A 3D spatial anchoring visualization method and system for BIM review inference chain

By constructing a directed acyclic graph for reasoning in BIM review and establishing an anchoring relationship with 3D space, the problem of the disconnect between the reasoning process and space in existing BIM review is solved, realizing an intuitive display and transparency of the reasoning process and improving review efficiency.

CN123118931APending Publication Date: 2026-11-17POWERCHINA HUADONG ENG CORP LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611613156.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing BIM review methods, the reasoning process is disconnected from the three-dimensional space and lacks logical connection, resulting in opaque review results. Reasoning steps cannot be accurately anchored in the three-dimensional scene, and the data structure of multi-engine reasoning steps is inconsistent, making it impossible to organize them into a unified, traversable, and navigable reasoning link.

Method used

By unifying the inference results of the rule engine, geometric calculation engine, and large language model into an inference chain, and establishing anchoring relationships with BIM components and spatial coordinates, a directed acyclic graph of inference is constructed, spatial anchoring parameters are calculated, camera flight animation, component visibility switching, and 3D annotation are executed, and inference data and explanatory information are displayed node by node.

Benefits of technology

It realizes the corresponding display of the reasoning process and three-dimensional space, improves the intuitiveness and transparency of the review process, facilitates the understanding and review of reviewers, supports the playback and traceability of the review process, and improves the review efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118931A_ABST
    Figure CN123118931A_ABST
Patent Text Reader

Abstract

This invention discloses a 3D spatial anchoring visualization method and system for BIM review inference chains. The method includes: acquiring the outputs of a rule engine, a geometric calculation engine, and a large language model and encapsulating them into a unified inference node data structure containing component IDs and spatial coordinates; constructing a directed acyclic graph of inference based on the input-output dependencies between inference nodes and generating a topologically ordered hierarchical sequence; calculating spatial anchoring parameters containing camera pose, component visibility status, and 3D annotation information based on the BIM component IDs and spatial coordinate ranges associated with the inference nodes; and performing camera flight, component visibility switching, and annotation emergence node by node in the 3D scene according to the hierarchical sequence, while simultaneously displaying the inference data. This invention achieves a high degree of integration and transparent visualization traceability of complex BIM review logic and 3D geometric space.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a three-dimensional spatial anchoring visualization method and system for BIM review reasoning chain. Background Technology

[0002] A key aspect of quality control in construction projects is the compliance review of BIM (Building Information Modeling) models. Currently, rule-based structured query methods can automatically check the component attributes, quantities, and spatial relationships within BIM models. In recent years, Large Language Models (LLMs) have been introduced into the specification interpretation and review reasoning stages. Their semantic understanding capabilities assist in parsing implicit constraints in specification clauses, and their internal chain-of-thought reasoning mechanism can decompose the judgment logic of complex specification clauses. Simultaneously, a geometric calculation engine handles precise measurements of spatial distances, areas, and volumes. The rule engine, geometric calculation engine, and LLM work collaboratively during the review process, each generating intermediate reasoning results.

[0003] However, existing BIM review methods have some shortcomings:

[0004] First, the review results are disconnected from the reasoning process. Existing review systems only output the final review conclusion (such as the area of ​​a fire compartment not meeting the standard) and mark the location of the non-compliant components in the 3D scene. However, the complete logical chain of reasoning—how the standard clauses are interpreted in a structured way, how model data is queried and extracted, how intermediate geometric calculations are performed, and how conditions are compared item by item—is completely hidden within the system. Reviewers are faced with the conclusion and the 3D model, but there is a lack of logical connection between the two. They cannot see in 3D space why the conclusion is valid or what reasoning steps were involved.

[0005] Second, the reasoning steps are disconnected from the three-dimensional spatial location. Even if some review systems can overlay review result annotations in a three-dimensional scene, what they present is a spatial visualization of the review conclusion, that is, displaying a label at the three-dimensional location of the problematic component, rather than a spatial visualization of the review reasoning process. Reviewers cannot see: which spatial range of the model corresponds to the interpretation of a certain standard, where the geometric start and end points of the area calculation are, where the spatial boundaries involved in the condition comparison are, and what specific spatial area the rectification suggestions point to. Each reasoning step lacks precise anchoring in three-dimensional space.

[0006] Third, the data structures for the reasoning steps of the multi-engine are not uniform. The output formats of the thought chain reasoning (natural language form) of the large language model, the query reasoning (structured query statement form) of the rule engine, and the quantitative reasoning (numerical calculation form) of the geometric calculation engine are different. They cannot be uniformly organized into a traversable and navigable reasoning link, and they cannot be presented step by step in the order of reasoning logic in a unified three-dimensional scene.

[0007] In summary, using the BIM model merely as a basis for review conclusions, rather than a spatial carrier for the reasoning process, leads to a lack of precise coordinate anchoring at each step of the review reasoning chain to the BIM 3D space. This prevents reviewers from tracing evidence layer by layer along the reasoning chain within the 3D scene. Summary of the Invention

[0008] Purpose of the invention: This invention provides a three-dimensional spatial anchoring visualization method and system for BIM review reasoning chain, in order to solve the problems of opaque logic in the existing BIM review process, decoupling of the reasoning chain from three-dimensional spatial geometry, and difficulty in intuitively tracing the review process.

[0009] The three-dimensional spatial anchoring visualization method of the present invention includes: During the review process, the input data, output data and logical dependencies of the reasoning steps are obtained from the outputs of the three types of execution engines: rule engine, geometric calculation engine and large language model. Each reasoning operation is encapsulated into a reasoning node data structure that includes node type, execution engine identifier, upstream dependent node ID list, associated BIM component ID list and three-dimensional spatial coordinate range. Based on the input-output dependencies between inference nodes, all inference nodes are constructed into a directed acyclic graph (DAG), and the DAG is topologically sorted to generate a sequence of inference levels. Based on the BIM component ID list and three-dimensional spatial coordinate range associated with each inference node in the inference directed acyclic graph, the spatial anchoring parameters of the inference node in the three-dimensional scene are calculated. The spatial anchoring parameters include the target camera pose, the set of target component visibility states, the spatial position of the three-dimensional annotation information, and the display content. Based on the spatial anchoring parameters, the current inference node executes camera flight animation, component visibility switching, and 3D annotation emergence in the 3D scene, and displays the inference data and explanation information of the current inference node node by node according to the inference hierarchy sequence.

[0010] Preferably, the method further includes: When the inference node is a canonical interpretation node, the spatial anchoring position is determined as a panoramic view of the model space range where the object under review is located; When the inference node is a model data extraction node, the spatial anchoring position is determined as the highlighted state of the bounding box of the target component of the extracted data; When the inference node is a geometric calculation node, the spatial anchoring position is determined as the visualization of the start and end points, boundary surfaces and intermediate geometries involved in the calculation in the three-dimensional scene; When the inference node is a condition comparison node, the spatial anchoring position is determined as the location of the component that does not meet the condition and a warning is displayed. When the inference node is a compliance determination node, the spatial anchoring position is determined as the three-dimensional position of the non-compliant component and the determination conclusion is superimposed.

[0011] Preferably, the execution order of the spatial anchoring parameters is as follows: first, the component visibility switching operation is performed in the three-dimensional scene; then, the camera flight animation is driven to the target camera pose; finally, the three-dimensional annotation information appears in the three-dimensional scene and the inference data and explanation information are displayed in the side information panel.

[0012] Preferably, it also includes: while displaying the current inference node, receiving a follow-up question instruction initiated by the user for the current inference node, combining the context data of the current inference node with the follow-up question text into prompt words and calling a large language model to generate supplementary explanations, while highlighting and flashing the positioning line or geometric coordinates of the component targeted by the follow-up question in the three-dimensional scene. After the review process is completed, the user's browsing path along the reasoning hierarchy sequence in the 3D scene, the dwell time at each reasoning node, follow-up questions, and manual annotations are recorded as a structured audit log, and the review process is replayed using the structured audit log.

[0013] Preferably, when displaying inference nodes in a directed acyclic graph in a 3D scene, different colors are used to distinguish rule engine nodes, geometric calculation engine nodes, and large language model nodes based on the execution engine identifier of each inference node.

[0014] Preferably, the method further includes: persistently storing the structured data of the inference directed acyclic graph in a database, wherein the inference node table is used to record the node type, execution engine identifier, upstream and downstream dependencies, spatial anchoring parameters and canonical clause references of the inference nodes, and the inference edge table is used to record the dependency relationship types between inference nodes.

[0015] The three-dimensional spatial anchoring visualization system of the present invention includes: The review reasoning chain structure module is used to obtain the input data, output data and logical dependencies of the reasoning steps from the outputs of the three types of execution engines: rule engine, geometric calculation engine and large language model during the review process. Each reasoning operation is encapsulated into a reasoning node data structure that includes node type, execution engine identifier, upstream dependent node ID list, associated BIM component ID list and three-dimensional spatial coordinate range. The inference DAG construction module is used to construct a directed acyclic graph of inference nodes based on the input-output dependencies between inference nodes, and to perform topological sorting on the inference DAG to generate an inference hierarchy sequence. The three-dimensional spatial anchoring mapping module is used to calculate the spatial anchoring parameters of the inference node in the three-dimensional scene based on the BIM component ID list and three-dimensional spatial coordinate range associated with each inference node in the inference directed acyclic graph. The spatial anchoring parameters include the target camera pose, the target component visibility state set, the spatial position of the three-dimensional annotation information, and the display content. The inference chain display module is used to execute camera flight animation, component visibility switching and 3D annotation emergence in the 3D scene according to the spatial anchoring parameters, and display the inference data and explanation information of the current inference node node by node according to the inference hierarchy sequence.

[0016] Preferably, it also includes a large model enhancement and explanation module, which is used to receive follow-up questions initiated by the user for the current inference node when the current inference node is displayed, combine the context data of the current inference node with the follow-up question text into prompt words and call the large language model to generate supplementary explanations, and at the same time highlight and flash the positioning lines or geometric coordinates of the component targeted by the follow-up question in the three-dimensional scene.

[0017] Preferably, it also includes an audit traceability module, which is used to record the user's browsing path along the reasoning hierarchy sequence in the 3D scene, the dwell time of each reasoning node, follow-up questions, and manual annotations as a structured audit log after the review process is completed, and to use the structured audit log to replay the review process.

[0018] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the three-dimensional spatial anchoring visualization method.

[0019] Compared with existing technologies, this invention has the following significant advantages: 1. By unifying the reasoning results of the rule engine, geometric calculation engine, and large language model into a reasoning chain and establishing an anchoring relationship with BIM components and spatial coordinates, the reasoning process is displayed in a corresponding three-dimensional space, improving the intuitiveness and transparency of the review process; 2. By linking the camera, component visibility, and 3D annotation according to the reasoning sequence, the review basis and spatial evidence are presented step by step, facilitating the understanding and review of reviewers; 3. Through interactive questioning and audit log recording, the review process is replayable and traceable, improving the efficiency of review and verification. Attached Figure Description

[0020] Figure 1 This is the overall system architecture diagram; among which, Figure 1 The solid arrows in the text represent data flow. Figure 1 The dashed arrows in the text represent feedback / requests;

[0021] Figure 2 This is a schematic diagram of the inference node data structure; where, Figure 2 The solid arrows in the text indicate data dependencies. Figure 2 The dashed arrows in the text indicate persistent storage;

[0022] Figure 3 This is a flowchart of the structured audit log playback process; among which, Figure 3 The dashed arrows in the diagram represent mappings, the dashed boxes represent calculation processes, the gray boxes represent input data, and the white boxes represent data entities.

[0023] Figure 4 It is a progressive display of the control flow diagram of the inference chain; among which, Figure 4 The solid arrows in the diagram represent control flow. Figure 4 In the diagram, dashed arrows indicate a return, ellipses indicate the start and end points, rectangles indicate processing, and diamonds indicate judgments.

[0024] Figure 5 This is a schematic diagram illustrating the step-by-step display of the reasoning chain in a 3D scene;

[0025] Figure 6 It is a user interaction flowchart; among which, Figure 6 Solid arrows in the text represent instructions. Figure 1 The dashed arrows in the text represent data;

[0026] Figure 7 It is a diagram showing the mapping relationship between inference nodes and 3D spatial anchoring; among which, Figure 7 The solid arrows in the diagram represent data flow. Detailed Implementation

[0027] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0028] Example 1

[0029] This embodiment provides a three-dimensional spatial anchoring visualization method for BIM review inference chain, including the following steps:

[0030] S1. During the review process, the input data, output data, and logical dependencies of the reasoning steps are obtained from the outputs of the three types of execution engines: rule engine, geometric calculation engine, and large language model. Each reasoning operation is encapsulated into a reasoning node data structure that includes node type, execution engine identifier, list of upstream dependent node IDs, list of associated BIM component IDs, and three-dimensional spatial coordinate range.

[0031] S101. Load the BIM model data to be reviewed and the corresponding review rule package, and start the review engine; the review engine includes three types of execution engines: rule engine, geometric calculation engine and large language model (LLM).

[0032] S102. During the review process, obtain the input data, output data, and logical dependencies of each inference operation in the output of each execution engine, and encapsulate each inference operation into a unified inference node data structure.

[0033] The inference node data structure should include at least the following: node type (standard interpretation node, model data extraction node, geometric calculation node, condition comparison node, compliance judgment node), the execution engine identifier that generated the node (rule engine identifier, geometric calculation engine identifier, large language model identifier), a list of upstream dependent node IDs, the output data of this node, a list of associated BIM component IDs, and the three-dimensional spatial coordinate range.

[0034] Through a unified reasoning node data structure, reviewers can see the full picture of collaborative reasoning of the three types of engines along the reasoning chain in a three-dimensional scene, which step is executed by which engine, and the corresponding evidence of the intermediate results of each engine in three-dimensional space, rather than just the thought chain text of the large language model.

[0035] S2. Based on the input-output dependencies between inference nodes, construct all inference nodes into a directed acyclic graph (DAG), and perform topological sorting on the DAG to generate an inference hierarchy sequence.

[0036] Specifically, based on the input-output dependencies between inference nodes, all inference nodes are constructed into a directed acyclic graph (DAG) for inference. In the node dependency construction, if the input of node B depends on the output of node A, a directed edge (A→B) is established from node A to node B, making the initial specification clause node the source node and the final review conclusion node the convergence node. Finally, the inference DAG is topologically sorted to generate a sequence of inference levels.

[0037] S3. Based on the BIM component ID list and 3D spatial coordinate range associated with each inference node in the directed acyclic graph, calculate the spatial anchoring parameters of the inference node in the 3D scene. These parameters include the target camera pose (viewpoint 3D coordinates, viewing direction vector, field of view), the target component visibility / hidden state set, the spatial location of the 3D annotation information, and the displayed content. The target component visibility / hidden state set specifies which components are fully displayed, which are semi-transparent (transparency is configurable), and which are completely hidden when displaying the current node. The displayed content of the 3D annotation information includes annotation text, guide lines, and the 3D coordinates of the annotation anchor points.

[0038] Furthermore, differentiated spatial anchoring mapping rules are adopted for inference nodes of different node types: When the inference node is a specification interpretation node, the spatial anchoring position is determined as a panoramic view of the model space range where the object under review is located; when the inference node is a model data extraction node, the spatial anchoring position is determined as the target component of the extracted data, and the component bounding box is highlighted; when the inference node is a geometric calculation node, the spatial anchoring position is determined as the start and end points, boundary surfaces, and intermediate geometries involved in the calculation, and their visualization in the 3D scene; when the inference node is a condition comparison node, the spatial anchoring position is determined as the location of the component that does not meet the conditions, and it is highlighted with a warning color; when the inference node is a compliance judgment node, the spatial anchoring position is determined as the 3D location of the non-compliant component and the judgment conclusion is superimposed.

[0039] For example, taking the composite spatial anchoring mapping rule that “the horizontal distance between the door and window openings on both sides of the firewall should not be less than 2.0m” as an example, this rule involves both semantic understanding (definition of firewall, door and window openings, and horizontal distance) and precise geometric calculation (distance between the nearest point pairs of the projection polygon of the opening).

[0040] N1 (Specification Interpretation Node): The large language model parses the semantics of the clauses. Firewalls are non-combustible walls with a fire resistance rating of ≥3.00h. Door and window openings include doors and windows. Horizontal distance is the shortest Euclidean distance between the bottom projection polygons of two openings. Output structured constraints. The anchoring strategy is a panoramic view of the entire building.

[0041] N2 (Model Data Extraction Node): Query all entities of type IfcWall = firewall, obtain the IfcWindow and IfcDoor on each firewall through the IfcRelVoidsElement relationship, and extract the bottom projection geometry of the opening. Anchoring strategy: All firewalls are highlighted with the first color, and the openings on them are marked with the second color.

[0042] N3 (Geometric Calculation Node): For each pair of projected polygons on the bottom surface of the openings, the nearest point pair algorithm is used. First, candidate pairs with a distance ≥2.0m from the bounding rectangle are eliminated. Then, the precise shortest distance is calculated for each edge of the remaining candidate pairs. Anchoring Strategy: For pairs of openings with a distance <2.0m, the shortest distance line between the two openings is drawn in the 3D scene and the distance value is labeled.

[0043] N4 (Condition Comparison Node): Compares the distance between each pair of openings with a 2.0m threshold to filter out pairs of illegal openings. The anchoring strategy is to highlight illegal opening pairs with a warning color border and mark them as: distance X.Xm < 2.0m.

[0044] N5 (Compliance Judgment Node): The large language model generates a description and rectification suggestions for each violation. Anchoring Strategy: Spatial annotations are drawn by combining the location of the violation and the suggested direction of movement.

[0045] The five nodes N1-N5 constitute a hybrid reasoning DAG, displayed sequentially in a 3D scene. Each node is generated by a large language model, a rule engine, a geometric calculation engine, another rule engine, and another large language model, with different colors used to distinguish engine types within the scene. Reviewers can follow the reasoning chain to see: how the large language model understands canonical semantics → how the rule engine queries components → how the geometric calculation engine measures distances → how the rule engine compares thresholds → how the large language model synthesizes and generates suggestions—each step corresponding to a specific component or geometric element in 3D space.

[0046] S4. Based on the spatial anchoring parameters, the current inference node executes camera flight animation, component visibility switching, and 3D annotation emergence in the 3D scene, and displays the inference data and explanation information of the current inference node node by node according to the inference hierarchy sequence.

[0047] Specifically, for the current inference node, based on its spatial anchoring parameters in the 3D scene, the component visibility switching and sectioning operations are first performed, then the camera flight animation is executed to the target pose, and finally the 3D annotation information and side information panel appear. At the same time, according to the inference hierarchy sequence, node by node, the information panel displays the normative clause reference, input / output data, and LLM-generated natural language explanation text for that inference node. The natural language explanation text is presented with a character-by-character typewriter effect.

[0048] S5. While displaying the current inference node, receive follow-up questions initiated by the user regarding the current inference node, combine the context data of the current inference node with the follow-up text to form prompt words, and call the large language model to generate supplementary explanations. At the same time, highlight and flash the positioning lines or geometric coordinates of the component targeted by the follow-up question in the 3D scene front-end (WebGL).

[0049] Specifically, the follow-up instructions include moving forward / backward / jumping along the inference chain to a specified node, and initiating follow-up questions on the large language model at the current node.

[0050] S6. After the review process is completed, the large language model records the user's browsing path along the reasoning hierarchy sequence in the 3D scene, the dwell time at each reasoning node, follow-up questions, and manual annotations as a structured audit log, and uses the structured audit log to replay the review process. Each review conclusion entry in the structured audit log includes a jump link; clicking it automatically switches the 3D scene to the spatial anchoring view of the reasoning node corresponding to that conclusion.

[0051] Example 2

[0052] This embodiment provides a 3D spatial anchoring visualization system for BIM review reasoning chains. It can be broadly divided into an execution engine layer, a core processing layer, and a display and interaction layer. Specifically, it includes a review reasoning chain structuring module, a reasoning DAG construction module, a 3D spatial anchoring mapping module, a reasoning chain display module, a large model enhancement and explanation module, and an audit traceability module, and can be configured with a side information panel. (Refer to...) Figure 1 During the compliance review process, the system establishes interfaces and data flow connections with multi-source execution engines, which include three types: rule engines, geometric calculation engines, and large language models (LLM). The rule engine is mainly responsible for performing deterministic attribute queries and logical comparisons, the geometric calculation engine is responsible for the accurate measurement of spatial geometric distances, areas, volumes, etc. in the BIM model, and the large language model is responsible for the semantic understanding of the specification clauses, the parsing of implicit constraints, and the decomposition of reasoning logic.

[0053] The core processing layer mainly performs multi-engine review extraction and structuring, inference DAG construction, 3D spatial anchoring mapping, progressive display of inference chain, LLM-enhanced explanation and interaction, and review tracing and playback.

[0054] During the review engine's operation, the structured review inference chain module intercepts the input data, output data, and logical dependencies of each inference operation from each execution engine in real time. It then encapsulates each inference operation into a standardized inference node data structure, achieving standardized organization of the intermediate derivation process across heterogeneous execution engines. Each inference node data structure can include: node type (specification interpretation node, model data extraction node, geometric calculation node, condition comparison node, compliance judgment node), the identifier of the execution engine that generated the node (rule engine identifier, geometric calculation engine identifier, large language model identifier), a list of upstream dependent node IDs, the node's output data, a list of associated BIM component IDs, and the 3D spatial coordinate range.

[0055] The inference DAG construction module constructs a directed acyclic graph (DAG) for all inference nodes based on the input-output dependencies between them. In node dependency construction, if the input of node B depends on the output of node A, a directed edge (A→B) is created from node A to node B, making the initial specification clause node the source node and the final review conclusion node the sink node. The inference DAG construction module further performs topological sorting on the inference DAG to generate a sequence of inference levels.

[0056] To provide a standardized data structure for knowledge accumulation during review and support cross-project retrieval, statistical analysis, and rule optimization, the system stores the structured data of the constructed inference directed acyclic graph (DAG) in a database. The database design primarily includes an inference node table, an inference edge table, and an inference DAG table. The inference node table records fields such as node type, execution engine identifier, upstream and downstream dependencies, spatial anchoring parameters, reference information to regulatory clauses, large language model interpretation text, and confidence level for each inference node. The inference edge table records the dependency types between inference nodes, which can include data dependencies, control dependencies, and semantic dependencies. The inference DAG table stores the review task identifier, model identifier, rule package identifier, and the node hierarchy sequence generated by topological sorting.

[0057] For example, the core fields of the reasoning_nodes table include: id (node ​​identifier, UUID primary key), dag_id (directed acyclic graph identifier, foreign key), node_type (node ​​type, enumeration: specification interpretation / model data extraction / geometric calculation / condition comparison / compliance judgment / result synthesis), engine (engine type, enumeration: rule engine / geometric calculation engine / LLM), input_refs (JSON array, upstream node IDs), output_data (JSON object), spatial_anchor (JSON object, containing a list of component IDs, 3D coordinate range, camera pose parameters, component explicit / implicit matrix, and annotation information list), regulation_ref (JSON object, specification reference, specification number, article section, and original text), llm_explanation (text), confidence (confidence, floating-point number), and created_at (creation time, timestamp). The reasoning edge table (reasoning_edges) includes id (edge ​​identifier), dag_id (directed acyclic graph identifier), from_node_id (starting node identifier), to_node_id (target node identifier), and dependency_type (dependency type: data dependency / control dependency / semantic dependency). The reasoning dag graph (reasoning_dags) includes id (graph identifier), review_task_id (review task identifier), model_id (model identifier), rule_package_id, topological_order (topological order, JSON array), root_nodes (root node set, JSON array), sink_nodes (sink node set, JSON array), and timestamps. This storage structure allows each inference chain and its 3D spatial anchoring information to be persistently retrieved, statistically analyzed, and reused across projects.

[0058] The 3D spatial anchoring mapping module calculates specific spatial anchoring parameters for each inference node in the directed acyclic graph (DAG) to establish a precise correspondence between the inference logic and the BIM 3D spatial geometric coordinates. Specifically, based on the BIM component ID list and 3D spatial coordinate range associated with the inference node, the module calculates spatial anchoring parameters including the target camera pose (viewpoint 3D coordinates, viewing direction vector, field of view, etc.), the target component visibility / hidden state set (fully displayed, semi-transparent with specified transparency, or completely hidden), and the spatial location and display content of the 3D annotation information (annotation text, guide lines, and 3D coordinates of annotation anchor points). Different spatial anchoring mapping rules are used for inference nodes of different node types.

[0059] When the inference node is a specification interpretation node, the spatial anchoring position is determined as a panoramic view of the model space range where the object under review is located; when the inference node is a model data extraction node, the spatial anchoring position is determined as the highlighted bounding box of the target component of the extracted data; when the inference node is a geometric calculation node, the spatial anchoring position is determined as the visualization of the start and end points, boundary surfaces, or intermediate geometries involved in the calculation in the 3D scene; when the inference node is a condition comparison node, the spatial anchoring position is determined as a warning display of the location of components that do not meet the conditions; when the inference node is a compliance judgment node, the spatial anchoring position is determined as the 3D location of the non-compliant component with an overlay of the judgment conclusion label. Furthermore, to intuitively distinguish the multi-engine collaborative inference process, when displaying the inference nodes in the inference DAG in the 3D scene, different colors can be used to distinguish rule engine nodes, geometric calculation engine nodes, and large language model nodes based on the execution engine identifier of each inference node. For example, rule engine nodes can use the first color identifier, geometric calculation engine nodes can use the second color identifier, and large language model nodes can use the third color identifier.

[0060] After determining the spatial anchoring parameters for each inference node, the inference chain display module can employ a specific control flow and execution sequence when applying the spatial anchoring parameters to the current inference node: First, it performs a component visibility switching operation (and a sectioning operation for specific spatial structures) in the 3D scene, making the target component prominent and hiding or semi-transparently occluding it; then, it drives the camera flight animation to the target camera pose, smoothly transitioning to the most suitable perspective for observing the evidence of this inference step; finally, it displays 3D annotation information in the 3D scene and simultaneously displays the inference data and explanation information (references to standard provisions, input / output data, and natural language explanation text generated by the large language model) of the inference node in the side information panel. This strict execution sequence allows reviewers to smoothly browse the inference hierarchy node by node in the 3D scene, intuitively observing the corresponding evidence and evolutionary state of each inference node in the BIM 3D space.

[0061] In some embodiments, the system, through the cooperation of the inference chain display module and the large model enhancement and interpretation module, receives and responds to various interactive commands initiated by the user in the 3D scene, achieving efficient navigation and in-depth questioning of the directed acyclic graph of inference. User interactive commands may include forward, backward, jump to a specified inference node, and questioning commands initiated at the current inference node. When the user clicks the forward or backward button in the 3D scene, the system can automatically obtain the next or previous inference node in the inference hierarchy sequence generated by topological sorting, and smoothly update the display content of the 3D scene view and side information panel according to the preset execution order. When the user initiates a jump command by clicking the node tree or view list in the graphical interface, the three-inference chain display module can quickly retrieve the spatial anchoring parameters corresponding to the target inference node, drive the camera to fly directly to the preset camera pose of the target inference node, and simultaneously complete the switching of the target component's visible / hidden state set and the emergence of 3D annotation information. In particular, in the display state of the current inference node, if the reviewer has questions about the derivation process or intermediate data of the current node, they can initiate questioning commands for the current inference node. Upon receiving a follow-up question instruction, the system can extract the contextual data of the current inference node (the associated BIM component ID list, the range of three-dimensional spatial coordinates, the input and output values, and the corresponding references to standard clauses), and combine this contextual data with the follow-up question text entered by the user to encapsulate it into structured prompt words. Then, it calls a large language model to perform in-depth analysis and generate targeted natural language supplementary explanations.

[0062] During the process of reviewers browsing and interacting with the 3D scene along the reasoning hierarchy, the system collects and structures the user's complete browsing path along the reasoning hierarchy, the time spent at each reasoning node, the initiated follow-up questions, and manually added annotations or markings in real time, automatically recording this as a structured review audit log. When subsequent quality review, dispute arbitration, or experience review is required, reviewers can load the specified structured review audit log and initiate the review process replay function. In the replay process, the audit traceability module parses the timestamps and operation sequences in the audit log, reconstructing the reviewer's 3D perspective switching trajectory, component highlighting status changes, reasoning data and supplementary explanations from the large language model displayed in the side information panel, as well as follow-up question interactions and flashing prompts in a timeline format. The audit traceability module provides a verifiable technical means for review quality review, making the reasoning process auditable and disputes traceable.

[0063] Example 3

[0064] This example uses the review of fire compartment area in a building project's BIM model as a scenario. The BIM model to be reviewed (IFC format) contains three fire compartments, and the loaded review rule package includes a maximum allowable building area of ​​2500m² for fire compartments of fire resistance ratings of Class I and II. 2 Equal constraint rules.

[0065] S1. Load the BIM model and review rules, and start the review engine. The rule engine, geometric calculation engine, and large language model work together to perform the review, extracting reasoning steps from the output of each engine and encapsulating them into a unified reasoning node.

[0066] Reference Figure 2 In this embodiment, 5 inference nodes are generated:

[0067] N1 (Specification Interpretation Node): LLM parses specification clauses and identifies key constraint entities (fire resistance rating equal to Class I, constraint attribute equal to area, operators = and ≤, threshold equal to 2500m). 2 Output structured constraints {fire compartment, area ≤ 2500m²} 2}. Related components: No specific components; spatial coordinate range = overall bounding box of the building model. Anchoring strategy: panoramic top view.

[0068] N2 (Model Data Extraction Node): The rule engine performs an IFC spatial query on the BIM model, traverses IfcSpace entities, filters spaces whose space type equals fire compartment and whose fire resistance rating equals Level 1, and returns a list of 12 boundary wall IDs for fire compartment 3. Associated Components: The IfcUniqueId of the walls [W001...W012] in compartment 3, with spatial coordinates equal to the union of the bounding boxes of each wall. Anchoring Strategy: When the camera is moved to close-up of compartment 3, the 12 boundary walls are highlighted sequentially with yellow borders.

[0069] N3 (Geometric Calculation Node): The geometric calculation engine obtains the projection of the bottom surface of the walls in partition 3, enclosing them as a closed polygon, and calculates the area of ​​the bottom polygon using the shoelace formula. Calculation result: Bottom surface 56.0m × 50.0m, area equal to 2800.00m². 2 Related component: Bottom surface IfcSlab-003 (ID=3kL9mN8pQ2rS), spatial coordinate range = X: 8.0~64.0, Y: 0.0~50.0, Z: 0.0~0.3, bounding box = 56.0×50.0×0.3m. Anchoring strategy is geometric calculation class → spatial geometric elements + start and end labels, i.e.: the camera is raised to a vertical top view directly above the bottom surface (36,25,30), the bottom surface is covered by the blue plane, label: 56.0m×50.0m=2800.00m 2 .

[0070] N4 (Conditional Comparison Node): Reference Figure 5 The rules engine will output an area of ​​2800m² for N3. 2 With N1 output threshold 2500m 2 The comparison showed that the requirement was not met, exceeding the limit by 300m. 2 (12%). Related components: Overall spatial range of Zone 3. Anchoring strategy: Camera lowered to eye level (36, 25, 6), bottom surface with red semi-transparent overlay and pulse warning animation, center 3D label: 2800m 2 >2500m 2 Exceeding 300m 2 (12%), warning of exceeding the standard. It should be noted that the spatial anchoring parameters for each step are as follows: Figure 7 The mapping rules are pre-calculated, and the progressive display modules drive the scene transitions in sequence.

[0071] N5 (Compliance Judgment Node): The large language model integrates information from N1-N4 and analyzes the reasons for exceeding the limit, namely, the excessive extension of the southwest corner boundary line of the zone and the inclusion of the corridor area into zone 3. It is recommended that the southwest corner corridor area be included in the adjacent fire compartment 2 (adjusted area approximately 2380m²). 2Related components: Zone 3 and suggested adjustment area. Anchoring strategy: Zoom out the camera (40, 25, 20), Zone 3 is marked in red, and the suggested adjustment area is marked in yellow dashed box.

[0072] S2. Construct a directed acyclic graph of N1-N5 according to their dependencies: N1→N2→N3→N4→N5, and the topological sorting yields [N1,N2,N3,N4,N5].

[0073] S3. Calculate and output the spatial anchoring parameters for each inference node. The anchoring mapping calculation process is as follows: Component bounding box → Camera pose C=Center(BBox), d=max(L,WH)x1.5; Component ID → Vis={id→1.0lidET)U{→0.2lidtT); Coordinate range →> 3D annotation position Pos=(Xc,Yc,Zmax+0.5), guide line]; Component boundary → Section plane parameters Z_plane=Zmax+t,Normal=(0,0,-1); Node type → Anchoring strategy Strategy=Lookup(node_type).

[0074] Taking N3 as an example, refer to Figure 7 Node type → Anchoring strategy: Specification interpretation → Panoramic view; Model extraction → Bounding box highlighting; Geometric calculation → Geometric elements + start and end labels; Condition comparison → Violation warning; Compliance judgment → Precise anchoring + judgment labeling.

[0075] Output target camera pose: viewpoint (36,25,30), viewing direction pointing to the center of the bottom surface (36,25,0), FOV=60°, animation is smooth flight; component visibility status: fully display the bottom surface IfcSlab-003 with blue highlight, irrelevant components are semi-transparent with alpha=0.2, roof / facade is hidden; 3D annotation: position (36,25,0.5), text "56.0×50.0=2800.00m" 2 "Blue background with white text and guide lines; horizontal section plane Z=0.5, exposing bottom projection; strategy is geometric calculation class → spatial geometric elements with start and end labels."

[0076] S4, Reference Figure 6The system is displayed node by node in a 3D scene according to the reasoning hierarchy. After the user clicks "Start Review," the following steps are observed: N1 → The camera flies to a panoramic overhead view (50, 30, 40), and the original text of the specification is displayed on the reasoning chain navigation panel, with the LLM explanation presented word by word; N2 → The camera moves to a close-up view of partition 3 (20, 30, 12), and the yellow borders of the 12 boundary walls flash sequentially; N3 → The camera rises to a top-down view (36, 25, 30), with the bottom surface covered in blue, and dimension labels and area tags appearing; N4 → The camera descends to (36, 25, 6), and the bottom surface changes from red to semi-transparent with a pulse animation, displaying the excess amount label; N5 → The camera zooms out to (40, 25, 20), and rectification suggestions are displayed with red markings and yellow dashed boxes. The spatial anchoring switching of each node (camera pose, component visibility, and label content) allows the reviewer to gradually see the spatial evidence of each reasoning step in 3D space.

[0077] Reference Figure 4 The progressive display control flow of the inference chain is as follows: Load the DAG and its topology sequence, initialize index i=0 and set N to the total number of DAG nodes; determine if i is less than N. If not, the process ends; if so, read the DAG[i] node, query the spatial anchoring parameters, and execute scene transition animations such as component visibility, camera flight, and annotation appearance. After displaying the information panel, wait for user interaction. Users can choose three types of index change operations: forward (i increments), backward (i decrements), and jump (i is assigned the value N). After the operation, a boundary check is performed (i is set to 0 when i < 0, and the process ends when i ≥ N). After updating the index, the process returns to the node judgment stage. Users can also choose to ask follow-up questions, enter the follow-up question sub-process to complete Prompt splicing, LLM call generation, panel and scene update. After the follow-up question is completed, the index i is not modified, and the process returns to wait for user interaction. It should be noted that after forward / backward / jump, a boundary check is performed (i < 0 → 0, i ≥ N → end); the follow-up question does not change i, and the process returns to wait for interaction after completion.

[0078] S5. User Follow-up Question. At node N3, the reviewer inputs "Start and end points of the bottom surface length of partition 3?". The system concatenates the N3 context (associated component geometric data) with the question into a Prompt, calls the large language model, generates supplementary explanations, and highlights the positioning lines and Y-coordinate labels of W003 (north exterior wall, Y=56.0) and W008 (south exterior wall, Y=0.0) in the 3D scene. Input "Area calculation process?", the answer is "Shoelace formula, 12 vertices → 2800m".

[0079] S6. Record the complete operation sequence. The audit log includes timestamp, node ID, operation type (forward / backward / jump / query), query text, follow-up text, return text, manual annotations, and large language model answers. It is stored in JSON format and supports complete playback.

[0080] Reference Figure 3 The review process involves N1 specification interpretation (15s), N2 model extraction (20s), N3 geometric calculation (45s), query 1, N4 condition comparison (12s), and N5 compliance determination (30s). Real-time recording is then performed to generate an audit log, including timestamps (14:05:32, 14:06:18, 14:06:30), nodes, operation types, query content (start and end points of bottom surface length), LLM return (W003→W008(56→0)), and annotations (partition 3 boundary lines). Replay control includes rewind, pause, fast forward, 1×, and jump to N:[jump], driving replay display, including scene view reproduction, component visibility reproduction, annotation content reproduction, and query answer reproduction. This can be applied to review quality verification, dispute tracing, training, and system debugging.

[0081] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A three-dimensional spatial anchoring visualization method for BIM review reasoning chain, characterized in that, include: During the review process, the input data, output data and logical dependencies of the reasoning steps are obtained from the outputs of the three types of execution engines: rule engine, geometric calculation engine and large language model. Each reasoning operation is encapsulated into a reasoning node data structure that includes node type, execution engine identifier, upstream dependent node ID list, associated BIM component ID list and three-dimensional spatial coordinate range. Based on the input-output dependencies between inference nodes, all inference nodes are constructed into a directed acyclic graph (DAG), and the DAG is topologically sorted to generate a sequence of inference levels. Based on the BIM component ID list and three-dimensional spatial coordinate range associated with each inference node in the inference directed acyclic graph, the spatial anchoring parameters of the inference node in the three-dimensional scene are calculated. The spatial anchoring parameters include the target camera pose, the set of target component visibility states, the spatial position of the three-dimensional annotation information, and the display content. Based on the spatial anchoring parameters, the current inference node executes camera flight animation, component visibility switching, and 3D annotation emergence in the 3D scene, and displays the inference data and explanation information of the current inference node node by node according to the inference hierarchy sequence.

2. The three-dimensional spatial anchoring visualization method according to claim 1, characterized in that, The method further includes: When the inference node is a canonical interpretation node, the spatial anchoring position is determined as a panoramic view of the model space range where the object under review is located; When the inference node is a model data extraction node, the spatial anchoring position is determined as the highlighted state of the bounding box of the target component of the extracted data; When the inference node is a geometric calculation node, the spatial anchoring position is determined as the visualization of the start and end points, boundary surfaces and intermediate geometries involved in the calculation in the three-dimensional scene; When the inference node is a condition comparison node, the spatial anchoring position is determined as the location of the component that does not meet the condition and a warning is displayed. When the inference node is a compliance determination node, the spatial anchoring position is determined as the three-dimensional position of the non-compliant component and the determination conclusion is superimposed.

3. The three-dimensional spatial anchoring visualization method according to claim 1, characterized in that, The execution order of the spatial anchoring parameters is as follows: first, the component visibility switching operation is performed in the three-dimensional scene; then, the camera flight animation is driven to the target camera pose; finally, the three-dimensional annotation information appears in the three-dimensional scene and the inference data and explanation information are displayed in the side information panel.

4. The three-dimensional spatial anchoring visualization method according to claim 1, characterized in that, Also includes: While displaying the current inference node, it receives follow-up questions from the user regarding the current inference node, combines the context data of the current inference node with the follow-up text to form prompt words, and calls the large language model to generate supplementary explanations. At the same time, it highlights and flashes the positioning lines or geometric coordinates of the component targeted by the follow-up question in the 3D scene. After the review process is completed, the user's browsing path along the reasoning hierarchy sequence in the 3D scene, the dwell time at each reasoning node, follow-up questions, and manual annotations are recorded as a structured audit log, and the review process is replayed using the structured audit log.

5. The three-dimensional spatial anchoring visualization method according to claim 1, characterized in that, When displaying inference nodes in a directed acyclic graph in a 3D scene, different colors are used to distinguish rule engine nodes, geometric calculation engine nodes, and large language model nodes based on the execution engine identifier of each inference node.

6. The three-dimensional spatial anchoring visualization method according to claim 1, characterized in that, Also includes: The structured data of the inference directed acyclic graph is persistently stored in the database. The inference node table is used to record the node type, execution engine identifier, upstream and downstream dependencies, spatial anchoring parameters and canonical clause references of the inference nodes, and the inference edge table is used to record the dependency relationship types between inference nodes.

7. A three-dimensional spatial anchoring visualization system for BIM review reasoning chain, characterized in that, include: The review reasoning chain structure module is used to obtain the input data, output data and logical dependencies of the reasoning steps from the outputs of the three types of execution engines: rule engine, geometric calculation engine and large language model during the review process. Each reasoning operation is encapsulated into a reasoning node data structure that includes node type, execution engine identifier, upstream dependent node ID list, associated BIM component ID list and three-dimensional spatial coordinate range. The inference DAG construction module is used to construct a directed acyclic graph of inference nodes based on the input-output dependencies between inference nodes, and to perform topological sorting on the inference DAG to generate an inference hierarchy sequence. The three-dimensional spatial anchoring mapping module is used to calculate the spatial anchoring parameters of the inference node in the three-dimensional scene based on the BIM component ID list and three-dimensional spatial coordinate range associated with each inference node in the inference directed acyclic graph. The spatial anchoring parameters include the target camera pose, the target component visibility state set, the spatial position of the three-dimensional annotation information, and the display content. The inference chain display module is used to execute camera flight animation, component visibility switching and 3D annotation emergence in the 3D scene according to the spatial anchoring parameters, and display the inference data and explanation information of the current inference node node by node according to the inference hierarchy sequence.

8. The three-dimensional spatial anchoring visualization system according to claim 7, characterized in that, It also includes a large model enhancement and explanation module, which, while displaying the current inference node, receives follow-up questions initiated by the user regarding the current inference node, combines the context data of the current inference node with the follow-up text into prompt words, calls the large language model to generate supplementary explanations, and simultaneously highlights and flashes the positioning lines or geometric coordinates of the component targeted by the follow-up question in the 3D scene.

9. The three-dimensional spatial anchoring visualization system according to claim 7, characterized in that, It also includes an audit traceability module, which is used to record the user's browsing path along the reasoning hierarchy sequence in the 3D scene, the dwell time of each reasoning node, follow-up questions, and manual annotations as a structured audit log after the review process is completed, and to use the structured audit log to replay the review process.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the three-dimensional spatial anchoring visualization method according to any one of claims 1 to 6.