A 3D Modeling Method and System for Construction Management Behavior Based on a Large Language Model

By using a large language model-based approach, the problem of converting fuzzy positions in natural language instructions at construction sites was solved, generating three-dimensional poses that conform to existing component geometric relationships and engineering constraints, thus enabling accurate model instance creation and updating in construction management.

CN123118952APending Publication Date: 2026-11-17CHINA CONSTR FOURTH ENG DIV CORP LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611574226.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-12
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The ambiguous relative positions in the natural language management instructions at the construction site are difficult to convert into BIM 3D scenes that can directly drive the creation or updating of model instances, and it is also difficult to meet the existing geometric relationships of components and engineering constraints.

Method used

A method based on a large language model is adopted. Through spatial hierarchical retrieval, candidate reference object disambiguation and reference sub-region determination, the relative orientation is converted into a local candidate position. Combined with the geometric information of the target entity and engineering constraints, a feasible 3D pose is generated for the creation or updating of model instances in digital twin 3D scenes.

Benefits of technology

It enables the accurate conversion of fuzzy positions in natural language management instructions into candidate 3D poses in the engineering world coordinate system, ensuring that 3D model instances conform to existing component geometric relationships and engineering constraints, thereby improving the accuracy and efficiency of construction management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118952A_ABST
    Figure CN123118952A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for 3D modeling of construction management behaviors based on a large language model. The method involves acquiring natural language management instructions from the construction site, parsing the target entity, execution actions, and spatial location descriptions using a large language model, and determining 3D model resources. Candidate reference objects are retrieved and disambiguated within a BIM-based 3D spatial knowledge base to determine the target reference object and its sub-region. Local candidate positions are determined based on the target reference object's local coordinate system and transformed to the engineering world coordinate system, generating a set of candidate 3D poses. Hard constraint verification and soft constraint evaluation are performed by combining the target entity's geometric information, existing component geometric information, and engineering constraints to determine the target's 3D pose. Model instances in a digital twin 3D scene are created or updated based on the execution actions and the target's 3D pose. This invention can convert fuzzy relative spatial descriptions into executable 3D poses that meet engineering constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of computer 3D model processing, building information modeling, and artificial intelligence, specifically to a 3D modeling method and system for construction management behavior based on a large language model. Background Technology

[0002] With the application of Building Information Modeling (BIM) and digital twin technology on construction sites, buildings, floors, rooms, components and their spatial locations on construction sites can be expressed through three-dimensional models. Construction managers can also use three-dimensional scenes to view or update the status of on-site objects.

[0003] Construction site management instructions, inspection records, and rectification requirements often include spatial descriptions in natural language, such as "staircase corner," "left side of component," or "near passageway." These descriptions typically use existing components or spatial areas as references but do not directly provide absolute coordinates and orientations usable by the 3D engine. When multiple similar components exist on the same floor, or when the referenced component itself contains multiple local areas, it is difficult to directly determine the specific pose of the object to be manipulated in the 3D scene based solely on entity names or general semantic similarity.

[0004] Existing BIM or digital twin systems can query and update scene objects based on manual input, structured status data, or known location data. There are also methods that utilize natural language to assist in retrieving existing objects or generating 3D scenes. However, when using fuzzy relative spatial descriptions referencing existing components to create or update the location of objects in actual construction scenes, further processing is needed to uniquely determine the reference object, locate local areas within the reference object, convert local directions to engineering world coordinates, and address the limitations imposed on candidate locations and orientations by the geometric relationships of existing components and engineering constraints. Without these processing steps, semantically reasonable locations may collide with existing components, fall on non-attachable surfaces, intrude into restricted access spaces, or fail to serve as direct spatial transformation parameters for 3D model instances.

[0005] Therefore, there is a need for a 3D modeling scheme that can convert the fuzzy relative positions in the natural language management instructions at the construction site into 3D poses that can directly drive the creation or updating of model instances, and keep the 3D poses consistent with the local space of the reference component and the existing geometric relationships and corresponding engineering constraints of the construction site. Summary of the Invention

[0006] The technical problem to be solved by this invention is: how to convert the fuzzy relative position in the natural language management instructions of the construction site into a three-dimensional pose that can directly drive the creation or updating of model instances in the existing BIM three-dimensional scene, and make the three-dimensional pose correspond to the local space of the reference component pointed to by the natural language, and also satisfy the geometric relationship of the existing component and the corresponding engineering constraints.

[0007] To address the aforementioned technical problems, this invention provides a method and system for 3D modeling of construction management behaviors based on a large language model. This solution takes natural language management instructions from the construction site as input, extracts target entities, execution actions, and spatial location descriptions from the large language model. Within a BIM-based 3D spatial knowledge base of the construction site, spatial hierarchy retrieval, candidate reference object disambiguation, and reference sub-region determination convert relative orientations into local candidate positions and transform them to the engineering world coordinate system. Candidate 3D poses are generated within the engineering world range corresponding to the reference sub-region or its neighborhood. Then, by combining the geometric information of the target entity, the geometric information of existing components, and engineering constraints, a set of feasible poses is formed, and the target 3D pose is determined. Finally, model instances in a digital twin 3D scene are created or updated based on the execution actions and 3D model resources.

[0008] The first aspect of this invention provides a three-dimensional modeling method for construction management behavior based on a large language model, the method comprising: Natural language management instructions from the construction site are obtained, and the target entity, the action to be executed, and the spatial location description including at least spatial hierarchy information, reference object information and relative orientation information are obtained through large language model parsing. The corresponding three-dimensional model resources are determined based on the target entity.

[0009] Based on the spatial hierarchy information, a spatial retrieval range is determined in a three-dimensional spatial knowledge base of the construction site constructed based on Building Information Modeling (BIM) and including at least spatial hierarchy information, component geometric information, and component local coordinate system, and candidate reference objects corresponding to the reference object information are obtained from the spatial retrieval range.

[0010] The candidate reference objects are disambiguated based on the limiting information in the spatial location description to determine the target reference object, and a reference sub-region is determined from the geometric information of the target reference object based on the relative orientation information.

[0011] Using the local coordinate system of the target reference object and the reference sub-region as a reference, the relative orientation information is converted into local direction information. Based on the local direction information, local candidate positions are determined and transformed to the engineering world coordinate system. Based on the transformed candidate positions, a set of candidate three-dimensional poses is generated within the engineering world range or its neighborhood corresponding to the reference sub-region.

[0012] Based on the target entity geometric information, component geometric information and engineering constraints corresponding to the 3D model resources, the candidate 3D pose set is subjected to hard constraint verification to obtain a feasible pose set that satisfies all hard constraints. The feasible pose set is then evaluated based on soft constraints to determine the target 3D pose.

[0013] Based on the executed action, the target 3D pose, and the 3D model resources, a model instance corresponding to the target entity is created or updated in the digital twin 3D scene, and the spatial transformation parameters of the model instance are determined based on the target 3D pose.

[0014] In one possible implementation, a structured pattern is pre-established that specifies the target entity, the action to be performed, and the corresponding fields of spatial location description. The output of the large language model is then constrained and decoded according to the structured pattern. The spatial location description is parsed into at least one of building, floor, grid, room, reference category, ordinal limit, orientation limit, and distance or range. The parsing results are then subjected to structural compliance verification and field integrity verification.

[0015] In one possible implementation, vector semantic retrieval and domain alias matching are performed based on the target entity and its context, and the model resource candidates returned by the two channels are fused and sorted to determine a unique model resource identifier. The corresponding 3D model file and model resource information for unifying the model coordinate reference are obtained from the model resource identifier. The model resource information includes at least one of size unit, coordinate origin and default orientation.

[0016] In one possible implementation, the spatial retrieval range is narrowed down step by step according to the spatial inclusion relationship between buildings, floors, grids or rooms and components, and a candidate reference set is formed by using the matching relationship between reference information and component semantic tags and the component geometric bounding box.

[0017] In one possible implementation, candidate references are comprehensively evaluated based on at least one of ordinal constraints, relative orientation, semantic tags, and topological relationships to determine the target reference; then, based on the relative orientation, reference sub-regions are determined from the geometric bounding box boundary, component outline edge, attachable surface, or semantic sub-component of the target reference.

[0018] In one possible implementation, the geometric center of the reference sub-region in the local coordinate system of the target reference object is used as a reference. Local candidate positions are obtained based on the orientation offset, local unit direction and offset distance. Initial candidate positions are obtained through the coordinate transformation relationship between the local coordinate system and the engineering world coordinate system. The position and orientation are sampled within the engineering world range or its neighborhood corresponding to the reference sub-region to form a set of candidate three-dimensional poses.

[0019] In one possible implementation, the geometric information of the target entity is obtained based on the 3D model resources, and engineering hard constraint verification is performed on the candidate 3D poses. The candidate 3D poses that satisfy all hard constraints are formed into a feasible pose set, and the feasible poses are evaluated according to soft constraints within the feasible pose set. The feasible 3D pose with the highest comprehensive score is determined as the target 3D pose. When there are no candidate 3D poses that satisfy all hard constraints, the search range is expanded, the attachable surface is switched, or the candidate region is output for manual confirmation.

[0020] In one possible implementation, for actions such as adding, placing, installing, or configuring, a model prototype object is generated based on the 3D model resource, and the world transformation matrix of the model instance is constructed based on the target 3D pose. For actions such as moving or rearranging, an existing model instance is located based on the original position description, and the target 3D pose after movement is calculated based on the target position description, and its world transformation matrix and bounding box are updated. For actions such as inspection, repair, dismantling, acceptance, or status change that do not involve position changes, the target model instance is determined from the object registry or spatial index based on the spatial position description, and its model attributes or visualization status is updated without recalculating the target 3D pose or changing its world transformation matrix. In cases where an instruction contains multiple actions, they are processed sequentially according to the order expressed by the instructions; actions requiring position changes undergo pose calculation, while actions not involving position changes undergo corresponding attribute or status updates.

[0021] In one possible implementation, a bidirectional association is established between structured management behavior records and model instances; when the execution status of management behavior, actual 3D pose, or model attributes change, a state event is generated, and the scene state of the structured management behavior records and model instances is updated synchronously based on the state event.

[0022] A second aspect of the present invention provides a three-dimensional modeling system for construction management behavior based on a large language model. The system includes a semantic parsing module, a three-dimensional model resource management module, a three-dimensional spatial knowledge base for construction sites, a spatial pose calculation module, and a three-dimensional scene operation module. Each module operates collaboratively according to the data transmission relationship of natural language instruction parsing, three-dimensional model resource determination, spatial pose calculation, and three-dimensional scene operation to achieve the aforementioned method.

[0023] Compared with existing processing methods, this invention maps the fuzzy relative spatial description in natural language to a unique target reference and its local geometric region step by step, and uses the local geometric region and the local coordinate system of the target reference as the spatial calculation benchmark, so that the natural language position description lacking absolute coordinates can be converted into a candidate three-dimensional pose in the engineering world coordinate system.

[0024] Furthermore, this invention utilizes the geometric information of the target entity, the geometric information of components in the existing 3D scene, and the corresponding engineering constraints to perform hard constraint filtering on candidate 3D poses, and only performs soft constraint evaluation within the set of feasible poses that satisfy the hard constraints. This avoids poses that do not satisfy the existing scene geometric relationships or corresponding engineering constraints from being directly used as the result of 3D scene operations, so that the obtained target 3D poses can be directly used for model instance creation or spatial state update in the digital twin 3D scene.

[0025] Furthermore, by writing the target 3D pose into the spatial transformation parameters of the model instance, the spatial semantic solution results can be directly applied to specific model objects in the digital twin 3D scene. In the implementation method of establishing a bidirectional association between structured management behavior records and model instances, the scene state of the model instance can also be updated synchronously when the management behavior state or the actual 3D pose changes. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method steps of the present invention.

[0027] Figure 2 This is a schematic diagram of the three-dimensional spatial knowledge base for the construction site of this invention.

[0028] Figure 3 This is a flowchart of the three-dimensional pose calculation process of the present invention.

[0029] Figure 4 This is a schematic diagram illustrating the bidirectional association between management behavior records and model instances in this invention.

[0030] Figure 5 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0031] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0032] To facilitate understanding of the embodiments of this application, the technical terms involved in the embodiments of this application will be explained below.

[0033] In this application, "3D model resources" refers to standardized model files and their metadata that can be repeatedly accessed; "model instances" refers to specific scene objects created from 3D model resources within a specific digital twin 3D scene and possessing independent spatial transformation parameters. "Target entities" are objects created, moved, inspected, repaired, or whose status is updated in natural language management instructions; "reference points" are existing BIM components or scene objects used to interpret the relative spatial position of the target entity, and these two can be different objects.

[0034] The reference sub-region referred to in this application refers to a local geometric region within the geometric range of the target reference object, further determined based on relative orientation in natural language. Determining the reference object can only answer "relative to which component for positioning," while the reference sub-region further answers "which local area of ​​this component serves as the spatial calculation reference," thereby avoiding directly equating the entire reference component with the specific placement position of the target entity.

[0035] The three-dimensional pose referred to in this application includes position and orientation, denoted as _____. ,in Represents the three-dimensional position in the engineering world coordinate system. This indicates the orientation of the target model. For 3D objects that need to be placed, moved, or rearranged, determining the position alone is insufficient to uniquely determine its spatial state. Therefore, subsequent constraint verification and soft constraint evaluation can both use the joint pose composed of position and orientation as the processing object.

[0036] In an implementation environment, the digital side pre-contains BIM data, a 3D model resource library, and a digital twin 3D scene corresponding to the construction project. The target 3D pose calculated by the system is used for the planned or physical model instance in the digital twin 3D scene; the actual physical operations such as installation, relocation, and maintenance are performed by on-site personnel, who can submit the execution results or confirm the actual pose after completion for updating the digital scene.

[0037] See Figure 1 This embodiment provides a three-dimensional modeling method for construction management behavior based on a large language model, including steps S1 to S6. Each step forms a continuous processing chain in the order of natural language parsing, spatial reference determination, candidate pose generation, pose constraint solving, and three-dimensional scene operation.

[0038] S1: Obtain natural language management instructions from the construction site, perform semantic parsing, and determine the 3D model resources corresponding to the target entity.

[0039] Natural language management instructions at the construction site can originate from text input via mobile terminals, speech-to-text transcription, or inspection records. The large language model performs semantic parsing on these instructions, outputting the target entity, the action to be performed, and a spatial location description; the spatial location description includes at least spatial hierarchy information, reference point information, and relative orientation information. To enable subsequent spatial processing to directly utilize the output, a structured schema can be pre-established, specifying the data type, mandatory / unmandable value range, and allowed value range for each field. During the large language model generation process, constrained decoding corresponding to this structured schema is employed.

[0040] Before inputting into the large language model, the original text can undergo speech-to-text error correction, synonym normalization, unit unification, and invalid character cleanup, while retaining the original instructions and the position indexes of each text segment. The preprocessed text serves as the input for semantic parsing, and the position indexes are used to locate the corresponding original semantic segments during field validation, manual confirmation, or subsequent tracing.

[0041] The structured schema can be defined using JSON Schema, specifying at least the data type, required fields, allowed value range, and nullable conditions for the fields corresponding to the target entity, execution action, and spatial location description. The spatial location description includes at least fields such as spatial hierarchy, reference point, and relative orientation. For actions involving movement or rearrangement, original location description and target location description fields are also provided, used to locate the existing target model instance and describe the spatial location after movement, respectively. For actions such as inspection, repair, demolition, acceptance, or status change that do not involve location changes, the spatial location description is used to locate the existing target model instance; it is not required to recalculate the target's 3D pose. During constrained decoding, a finite state constraint or syntax parser corresponding to the JSON Schema can determine the legality of the currently generated character sequence in real time and mask the next candidate token, retaining only those tokens that can continue to form a legal structure and conform to the field name, data type, and enumeration range, thus reducing missing fields, misaligned fields, incorrect types, or illegal enumeration values.

[0042] For example, for the query "Please have the safety officer place a fire extinguisher at the corner of the stairs on the 3rd floor of Building 2", the following can be obtained: target entity "fire extinguisher", action "place", building "Building 2", floor "3rd floor", reference object "stairs", and relative location "corner". The safety officer's information as a participant can be stored in the structured management behavior record, but it is not a necessary spatial field for subsequent 3D pose calculation.

[0043] The large language model can complete semantic role labeling and field normalization based on the original instruction context; among them, information related to subsequent spatial pose calculation, such as building, floor, axis grid, room, reference object category, ordinal limit, orientation limit, and distance or range, is written into the spatial location description field, and other management information can be written into the structured management behavior record.

[0044] After the structured results are generated, they undergo grammatical compliance, field completeness, and domain rule validation. When key fields are missing, there are conflicts between fields, or there are multiple interpretations of target entities or spatial locations, the error information and the original instructions can be re-entered into the large language model for a limited number of error correction iterations. If the result still cannot be uniquely determined after error correction, the fields to be confirmed and candidate values ​​are output for management personnel to confirm. The large language model can be a general model or a fine-tuned model using construction domain corpus. In this embodiment, whether or not it has undergone domain fine-tuning is not a necessary condition for subsequent spatial solution calculation.

[0045] For the 3D model resources corresponding to the target entity, a combination of vector semantic retrieval and domain alias dictionary matching can be used. The canonical name, alias, category, purpose, and attribute tags of each model resource record in the resource repository are encoded to form a resource vector. The same embedding method is used for the target entity and its context to obtain the query vector. Let the target entity query vector be related to the first... The original cosine similarity of the resource record vectors is Then it can be linearly normalized into a vector score. .

[0046] (1) The domain alias dictionary establishes a mapping based on canonical entity names, alias sets, and model resource identifiers. For alias candidates that do not match precisely, normalized edit distance can be used as a reference. Jaccard similarity of terms and semantic similarity of aliases Calculate dictionary score Among them, normalization Edit distance, term Jaccard similarity, and alias semantic similarity were all normalized to the 0-1 range using a uniform scale.

[0047] (2) in, , and The weights are non-negative and sum to 1. Candidates returned from the vector and dictionary channels are merged by model resource identifier and combined with consistency scores based on resource category, action, and spatial context. Confidence of Computational Resource Fusion For a candidate returned by only one channel, the score of the channel that did not return the candidate can be set to zero or a preset default value; the consistency score is determined based on the degree of matching of resource category, execution action and spatial context, and normalized to the range of 0 to 1.

[0048] (3) in, , and The weights are non-negative and sum to 1. When the highest fusion confidence reaches the acceptance threshold and the score difference with the second-ranked candidate reaches the distinction threshold, the corresponding model resource identifier is determined; otherwise, several candidate resources are output for manual selection. The resource record corresponding to the model resource identifier may include information such as the 3D model file address, format, version, size unit, coordinate origin, default orientation, and connection anchor points. Before subsequent hard constraint verification, model geometry information, bounding box, or collider data for spatial verification can be obtained based on the 3D model file corresponding to the model resource identifier; and based on one or more of the aforementioned model resource information actually obtained in the resource record, one or more of the corresponding items in the model's size scale, local coordinate origin, and initial orientation are unified so that subsequent geometric verification and spatial transformation use a consistent coordinate reference.

[0049] S2, determine the spatial retrieval range based on spatial hierarchy information and obtain candidate reference objects.

[0050] Before performing fuzzy location calculations, it is necessary to incorporate spatial hierarchy information such as building, floor, and room from natural language into the existing BIM 3D spatial data to limit the actual retrieval range of reference objects. See also Figure 2 The three-dimensional spatial knowledge base of the construction site can be formed by parsing BIM or Industry Foundation Classes (IFC) data. Nodes represent buildings, floors, grids, rooms, components or existing scene entities, and records relationships such as inclusion, adjacency, connection, location or attachment between nodes.

[0051] In one implementation, building, floor, space, and component entities are read from BIM or IFC data, and hierarchical nodes of building-floor-room / grid-component are formed according to the original hierarchical relationship; the geometric bounding box and geometric center are extracted based on the geometric expression of the components, and a local coordinate system is established by reading or based on the component's own orientation, grid direction, and component principal axis direction; for surfaces that can be used for installation or placement of target objects, the corresponding attachable surfaces or surface attributes are saved. The above analysis belongs to the basic data preparation process for forming a spatial knowledge base.

[0052] The spatial pose calculation module first queries along the containment relationships of building nodes, floor nodes, and grid or room nodes. Within a limited space, it obtains similar components based on the matching relationship between the reference object category and the component semantic label. Then, it uses the component geometric bounding box to eliminate components that are not located within the target floor, grid range, or room boundary, forming a candidate reference object set. Each candidate reference should carry at least a component identifier, semantic label, geometric center, geometric bounding box, and local coordinate system for subsequent disambiguation and coordinate transformation.

[0053] S3, disambiguates candidate reference objects to determine the target reference object and reference sub-region.

[0054] When multiple similar reference components exist for a target floor, category information such as "stairs" alone cannot uniquely determine the specific object referred to by natural language. Therefore, the ordinal conformity, location conformity, semantic label conformity, and topological relation conformity of candidate references are compared. Let the set of matching factors be... Including ordinal, orientation, semantic, and topological factors, the first The basic weight of the class factor is And the basic weight is a non-negative value; whether this factor is available in the current instruction is determined by... It means that the first The conformity of each candidate under this factor is: If the conformity of each factor is normalized to the range of 0 to 1, then a comprehensive matching score can be expressed as: (4) in, Indicates the first The overall matching score of each candidate reference. Only the information actually possessed by the current instruction is set. When the current instruction does not contain the corresponding matching factor, the corresponding available identifier is set to 0. The conformity of each factor is determined based on the degree of matching between the candidate reference and the corresponding limiting information and normalized to the range of 0 to 1. Among them, the orientation, semantic, and topological factors are determined based on spatial orientation relationship, semantic label matching relationship, and topological relationship, respectively. For valid factors participating in the current matching calculation, the sum of their corresponding basic weights is guaranteed to be greater than 0. When the candidate with the highest comprehensive score reaches the reference acceptance threshold and the difference between its score and the second-ranked candidate reaches the discrimination threshold, the candidate is determined as the target reference. Otherwise, output candidate references and their matching information for manual confirmation. Weights and thresholds are configurable parameters that can be set based on project BIM accuracy, number of candidates, and verification results.

[0055] For the ordinal factor in equation (4), "first", "second", "closest to the entrance", etc. can be converted into ordinal numbers. Alternatively, a sorting rule can be established, and candidate sorting axes can be determined based on the instruction context. If a grid direction exists, the positive grid direction can be used. If a room entrance or personnel movement direction exists, the entrance point can be used as the sorting origin, and the corresponding movement direction as the sorting direction. If neither is specified, the sorting direction can be determined according to the preset order of the X, Y, and Z axes of the project coordinate system. Let the first... The geometric center of each candidate reference is The origin of the sort is The unit sorting direction vector is Then the candidate sorting value Press Confirmed; the ascending or descending order of the sorting results is determined by the directional semantics in the instruction or the preset sorting rules, and is based on the sequence number. Select the corresponding candidates. If the ranking values ​​are the same and cannot be distinguished by combining other existing matching factors, the corresponding candidates are retained as parallel candidates and enter the aforementioned candidate confirmation process; the ranking results are used for subsequent ordinal compliance calculation and candidate reference disambiguation.

[0056] After determining the target reference point, it is necessary to further determine the reference sub-region based on relative orientations such as "corner," "left side," and "nearby." To ensure that relative orientation corresponds to an executable local geometric region, a mapping relationship between relative orientation and the extraction type of the reference sub-region can be established in advance. Based on the relative orientation, the corresponding extraction type can be determined from the geometric bounding box boundary, component outline edge, attachable surface, or semantic sub-component. For example, "corner" can be mapped to the intersection of the two boundary surfaces of the target staircase bounding box, the two edges of the component outline, or the semantic sub-component; directions such as "left" and "right" are interpreted according to the corresponding axes of the target reference object's local coordinate system.

[0057] Reference sub-region Candidate sub-regions can be extracted from the geometric bounding box, contour boundary, attachable surface, or semantic sub-component of the target reference object according to the mapping relationship. When multiple candidate sub-regions are obtained and cannot be uniquely determined based on the current relative orientation and existing limiting information, the candidate sub-regions are output for management personnel to confirm, and the confirmed candidate sub-regions are used as reference sub-regions. The determined reference sub-regions serve as the initial spatial range for generating candidate 3D poses in the next step, and can also serve as the spatial boundary for subsequent neighborhood sampling and engineering constraint verification. When the geometric bounding box boundary is used to form candidate sub-regions or coarse positioning ranges, it is not directly equivalent to the final executable installation position; the final executability is still determined by combining the actual component geometry, attachable surface information, and subsequent hard constraint verification results.

[0058] S4. Based on the local coordinate system of the target reference object and the reference sub-region, determine the local candidate positions and transform them to the engineering world coordinate system to generate a set of candidate 3D poses.

[0059] See Figure 3 In natural language, relative orientation typically lacks absolute coordinate meaning and needs to be interpreted relative to the target reference object's own orientation. Let the rotation matrix from the target reference object's local coordinate system to the engineering world coordinate system be... The translation vector is Reference sub-region The center in the local coordinate system is The fixed offset corresponding to the azimuth description is The local unit direction vector is The offset distance is Wherein, the fixed offset is a three-dimensional offset vector in the local coordinate system of the target reference object, and its components use a length unit consistent with the center and offset distance of the reference sub-region; the local unit direction vector is a dimensionless unit vector. The local unit direction vector is determined according to a preset mapping relationship between the relative orientation and the local axis of the target reference object, the orientation offset is determined according to a preset offset rule corresponding to the relative orientation, and the offset distance is preferentially obtained from the distance or range field in the spatial position description; if not explicitly given, a preset offset distance corresponding to the relative orientation is used; then the local candidate position... It can be represented as: (5) After transforming the local candidate positions to the engineering world coordinate system, the initial candidate positions are obtained. : (6) When the engineering world coordinate center of the reference sub-region is denoted as In this case, the above formula can also be written as The initial orientation of the target model can be determined based on at least one of the following: the normal of the attachable surface of the reference sub-region, the local coordinate axis of the target reference object, or the approach direction of the personnel. The approach direction of the personnel can be determined based on the location of the room entrance, the direction of personnel movement, or the passable area corresponding to the candidate location, for example, making the front of the target model face the passable area.

[0060] Initial candidate position It's not the final position output directly, but rather the center of the candidate pose search. Within the reference sub-region... Within the corresponding engineering world or its neighborhood, multiple candidate locations can be generated using conventional spatial sampling. When the reference sub-region corresponds to an attachable surface, two-dimensional sampling can be performed along the attachable surface. For each candidate location, one or more candidate orientations are formed based on the attachable surface normal, local coordinate axes, or the direction of personnel approach, thereby obtaining a set of candidate three-dimensional poses. The sampling step size, number of candidates, and orientation discretization method can be configured according to the project's spatial scale and model size; this embodiment does not limit specific values.

[0061] S5 performs hard constraint verification and soft constraint evaluation on the candidate 3D pose to determine the target 3D pose.

[0062] Candidate 3D poses may contain semantically correct positions but are not executable in engineering space, such as collisions with existing components, locations on non-attachable surfaces, or intrusions into restricted access spaces. Therefore, this embodiment first uses hard constraints to form a feasible region, and then performs soft constraint evaluation only within the feasible region, so that mandatory engineering conditions and preferred evaluation conditions have different processing levels.

[0063] The candidate poses are uniformly represented as Before performing hard constraint verification, the geometric information of the target entity is obtained based on the 3D model resources determined in S1, and its coordinate reference is unified according to the aforementioned rules. Then, combined with the geometric information of existing components, attachable surfaces, and construction area information in the 3D spatial knowledge base of the construction site, hard constraint verification is performed on each candidate pose. Hard constraints may include at least one of the following: installation location range, attachability, clearance, collision or occlusion with existing components or facilities, and permitted construction area. Engineering constraints may originate from national standards, industry specifications, project technical requirements, or pre-configured engineering rules; specific thresholds can be configured according to the target entity category and project requirements, and this embodiment does not limit specific values.

[0064] When performing hard constraint verification, the geometric bounding boxes, attachable surfaces, and semantic tags of floor slabs, walls, passageways, fire protection facilities, and other existing components in the 3D spatial knowledge base of the construction site can be read, and the hard constraints can be represented as geometric inequalities, equality expressions, or Boolean decision conditions. For candidate 3D poses, it can be determined in sequence whether the target entity is in the allowed construction area, whether its installation position meets the pre-configured height range, whether the geometry of the target entity in the candidate pose intersects with existing components or facilities, whether it intrudes into the restricted passage clearance, and whether the corresponding attachable surface meets the load-bearing or installation conditions; if any hard constraint is not met, the corresponding candidate 3D pose is removed from the feasible candidates.

[0065] Let the set of hard constraint indices be , No. The decision function for a hard constraint is: When the corresponding hard constraint is satisfied, ,otherwise Then the set of feasible poses that satisfy all hard constraints. It can be represented as: (7) In one implementation, for The feasible pose in the image is evaluated based on at least one soft constraint among visibility, accessibility, personnel accessibility, maintenance convenience, offset from the initial candidate position, and impact on traffic flow. To avoid subjective management rules being used directly as the judgment conclusion, soft constraints can be characterized by quantities that can be calculated from 3D scene data, such as visibility distance, degree of occupancy, spatial distance from passable areas or areas accessible to personnel, offset from the initial candidate position, and degree of occupancy of existing traffic space. Let the first... The normalization utility of soft constraints is Weights are Each weight is non-negative and the sum of the weights is 1. Each normalized utility value is between 0 and 1. For indicators where smaller values ​​are preferred (e.g., offset from the initial candidate position, degree of occupancy of existing passageway space), normalization ensures that smaller original values ​​correspond to higher utilities. For indicators where larger values ​​or evaluation levels are preferred (e.g., visibility, accessibility), normalization ensures that larger original values ​​correspond to higher utilities. The comprehensive score can then be expressed as: (8) in, This represents the index of the soft constraints participating in the current evaluation, and the summation range includes all soft constraints participating in the current evaluation. Represents feasible 3D pose The soft constraint comprehensive score is used. The pose with the highest comprehensive score within the feasible pose set is selected as the target 3D pose. Therefore, the target location and target orientation It was obtained from the same joint pose evaluation.

[0066] In this context, occlusion in the hard constraints is used to determine whether a candidate pose would obstruct existing facilities with mandatory functional requirements, such as fire-fighting equipment or evacuation signs. Visibility in the soft constraints is used to compare candidate poses that satisfy all hard constraints; it can be characterized by quantities that can be calculated from 3D scene data, such as the target entity's visibility distance and degree of unobstructedness. These two constraints are used for feasibility filtering and feasible pose selection, respectively.

[0067] In another implementation, after hard constraint verification of candidate points and their orientations is completed, the final position can also be determined first based on soft constraint evaluation. When the final location corresponds to multiple feasible orientations, the multiple feasible orientations are evaluated based on at least one of visibility, operating surface orientation, personnel approach direction, and degree of obstruction to determine the final orientation. In this step-by-step implementation method, the final determination is... Then, the soft constraint comprehensive score of the final three-dimensional pose can be calculated according to equation (8). The comprehensive score is only used to calculate the comprehensive evaluation value of the spatial solution according to formula (9), and is not used to reselect the final position or final orientation.

[0068] when When empty, the search range of the initial candidate position can be expanded in a preset order, other attachable surfaces of the target reference can be switched, or the next candidate can be selected from the candidate reference set to re-execute the reference determination and pose solution; if there is still no feasible pose, the output will show no automatic feasible solution, conflict constraint terms and candidate regions, which will be confirmed by the administrator.

[0069] exist In the case of non-empty spaces, the comprehensive evaluation value of spatial solution can also be calculated. The aforementioned reference object matching score and soft constraint comprehensive score are both in the range of 0 to 1. The target reference object matching score in equation (9) represents the comprehensive matching score corresponding to the finally selected target reference object. For example, the smaller value between the target reference object matching score and the soft constraint comprehensive score of the target's three-dimensional pose is taken as the comprehensive evaluation value of the spatial solution: (9) in, This represents the comprehensive evaluation value of the spatial solution; This represents the final, determined target reference point. The comprehensive matching score corresponding to the target reference object is the comprehensive matching score of the candidate corresponding to the target reference object in equation (4); and These represent the final position and final orientation of the target in three dimensions, respectively. This represents the soft-constraint comprehensive score of the target's three-dimensional pose. The spatial solution comprehensive evaluation value is an internal comprehensive evaluation index used for automatic threshold determination and does not represent a statistical probability. When the target's position is not lower than a preset threshold, the spatial pose calculation module outputs the target's 3D pose; when... When the value is below a preset threshold, even if a feasible pose that satisfies the hard constraints exists, automatic scene updates can be paused, and candidate reference objects, candidate poses, and comprehensive evaluation values ​​can be output for management personnel to confirm. For the aforementioned cases where manual confirmation is required due to ambiguity, insufficient comprehensive evaluation values, or the absence of an automatic feasible solution, the management personnel's confirmation results are written back to the corresponding processing record, and subsequent processing continues based on the confirmed object, region, or pose.

[0070] [S6, Create a model instance or update an existing model instance based on the executed action, the target's 3D pose, and 3D model resources.]

[0071] For actions such as adding, placing, installing, or configuring, the 3D engine reads the corresponding resource record based on the model resource identifier determined in S1, loads the 3D model file corresponding to the resource record, parses the complete model geometry, material, hierarchy, and connection anchor points, and performs corresponding unification on the parsed model data according to the same coordinate reference unification rules as mentioned above, based on one or more of the actual obtained size units, coordinate origin, and default orientation, forming a repeatable instantiable model prototype object. The model resource identifier is used to locate resource records and is not equivalent to a specific model instance in the scene.

[0072] When creating a scene instance based on the model prototype object, an instance identifier (instanceId) is assigned to the instance, and the instance is then determined based on the target's 3D pose. Forming the world transformation matrix The world transformation matrix, based on the unified local coordinate reference of the aforementioned model, is used to transform the vertices and anchor points in the local coordinate system of the model to the engineering world coordinate system.

[0073] (10) in, The world transformation matrix represents the model instance; Indicates the direction from the target The converted Rotation matrix; This represents the target position vector of the local coordinate origin of the model in the engineering world coordinate system after the model coordinate datum unification is completed, with dimension 1. ; in the matrix express Zero vector, This represents the homogeneous coordinate term. The connection anchor points in the resource record, as model data in the local coordinate system, are transformed to the engineering world coordinate system along with the world transformation matrix. The created model instance can save the instanceId, model resource identifier, model file version, entity category, world transformation matrix, bounding box, visualization status, and business attributes, and is added to the scene tree, spatial index, and object registry of the digital twin 3D scene. Thus, the target 3D pose obtained in S5 is directly converted into the spatial transformation parameters of the model instance, completing the closure from spatial semantic calculation to 3D scene operation.

[0074] For movement or rearrangement actions, when the spatial location description includes both the original location description and the target location description, the system first filters matching scene instances in the object registry based on the resource category of the 3D model resource. Then, it narrows down the candidate range using the reference objects, reference sub-regions, and spatial indexes corresponding to the original location description. Finally, it performs a comprehensive matching based on type consistency, distance proximity, reference sub-region overlap, and command context consistency to determine the target model instance. After determining the target model instance, it performs the aforementioned S2 to S5 spatial pose calculations on the target location description to obtain the moved target 3D pose. For inspection, repair, dismantling, acceptance, or status change actions within a command, if the action itself does not involve a change in location, the target model instance can be determined from the object registry or spatial index based on the resource category of the 3D model resource and at least one of the target reference objects or reference sub-regions corresponding to the spatial location description. When the action does not involve a change in location, the world transformation matrix of the target model instance is not changed by the action. For actions that do not involve changes in spatial pose, the corresponding processing can be performed directly based on the determined target model instance, without needing to re-execute the target 3D pose calculation for the action.

[0075] Within the existing candidate model instances, location matching can also be determined based on one or more of the following conditions: whether the bounding box of the candidate instance intersects with the reference sub-region, and whether the candidate instance and the target reference have a specified topological relationship. After comprehensive matching is completed, when the highest comprehensive matching score reaches the target object acceptance threshold, and the difference between its score and the second-ranked candidate reaches the discrimination threshold, the scene instance with the corresponding instanceId is determined as the target model instance; otherwise, the instanceId, location, type, and preview information of the candidate instance are output for administrator confirmation. The target object acceptance threshold and discrimination threshold are both configurable parameters.

[0076] Once an existing model instance is identified, at least one of its world transformation matrix, bounding box, model attributes, or visualization state is updated based on the action performed. For movement or rearrangement actions, the world transformation matrix and bounding box are updated based on the target's 3D pose after the movement. The scene tree, spatial index, or object registry is updated synchronously after each modification. Physical movement or maintenance in the real-world environment is performed by on-site personnel. When on-site personnel confirm that the actual 3D pose differs from the suggested pose, the system can save the deviation information and update the model instance's world transformation matrix based on the confirmed actual 3D pose.

[0077] Further, see Figure 4 After a model instance is created or updated, a structured management behavior record corresponding to that model instance can be generated, and a two-way association can be established between the structured management behavior record and the model instance.

[0078] Structured behavior records can include behavior identifiers (behaviorId), original instruction text, target entity, executed action, participants, target 3D pose, time information, and behavior state. Model instances can include instance identifiers (instanceId), model resource identifiers, spatial transformation parameters, scene state, associated behaviorId, and update time. By writing instanceIds into behavior records and one or more behaviorIds into model instances, a bidirectional index is formed, pointing from structured behavior records to model instances and from model instances to structured behavior records.

[0079] To differentiate between planned changes in a digital twin 3D scene and the actual execution of tasks in the real world, structured management behavior records can be configured with pending status, on-site execution status, and scene operation results. The pending status records task stages such as pending assignment, pending execution, pending acceptance, or closed. The on-site execution status records implementation results such as not executed, executing, successful execution, failed execution, or canceled. Scene operation results record whether model instance creation, attribute updates, or scene index updates were successful. Successful creation of a planned model instance only indicates that the corresponding scene operation has been completed; it does not equate to the completion of the on-site behavior.

[0080] For objects that have not yet been actually installed but only represent pending locations, a planned model instance can be created first. After the on-site execution is completed, the planned model instance can be converted into a physical model instance based on the on-site confirmation results, or merged with the existing model instance confirmed on-site, while maintaining the association between the structured management behavior records and the corresponding model instances.

[0081] When the execution status of a management action, the actual 3D pose of the site confirmation, or the model attributes change, the system generates new status events and appends them without overwriting existing historical events. Status events can trigger updates to the visualized status, business attributes, or world transformation matrix of model instances, ensuring that the status changes in the management action records correspond to the specific model objects in the digital twin 3D scene. Management statuses such as pending, execution, acceptance, and rectification are only used to describe the management process; their technical role is reflected through changes in data records and changes in the scene status of model instances triggered by status events.

[0082] Based on the bidirectional index, when a query is initiated from a model instance in a digital twin 3D scene, one or more behaviorIds associated with it can be obtained according to the instanceId, and the original instructions, executed actions, participants, target 3D pose, actual 3D pose confirmed on-site, constraint verification results and state change information can be read according to the time information of the corresponding state events to form a behavior history corresponding to the model instance.

[0083] When a query is initiated from the structured management behavior record, the corresponding instanceId and status event can be obtained based on the behaviorId, and the corresponding model instance and reference sub-region can be located in the digital twin 3D scene by combining the object registry or spatial index; for a management behavior that is associated with multiple model instances, the corresponding scene objects can be located respectively based on the associated multiple instanceIds.

[0084] When it is necessary to replay the historical scene state, the model position, model attributes and visualization state that have been effective before the target time node can be read in the order of the timestamp of the state event, and the 3D scene state of the target time node can be restored or simulated accordingly. The historical state events are saved in an appending manner so that the current state update does not overwrite the existing event data used for replay.

[0085] The following example illustrates the continuous processing of the above method using a scenario involving the addition of a new fire extinguisher. The project manager inputs via a mobile terminal, "Please have the safety officer place a fire extinguisher at the corner of the stairs on the 3rd floor of Building 2." S1 resolves the target entity as the fire extinguisher, the action as placement, the building as Building 2, the floor as the 3rd floor, the reference point as the stairs, and the relative location as the corner, and then determines the corresponding 3D model resource of the fire extinguisher.

[0086] In S2, the spatial pose calculation module searches for stair components within the three-dimensional spatial knowledge base of the construction site along the inclusion relationship of "Building 2 - Floor 3 - Component", forming a set of candidate reference objects. In S3, when multiple staircases exist on a floor, disambiguation is performed based on the ordinal number, orientation, semantic, and topological information already provided in the instruction; if a staircase can be uniquely identified, the target staircase is determined; if it cannot be uniquely identified, candidate staircases and their matching information are output for management personnel to confirm, and the confirmation result is taken as the target staircase. Subsequently, based on the "corner", the reference sub-region corresponding to the staircase corner is extracted from the geometric boundary, attachable surface, or semantic sub-component of the target staircase.

[0087] In S4, the geometric center of the reference sub-region and the local coordinate system of the target staircase are used as the calculation reference. The orientation offset, local direction and offset distance are determined according to the preset orientation rules corresponding to the "corner". The initial candidate position is obtained through the transformation relationship from the local coordinate system to the engineering world coordinate system. Multiple candidate three-dimensional poses containing position and orientation are generated in the engineering world range or its neighborhood corresponding to the reference sub-region.

[0088] In S5, the geometric information of the target entity for spatial verification is obtained from the 3D model resources of the fire extinguisher. Each candidate 3D pose is then verified to meet hard constraints such as allowable area, attachability, installation location, clearance for passage, and collision or obstruction with existing components, forming a set of feasible poses. Further evaluation is conducted based on soft constraints such as visibility distance, accessibility, offset from the initial candidate position, and impact on passageways. The feasible 3D pose with the highest comprehensive score is determined as the target 3D pose. If no feasible pose exists or the comprehensive evaluation value of the spatial solution is insufficient, a candidate region or candidate pose is output for management personnel to confirm.

[0089] In S6, a model instance is created based on the 3D model resource of the fire extinguisher. The target's 3D pose is written into the instance's world transformation matrix, and the model instance is added to the digital twin 3D scene. The model instance can then represent the planned position to be executed. After the safety officer completes the fire extinguisher configuration on-site, he can submit the execution result or confirm the actual position via a mobile terminal. The system then updates the management behavior record and the scene state of the model instance accordingly.

[0090] In the second embodiment, taking the movement and repair of an existing fire extinguisher as an example, the safety officer inputs, "The fire extinguisher at the corner of the stairs on the 3rd floor of Building 2 has insufficient pressure. Move it to the east side passageway and report it for repair." The target object in this scenario is already registered as a model instance in the object registry. Therefore, creating a new instance is no longer the primary processing objective. Instead, the existing instance to be operated on is first determined, and then the 3D pose of the moved target is calculated.

[0091] Semantic parsing identifies the target entity as a fire extinguisher, with actions including moving and reporting for repair. Spatial descriptions of the original and target locations are extracted. The system filters fire extinguisher scene instances based on model resource categories, then narrows down the candidate pool using reference objects, reference sub-regions, and spatial index information corresponding to the original location. If a candidate cannot be uniquely determined, the administrator confirms the target object from among the candidate instances.

[0092] After determining the target model instance, the same spatial pose calculations as those described in S2 to S5 are performed on the "East Passage Entrance" to obtain the moving target pose that satisfies the existing component geometric relationships and engineering constraints. Subsequently, the world transformation matrix and bounding box of the model instance are updated, and the spatial index is updated synchronously. The repair status is written as a management behavior status event into the structured management behavior record, and the visualization status of the model instance can be further changed, but the aforementioned 3D pose calculation rules remain unchanged.

[0093] Corresponding to the aforementioned method, this embodiment also provides a 3D modeling system for construction management behavior based on a large language model. See [link to documentation]. Figure 5 The system includes a semantic parsing module, a 3D model resource management module, a 3D spatial knowledge base for construction sites, a spatial pose calculation module, and a 3D scene operation module.

[0094] The semantic parsing module is used to obtain natural language management instructions from the construction site, and through a large language model, it obtains the target entity, the action to be performed, and a spatial location description including spatial hierarchy information, reference object information, and relative orientation information. The 3D model resource management module is used to determine the corresponding 3D model resources based on the target entity and can access the 3D model resource library.

[0095] The 3D spatial knowledge base for the construction site is built based on BIM, storing at least the spatial hierarchy information, component geometric information, and component local coordinate systems of the construction site. The spatial pose calculation module obtains candidate reference objects based on the spatial hierarchy information, determines the target reference object and reference sub-region based on the spatial location description, determines local candidate positions using the target reference object's local coordinate system and the reference sub-region as a reference, and transforms them to the engineering world coordinate system. A set of candidate 3D poses is generated within the engineering world range corresponding to the reference sub-region or its neighborhood. Hard constraint verification and soft constraint evaluation are performed based on the target entity's geometric information, existing component geometric information, and engineering constraints corresponding to the 3D model resources to determine the target 3D pose.

[0096] The 3D scene operation module creates or updates a model instance corresponding to the target entity in the digital twin 3D scene based on the executed action, the target's 3D pose, and 3D model resources, and determines the spatial transformation parameters of the model instance based on the target's 3D pose. Data is transferred between modules according to structured instruction objects, 3D model resource information, spatial calculation results, and scene operation results; specific processing rules identical to those described above can be found in the relevant explanations of steps S1 to S6.

[0097] The data flow between modules can be specifically as follows: the semantic parsing module outputs structured instruction objects; the 3D model resource management module outputs model resource identifiers and corresponding resource information; the 3D spatial knowledge base of the construction site provides the spatial pose calculation module with spatial hierarchy information, component geometric information, component local coordinate system, attachable surface information, and spatial relationship data; the spatial pose calculation module outputs target reference objects, reference sub-regions, target 3D pose, constraint verification results, and optional comprehensive evaluation values ​​for spatial calculation; the 3D scene operation module outputs model instance identifiers, spatial transformation parameters, and scene operation results. The above data establishes a correspondence through identifiers such as model resource identifiers, component identifiers, instanceId, and behaviorId.

[0098] In a further embodiment, the system may also include a digital twin construction and visualization interaction module, which is used to bidirectionally associate structured management behavior records, model instances and status events; query the associated behaviorId and status events according to instanceId, or locate the associated model instance according to behaviorId; and restore the model position, model attributes or visualization status of the corresponding time node according to the time information in the status event.

[0099] In one variant, the 3D model resource library, the 3D spatial knowledge base of the construction site, the object registry, and the behavior event library can be deployed on the same server or different data nodes, as long as they can provide the corresponding data to the aforementioned processing module and maintain the correspondence between identifiers such as model resource identifiers, component identifiers, instanceId, and behaviorId.

[0100] This document uses specific embodiments to illustrate the principles and implementation methods of the present invention. The above embodiments are only used to help understand the method and core ideas of the present invention; those skilled in the art can make corresponding adjustments without departing from the technical concept of the present invention, and the content of this specification should not be construed as a limitation on the scope of protection of the present invention.

Claims

1. A three-dimensional modeling method for construction management behavior based on a large language model, characterized in that, include: Natural language management instructions from the construction site are obtained, and the target entity, the action to be executed, and the spatial location description including at least spatial hierarchy information, reference object information and relative orientation information are obtained through large language model parsing. The corresponding three-dimensional model resources are determined based on the target entity. Based on the spatial hierarchy information, a spatial retrieval range is determined in a three-dimensional spatial knowledge base of the construction site constructed based on Building Information Modeling (BIM) and including at least spatial hierarchy information, component geometric information, and component local coordinate system, and candidate reference objects corresponding to the reference object information are obtained from the spatial retrieval range. The candidate reference objects are disambiguated based on the limiting information in the spatial location description to determine the target reference object, and a reference sub-region is determined from the geometric information of the target reference object based on the relative orientation information. For each action in the execution action, determine whether the action needs to change the spatial position of the target entity; for actions that need to create or change the spatial position, based on the local coordinate system of the target reference object and the reference sub-region, convert the relative orientation information into local direction information, determine local candidate positions according to the local direction information and transform them to the engineering world coordinate system, and generate a set of candidate three-dimensional poses based on the transformed candidate positions within the engineering world range or its neighborhood corresponding to the reference sub-region. For actions that require creating or changing spatial positions, hard constraint verification is performed on the candidate 3D pose set based on the target entity geometry information, component geometry information, and engineering constraints corresponding to the 3D model resource to obtain a feasible pose set that satisfies all hard constraints. The feasible pose set is then evaluated based on soft constraints to determine the target 3D pose. For actions that do not involve changes in spatial position, an existing target model instance is determined from the object registry or spatial index based on the target entity and the spatial position description, without re-evaluating the target 3D pose. Based on the sequence of the actions, a model instance corresponding to the target entity is created or updated in the digital twin 3D scene; for actions that require creating or changing the spatial position, the spatial transformation parameters of the model instance are determined based on the target 3D pose; for actions that do not involve changes in spatial position, the model attributes or visualization state of the target model instance are updated, while keeping its spatial transformation parameters unchanged.

2. The three-dimensional modeling method for construction management behavior based on a large language model according to claim 1, characterized in that, The process of parsing the natural language management instructions using a large language model includes: A structured schema is pre-established that defines the target entity, the action to be performed, and the corresponding fields for spatial location description. The action to be performed field is used to represent a single action or multiple actions arranged in the order of instructions. The output of the large language model is subjected to constrained decoding based on the structured pattern, and semantic fragments corresponding to the target entity, the action to be performed, and the spatial location description are extracted from the natural language management instructions. The spatial location description is further analyzed into at least one of building, floor, grid, room, reference object category, ordinal limit, orientation limit, and distance or range; when the execution action includes moving or rearranging, the spatial location description also includes an original location description and a target location description, the original location description is used to locate an existing target model instance, and the target location description is used to determine the three-dimensional pose of the target after the movement; Perform structural compliance and field integrity checks on the parsed results, and generate or output candidates to be confirmed when there are missing fields, field conflicts, or multiple candidate interpretations.

3. The three-dimensional modeling method for construction management behavior based on a large language model according to claim 1, characterized in that, The step of determining the corresponding 3D model resource based on the target entity includes: The target entity and its context are vector-encoded, and vector semantic retrieval is performed in the 3D model resource library to obtain a first resource candidate set; A second set of resource candidates is obtained based on the matching relationship between the target entity and the domain alias; The first resource candidate set and the second resource candidate set are merged and sorted to determine the model resource identifier corresponding to the target entity. The corresponding three-dimensional model file and model resource information for unifying the model coordinate reference are obtained based on the model resource identifier. The model resource information includes at least one of size unit, coordinate origin and default orientation.

4. The three-dimensional modeling method for construction management behavior based on a large language model according to claim 1, characterized in that, The process of obtaining candidate references includes: Based on the spatial inclusion relationship between buildings, floors, axes, or rooms and components, the spatial retrieval range is narrowed level by level according to the spatial hierarchy information; Based on the matching relationship between the reference object information and the component semantic tags, similar components are obtained; By using the geometric bounding boxes of the same type of components to eliminate components located outside the spatial retrieval range, a set of candidate reference objects is obtained, wherein each candidate reference object is associated with at least a component identifier, a geometric center, a geometric bounding box, and a local coordinate system.

5. The three-dimensional modeling method for construction management behavior based on a large language model according to claim 4, characterized in that, The determination of the target reference object and the reference sub-region includes: Based on at least one of the ordinal constraints, relative orientation, semantic tags, and topological relationships in the spatial location description, candidate reference objects are sorted and matched to determine the target reference object from the candidate reference objects; Map the relative orientation to the local coordinate system of the target reference object; Based on the relative orientation, the reference sub-region is determined from the geometric bounding box boundary, component outline edge, attachable surface, or semantic sub-component of the target reference.

6. The three-dimensional modeling method for construction management behavior based on a large language model according to claim 5, characterized in that, The generated candidate 3D pose set includes: Obtain the geometric center of the reference sub-region in the local coordinate system of the target reference object, and determine the local candidate position based on the geometric center, the azimuth offset corresponding to the relative orientation, the local unit direction, and the offset distance; By utilizing the coordinate transformation relationship between the local coordinate system and the engineering world coordinate system of the target reference object, the local candidate position is converted into an initial candidate position in the engineering world coordinate system; The initial orientation is determined based on at least one of the attachment surface normal of the reference sub-region, the local coordinate axis of the target reference object, or the personnel approach direction. Using the initial candidate position as the search center, multiple candidate positions are generated within the engineering world range or its neighborhood corresponding to the reference sub-region, and at least one candidate orientation is generated for each candidate position to form the candidate three-dimensional pose set.

7. The three-dimensional modeling method for construction management behavior based on a large language model according to claim 6, characterized in that, Determining the target's three-dimensional pose includes: For each candidate 3D pose in the candidate 3D pose set, hard constraint verification is performed based on the target entity geometry information corresponding to the 3D model resource, the existing component geometry information, the attachable surface information, and the construction area information in the 3D spatial knowledge base of the construction site. The hard constraints include at least one of the following: installation position constraints, attachability constraints, passage clearance constraints, existing component collision or occlusion constraints, and allowable construction area constraints. The candidate 3D poses that satisfy all the hard constraints are combined into a set of feasible poses; Based on at least one soft constraint among visibility, accessibility, personnel accessibility, maintenance convenience, offset from the initial candidate position, and degree of influence on traffic flow, each feasible 3D pose in the feasible pose set is evaluated, and the feasible 3D pose with the highest comprehensive score is determined as the target 3D pose. If no feasible 3D pose satisfies all the hard constraints, expand the candidate pose search range, switch the attachable surface of the target reference, or output a candidate region to be manually confirmed.

8. The three-dimensional modeling method for construction management behavior based on a large language model according to claim 1, characterized in that: When the execution action is to add, place, install or configure, a model prototype object is generated based on the three-dimensional model resource, and a world transformation matrix of the model instance is constructed based on the target three-dimensional pose. The model instance generated based on the model prototype object is then added to the digital twin three-dimensional scene. When the execution action is to move or rearrange and the spatial location description includes the original location description and the target location description, the target model instance is determined from the object registry or spatial index of the digital twin 3D scene according to the resource category of the 3D model resource and the reference object and reference sub-region corresponding to the original location description. The target 3D pose after movement is determined according to the target location description, and the world transformation matrix and bounding box of the target model instance are updated accordingly. For the inspection, repair, dismantling, acceptance, or status change actions in the execution actions, when the action does not involve changes in spatial location, based on the resource category of the 3D model resource and the spatial location description, an existing target model instance is determined from the object registry or spatial index, and at least one of its model attributes or visualization status is updated, while keeping the world transformation matrix of the target model instance unchanged; after the creation or update of the model instance is completed, the scene tree, spatial index, or object registry of the digital twin 3D scene is updated synchronously.

9. The three-dimensional modeling method for construction management behavior based on a large language model according to any one of claims 1 to 8, characterized in that, Also includes: Generate a structured management behavior record corresponding to the model instance, write the instance identifier of the model instance into the structured management behavior record, and write the corresponding behavior identifier into the model instance to establish a bidirectional association between the structured management behavior record and the model instance; When the execution status of management behavior, actual 3D pose, or model attributes change, a status event is generated, and the scene status of the structured management behavior record and the model instance is updated synchronously according to the status event. When the actual 3D pose confirmed on-site is different from the target 3D pose, the deviation information between the actual 3D pose and the target 3D pose is saved, and the world transformation matrix of the model instance is updated according to the actual 3D pose.

10. A three-dimensional modeling system for construction management behavior based on a large language model, characterized in that, include: The semantic parsing module is used to obtain natural language management instructions from the construction site and to obtain target entities, execution actions, and spatial location descriptions containing spatial hierarchy information, reference information, and relative orientation information through a large language model. The 3D model resource management module is used to determine the corresponding 3D model resources based on the target entity. A three-dimensional spatial knowledge base for the construction site, built on Building Information Modeling (BIM), and storing at least the spatial hierarchy information of the construction site, the geometric information of the components, and the local coordinate system of the components; The spatial pose calculation module is used to obtain candidate reference objects based on the spatial hierarchy information, determine the target reference object and reference sub-region based on the spatial position description, determine the local candidate position using the local coordinate system of the target reference object and the reference sub-region as a spatial reference and transform it to the engineering world coordinate system, generate a set of candidate three-dimensional poses in the engineering world range or its neighborhood corresponding to the reference sub-region, and perform hard constraint verification and soft constraint evaluation on the set of candidate three-dimensional poses based on the target entity geometric information, existing component geometric information and engineering constraints corresponding to the three-dimensional model resources, so as to determine the target three-dimensional pose. The 3D scene operation module is used to create or update a model instance corresponding to the target entity in the digital twin 3D scene according to the execution action and the 3D model resources; for execution actions that require creating or changing the spatial position, the spatial transformation parameters of the model instance are determined according to the target 3D pose; for execution actions that do not involve changes in spatial position, the model attributes or visualization state of the existing target model instance are updated, while keeping its spatial transformation parameters unchanged.