Building information model compliance review method based on IFC-Graph and large language model

Abstract

The invention discloses a building information model compliance examination method based on IFC-Graph and a large language model, and the method achieves the precise structured analysis of supervision provisions through BILUO annotation and self-correction, completely reserves the BIM semantics and relation through an IFC-Graph knowledge graph, achieves the semantic consistency of a supervision concept entity and an IFC concept entity through semantic alignment, and achieves the compliance examination of the supervision concept entity and the IFC concept entity. And finally, realizing multi-hop retrieval and interpretable automatic review by utilizing a large language model in combination with GraphRAG. An IFC-Graph knowledge graph is constructed based on a BIM model, entities in a supervision text and IFC class tags are subjected to semantic alignment, reasoning and automation are performed by fusing a large language model in an examination process, and meanwhile, the accuracy of LLM text data generation is improved through GraphRAG, so that the accuracy of BIM compliance examination is improved, and the efficiency of BIM compliance examination is improved. The problems of difficulty in semantic alignment of the building supervision provisions and the BIM model and low efficiency of automatic compliance review are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent agents, and particularly relates to a building information model compliance review method based on IFC-Graph and a large language model. BACKGROUND

[0002] In the context of accelerating digitalization, the construction industry has gradually recognized the importance of "design intelligence". "Design intelligence" refers to the ability to extract valuable insights from the vast amount of data stored in building information models (BIM) and make more intelligent and efficient design decisions accordingly. As the industry's demand for digital transformation becomes increasingly urgent, automated compliance checking (ACC) based on BIM has become a core research content for the construction industry to move towards digitalization and intelligentization, and is a fundamental and key work for the future development and transformation of the construction industry, becoming a global focus of research.

[0003] As the application of BIM technology gradually deepens, BIM, as the core data carrier of the modern construction industry, is crucial for interoperability and accessibility, as different software and systems need to share and exchange building information. The Industry Foundation Classes (IFC) standard is widely used for storing and exchanging building information, as it provides a unified data format, promoting interoperability between different software. However, existing technical solutions often fail to completely preserve the complex relationships and hierarchical structures of the original data when converting IFC data into other formats for storage, making it difficult to handle the complexity of BIM data. This not only weakens the integrity and accuracy of the data, but also may affect in-depth analysis and research based on these data.

[0004] At the same time, although the BIM compliance review method based on large language models (LLM) has made progress, there are still some technical bottlenecks. Although more accurate and comprehensive answers can be obtained through LLM interaction, the model still has deficiencies in understanding and analyzing professional terms in the field of architecture, complex specifications, and specific attributes and relationships in IFC data, especially topological relationships, geometric structures, and implicit relationships.

[0005] Therefore, how to solve the problems of difficulty in aligning building regulatory provisions with BIM model semantics and low efficiency of automated compliance review based on the building information model compliance review method based on IFC-Graph and a large language model has become a technical problem that needs to be solved by technical personnel in this field. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a building information model compliance review method based on IFC-Graph and a large language model, constructs an IFC-Graph knowledge graph based on a BIM model, and performs semantic alignment on entities in regulatory text and IFC class labels, fuses a large language model for reasoning and automation during the review process, improves the accuracy of LLM generated text data through GraphRAG, improves the accuracy of BIM compliance review, and solves the problems of difficulty in semantic alignment of building regulatory provisions and BIM models and low efficiency of automatic compliance review.

[0007] To solve the above technical problems, the present application adopts the following technical solutions:

[0008] A building information model compliance review method based on IFC-Graph and a large language model, comprising the following steps:

[0009] S1, obtaining regulatory text of building design regulatory provisions and a BIM model to be reviewed;

[0010] S2, using a large language model, performing concept entity recognition on the regulatory text based on a BILUO labeling system to obtain each preliminary recognized entity;

[0011] S3, verifying and optimizing each preliminary recognized entity through a preset self-correction method to obtain each corrected regulatory concept entity;

[0012] S4, extracting structured attribute information in the BIM model to be reviewed, converting the structured attribute information into graph structure data in the form of a label property graph LPG, and constructing an IFC-Graph knowledge graph;

[0013] S5, matching and mapping each corrected regulatory concept entity with an IFC class in a predefined IFC class label set of the BIM model through a large language model to obtain a semantic alignment result of the regulatory concept entity and the IFC class in the BIM model;

[0014] S6, converting a review request input by a user into a query statement suitable for the IFC-Graph knowledge graph based on the semantic alignment result through a large language model;

[0015] S7, executing the query statement in the IFC-Graph knowledge graph and performing multi-hop retrieval combined with GraphRAG to obtain a retrieval result; comparing and analyzing the retrieval result with regulatory provisions and performing compliance review to obtain a compliance review result of the BIM model to be reviewed.

[0016] As a preferred solution, in the step S3, the preset self-correction method comprises:

[0017] 1) Introduce the knowledge graph in the field of architecture, cross-verify the preliminary identified entities through the relationship information in the knowledge graph, identify the error entities or entities with inconsistent attribute values in the preliminary identified entities;

[0018] 2) Review the preliminary identified entities through another large language model, and optimize the preliminary identified entities through multiple rounds of feedback loops, correct the entities with errors, omissions or non-standard in the preliminary identified entities, and finally obtain the corrected regulatory concept entities.

[0019] As a preferred solution, in step S4, after extracting the structured attribute information in the BIM model, it further includes: reclassifying the structured attribute information into literal attributes, entity attributes and relationship attributes according to the IFC attribute values, and converting the literal attributes, entity attributes and relationship attributes into graph structure data in the form of label property graph LPG.

[0020] As a preferred solution, in step S4, the structured attribute information in the BIM model to be reviewed is extracted through IfcOpenShell; the structured attribute information includes building elements, attributes and relationships.

[0021] As a preferred solution, in step S4, when extracting the relationship in the BIM model, the implicit relationship between the building elements is obtained through geometric analysis and topological reasoning by TopologicPy; at the same time, the class hierarchical inheritance relationship of IFC standard is extracted to obtain the inheritance chain information of IFC class, and the inheritance chain information of IFC class is stored in the graph database in the form of node label together with the graph structure data.

[0022] As a preferred solution, in step S5, the IFC concept entity semantic information of IFC class is obtained from the IFC standard by calling bsDDAPI to assist the large language model to realize the matching and mapping of regulatory concept entities and IFC classes.

[0023] As a preferred solution, in step S5, the matching and mapping of regulatory concept entities and IFC class labels includes mapping the natural language concepts in the regulatory provisions to the corresponding node labels or attribute keys in the IFC-Graph knowledge graph.

[0024] As a preferred solution, in step S6, it specifically includes the following steps:

[0025] S601, obtain the review request and Prompt template and user question and answer examples; wherein the Prompt template is set according to the structure and format of the graph structure data in the IFC-Graph knowledge graph;

[0026] S602, entity recognition is performed on the review request, and the recognized entities are mapped in combination with the node labels and attribute keys in the IFC-Graph knowledge graph;

[0027] S603, the semantic alignment result is filled into the Prompt template as a variable, and the mapped review request, the filled Prompt template, and the user question and answer examples are combined to generate a final prompt Prompt;

[0028] S604, input the final prompt Prompt into a large language model to generate a Cypher language query, and the generated Cypher language query is used as a query statement applicable to the IFC-Graph knowledge graph.

[0029] As a preferred scheme, in the step S7, the compliance review result includes: if no object related to the review request is found in the search result, outputting preset information of not found review object; otherwise, comparing the search result with the regulatory text to obtain a compliance conclusion, generating an explanation chain based on the associated path obtained from the IFC-Graph knowledge graph, and then converting the explanation chain into a natural language description to form the compliance review result.

[0030] As a preferred scheme, in the step S7, it further includes: in response to a follow-up request initiated by a user based on the compliance review result, automatically matching entities in the follow-up request, and triggering subgraph retrieval and incremental information generation of the IFC-Graph knowledge graph.

[0031] Compared with the prior art, the present application has the following technical effects:

[0032] 1、The present application firstly performs concept entity recognition and classification based on the BILUO labeling system and a large language model, more accurately defines the boundaries and types of entities, and through a self-correction method, knowledge-driven verification, and multi-model cross-validation self-correction mechanism, can effectively identify and correct errors, omissions, or inconsistencies in preliminary identification, greatly improve the quality of entity data, reduce the bias that may be generated by a single model through multi-model cross-validation, improve the stability and semantic coherence of the entity recognition result, and through multiple rounds of feedback cycles, continuously optimize the recognition result, correct errors, omissions, or entities that do not meet the standards, realize high-precision analysis and semantic disambiguation of complex and ambiguous natural language regulatory provisions, and significantly improve the accuracy and reliability of complex entity recognition in regulatory provisions.

[0033] 2、The application converts the structured attribute information of the BIM model into a graph structure data in the form of a label property graph (LPG) and stores it in a graph database to construct an IFC-Graph knowledge graph; in this way, on the one hand, the complex topological, hierarchical and spatial relationships between building elements are completely retained by using the graph structure, the information loss problem when converting the IFC data into a traditional relational database or file is avoided, and the integrity and richness of the data are ensured; on the other hand, when performing compliance checking based on IFC-Graph, multi-hop retrieval can be performed through the graph database, related entities and their associated paths can be quickly traversed and retrieved, and the review efficiency is improved.

[0034] 3、The application accurately maps the regulatory concepts and the pre-defined IFC class label set, solves the semantic inconsistency problem between the BIM model and the specification text, ensures the accuracy of the review operation, and fundamentally avoids the misjudgment caused by concept ambiguity; in the review process, the natural language review request of the user is converted into an accurate graph query statement by using a large language model, so that the user can initiate complex review requirements without mastering complex query syntax, while ensuring the accuracy and executability of the converted query statement.

[0035] 4、The application combines GraphRAG and a large language model to realize joint reasoning of structured graph data and unstructured knowledge text data, GraphRAG (knowledge graph and retrieval enhancement RAG technology combination) can inhibit the preference of LPG and the hallucination generated by the large model, apply the GraphRAG method based on the large model to BIM compliance review, improve the integrity and accuracy of the review information, reduce manual intervention, and improve the review efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings, in which:

[0037] Figure 1 A flow chart of a building information model compliance review method based on IFC-Graph and a large language model is disclosed.

[0038] Figure 2 A processing process example diagram of the compliance result analysis of the BIM model of the embodiment. DETAILED DESCRIPTION

[0039] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0040] The present application will be further described in detail below with reference to the drawings.

[0041] Under the background of the current acceleration of the digital transformation of the construction industry, building information modeling (BIM) has become the core data carrier of the industry, but its compliance review still heavily relies on manual operation, which is inefficient and prone to errors. Although there are some automated review methods, they are mostly based on rule libraries or shallow semantic analysis, which are difficult to handle complex logical relationships, implicit constraints in regulatory provisions, and rich semantic information and topological associations in BIM models. In addition, there is a huge semantic gap between natural language described regulatory provisions and standardized BIM models, making it difficult to achieve precise alignment between the two, and existing technologies cannot simultaneously meet the requirements of review accuracy, efficiency and explainability.

[0042] In view of the above problems and deficiencies, the present application proposes a building information model compliance review method based on IFC-Graph and large language model. This method improves the accuracy of structured processing of regulatory text through the BILUO annotation system and double-model self-correction method. By converting IFC data into label property graph LPG and constructing IFC-Graph knowledge graph, it realizes the complete preservation and efficient query of BIM data relationships. Finally, through the GraphRAG technology, the natural language understanding ability of the large language model is combined with the multi-hop retrieval ability of the graph database, realizing an end-to-end automated process from "natural language review request" to "executable graph query" and then to "explainable review results".

[0043] As shown in Figure 1 , Figure 2 The present application proposes a building information model compliance review method based on IFC-Graph and large language model.

[0044] It should be noted that Figure 2 The implementation process in the present application is to encode the specific process of the present application into an intelligent program module for processing. Figure 2The concept entity recognition module, the self-correction module, the regulatory concept and IFC concept alignment mapping module, and the GraphRAG-based reasoning query module in the method are functional modules corresponding to the processes, and the functions are integrated on the software module, which is a common means of intelligent system, and will not be described here.

[0045] Specifically, the method of the embodiment includes the following steps:

[0046] S1, obtaining a regulatory text of a building design regulatory provision and a BIM model to be examined;

[0047] S2, using a large language model, performing concept entity recognition on the regulatory text based on a BILUO annotation system to obtain each preliminary recognized entity; the preliminary recognized entity includes an attribute constraint type entity and a relationship constraint type entity;

[0048] In specific implementation, to perform automatic BIM compliance review, it is first necessary to clarify how to convert the regulatory provisions of bridges, buildings, etc. from natural language into a machine-readable format. In this embodiment, the regulatory provisions are divided into attribute constraints and relationship constraints in combination with classes and attributes. The attribute constraint type provision refers to an attribute value that can be directly represented in the BIM model. These attribute values are usually specific and countable, such as size, material type, fire resistance rating, etc. The relationship constraint type provision refers to the association between components, such as connection relationship, spatial distance, etc., which is expressed by extending the IFC standard.

[0049] After completing the concept division, the regulatory text entity is then structured. Specifically, the regulatory text is input into the model to perform a standard semantic annotation task, identify the entities in the provision sentence, determine the semantic element type corresponding to each entity, and finally annotate the semantic element label to the sentence, aiming to convert the key information and concepts in the regulatory text into a machine-readable format to support automated processing and intelligent decision-making. In order to ensure the operability and consistency of entity recognition, this embodiment introduces the BILUO annotation system (Begin, In, Last, Unit, Out), and constructs a semantic element classification system of provision composition form according to the semantic definition of the BIM model, so that the entity recognition result strictly follows the IFC class and attribute system. This lays a foundation for subsequent entity and IFC model semantic alignment, ensuring semantic consistency.

[0050] S3, verifying and optimizing each preliminary recognized entity through a preset self-correction method to obtain each corrected regulatory concept entity;

[0051] In specific implementation, after the defined different semantic element entities are preliminarily identified, a verification mechanism is introduced for checking, the context information of the preliminarily identified entities is cross-verified by using the relationship information in the domain knowledge graph (such as the ontology of the IFC standard), so that potential error entities or inconsistent attribute values are identified; meanwhile, another large language model is adopted, and the preliminarily identified entities are critically reviewed and cross-verified through cooperation of two language models (one is responsible for extraction, and the other is responsible for critical review), so as to reduce the single-model bias and improve the stability and semantic coherence of entity identification; finally, the above correction process is executed through multiple iterations to optimize and correct the identification results, and finally the corrected regulatory concept entities are obtained.

[0052] S4, extracting the structured attribute information in the BIM model to be reviewed, converting the structured attribute information into graph structure data in the form of a label property graph (LPG), and constructing an IFC-Graph knowledge graph;

[0053] In specific implementation, in the IFC, an attribute is defined as a unit of information within an entity, is described through a specific type or a reference to a specific entity, and is divided into three categories: a direct attribute, a reverse attribute, and a derived attribute. Such attribute classification has significant limitations in the application of a model-driven method. Therefore, in step S4, after the original structured attribute information is extracted from the BIM model, the attribute classification is reclassified into three groups according to the type of the IFC attribute, that is, a literal attribute (storing text or numerical data), an entity attribute (directly referencing other entities), and a relationship attribute (explicitly describing the logical or spatial association between entities). The reclassification method simplifies the attribute structure, clarifies the data semantics, and significantly improves the model's analyzability and application compatibility, thereby facilitating the automatic processing and analyzability improvement of BIM data.

[0054] On this basis, the embodiment extracts and parses IFC data through two tools of IfcOpenShell and TopologicPy. Specifically, the embodiment includes the following steps:

[0055] Firstly, the structured attribute information in the BIM model under review is extracted through IfcOpenShell; the structured attribute information includes building elements, attributes, and relationships. More importantly, this step also builds and saves the complete IFC class hierarchy, that is, the inheritance relationship of all classes and their parent classes in the IFC dataset is saved, and a clear inheritance chain is established, for example, IfcWall inherits from IfcBuildingElement, IfcBuildingElement inherits from IfcProduct, and IfcProduct inherits from ifcObject. This inheritance chain information will be applied in the form of tags in the subsequent graph data construction (for example, a wall node can be marked with multiple level tags such as IfcWall:IfcBuildingElement:IfcProduct:ifcObject), which greatly optimizes the query efficiency in the graph database, especially when performing fast filtering and semantic reasoning based on categories.

[0056] Then, with the help of TopologicPy, the IFC data can be analyzed and the function of advanced topology can be used to mine implicit relationships, so as to identify deeper building information such as "the spatial nesting between rooms and walls" or "the topological relationship between doors and windows and openings", and to establish display relationships through the identified relationships. After the extraction of explicit and implicit relationships is completed, customized algorithms can be developed according to specific needs to handle complex scenarios. For example, in the aspect of layered mapping of composite wall materials, the following processing flow is designed: geometry reconstruction: extract the wall section profile and analyze the material layer structure defined by IfcMaterialLayerSet; data mapping: bind the properties of each layer of material to its corresponding geometric layer (such as the outermost layer of gypsum board); sequence verification: check whether the arrangement of the material layers meets the requirements of the construction specification (such as the waterproof layer should be located on the inside).

[0057] Finally, based on the above data, a knowledge graph is created in Neo4j software, the building elements are mapped as nodes in the graph structure, the attributes are mapped as attributes of the corresponding nodes, and the relationships are mapped as edges connecting the nodes, thereby constructing a graph structure data stored in the form of a labeled property graph (LPG).

[0058] In this embodiment, the traditional method can only process the data explicitly declared in the IFC file, while this embodiment reveals the essential associations between building elements that are not directly defined in the model by mining implicit relationships such as spatial nesting and topological connections. At the same time, by saving and utilizing the complete inheritance chain of IFC classes and converting it into multi-layer labels in the graph database, the problem of semantic information loss in traditional data processing is solved, so that the system can not only query specific "IfcWall", but also efficiently filter all "IfcBuildingElement" or even all "IfcProduct". This semantic-based filtering and querying capability greatly improves the efficiency and accuracy of the review.

[0059] In addition, IfcOpenShell as IFC parsing can accurately extract explicit information in the model, such as the defined properties of building elements, direct association relationships, and the most important IFC class inheritance chain, as well as pure data and semantic information; while TopologicPy as a geometry calculation engine processes the geometric data in IFC, and through spatial and topological operations, it mines the implicit relationships in the model that are not explicitly declared, such as spatial containment and component adjacency; through the cooperation of IfcOpenShell and TopologicPy, a knowledge graph is constructed that contains complete standard semantics and rich deep spatial and topological relationships, and is extremely rich in information, greatly enhancing the practicality and robustness of this embodiment.

[0060] S5, match and map each corrected regulatory concept entity with the IFC classes in the pre-defined IFC class label set of the BIM model through a large language model, to obtain a semantic alignment result of the regulatory concept entity and the IFC class in the BIM model; the IFC class label set includes a plurality of candidate IFC classes of the BIM model, used to identify and classify different building components or physical objects;

[0061] In specific implementation, the aligning to superclasses method is adopted, which can make the IFC concepts and regulatory concepts more flexibly matched and docked. The superclass refers to a higher class or a more general class in a class hierarchy. For example, if the regulatory concept describes a certain class of properties of a building (such as structural safety, energy efficiency, etc.), and these properties do not have the same representation in IFC, it may be necessary to dock the IFC concept to a higher level class (such as “IfcBuildingElement” or “IfcProduct”). Through this docking method, it can be better combined with the IFC class tagging processing form in step S4. Since the nodes in the figure have saved the complete class inheritance chain tags (such as an IfcWall and IfcBuildingElement tag for a wall node), the system can efficiently retrieve all corresponding instances in the graph regardless of which level the entity is mapped to, making the regulatory concept and IFC entity concept more accurate, consistent and available.

[0062] In specific implementation, in step S5, the IFC concept entity semantic information of the IFC class is obtained from the IFC standard by calling the bsDD API to assist the large language model to realize the matching and mapping of the regulatory concept entity and the IFC class; wherein the matching and mapping of the regulatory concept entity and the IFC class tag includes mapping the natural language concept in the regulatory provision to the corresponding node tag or attribute key in the IFC-Graph knowledge graph. This step specifically includes:

[0063] First, align the entities extracted by the LLM to match each entity with the candidate classes in the predefined IFC class tag set; at the same time, call the bsDD API, which provides a terminology library of IFC standards, can return the natural language definition, explanation and association of the candidate IFC class for the LLM, provide rich domain context for the LLM, thereby improving the reliability of semantic similarity; finally, the alignment result is output in a structured JSON format, clearly recording each regulatory entity and its mapped IFC class tag. For example, the input provision “the factory building with a first fire resistance rating has a fire resistance limit of no less than 3h for its load-bearing wall”, will get the following mapping results: the factory building is mapped to IfcBuilding, the load-bearing wall is mapped to IfcWall, and the fire resistance limit is mapped to FireRating. This process not only ensures that the building entities in the regulatory text are consistent with the component types in the BIM model, but also provides structured input for data processing in subsequent steps.

[0064] S6, converting the review request input by the user into a query statement suitable for the IFC-Graph knowledge graph based on the semantic alignment result by the large language model;

[0065] In specific implementation, the step S6 specifically includes the following steps:

[0066] S601, obtaining an examination request and a Prompt template and user question and answer examples; wherein the Prompt template is set according to the structure and format of the graph structure data in the IFC-Graph knowledge graph;

[0067] S602, performing entity recognition on the examination request, and mapping the recognized entities in combination with the node labels and attribute keys in the IFC-Graph knowledge graph;

[0068] S603, filling the semantic alignment result into the Prompt template as a variable, and combining the mapped examination request, the filled Prompt template and the user question and answer examples to generate a final prompt Prompt;

[0069] S604, inputting the final prompt Prompt into a large language model to generate a Cypher language query, and taking the generated Cypher language query as a query statement applicable to the IFC-Graph knowledge graph.

[0070] Specifically, the user inputs an examination request in natural language or structured form (such as “check whether the fire resistance of the factory building load-bearing wall is ≥ 3 hours”), the examination request of the user is analyzed, the key entities (such as “factory building”, “load-bearing wall”, “fire resistance”) therein are recognized, and the semantic alignment result of step S5 is used to map them to the corresponding node labels (such as IfcBuilding, IfcWall) and attribute keys (such as Loadbearing, FireRating) in the IFC-Graph; then, the mapped entities, the pre-defined Prompt template, and a small amount of example questions and answers are combined into a structured prompt, instructing the large model to strictly generate a query statement conforming to the Cypher syntax (for example, “your answer must be in the form of an executable Cypher query. Any explanation, additional information, etc. that does not belong to the Cypher query syntax should be completely omitted”); the generated query statement is used for subsequent retrieval in the IFC-Graph.

[0071] That is, the embodiment combines the user question, the Prompt template and the user question and answer examples into a complete prompt according to these information, and sends it to the large language model, which takes it as a context and instruction to understand the task it needs to perform. Through the entities mentioned in the question, such as (wall, door, column), the corresponding regulatory entities and IFC model components are found, and then the example form is referred to for reply.

[0072] For the Prompt template used in this embodiment, the generation of the Prompt template is a pre-designed process, and the template generally includes the following parts:

[0073] System role instruction: specifies the role and behavior of the LLM; for example: "You are an assistant responsible for converting natural language questions into Cypher query statements." Graph structure description: clearly describe the database schema obtained in the first step of analysis as the knowledge context of the LLM; output format requirement: strictly require the LLM to output only Cypher code, and cannot have any other explanations or text.

[0074] For example, user question -> Cypher query, the LLM completes the task through the following steps:

[0075] Placeholder: a position is left in the template for inserting the current specific user question and the mapped entity information at runtime. The placeholder is filled at runtime, and when a specific user question needs to be processed, the system fills the results of the previous steps (especially the semantic alignment results obtained after the "regulatory concept and IFC concept alignment" mapping, such as "load-bearing wall" -> IfcWall) into the placeholders of the above pre-set template to form a complete Prompt, which is finally sent to the LLM.

[0076] In this way, by pre-customizing the Prompt template based on the structure (node label, attribute key, relationship type) of the base IFC-Graph knowledge graph, it is ensured that the large language model strictly follows the logical structure and format requirements of the regulatory provisions when converting user review requests. This templated guidance effectively constrains the output space of the model, avoiding syntax errors, format chaos, or semantic deviations caused by the model's free play, significantly improving the accuracy of the generated graph query statements; in addition, the pre-defined Prompt template serves as a reusable tool, eliminating the need to rebuild complete instructions every time a user request is processed, thereby speeding up the query conversion and reducing computational overhead.

[0077] In this way, the user does not need to learn complex graph query syntax (such as Cypher) or understand the schema structure of the underlying database, and can directly describe the review intent in natural language (such as "check whether the fire resistance limit of the factory load-bearing wall is ≥ 3 hours") to obtain the result; at the same time, through entity mapping based on semantic alignment results, it is ensured that terms such as "load-bearing wall" and "fire resistance limit" in the user query can be accurately mapped to node labels (IfcWall) and attribute keys (FireRating) in the graph database, thereby avoiding query errors caused by concept ambiguity from the root.

[0078] S7. Execute the query statement in the IFC-Graph knowledge graph and perform multi-hop retrieval using GraphRAG to obtain the retrieval results; compare and analyze the retrieval results with the regulatory provisions and conduct compliance review to obtain the compliance review results of the BIM model to be reviewed.

[0079] In practice, compliance review results include: calling a graph database (such as Neo4j) via API, executing generated Cypher query statements, and dynamically responding to user queries using GraphRAG technology, including attribute retrieval, and retrieving nodes, attributes, and their relationships that meet the conditions (e.g., retrieving all IfcWall nodes with the Loadbearing attribute set to True and returning their FireRating attribute values); for example, if a user query is "check whether the fire resistance rating of the load-bearing walls of the factory building is ≥3 hours", the LLM outputs a structural query:

[0080] MATCH(b:IfcBuilding{Name:'Factory'})<-[:PART_OF]-(w:IfcWall{Loadbearing:true})

[0081] Where w.FireRating>=3

[0082] RETURN w.Name, w.FireRating

[0083] The retrieved attribute data (such as fire resistance limit value) is automatically compared with the specific requirements in the regulatory provisions (such as "greater than or equal to 3 hours") to determine compliance or non-compliance and generate the final review conclusion.

[0084] Meanwhile, GraphRAG works in conjunction with the Language Model. While LLM is a general knowledge model, it has limited understanding of specialized, structured domain knowledge (such as BIM and IFC standards) and is prone to "illusions." However, GraphRAG leverages the graph structure (nodes, relationships, and paths) of the IFC-Graph knowledge graph to retrieve the most relevant and accurate structured facts and data, providing this information as context to LLM. This greatly enhances the accuracy and depth of LLM's answers to questions within specialized domains.

[0085] When the search results are obtained, the semantic alignment results in step S5 are used. Due to the form of natural language of the user question, the alignment results are needed to align the results to the model to find accurate entities. Then the retrieved information is input to the large model, and the large model generates a natural language description text. The core of the compliance review is to compare the actual attribute values retrieved from the BIM model with the standard values specified in the regulatory provisions, and output the results of non-compliance.

[0086] When the embodiment is applied, the compliance review process of this step is as follows:

[0087] Obtain search results: After executing the Cypher query, the structured data returned from the Neo4j graph database is, for example, the fire resistance limit of the factory building, which may return a number 2.5 (hours);

[0088] Analyze the specification requirements in the user question: The system also needs to understand the specification values implied in the user question, for example, the review request is “whether it meets the specification requirements”, at this time the LLM will check the previous specification provisions;

[0089] Logical comparison and judgment: The system compares the search results (actual values) with the specification requirements (standard values) logically, for example, actual value (2.5) < specification value (3) → not compliant;

[0090] Generate review conclusion: According to the comparison result, output the clear compliance judgment. As the document states: “Compare the search data with the specification requirements, and mark compliance or non-compliance”.

[0091] For example: The specification requirement is that the fire resistance limit of the firewall of the factory building with a three-level fire resistance rating should not be less than 3 hours (this is the specification requirement); the factory building in the component group description has a three-level fire resistance rating, which meets the specification requirement type, and the fire resistance limit of the firewall is 2.5 hours (this is the information retrieved through the question and then converted into a text description); However, the fire resistance limit of the firewall mentioned in the description is only 2.5 hours, which is obviously lower than the requirement of 3 hours, so it can be directly judged as non-compliant.

[0092] In addition, if the search results do not find the object related to the review request, the preset information of not finding the review object is output; otherwise, compare the search results with the regulatory text to get the compliance conclusion, and generate an explanation chain based on the associated path obtained from the IFC-Graph knowledge graph (such as “the fire resistance limit of the load-bearing wall A is 2.5h < 3h”), and then convert the explanation chain into a natural language description, thereby forming the compliance review result.

[0093] At the same time, in response to the user's follow-up request initiated based on the compliance review results (such as "which walls are connected to this load-bearing wall?"), the entity in the follow-up request is automatically matched, and the subgraph retrieval and incremental information generation of the IFC-Graph knowledge graph are triggered.

[0094] The final compliance review results are output in JSON format, ensuring data structuring, machine readability, and integrability. By converting the JSON format into natural language form through a large model and formatting it into a human-readable format, the goal is to improve data accessibility through clear and concise presentation, generating a compliance report that users can understand. The entire process takes user questions as input, combines the semantic network and domain knowledge base of IFC-Graph, prompts, and examples to achieve end-to-end BIM compliance review reasoning from data retrieval to natural language generation.

[0095] In summary, compared with the prior art, the present application has the following technical effects:

[0096] 1、The present application first identifies and classifies the concept entity by combining the BILUO labeling system with a large language model, more accurately defines the boundaries and types of entities, and through self-correction methods such as knowledge-driven verification and multi-model cross-validation, effectively identifies and corrects errors, omissions, or inconsistencies in preliminary identification, significantly improving the quality of entity data. Multi-model cross-validation reduces the bias that may be generated by a single model, improves the stability and semantic coherence of the entity recognition results, and through multiple rounds of feedback cycles, continuously optimizes the identification results, corrects errors, omissions, or entities that do not meet standards, achieving high-precision analysis and semantic disambiguation of complex and ambiguous natural language regulatory provisions, significantly improving the accuracy and reliability of complex entity identification in regulatory provisions.

[0097] 2、The present application converts the structured attribute information of the BIM model into graph structure data in the form of a label attribute graph (LPG) and stores it in a graph database to construct an IFC-Graph knowledge graph. In this way, on the one hand, the graph structure completely retains the complex topological, hierarchical, and spatial relationships between building elements, avoiding information loss when converting IFC data into traditional relational databases or files, ensuring data integrity and richness. On the other hand, when conducting compliance checks based on IFC-Graph, the graph database can perform multi-hop retrieval, quickly traversing and retrieving related entities and their associated paths, improving review efficiency.

[0098] 3、The application solves the problem of semantic inconsistency between BIM models and specification texts by accurately mapping the regulatory concept with the pre-defined IFC class label set, ensures the accuracy of the review operation, and fundamentally avoids misjudgment caused by concept ambiguity; in the review process, the user's natural language review request is converted into an accurate graph query statement by using a large language model, so that the user can initiate complex review requirements without mastering complex query syntax, while ensuring the accuracy and executability of the converted query statement.

[0099] 4、The application realizes the joint reasoning of structured graph data and unstructured knowledge text data by combining GraphRAG with a large language model, GraphRAG (knowledge graph and retrieval enhanced RAG technology combination) can inhibit the preference of LPG and the illusion produced by the large model, apply the GraphRAG method based on the large model to BIM compliance review, improve the completeness and accuracy of the review information, reduce manual intervention, and improve the review efficiency.

[0100] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting, although the present application has been described by referring to the preferred embodiments of the present application, those skilled in the art should understand that various changes can be made in form and details without departing from the spirit and scope of the present application defined in the appended claims.

Claims

1. A compliance review method for building information models based on IFC-Graph and large language models, characterized in that, Includes the following steps: S1. Obtain the regulatory text of the building design regulatory provisions and the BIM model to be reviewed; S2. Using a large language model, based on the BILUO annotation system, perform concept entity recognition on the regulatory text to obtain the entities that are initially identified; S3. Through a preset self-correction method, each initially identified entity is verified and optimized to obtain each corrected regulatory concept entity. S4. Extract the structured attribute information from the BIM model to be reviewed, convert the structured attribute information into graph structure data in the form of a tag attribute graph (LPG), and construct the IFC-Graph knowledge graph. S5. Match and map each corrected regulatory concept entity with the IFC class in the predefined IFC class label set of the BIM model through the large language model to obtain the semantic alignment result between the regulatory concept entity and the IFC class in the BIM model. S6. Based on the semantic alignment results, the user-input review request is converted into a query statement suitable for the IFC-Graph knowledge graph using a large language model. S7. Execute the query statement in the IFC-Graph knowledge graph and perform multi-hop retrieval using GraphRAG to obtain the retrieval results; compare and analyze the retrieval results with the regulatory provisions and conduct compliance review to obtain the compliance review results of the BIM model to be reviewed.

2. The compliance review method for building information models based on IFC-Graph and large language models according to claim 1, characterized in that, In step S3, the preset self-calibration method includes: 1) Introduce a knowledge graph in the field of architecture, and cross-validate the initially identified entities through the relational information in the knowledge graph to identify erroneous entities or entities with inconsistent attribute values ​​in the initially identified entities; 2) The initially identified entities are reviewed through another large language model, and the initially identified entities are optimized through multiple rounds of feedback loops to correct erroneous, missing or non-compliant entities in the initially identified entities, and finally the corrected regulatory concept entities are obtained.

3. The compliance review method for building information models based on IFC-Graph and large language models according to claim 1, characterized in that, In step S4, after extracting the structured attribute information from the BIM model, the method further includes: reclassifying the structured attribute information into text attributes, entity attributes, and relational attributes based on the IFC attribute values, and converting the text attributes, entity attributes, and relational attributes into graph structure data in the form of a label attribute graph (LPG).

4. The compliance review method for building information models based on IFC-Graph and large language models according to claim 1, characterized in that, In step S4, the structured attribute information of the BIM model to be reviewed is extracted using IfcOpenShell; the structured attribute information includes building elements, attributes, and relationships.

5. The compliance review method for building information models based on IFC-Graph and large language models according to claim 4, characterized in that, In step S4, when extracting relationships in the BIM model, implicit relationships between building elements are obtained through geometric analysis and topological reasoning using TopologicPy. At the same time, the class hierarchy inheritance relationship of the IFC standard is extracted to obtain the inheritance chain information of the IFC class, and the inheritance chain information of the IFC class is stored in the graph database together with the graph structure data in the form of node labels.

6. The compliance review method for building information models based on IFC-Graph and large language models according to claim 1, characterized in that, In step S5, the semantic information of the IFC concept entity of the IFC class is obtained from the IFC standard by calling bsDDAPI, which assists the large language model in realizing the matching and mapping of regulatory concept entities and IFC classes.

7. The compliance review method for building information models based on IFC-Graph and large language models according to claim 6, characterized in that, In step S5, the matching and mapping of regulatory concept entities and IFC class tags includes mapping the natural language concepts in regulatory provisions to the corresponding node tags or attribute keys in the IFC-Graph knowledge graph.

8. The compliance review method for building information models based on IFC-Graph and large language models according to claim 1, characterized in that, Step S6 specifically includes the following steps: S601. Obtain the review request, as well as the Prompt template and user Q&A examples; wherein, the Prompt template is set according to the structure and format of the graph structure data in the IFC-Graph knowledge graph; S602. Perform entity identification on the review request and map the identified entities using the node labels and attribute keys in the IFC-Graph knowledge graph; S603. Fill the Prompt template with the semantic alignment result as a variable, and merge the mapped review request, the filled Prompt template and the user Q&A example to generate a final prompt. S604. Input the final prompt into the large language model to generate a Cypher language query, and use the generated Cypher language query as a query statement applicable to the IFC-Graph knowledge graph.

9. The compliance review method for building information models based on IFC-Graph and large language models according to claim 1, characterized in that, In step S7, the compliance review result includes: if no object related to the review request is found in the search results, a preset "no review object found" information is output; otherwise, the search results are compared with the regulatory text to obtain a compliance conclusion, and an explanation chain is generated based on the association path obtained from the IFC-Graph knowledge graph. The explanation chain is then converted into a natural language description to constitute the compliance review result.

10. The compliance review method for building information models based on IFC-Graph and large language models according to claim 9, characterized in that, Step S7 further includes: responding to follow-up questions initiated by users based on the compliance review results, automatically matching the entities in the follow-up questions, and triggering subgraph retrieval and incremental information generation of the IFC-Graph knowledge graph.

Images