplicate identifiers or units that cannot be converted are found, the node is written to the graph construction isolation table and is not included in the set of updatable nodes.

[0061] The module establishes typified edges according to the connection, support, containment, assembly, and construction sequence relationships in the initial BIM. Each edge records the starting component identifier, ending component identifier, relationship type, direction, source IFC relationship identifier, relationship version, and validity status. Connection edges and support edges also record the design contact surface and allowable connection gaps. Assembly edges and construction sequence edges record the previous and subsequent states. Containment edges record the parent space or composite component identifier. Repeated relationships are merged according to the starting point, ending point, and relationship type, and all source identifiers are saved.

[0062] For spatial collision-related edges, the module first establishes an R-tree index by aligning the bounding box with the axis after expanding by 20 mm. Only when two components intersect in the projection intervals of the three coordinate axes of the field coordinate system will they enter the precise intersection test. The precise test uses a triangular mesh bounding body hierarchical structure to calculate the minimum surface distance. When the minimum surface distance is not greater than the sum of the modeling tolerances of the corresponding components, a spatial collision-related edge is established. The edge records the minimum distance, candidate intersection area, calculation mesh version, and allowed intersection flag. Components that belong to the design contact or reserved hole relationship are distinguished from the prohibited collision relationship by the allowed intersection flag.

[0063] Axis-aligned bounding boxes on coordinate axes The amount of overlap on according to Calculate, where, Representing components With components On the coordinate axes The amount of overlap, in millimeters. Representing components On the coordinate axes The upper bound of the interval above, Representing components On the coordinate axes The upper bound of the interval above, Representing components On the coordinate axes The lower bound of the interval above, Representing components On the coordinate axes The lower bound of the interval is given above, and all four boundary quantities are in millimeters. and This indicates that the component indices obtained from the current BIM constraint diagram's active component index set are all distinct and are not applicable to engineering units as indexes. This indicates the coordinate axis index obtained from the three-axis index set of the field coordinate system (horizontal, vertical, and vertical axes). As an index, it is not applicable to engineering units. The exact intersection test is only performed when all three overlap values ​​are not less than zero.

[0064] For attribute derivation relationships, the module parses the abstract syntax tree of the typed attribute calculation rules, parses each source field path into source component identifier and source attribute key, and parses the target field path into target component identifier and target attribute key. Directed attribute derivation edges are established only when the data types of the source field and the target field satisfy the calculation operator signature. The edge records the source field path, target field path, calculation rule summary, unit conversion factor and rule version. When the source field cannot be parsed or a circular derivation occurs, the corresponding rule is written into the isolation table and prevented from entering the automatic update chain.

[0065] The module establishes a field constraint record for each updatable field. The record includes the component's globally unique identifier, field path, field data type, allowed value set, engineering unit, lower and upper limits of the allowed range, field storage precision, initial BIM modeling tolerance, allowed deviation for project acceptance, project acceptance rule version, constraint level flag, constraint source and effective version. The component's globally unique identifier and field path form a composite key consistent with the component node table, and the project acceptance rule version forms a foreign key pointing to the project acceptance rule table. The constraint level flag has three levels: non-restrictible, project adjustable, and prompt. The non-restrictible level is used for data types, legal codes, and explicitly prohibited construction sequences. The boundary of the project adjustable level is given by the project acceptance rule version. The prompt level only generates review information and does not block submission separately.

[0066] In this specific implementation, one dimension field record in the field constraint library uses the component identifier and field path as a joint key, the data type is decimal specified point number, the unit is millimeters, the allowable range is 10 millimeters above and below the design value, the constraint level is project adjustable, and the source is project acceptance rule V3.2. Another construction status record uses an enumeration data type, allows a set of five finite construction states, and the constraint level is non-relaxable. The module uses this joint key to obtain a unique constraint in subsequent gating and shadow verification.

[0067] After the graph is constructed, the module performs reference integrity checks on the node table and various edge tables, performs directed cycle checks on construction sequence edges, performs endpoint existence checks on connection edges and support edges, and performs field existence and unit conversion checks on attribute derived edges. Edges that fail the checks are moved to the graph construction isolation table and the reasons for failure are retained. Nodes and edges that pass the checks are written to the same BIM constraint graph version. The graph version is generated by the initial BIM version, relation summary, and constraint library version.

[0068] To ensure the reproducibility of the graph version, the module sorts node records by component identifier and field path, and edge records by relation type, start identifier, and end identifier. Then, it calculates the content summary for the normalized sequence. The graph version record saves the initial BIM file summary, node summary, edge summary, field constraint summary, construction program version, and construction time. If the graph version held is found to be inconsistent with the record summary at any subsequent stage, the automatic update of this round will be terminated and the generated observation events will be retained.

[0069] The module combines the component node table, relation edge table, field constraint table, spatial index, and attribute derivation index into a BIM constraint diagram, and establishes query indexes for component identifier, component category, site grid code, field path, and relation type. The BIM constraint diagram version output by S2 remains read-only until the current update transaction is completed, and is used by S3 to generate candidate relations, S4 to perform field evidence gating, S5 to retrieve adjacency features, and S6 to extend affected dependency closures.

[0070] In this specific embodiment, S3 includes:

[0071] The candidate solution module reads the set of observation events output by S1 and the BIM constraint diagram output by S2. For observation events carrying component identifiers and with a decoding confidence of not less than 0.95, it uses the component identifier inverted index to locate component nodes with the same globally unique identifier and establishes a deterministic candidate relationship. If the identifier does not exist in the diagram, has been isolated, or hits multiple active versions at the same time, the deterministic state is revoked and the non-deterministic candidate generation process is entered.

[0072] For events where no deterministic candidate relationship has been established, the module extends the bounding box of the event site coordinates by 500 mm outward to query the spatial index of the components. It retains nodes where the geometric bounding volumes intersect or the minimum distance is no more than 500 mm. Then, it filters layer by layer based on the following conditions: the angle between the main direction of the point cloud and the main direction of the component design is no more than 15 degrees; the image category is consistent with the component category; the adjacent components are known to be reachable in one jump in the BIM constraint diagram; and the observation time is from two hours before the start of the construction plan to four hours after the end of the construction plan. Only when the spatial conditions are met and two of the other four conditions are met can the component be included in the candidate component set.

[0073] Candidate records include observation event identifier, candidate component identifier, candidate source, geometric residual, adjacency consistency count, planning window, quality fields of each modality, and retention reason. When the candidate set is empty, the module writes the event to the unassociated event queue and waits for subsequent observations. It does not use the nearest node to force matching. When the number of candidates exceeds twenty, it is first pruned to twenty in ascending order of geometric residual, descending order of adjacency consistency count, and lexicographical order of component identifier to limit the joint solution scale and maintain determinism.

[0074] The module for each observed event With candidate components Calculate the dimensionless joint cost ,in, Represents observed events Related candidate components The joint cost and its value is not less than zero. , , , and The weights represent the geometric, topological, construction sequence, sensing quality, and typified attribute residuals, respectively. Each weight is a dimensionless non-negative number in the range of zero to one, and their sum is one. , , , and The following are represented, in order: geometric residuals normalized to component dimensional references; the ratio of the number of adjacent inconsistent edges to the number of edges to be checked; the ratio of the number of violations of prior state relations to the number of construction sequence relations; the ratio of the measurement covariance trace to the upper bound of the calibration covariance; and the ratio of the number of out-of-bounds candidate attributes to the total number of candidate attributes. Each residual is a dimensionless number trimmed to the range of zero to one. Category subscripts are used. , , , and These represent geometric, topological, ordinal, sensing, and attribute categories, respectively. Their values ​​are derived from a fixed set of categories in the candidate solution parameter table and are used as category labels, not in engineering units. This represents the index of the observation event obtained from the current set of observation events. This indicates the index of the candidate component obtained from the set of candidate components for the corresponding event. Neither of these indices is applicable to engineering units.

[0075] , , , and The value of is obtained by minimizing the error correlation rate of independent calibration events with component identification. In this specific implementation, the values ​​are 0.35, 0.2, 0.15, 0.15, and 0.15 in sequence, and are written into the candidate solution parameter table V2.1 along with the calibration event summary; Calculation When the number of edges to be checked is zero, calculate When the construction sequence relation coefficient is zero or calculated When the total number of candidate attributes is zero, the corresponding residuals are marked as uncalculable and not assigned a zero value; when any residual is uncalculable because the corresponding count base is equal to zero or there is a lack of available input, the item is deleted and the remaining weights are renormalized according to the sum of the remaining weights; when there are fewer than three remaining valid items, the candidate is marked as insufficient evidence.

[0076] The module prioritizes the candidate with the lowest joint cost and employs mutually exclusive and exhaustive decision branches: if the candidate set contains only one candidate and its cost is no greater than 0.65, it is marked as a single correspondence; if it contains only one candidate but its cost is greater than 0.65, it is marked as insufficient evidence and no field candidate is output; if there are at least two candidates, if the difference between the second smallest cost and the smallest cost is no less than 0.1 and the smallest cost is no greater than 0.65, it is marked as a single correspondence; if the difference is less than 0.1, all candidates within the range of joint cost difference are retained and marked as mutually exclusive candidates; if the difference is no less than 0.1 but the smallest cost is greater than 0.65, it is marked as insufficient evidence and written to the unassociated event queue; if multiple candidates have the same joint cost, the preferred order is given according to deterministic identifier hit, adjacency consistency count, and component identifier lexicographical order, but the mutually exclusive candidate set is still preserved without masking ambiguity;

[0077] For pose parameters, the module performs iterative nearest-point registration from point to plane using the candidate component design surface and event point cloud. The iteration terminates when the change in the root mean square residual between two adjacent iterations does not exceed 0.5 mm or reaches 50 iterations. The module outputs translation vector, unit quaternion, and 6x6 pose covariance. The translation variance and translation covariance of the pose covariance are in square millimeters, the rotation variance and rotation covariance are in square radians, and the translation and rotation intersection terms are in millimeters multiplied by radians. For size parameters, the module projects the registered point cloud along the local axis of the component and forms candidate values ​​for length, width, and height using 2.5% and 97.5% quantile differences. When the effective coverage of the occlusion direction is less than 40%, the candidate size for that direction is not output.

[0078] In each round of iterative nearest point registration, corresponding points are first established according to the design surface normal angle not exceeding 20 degrees and the absolute distance from the point to the surface not exceeding 50 millimeters. Then, the measurement covariance inverse matrix, which is in square millimeters for translation observation components, in square radians for rotation observation components, and in millimeters multiplied by radians for translation-rotation cross components, is used to weight the residuals of the corresponding points and solve for the six-degree-of-freedom increment. When the incremental translation norm exceeds 20 millimeters or the rotation angle exceeds 5 degrees, the round is marked as divergent and the results of the previous round are retained. When divergence occurs for three consecutive rounds, no pose candidate is output and the registration failure reason is written into the corresponding component result.

[0079] For finite construction states, the module forms state evidence based on component identification, equipment status, and point cloud visibility. When the QR code first enters the installation area and the hook height decreases, it is recorded as being hoisted. When the pose is within the design tolerance for 30 consecutive seconds, it is recorded as being in place. When the connection test results meet the design relationship for 60 consecutive seconds, it is recorded as being fixed. When the quality acceptance record is signed and valid, it is recorded as being accepted. State transitions must proceed along the allowed edges of the finite state machine. When a reverse observation occurs, the current state is retained and the reverse observation is written into the abnormal evidence field.

[0080] For categorized attributes, the module obtains candidate values ​​from image recognition labels, quality sensor sequence statistics, and construction record fields. First, it performs data type conversion and unit conversion according to the S2 field constraints. Then, it forms a weighted median or weighted majority result based on the number of sources within the same unified time window, the quality of each source, and the freshness of time. Sources that cannot complete the type conversion are recorded as invalid evidence. When cross-modal candidate values ​​are inconsistent, each source value and its measurement interval are retained for use by S4 gating.

[0081] The module writes the preferred component identifier, mutually exclusive candidate set, joint cost decomposition, pose parameter candidate value, dimension parameter candidate value, finite construction state candidate value, typified attribute candidate value, each candidate measurement interval, evidence source index, and solution parameter version into the corresponding result table of the component. The corresponding result of the component and the field candidate value are passed to S4 with the observation event identifier and the component identifier as the joint key, and S5 is allowed to add new observations according to the original event identifier and recalculate on the same solution record.

[0082] In this specific embodiment, S4 includes:

[0083] The field gating module reads the component-related results and field candidate values ​​output by S3. It retrieves the current field value from the current read-only version of the S2 component node table based on the component's globally unique identifier and field path. It also retrieves the data type, unit, field storage precision, initial BIM modeling tolerance, and constraint level from the S2 field constraint table. Finally, it retrieves the project acceptance allowable deviation from the project acceptance rule table based on the project acceptance rule version saved in the field constraint record. The component identifier, field path, drawing version, effective version, and project acceptance rule version of the three read results must be consistent. If any composite key is not hit or the versions are inconsistent, the field is marked as a field to be verified, and a constraint input missing reason code is written. Pose translation and dimension fields use millimeter differences, attitude fields use quaternion angles, finite construction states use finite state machine transition distances, and categorized attributes are compared according to field data types using numerical differences, set inequalities, or string normalization inequalities.

[0084] Each numeric field Normalized differences according to Calculate, where, Representation field Dimensionless normalized differences This indicates that the field candidate values ​​output by S3 are in the same unit as the current field. This indicates the current field value in the BIM master version, and the unit is consistent with the candidate value. Representation field The update threshold is set and the unit is consistent with the current field. Representation field The positive value record increments and the unit is consistent with the current field. This represents the field path index obtained from the currently valid record in the S2 field constraint table. As an index, it is not applicable to engineering units. The update threshold is the maximum value specified by the project rules among the initial BIM modeling tolerance, the corresponding sensor 95th percentile calibration error, and the project acceptance allowable deviation. The record increment is half of the field storage precision. Field constraint records are only updated when... Not less than If the condition is not met, the field will be marked as invalid for the threshold configuration and automatic generation of change events will be prohibited, so that a normalized difference greater than one is always equivalent to the difference exceeding the update threshold.

[0085] When the normalization difference of a numerical field is greater than one, the candidate state of an enumerated field is different from the current state and the transformation edge is valid, or the set field has member changes or the normalized string is different from the current value, the module generates a component field change event. The event record includes the event's globally unique identifier, the component's globally unique identifier, the field path, the value before the change, the candidate value, the candidate measurement interval, the evidence quality information, the timestamp, the original data index, the association constraint identifier, the mutually exclusive candidate set, the solution parameter version, and the gating status. Candidates that do not exceed the update threshold are only written to the observation history and no field change event is generated.

[0086] For each change event, the module generates repeatable measurement uncertainties and cross-modal discretizations based on the field data type: For numerical, pose, and size fields, the measurement uncertainty is calculated as the upper bound minus the lower bound of the candidate measurement interval, and the cross-modal discretization is calculated as the maximum absolute deviation of the candidate values ​​of each modality relative to the weighted median, with both units consistent with the field; For enumerated state and normalized string fields, the measurement uncertainty is calculated as one minus the ratio of the source quality-weighted votes of the highest candidate category to all valid votes, and the cross-modal discretization is calculated as one minus the consistency rate of each modality with respect to the highest candidate category, both being dimensionless numbers in the range of zero to one; For set fields, the measurement uncertainty is calculated as the quality-weighted average of the pairwise Jaccard distances of each source candidate set, with the largest Jaccard distance among them being the cross-modal discretization, and the Jaccard distance between empty sets is defined as zero, both being dimensionless numbers in the range of zero to one;

[0087] The module also obtains the effective observation coverage rate in the range of zero to one by dividing the observed effective surface area of ​​the component by the BIM visible surface area, and obtains the occlusion rate in the range of zero to one by dividing the projected area obscured by other components or construction equipment by the theoretical visible projected area. The continuous observation duration is obtained by the time difference between the first and last observations of the candidate intervals of the numerical field that are continuously overlapping or the highest candidate category of the discrete field that is continuously consistent. The unit is seconds. The quality weight of each source is the inverse variance normalized value of the S1 quality field within the effective range. Discrete sources without effective variance use the weight normalized by the accuracy of the calibrated samples. When the total weight of the effective source is zero, the corresponding index is marked as uncalculated.

[0088] The constraint residuals are selected by field type from the connection or support topology residuals, construction sequence violation quantities, and typified attribute value violation quantities in S3. The pose and dimension fields use topology residuals, the construction status uses sequence violation quantities, and the typified attributes use value violation quantities. All residuals are in the range of zero to one. The module writes the six gating indicators and their original record keys into the evidence quality information, so that the gating conclusions can be recalculated from the original data, calibration parameters, and constraint diagrams.

[0089] The upper bound for measurement uncertainty of numerical fields is obtained from the 95th percentile of the interval width calibration sample with the same sensor and the same field type. The upper bound for measurement uncertainty of enumerated state, string, and set fields is obtained from the 95th percentile of the corresponding consistency rate or Jaccard distance calibration sample. The lower bound for the confidence level applicable to mutually exclusive candidates of enumerated state and string fields is the 95th percentile of the corresponding category accuracy in the pre-deployment calibration sample with true field value labels. The lower bound for effective coverage and the upper bound for occlusion rate are obtained from the simulation results of the visible surface range of the component BIM and the sensor field of view. The upper bound of the cross-modal discrete quantity of numerical fields is taken from the corresponding field update threshold. The upper bound of the cross-modal discrete quantity of discrete fields is taken from the 95th percentile of the corresponding type calibration error rate. The lower bound of the continuous holding time is obtained from the construction plan time window and sampling period. The upper bound of the constraint residual is obtained from the relationship type and constraint level in the BIM constraint diagram. In this specific implementation, the gate boundary is written into the evidence acceptance boundary table V4.0. The joint key is the field type, component category and sensor combination. When the joint key is not hit, the field is marked as the field to be verified and the boundary missing reason code is written. No arbitrary default boundary is used.

[0090] The evidence acceptance boundary table is generated jointly from pre-deployment calibration data, visibility simulation of similar components, and project acceptance rules. The table fields include composite key, applicable data type, upper bound and unit of measurement uncertainty, lower bound of coverage, upper bound of occlusion rate, upper bound and unit of cross-modal discrete quantity, lower bound of confidence level for enumerated state or string, confidence level calibration sample size, confidence level calibration summary, lower bound of hold duration and unit of seconds, upper bound of constraint residual, general sample size, sensor calibration version, project acceptance rule version, boundary version, and effective time. The boundary unit of numerical fields is consistent with the field. The boundaries of enumerated state, string, and set fields are dimensionless. If any applicable sample size is less than thirty, the corresponding record is not effective and a boundary sample insufficient reason code is written. When the project acceptance rules or sensor calibration version change, the affected records are recalculated and the old version is retained.

[0091] In this specific implementation, the boundary record of a precast wall panel size field uses the size, precast wall panel, point cloud, and total station as the joint key. The upper limit of the measurement interval width is six millimeters, the lower limit of the effective coverage rate is 0.6, the upper limit of the occlusion rate is 0.3, the upper limit of the cross-modal discrete quantity is eight millimeters, the lower limit of the continuous holding time is ten seconds, and the upper limit of the constraint residual is 0.15. The record source includes forty calibration samples and project acceptance rule V3.2. The field gating module only uses this record when the sensor combination and version are both matched.

[0092] The module marks the field data type, allowed value set, range, and construction sequence conditions as non-relaxable field value constraints. If a candidate value violates any of these conditions, it is marked as a conflicting field. For mutually exclusive candidates, a numerical field is marked as a conflicting field when the candidate measurement intervals do not overlap. A field is marked as a conflicting field when the highest candidate categories of enumerated states or string fields are different and their respective weighted confidence scores are not lower than the applicable lower bound of confidence scores. A field is marked as a conflicting field when the maximum Jaccard distance of a set exceeds the applicable upper bound of discrete quantities and there is no common member combination that satisfies the allowed set constraints. Other mutually exclusive cases are reserved as fields to be verified and the sources of each candidate are saved.

[0093] When there are no mutually exclusive candidates, the constraints on the value of non-relaxable fields are not violated, the measurement uncertainty calculated by field type is not greater than the corresponding upper bound, the effective observation coverage is not less than the lower bound, the occlusion rate is not greater than the upper bound, the cross-modal discrete quantity calculated by field type is not greater than the corresponding upper bound, the continuous observation holding time is not less than the lower bound and the constraint residual is not greater than the upper bound, the module marks the field as a field that can be submitted. When any applicable indicator cannot be calculated or is not satisfied but has not reached the conflict condition, it is marked as a field to be verified, and the data type, calculation rule identifier, unsatisfied indicator, corresponding boundary value, reason code and evidence gap are listed in the gating results.

[0094] The module writes the initial gating results into the field gating result table. The submitable fields carry candidate values ​​and all evidence indexes, the fields to be verified carry unmet indicators and mutually exclusive candidate information, and the conflict fields carry the judgment criteria for non-relaxable constraints or non-overlapping intervals. The table records maintain the same event identifier as the original component field change events. The fields to be verified are sent to S5 to generate a supplementary data collection task, and the submitable fields and conflict fields, along with the BIM constraint diagram version, are sent to S6 for processing.

[0095] In this specific embodiment, S5 includes:

[0096] The supplementary sampling scheduling module reads the fields to be verified, evidence gaps, reason codes, mutually exclusive candidate sets, original component field change event identifiers of S4, and adjacent nodes of the BIM constraint diagram of S2. It first classifies the reasons to be verified into three categories: mutually exclusive candidate to be verified, insufficient single candidate evidence, and missing acceptance boundaries. Then, it obtains the observation object type, spatial resolution, field of view, reachable pose, available measurement period, calibration version, and current task status of each sensor from the available sensor catalog.

[0097] For fields to be verified with mutually exclusive candidates, the module compares the geometric surface, component category, adjacency relationship, QR code location, limited construction status and typological attributes of each candidate. Only when the predicted observations between candidates are different, can be measured by currently available sensors and do not depend on inaccessible areas, can be formed as supplementary features. The feature record includes feature identifier, predicted value, prediction interval, applicable sensing mode, observation surface, adjacent component identifier and field accessibility. Duplicate features are merged according to feature identifier and observation surface.

[0098] For numerical features that can be supplemented, the module compares candidate one and candidate two in terms of features. The discrimination on Calculate, where, Representing numerical features The dimensionless discrimination index is given, and a larger value indicates that it is easier to distinguish. Represents a candidate pair of features The predicted value, Indicates candidate two-pair features The predicted values, both using the same engineering units. This represents the half-width of the prediction interval for candidate one. This represents the half-width of the prediction interval for candidate two, and both units are consistent with the predicted values. Representation of features The positive zero protection value is determined by the sensor resolution, and its unit is consistent with the predicted value. This represents the feature index obtained from the feature table that can be supplemented from the current event. As an index, it is not applicable to engineering units. When there are more than two candidates, the minimum discriminant among all candidate pairs is taken as the conservative discriminant of the feature.

[0099] For non-numerical features that can be supplemented, the module uses a dimensionless discriminant in the range of zero to one: enumerated states, component categories, and normalized strings take zero when candidate predicted categories are the same, and take the lowest value of the two candidate prediction confidences when predicted categories are different; adjacency relations and other set features take the Jaccard distance between two candidate prediction sets, and the distance between two empty sets takes zero; candidate prediction categories and prediction sets directly read the corresponding results of S3 components and the candidate attribute, state, and relation records in the S2 constraint graph without calling the learning model, and the prediction confidence is determined by the known component identifier. The correct number of predictions for the same category in the calibration sample is divided by the total number of predictions for that category to obtain and record the calibration sample summary and version. The uncertainty of the set prediction is obtained by the 95th percentile of the Jaccard distance of the calibration samples. The discrimination is then divided by one and the uncertainty is added to complete the conservative reduction. If there is a lack of candidate records, calibration samples, or applicable versions, the feature is marked as uncomputable. If the expected collection time exceeds the mission validity period or the expected collection cost exceeds the project supplementary collection limit, it is excluded. If there are more than two candidates, the minimum value among all candidate pairs is taken, and the values ​​are processed according to the lexicographical order of the feature identifier.

[0100] The module uses achievable independent discrimination thresholds based on feature type: the conservative discrimination threshold for numerical features is 1.2, and the conservative discrimination threshold for non-numerical features in the range of 0 to 1 is 0.65. Both thresholds are selected by pre-deployment supplementary calibration samples with known component identifiers under the constraint that the error association rate does not exceed 5%, and are written into the supplementary sampling discrimination parameter table along with the sample summary, sample size, and version. When the sample size is less than 30, no automatic supplementary sampling task is generated for the corresponding type. The module obtains the dimensionless discrimination margin by dividing the feature's conservative discrimination margin by the corresponding type threshold. Features with a discrimination margin less than 1 are first excluded. Then, unique features to be supplemented are selected in descending order of discrimination margin, ascending order of expected acquisition duration, and lexicographical order of feature identifier. When there are ties, the unique result is determined by the expected acquisition duration and feature identifier in turn. When the supplementable feature set is empty or all are below their respective type thresholds, the corresponding field to be verified is relabeled as a conflict field, the reason for not being able to form a valid supplementary observation is written, and the generation of supplementary sampling tasks is stopped.

[0101] For fields to be verified where there are no mutually exclusive candidates and the acceptance boundary is valid, the module reads the current observed feature values ​​and boundary values ​​that do not meet the acceptance boundary item by item, and then follows the... Calculate the evidence gap, where, Indicates observation characteristics The dimensionless evidence gap, and the larger the value, the more priority should be given to supplementary sampling. Indicates observation characteristics The current value, This represents the acceptance boundary value corresponding to the observed feature; both are expressed in the same unit. Indicates observation characteristics The non-zero normalized scale has the same unit as the current value. This indicates the observation feature index obtained from the current gating result that does not meet the indicator set. As an index, it is not applicable to engineering units. The normalization scale is the difference between the upper and lower bounds of the sensor calibration sample. When the difference is zero, the minimum recording increment that is pre-configured by the data acquisition system and is greater than zero is taken.

[0102] For fields to be verified with the cause code being missing boundaries or insufficient boundary samples, the module does not read non-existent boundary values ​​and does not apply the evidence gap formula. Instead, it generates boundary calibration tasks or project rule configuration tasks: when boundaries are missing, the person responsible for the project rules supplements boundary records with applicable field types, component categories, and sensor combinations; when boundary samples are insufficient, it supplements calibration samples of the same type of components until there are no fewer than thirty. The task record saves the cause code, the missing joint key, the required sample size, the responsibility queue, and the original event identifier; the BIM master version field value remains unchanged until the new boundary record is version-signed and reaches the effective time. After the record becomes effective, the S4 gating is re-executed using the same event identifier. If there are still no valid boundaries, it is transferred to the manual review queue.

[0103] The module determines the observation features to be supplemented by descending order of evidence gaps, ascending order of expected acquisition duration, and dictionary order of observation feature identifiers. For insufficient coverage, it selects viewpoints that can increase the number of unobserved surfaces. For excessive occlusion, it selects viewpoints whose line of sight does not pass through the occlusion enclosure. For excessively wide measurement intervals or excessive cross-modal discrete quantities, it selects sensors with different modes than existing evidence by ascending order of calibration error. For insufficient holding time, it extends the continuous acquisition window of the same viewpoint. For excessive constraint residuals, it selects features that can observe the connection surface of adjacent components or the construction status.

[0104] After identifying the features to be observed, the module uses the camera intrinsic and extrinsic parameters, scanner extrinsic parameters, or total station station parameters saved in S1 to back-project the feature observation surface onto the field coordinate system. It then expands the design surface boundary of the candidate component by 100 mm along each of the three coordinate axes to form a three-dimensional perception area. Subsequently, it filters out devices from the available sensor catalog that have mismatched observation object types, spatial resolution that does not reach half of the minimum feature size, field of view that cannot cover the three-dimensional perception area, target pose that is unreachable, or measurement period that overlaps with construction restricted period.

[0105] The module selects the execution device from the reserved devices in ascending order of calibration error, ascending order of arrival time, and dictionary order of device number, and generates a supplementary sampling task record. The record includes the globally unique identifier of the task, the globally unique identifier of the component, the field path, the 3D sensing area, the target surface mesh identifier and mesh version, the features to be supplemented for observation, the required sensing mode, the spatial resolution requirement, the target viewpoint or measurement window, the device number, the original component field change event identifier, the original evidence index, the lower bound of the target area overlap rate and its project rule version, the task validity period, and the completion criteria. If no device meets the conditions, the task is kept in the pending scheduling state and the advancement of the original field is prohibited.

[0106] The completion criteria for the supplementary sampling task are written according to the features to be supplemented. The coverage task requires the newly added effective surface area to reach the gap area. The interval width task requires at least five new observations with a time span of at least ten seconds. The mutually exclusive candidate differentiation task requires the target feature measurement interval to intersect with only one candidate prediction interval. The target area overlap rate is calculated using the area of ​​the target surface grid effectively covered by the returned data in the field coordinate system as the numerator and the total area of ​​the target surface grid of the same grid version in the task record as the denominator. Both areas are in square millimeters, and the ratio is a dimensionless number ranging from zero to one. The qualified lower bound for the specific implementation method is 0.8. This lower bound is obtained by the field calibration samples of the project supplementary sampling rules under the condition that the effective feature detection rate is not less than 95%, and is written into the task along with the rule version. After the equipment uploads the data, the equipment number, calibration version, target surface grid version, target area overlap rate and task validity period are checked in sequence. If the overlap rate is lower than the qualified lower bound, the data is retained as ordinary observation, the task is kept in the pending sampling state and the viewpoint or equipment is reselected according to the original scheduling rules. If any other item is not met, the data is retained as ordinary observation but is not used as evidence of the completion of the task.

[0107] The available sensor catalog is refreshed by the device management service every ten minutes. The recorded spatial resolution and calibration error are obtained from standard calibration board tests. The reachable pose is obtained from the intersection of the device motion envelope and the construction restricted area. The available measurement time period is obtained from the construction plan and the device task queue. The catalog version and the supplementary sampling task are saved together. When the catalog changes cause the selected device to become invalid, the device allocation is cancelled and the same sorting rules are used to reselect. The original component field change event identifier remains unchanged.

[0108] After the device returns a new observation, the system adds the original record key of the new observation to the original data index according to the original component field change event identifier. It retains the original event's pre-change value, event identifier, and collected evidence. It does not create alternative events that would lose historical lineage. Then, it updates the corresponding results of the component using the same candidate generation rule (S3), joint cost parameters, and field candidate calculation method. Finally, it updates the measurement interval, coverage, occlusion rate, cross-modal discrete quantity, hold duration, and constraint residuals using the same evidence acceptance boundary (S4).

[0109] When an updated field is marked as a submittable field, the module closes the supplementary data collection task and sends the field candidate value, the updated gating result, and the complete evidence index to the affected dependency closure entry point of S6. If it is still marked as a field to be verified, the current field value of the BIM master version is maintained and the same supplementary data collection task is updated with the remaining evidence gaps. If the boundary is not met after three supplementary data collection attempts, it is transferred to the manual review queue. When a field is marked as a conflict field, automatic writing is stopped, unexecuted tasks are closed, and the reasons for mutually exclusive candidates, non-relaxable constraints, or lack of effective supplementary data collection features are written into the conflict record.

[0110] In this specific embodiment, S6 includes:

[0111] The transaction update module reads the initial gating result of S4, the updated gating result of S5, and the read-only BIM constraint diagram version of S2 simultaneously. It only uses fields with a gating status of "can be submitted" as seed fields. Fields to be verified and conflict fields keep the current value of the BIM master version unchanged and retain the supplementary collection task identifier or conflict record identifier respectively. If the same component field exists in both gating results, the updated gating result with more original evidence and a newer timestamp is used as the seed.

[0112] For pose or size parameter seeds, the module first constructs a motion sweep query body by the union of the bounding body before updating and the bounding body after writing candidate values, and uses the query body to retrieve the full graph spatial index of S2. All potential intersecting components and their connection, support and collision constraints are added to the temporary candidate relation set. Then, a breadth-first traversal is performed from the seed component along the connection relation, support relation, existing spatial collision associated edge and temporary candidate relation. For finite construction state seeds, the traversal is along the assembly relation and construction sequence relation. For typed attribute seeds, the traversal is along the containment relation and attribute derivation relation. The traversal is deduplicated by the globally unique identifier of the component and the field path. When the associated field of the adjacent node is not affected by the change of the seed candidate value, the expansion along the branch stops.

[0113] The impact determination adopts a fixed mapping from field to relation type. Changes in pose and size trigger recalculation of contact surface distance, support direction and collision enclosure. Changes in construction status trigger recalculation of pre- and post-construction status and assembly completion. Changes in typed attributes trigger recalculation of derivation rules and parent summary attributes that reference the field. The module writes the seed field, triggered relation, adjacent field, extension level and stopping reason into the dependency closure record, thereby forming a repeatable and verifiable affected dependency closure.

[0114] The module copies the component nodes, related edges, and constraint records from the affected dependency closures from the BIM master version to the memory shadow subgraph. The shadow record retains the master version number and the field values ​​before the change. Then, it writes only the candidate values ​​of the submittable fields into the shadow subgraph and recalculates the derived attributes from the upstream to the downstream of the topological order according to the attribute derivation rules of S2. When a derivation cycle, missing source field, or incommensurable unit is detected, the shadow writing stops and a failure check item is generated.

[0115] The geometric verification first reconstructs the bounding volume of the component based on the updated pose and dimensions, and compares the spatial query results before and after the update in the shadow subgraph to form three types of collision candidate relationships: newly added, retained, and disappeared. For each newly added and retained candidate pair, an intersection judgment is performed. For candidate pairs that are not allowed to intersect, a triangular mesh intersection test is performed, requiring a penetration depth of zero. The disappeared candidate relationship is removed only when it is confirmed that the bounding volume does not intersect after the update and the minimum surface distance is greater than the sum of the corresponding modeling tolerances. For the design contact relationship, a continuity check is performed to ensure that the included angle of the normal of the connection surface is not greater than two degrees and the connection gap is not greater than the design allowable gap. For the support relationship, the supported projection area is required to fall into the effective boundary of the support surface. If any condition is not met, the component identifier, relationship identifier, measurement value, allowable boundary, and failure evidence are recorded.

[0116] Construction sequence verification checks whether the status of the preceding components has been reached and records the preceding status according to the construction sequence. It also checks whether the completion of the sub-components supports the target status of the parent component according to the assembly. Status rollback is only allowed when the change event explicitly carries an authorized error correction flag. Type and value verification checks the data type, unit, allowed set and allowed range according to the field constraint table. Information delivery requirement verification checks whether the component identifier, version number, modification subject, modification time and evidence index are complete.

[0117] All verification items are either passed or failed. The module generates recovery patches according to the shadow write order. The recovery patch records the component identifier, field path, candidate value, value before change, relationship addition / deletion items and major version number in the reverse order of field writing, and calculates the content summary. The recovery patch is only executed when the shadow verification fails, the database transaction commit fails, or the integrity check after commit fails. Fields to be verified and conflicting fields are not included in the recovery patch because their values ​​have never been written to the shadow candidate.

[0118] When all checks for geometric collision and connection continuity, construction sequence, data type and value range, and information delivery requirements pass, the module compares the shadow subgraph with the BIM master version, extracts only the fields and relationships that have changed within the affected dependency closure, and forms a local incremental patch. The patch record includes the base version, target version, component identifier, field path, value before change, value after change, relationship operation, gating event identifier, and patch summary. Components and relationships outside the closure are not included in the patch.

[0119] When shadow verification begins, the module generates a complete read set version list for all component nodes, relation edges, field constraints, spatial index shards, and derived rules actually read in this verification. The list records the object type, global identifier, base version, and content summary for each item. Database transactions lock or review all version records in the read set in a fixed order of object type, project number, and global identifier, not just the patch write set. If the current version of any component, relation, constraint, index shard, or rule is inconsistent with the shadow base version, the current commit is rolled back and the dependency closure and shadow subgraph are rebuilt with the new major version.

[0120] After all read set versions are verified to be consistent, the transaction writes the local incremental patch, the incremented component version number, the modification history consisting of the modification subject and the modification time, and the evidence index of the component field change event to the original data within the same serializable isolation range. All writes are successful, the patch digest is verified to be consistent, and the read set version is checked again before committing. If version inconsistency, constraint verification failure, digest inconsistency, or any database write failure occurs, the entire transaction is rolled back.

[0121] Transactions adopt a serializable isolation level and acquire version row locks in the fixed order. If the lock wait time exceeds five seconds, the current transaction is judged as a failed commit. The module does not change the gating state but puts the local incremental patch back into the retry queue. Before retrying, the BIM master version is reread, the full drawing spatial index is re-queried, and the dependency closure and shadow subgraph are rebuilt. The retry is at most two times. If it still fails, it is written to the conflict isolation area, thereby avoiding the overwriting of concurrent commit results with old base version or expired verification read set.

[0122] When a recovery patch is executed, it first checks whether the candidate value summary of the target cache or temporary record is consistent with the candidate value summary recorded in the patch. Then, it restores the field values, relation items and version flags in reverse order. For each item restored, it reads and compares the value summary before the change. When all are consistent, the recovery status is recorded as complete. If the summary is inconsistent, the automatic recovery is stopped, the corresponding component version is frozen and written to the recovery exception queue to prevent the recovery operation from overwriting the data that has been committed by another transaction.

[0123] After a transaction is rolled back, the module application recovery patch cleans up any possible shadow caches and uncommitted relationship states, and writes failed component field change events, failed validation items, database error codes, recovery patch summaries, and original data indexes into the conflict isolation zone. After a transaction is successfully committed, the local incremental patch identifier, modification history identifier, and evidence index identifier are written back to the gating result, the component version lock is released, and the BIM master version update event is published, so that subsequent update cycles only read the new version that has been atomically committed and whose evidence is traceable.

[0124] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0125] This invention organizes on-site multimodal observations, BIM component fields, and component engineering relationships into a continuous data processing chain. This allows observed events to be transformed into verifiable component field change events through candidate association, joint solution, and field-level difference determination. Consequently, the model update object is transformed from manual modification of the entire model to local field patches with clear evidence sources and dependencies.

[0126] This invention sets a field evidence sufficiency gating before writing candidate fields into the shadow subgraph and keeps the re-collection task associated with the original component field change event. This structure allows new observations to re-participate in the corresponding solution and gating in the same event, and only enters the main version when the evidence conditions and dependency checks are met, thus taking into account continuous updates, conflict isolation, and evidence traceability.

Claims

1. A BIM twin dynamic update method, characterized in that, include: S1. Acquire multimodal sensing data from the construction site and generate observation events carrying spatiotemporal references, sensor sources, and measurement quality. S2. Parse the initial BIM and generate a BIM constraint diagram with component identifiers and component fields as node data and component engineering relationships as edges; S3. Establish candidate relationships between observations and components based on observation events and the BIM constraint diagram, jointly solve the observation matching residuals and constraint violation quantities, and obtain the corresponding results of components and candidate values ​​of component fields. S4. Convert the difference between the candidate values ​​of the component field and the current field value into component field change events that are related to the observation evidence. Implement field evidence sufficiency gating on the component field change events to obtain the initial gating results including submittable fields, fields to be proven, and conflicting fields. S5. Generate a supplementary observation task for the original component field change event based on the evidence gap of the field to be verified. Backfill the new observations obtained from the supplementary observation to the original component field change event. Re-execute the joint solution and field evidence sufficiency gating to obtain the updated gating results. S6. Extract the affected dependency closures of the submitable fields based on the initial gating results, the updated gating results, and the BIM constraint diagram. Write the field candidate values ​​in the shadow subgraph and perform constraint verification. If the verification passes, write the local incremental patch, modification history, and evidence index to the BIM main version in the same transaction. If the verification fails or the field belongs to the field to be verified or the conflicting field, keep the corresponding field value unchanged.

2. The BIM twin dynamic update method according to claim 1, characterized in that, S1 includes: Acquire at least two types of sensing data from point cloud, construction images, positioning measurements, component identification, construction equipment operating conditions, and environmental or quality sensing sequences. Unify all sensing data into the field coordinate system and time axis; encapsulate the observed events according to the data acquisition time, field coordinates, sensor source identifier, measurement covariance or sensor quality label, and raw data index.

3. The BIM twin dynamic update method according to claim 1, characterized in that, S2 include: For each component in the initial BIM, establish component nodes containing a globally unique component identifier, pose parameters, dimensional parameters, finite construction status, typified attributes, and version number; establish the edges based on the connection relationships, support relationships, inclusion relationships, assembly relationships, and construction sequence relationships in the initial BIM; establish spatial collision-related edges based on the intersection relationships of the component's geometric bounding body in the site coordinate system; and establish attribute derivation relationships based on the source field references in the typified attribute calculation rules. The BIM constraint diagram is obtained by associating the field data type, allowed value set, unit, allowed range, and constraint level flag indicating whether deviation from the corresponding constraint is allowed to the corresponding component node.

4. The BIM twin dynamic update method according to claim 3, characterized in that, S3 includes: establishing deterministic candidate relationships between observation events carrying component identifiers and component nodes with the same component identifiers; for observation events without established deterministic candidate relationships, generating a candidate component set based on spatial neighborhood, geometry, component category, adjacency relationship, and construction time window; within the candidate component set, weighting the observation matching residuals based on the measurement quality of each sensing mode, and jointly solving the problem by combining the topological residuals of connection or support relationships, the construction sequence violation amount of finite construction states, and the value violation amount of typified attributes, outputting the corresponding component results and candidate values ​​of component fields for pose parameters, size parameters, finite construction states, and typified attributes.

5. The BIM twin dynamic update method according to claim 4, characterized in that, S4 includes: For pose parameters, size parameters, finite construction states, and typified attributes, field update thresholds are set respectively. The field update thresholds are determined by at least one of the modeling tolerance, sensor calibration error, and project acceptance allowable deviation of the corresponding field in the initial BIM. When the difference between the candidate value of a component field and the current field value exceeds the corresponding field update threshold, a component field change event is generated, which includes the component's globally unique identifier, field path, previous value, candidate value, evidence quality information, timestamp, original data index, and association constraints.

6. The BIM twin dynamic update method according to claim 5, characterized in that, The field evidence sufficiency gating includes: for each component field change event, obtaining the measurement interval width of candidate values, effective observation coverage, occlusion rate, cross-modal candidate value dispersion, continuous observation retention time, and constraint residuals; determining the acceptance boundaries of the aforementioned items based on sensor calibration samples, visible surface range of components, corresponding field update thresholds, construction plan time windows, and BIM constraint diagrams; defining field value constraints that are not allowed to deviate from the constraint level as non-relaxable field value constraints; marking the corresponding field as a submitable field when all items meet the corresponding acceptance boundaries, there are no mutually exclusive candidates, and the non-relaxable field value constraints are not violated; marking the corresponding field as a conflicting field when there are mutually exclusive candidates and their measurement intervals do not overlap, or when the non-relaxable field value constraints are violated; and marking the remaining fields as fields to be verified.

7. The BIM twin dynamic update method according to claim 6, characterized in that, S5 includes: For fields to be verified with mutually exclusive candidates, features that differ in predicted observations and can be observed by available sensors are extracted from the field differences of the mutually exclusive candidates and the adjacent nodes of the BIM constraint diagram to form a supplementary feature set. The discriminative power of each feature in the supplementary feature set against the mutually exclusive candidates is calculated, and the feature with the highest discriminative power is identified as the supplementary observation feature. When the supplementary feature set is empty, the corresponding field is relabeled as a conflicting field, and the generation of supplementary observation tasks is stopped. For fields to be verified without mutually exclusive candidates, the absolute value of the difference between each observation feature that does not meet the acceptance boundary and the corresponding acceptance boundary is calculated. The absolute value of the difference is divided by the non-zero normalized scale of the corresponding observation feature, and the observation feature with the largest ratio is identified as the supplementary observation feature. The measurement features are defined as follows: the non-zero normalized scale is the difference between the upper and lower bounds of the corresponding observation feature in the sensor calibration sample, and when the difference is zero, the data acquisition system takes the pre-configured record increment of the observation feature that is greater than zero; for the field to be verified whose observation features have been determined to be supplemented, the observation features to be supplemented are back-projected onto the field coordinate system to determine the three-dimensional perception area, and a supplementary acquisition task is generated based on the observation object type, spatial resolution, field of view, reachable pose and available measurement time period of the available sensors. The supplementary acquisition task includes the globally unique identifier of the component, the field path, the three-dimensional perception area, the required sensing mode, the spatial resolution requirement, the target viewpoint or measurement time window and the original component field change event identifier.

8. The BIM twin dynamic update method according to claim 7, characterized in that, The new observations obtained from the supplementary collection task are added to the original data index according to the original component field change event identifier. The original value and event identifier of the component field change event are retained. The corresponding results of the component, the candidate values ​​of the component field, and the items required for field evidence sufficiency gating are updated using the new observations. When the updated corresponding field is marked as a submittable field, it is sent to the affected dependency closure. When it is still marked as a field to be verified, the current field value is maintained and the supplementary collection task is updated. When it is marked as a conflict field, the automatic writing of the field is stopped and the conflict record is saved.

9. The BIM twin dynamic update method according to claim 6, characterized in that, S6 include: For submittable fields of pose or size parameters, extend the affected dependency closures along the edges of connectivity, support, and spatial collision. For committable fields with finite construction states, extend the affected dependency closures along assembly and construction sequence relationships; for committable fields with typed attributes, extend the affected dependency closures along containment and attribute derivation relationships. Perform geometric collision and connectivity continuity checks, construction sequence checks, data type and value range checks, and information delivery requirement checks on the shadow subgraph, and output the check results and a recovery patch opposite to the field writing direction.

10. The BIM twin dynamic update method according to claim 9, characterized in that, When the verification results show that all verifications pass, the difference between the shadow subgraph and the BIM master version is converted into a local incremental patch containing only the components and relationships within the affected dependency closure. In a database transaction, the local incremental patch, component version number, modification subject, and modification time constitute the modification history, as well as the evidence index of component field change events to the original data. If any verification fails or the database transaction commit fails, the database transaction is rolled back or the recovery patch is applied, and the component field change events that fail the verification are written as exception records to the conflict isolation area.