An artificial intelligence-based building information model compliance processing method, storage medium, device, and equipment

Abstract

The application provides an artificial intelligence-based building information model compliance processing method and device, electronic equipment and storage medium. A structured compliance knowledge graph containing mandatory parameters, naming rules, geometric constraints and data mapping relationships is constructed by analyzing standard documents through a natural language processing model. When the component is loaded, its geometric feature vector is extracted and a nested hierarchical topology graph is modeled to generate a digital twin archive containing geometric properties, parameter lists and risk scores. The archive is matched and compared with the knowledge graph to dynamically generate a processing strategy sequence and schedule a geometric optimization and data encapsulation module to perform adaptive processing. The geometric optimization uses a deep learning-driven non-uniform mesh simplification algorithm to preserve key area details and repair topological defects. Data encapsulation achieves intelligent matching and binding of custom parameters to standard attribute sets through semantic mapping. Through the above method and device, the efficiency, accuracy and automation level of building information model cross-platform compliance conversion can be significantly improved.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent building information processing and artificial intelligence collaborative optimization, specifically to a building information model compliance processing method, storage medium, device, and equipment based on artificial intelligence. Background Technology

[0002] Current BIM model cross-platform compatibility handling primarily relies on pre-set static rule engines, using fixed mapping tables or scripts to achieve parameter matching and geometric simplification. This approach struggles to adapt to the dynamic updates of multi-source heterogeneous standards. Traditional methods require manual configuration of verification conditions and conversion logic across different cloud platforms, resulting in low efficiency and a tendency to overlook implicit constraints. Especially when standard documents are published as unstructured PDFs or web pages, critical requirements such as "doors and windows must contain a FireRating field" or "naming format is [type][number]" cannot be automatically recognized and parsed by the system, causing compliance checks to lag behind actual specification changes. Furthermore, existing tools lack the ability to jointly understand the semantic and geometric features of components, failing to accurately assess the risks of complex families exhibiting excessive facet counts, nested loops, or zero thickness. The processing often employs a uniform simplification strategy, ignoring differences in component functional areas and distorting engineering information. In addition, the system lacks the ability to continuously learn from historical uploads, leading to recurring errors and hindering the formation of an experience-accumulating and intelligent optimization loop.

[0003] Figure 10 For models of heterogeneous building walls generated using Revit software, importing these models into domestic BIM sharing platforms requires manual processing of model configurations and parameters, severely reducing migration efficiency. There is an urgent need for systems based on big data processing and image processing capabilities to automate the migration and adaptation of these models. Summary of the Invention

[0004] Based on this, in order to address the technical problems of relying on manual interpretation, low efficiency, error-proneness, and difficulty in adapting to dynamic updates of multiple versions of specifications during the migration of Building Information Models (BIM) between different software, an artificial intelligence-based BIM compliance processing method, storage medium, device, and equipment are proposed. This invention protects an AI-based compliance processing method for Building Information Modeling (BIM). The method receives standard documents published by a target platform, uses a natural language processing model to semantically parse the unstructured text in the standard documents, and generates a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints, and data mapping relationships. In response to component loading operations in the BIM, the method graphically represents the three-dimensional geometry of the components and extracts geometric feature vectors. Simultaneously, it models the nested hierarchical structure of the components as a topological graph composed of nodes and edges, identifies abnormal reference relationships and structural complexity indicators, and generates a digital twin profile of the component containing geometric attributes, a parameter list, and a risk score. The method matches and compares the component's digital twin profile with the compliance knowledge graph to determine potential compliance conflict points for the component on the target platform. Based on the comparison results and a preset decision logic, it dynamically generates a processing strategy sequence for the component. According to the processing strategy sequence, it schedules the corresponding geometric optimization module and data encapsulation module to perform adaptive processing.

[0005] Furthermore, the adaptive processing of the corresponding geometry optimization module and data encapsulation module according to the processing strategy sequence includes: the geometry optimization module uses a deep learning-driven mesh simplification algorithm to perform non-uniform simplification of high-face-count components by preserving key areas, and calls a 3D geometry repair tool to correct zero-thickness or self-intersecting defects; the data encapsulation module uses a semantic mapping model to intelligently match the component's custom parameters to the standard attribute set of the target platform and bind the corresponding entity and attribute set.

[0006] Furthermore, the geometry optimization module employs a deep learning-driven mesh simplification algorithm to perform non-uniform simplification of high-face-count components while preserving key areas. It then calls upon a 3D geometry repair tool to correct zero-thickness or self-intersecting defects, including: identifying key area features through a trained neural network model, preserving mesh details corresponding to fillets, bolt holes, and load-bearing edges; performing vertex clustering and face merging operations on meshes corresponding to flat areas to reduce the face count; and using a trained small discriminative model to detect zero-thickness surfaces and self-intersecting geometries, triggering the 3D geometry repair tool to call the target interface for defect correction.

[0007] Furthermore, the data encapsulation module utilizes a semantic mapping model to intelligently match component-defined parameters to the standard attribute set of the target platform and bind the corresponding entities and attribute sets. This includes: parsing the semantic meaning of component-defined parameters using a natural language processing model, mapping non-standard parameter names to the standardized attribute fields corresponding to the target platform; matching suitable container family templates from a preset template library based on the complexity index of component type and nested hierarchy; binding the component to the selected container family template, and associating it with the corresponding entity type and attribute set.

[0008] Furthermore, upon receiving standard documents published by the target platform, the system utilizes a natural language processing model to perform semantic parsing on the unstructured text within the standard documents, generating a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints, and data mapping relationships. This includes: extracting text blocks and their hierarchical structure from the standard documents, preserving the contextual relationships between chapters, clauses, and tables; applying a fine-tuned multimodal natural language processing model to perform joint semantic parsing on the extracted text, identifying the subject objects, constraint types, and target attributes in normative statements; constructing a heterogeneous graph containing standard clauses, component types, parameter names, entity and constraint condition nodes, connecting each node through semantic relationship edges, and storing them in a graph database in the form of triples; and triggering a comparison process when a new version of the standard document is received to identify newly added, modified, or obsolete clauses and update the graph node status and dependency paths.

[0009] Furthermore, in response to the loading operation of components in the building information model, the three-dimensional geometry of the components is represented graphically and geometric feature vectors are extracted. At the same time, the nested hierarchical structure of the components is modeled as a topological graph composed of nodes and edges, including: rendering the components as lightweight three-dimensional images, using convolutional neural networks to analyze the images to identify the component type and estimate the number of faces; representing the components and their sub-components as nodes in the graph, and the reference relationship as edges, and using graph neural networks to detect reference cycles and nesting depth.

[0010] Furthermore, in response to the execution result of the component upload operation, the processing strategy sequence corresponding to the component and the success or failure status of the upload are recorded, and the encrypted processing log is uploaded to the cloud training platform under the user's authorization. The processing log of the failed upload is used as a new training sample and input into the joint training process of natural language processing model and graph neural network. The call weight of each module in the processing strategy sequence is adjusted based on the processing log of the successfully uploaded component. After receiving a new version of the standard document or accumulating a certain number of failure samples, the incremental update of compliance knowledge graph and decision logic is triggered.

[0011] This invention protects an AI-based building information model (BIM) compliance processing device, comprising: a standard document parsing module, used to receive standard documents published by a target platform, perform semantic parsing on unstructured text in the standard documents using a natural language processing model, and generate a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints, and data mapping relationships; a component file generation module, used to respond to the loading operation of components in the BIM, to perform image representation of the three-dimensional geometric shape of the components and extract geometric feature vectors, and simultaneously model the nested hierarchical structure of the components as a topological graph composed of nodes and edges, identify abnormal reference relationships and structural complexity indicators, and generate a component digital twin file containing geometric attributes, parameter lists, and risk scores; a compliance comparison and decision module, used to match and compare the component digital twin file with the compliance knowledge graph, determine the potential compliance conflict points of the component on the target platform, and dynamically generate a processing strategy sequence for the component based on the comparison results and a preset decision logic; and an adaptive processing execution module, used to schedule the corresponding geometric optimization module and data encapsulation module to perform adaptive processing according to the processing strategy sequence.

[0012] This invention protects an electronic device, comprising: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate via the bus. The machine-readable instructions are executed by the processor to perform the steps of a building information model compliance processing method based on artificial intelligence.

[0013] This invention protects a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of an artificial intelligence-based building information model compliance processing method.

[0014] This invention protects an AI-based compliance processing method for Building Information Modeling (BIM). It achieves adaptive compliance conversion of building components through intelligent processing strategy sequence scheduling of geometric optimization and data encapsulation modules, significantly improving the processing efficiency and accuracy of complex building models during cross-platform migration and standard adaptation. The geometric optimization module introduces a deep learning-driven mesh simplification algorithm, performing non-uniform simplification on high-face-count components while preserving details in key areas such as fillets, bolt holes, and load-bearing edges. Combined with a 3D geometric repair tool, it automatically corrects common topological defects such as zero thickness or self-intersections, ensuring the geometric integrity of the simplified model while reducing computational costs. The data encapsulation module intelligently matches custom parameters of components to the standard attribute set of the target platform through a semantic mapping model, automating the binding of entities and attributes. This solves the semantic loss and attribute mismatch problems caused by traditional manual mapping, greatly enhancing the interoperability and standardization of model data. Furthermore, it accurately identifies key region features through a trained neural network, combined with vertex clustering and facet merging techniques. The technology efficiently reduces the surface area of ​​flat regions, resulting in a lighter model without loss of visual or functional features. Simultaneously, a small discriminative model is introduced to detect geometric defects in real time and trigger repair tool interface calls, achieving closed-loop automation of defect identification and repair, thus improving the robustness and stability of the processing flow. Furthermore, during system operation, the processing strategy sequence and execution status are dynamically recorded in response to component upload results. With user authorization, encrypted logs are uploaded to the cloud training platform. Logs from failed uploads are used as new samples to continuously optimize the joint training effect of the natural language processing model and graph neural network, improving the system's understanding and response to abnormal scenarios. The call weights of each processing module are adjusted based on successful cases to achieve personalized and adaptive optimization of strategy execution. When a new version of the standard document is received or a sufficient number of failed samples are accumulated, an incremental update mechanism for the compliance knowledge graph and decision logic is automatically triggered, ensuring the system can quickly respond to changes in industry standards and continuously evolve, thereby constructing an intelligent BIM compliance processing system with self-learning and iterative upgrade capabilities. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This application provides a flowchart of a building information model compliance processing method. Figure 2 This application provides a flowchart of a heterogeneous map construction method based on BIM standard documents. Figure 3 A flowchart of a component analysis method based on a neural network provided in this application embodiment; Figure 4 This application provides a flowchart of a geometry optimization and data encapsulation process. Figure 5 This application provides a flowchart of a neural network-based mesh simplification and geometric repair method. Figure 6 This application provides a flowchart of a component parameter mapping and template matching embodiment; Figure 7 This application provides a diagram of a building information model compliance processing device. Figure 8 This application provides an embodiment of an electronic device for compliance processing of Building Information Modeling (BIM). Figure 9 This application provides a sample specification for a building information sharing platform. Figure 10 This application requires 3D modeling of heterogeneous wall structures that need to be transferred to other BIM platforms. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.

[0017] Research has revealed that traditional Building Information Modeling (BIM) compliance processing methods primarily rely on manual interpretation of platform standard documents and preset rule scripts. This approach suffers from weaknesses such as poor understanding of unstructured specification texts, difficulty adapting to changing platform requirements, and inability to identify potential risks in complex components. Existing tools typically employ fixed mapping logic for parameter transformation and geometric simplification, lacking in-depth analysis of component semantic hierarchy and topology. This leads to processing results that frequently deviate from actual engineering needs when dealing with high-surface-count free-form surfaces, deeply nested families, or non-standardized parameter naming, sometimes even resulting in upload failures or data loss. Furthermore, traditional solutions lack learning and evolution capabilities, failing to accumulate experience from historical cases, leading to recurring issues and severely hindering the efficiency and quality of BIM models in cross-platform collaboration. Using traditional methods to process... Figure 10When dealing with heterogeneous wall structures, a significant amount of time is required, necessitating continuous manual adjustments and adaptation to the shared platform, which severely impacts the efficiency of model sharing.

[0018] Based on this, please refer to Figure 1 This application provides a building information model (BIM) compliance processing method based on artificial intelligence. This method uses artificial intelligence to automatically understand platform standards and specifications and intelligently diagnose and optimize component compliance. Specifically, it includes: S101 receives a standard document published by the target platform, uses a natural language processing model to perform semantic parsing on the unstructured text in the standard document, and generates a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints, and data mapping relationships. S102 responds to the loading operation of the component in the building information model, performs image representation of the three-dimensional geometric shape of the component and extracts geometric feature vectors, and models the nested hierarchical structure of the component as a topological graph composed of nodes and edges, identifies abnormal reference relationships and structural complexity indicators, and generates a digital twin file of the component containing geometric attributes, parameter list and risk score. S103 Match and compare the digital twin file of the component with the compliance knowledge graph to determine the potential compliance conflict points of the component on the target platform, and dynamically generate a processing strategy sequence for the component based on the comparison results and preset decision logic. S104 According to the processing strategy sequence, the corresponding geometry optimization module and data encapsulation module perform adaptive processing.

[0019] This application provides an AI-based compliance processing method for Building Information Modeling (BIM), improving the efficiency and accuracy of compliance processing in cross-platform applications. By automatically parsing standard documents and constructing a structured compliance knowledge graph using a natural language processing model, dynamic learning and continuous updating of platform specifications are achieved. Through image-based representation of the three-dimensional geometric morphology of the components to be processed and graph neural network modeling of nested structures, the geometric complexity and topological anomalies of the components are obtained, and digital twin archives of the components are generated, enhancing the predictive ability for potential migration problems. Based on intelligent comparison of the archives and the knowledge graph, and combined with preset decision logic, a processing strategy sequence is dynamically generated, enabling the system to adaptively schedule geometric optimization and data encapsulation modules for different component characteristics, achieving precise repair and efficient lightweighting. The overall solution relies on artificial intelligence to automatically identify and process problems of the components applied to new platforms, improving the efficiency of component compatibility across different platforms.

[0020] The following is a detailed explanation of the specific process for compliance processing of Building Information Modeling (BIM): S101 receives a standard document published by the target platform, uses a natural language processing model to perform semantic parsing on the unstructured text in the standard document, and generates a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints, and data mapping relationships.

[0021] The process begins with the system retrieving standard documents in PDF, HTML, or Word format from the target BIM platform via a cloud or local interface. These documents typically contain a large amount of unstructured natural language descriptions, technical clause tables, and nested chapter structures. To achieve machine-readable conversion of the specification content, the system first performs a document preprocessing workflow. Using an OCR engine combined with layout analysis algorithms, it extracts the original text blocks and their hierarchical structure information, identifying logical units such as titles, sub-clauses, list items, and data tables, preserving their contextual position in the original document to ensure that subsequent semantic parsing is based on the complete context. Figure 9 This is a screenshot of a standard document from a domestic BIM sharing platform vendor. The standard document clearly defines the component specifications.

[0022] See Figure 2 The process of building a structured compliance knowledge graph for components is as follows: S1011 In the above steps, extract text blocks and their hierarchical structure from the BIM standard document, and preserve the contextual relationships of chapters, clauses and tables.

[0023] In practice, a deep learning-based document layout detection model is used to segment each page of the document into different regions, such as main text paragraphs, numbered clauses, comment boxes, and attribute tables. A document structure tree is also constructed to record the parent-child and sibling relationships between text blocks.

[0024] For example, Figure 9 In a standard document from a well-known domestic BIM vendor, Chapter 5, "Model Organization," serves as the parent node, containing child nodes such as "5.1 General Provisions" and "5.2 Model Architecture." Each child node is associated with corresponding text content and coordinates. For tables spanning multiple pages, the system uses column alignment and header continuity checks to ensure complete restoration, thereby guaranteeing the integrity of parameter constraints and mapping rules.

[0025] S1012 applies a finely tuned multimodal natural language processing model to perform joint semantic parsing on the extracted text, identifying the subject, constraint type, and target attribute in normative statements.

[0026] A multimodal Transformer model finely tuned with a corpus from the architectural component model domain is used. In this implementation, the VisualBERT model is used, but other similar models can also be used. The text content is jointly encoded with the visual layout features of the area in which it is located, thereby improving the ability to understand the semantics under complex typography.

[0027] The model performs sensitive identification of keyword patterns. In one optional implementation, it identifies words such as "must," "should," and "must not" to determine the mandatory level of the statement. Simultaneously, it extracts four core elements through a named entity recognition module: subject object, constraint type, named regular expression, and IFC mapping path. The subject object is the entity corresponding to the model; in one implementation, it could be a curtain wall system or electromechanical piping. The constraint type is the restriction condition of certain structures of the subject object; in one implementation, the constraint condition is a maximum number of faces ≤ 100,000. The named regular expression explicitly defines the naming method; in one implementation, the naming method can be [component type][two-digit number]). The IFC mapping path is the data location path; in one implementation, the location path is expressed as IfcWindow → Pset_WindowCommon.GlobalId. In a general example, in the sentence "All fire doors must be marked with a fire resistance rating of not less than 1.5 hours," the model outputs a triple: (fire door, mandatory parameter, fire resistance rating ≥ 1.5h), and marks its source clause number as "Sec.7.3.2".

[0028] S1013 constructs the extraction results into a heterogeneous graph containing standard clauses, component types, parameter names, entity and constraint nodes, connects each node through semantic relationship edges, and stores them in the graph database in the form of triples.

[0029] The constructed compliance knowledge graph adopts a multi-layered heterogeneous structure. Nodes include categories such as standard document version, chapter entry, component classification, parameter name, IFC entity type, and geometric constraints. Edges represent various semantic relationships such as requirement, belonging to, mapping to, restricted to, and example. The graph data is organized in triple format and stored in the graph database Neo4j or JanusGraph. Each node is appended with metadata fields, including a confidence score derived from the model output probability, an applicability label (e.g., "applicable only to high-rise building projects"), an effective date, and citation frequency statistics.

[0030] To further enhance the intelligent reasoning capabilities of the graph, a knowledge embedding algorithm is introduced to map nodes and relationships to a low-dimensional vector space, supporting similarity calculation and latent rule completion. For example, based on the proximity of rooftop photovoltaic panels and curtain wall photovoltaic modules in the embedded space, it can be inferred that the former may also need to meet the light transmittance requirements.

[0031] S1014 When a new version of the standard document is received, a comparison process is triggered to identify newly added, modified, or obsolete clauses and update the graph node status and dependency paths.

[0032] Whenever the target platform releases an updated specification document, the system automatically initiates a difference analysis process: First, it uses the system's embedded model to compare the semantic changes between the new and old versions of the clauses one by one to determine whether there are any additions, revisions, or repeals; then, it locates the corresponding node in the graph, marks its status as "pending verification", "updated", or "deprecated", and simultaneously adjusts its dependency relationship links with other nodes.

[0033] In one specific implementation, if the new specification adds "all device families must include carbon footprint parameters," the system creates a new mandatory parameter node in the graph and establishes requirement relationship edges between it and the category. Finally, the updated compliance knowledge graph is pushed to the edge plugin cache to ensure that local processing always follows the latest specifications.

[0034] S102 responds to the loading operation of the component in the building information model, performs image representation of the three-dimensional geometry of the component and extracts geometric feature vectors, and models the nested hierarchical structure of the component as a topological graph composed of nodes and edges, identifies abnormal reference relationships and structural complexity indicators, and generates a digital twin file of the component containing geometric attributes, parameter list and risk score.

[0035] In this process, users load BIM components (optionally Revit family files) into the project environment, which automatically triggers a multi-dimensional diagnostic process in the system background. This process constructs a complete digital profile of the component from both geometric and structural perspectives. The generated digital twin profile of the component not only records its explicit parameters but also includes risk scores and optimization suggestions derived from the model, serving as the basic input for subsequent compliance comparisons and the generation of processing strategies.

[0036] See Figure 3 The specific process for generating a digital twin archive of components is as follows: S1021 renders the component as a lightweight 3D image and uses a convolutional neural network to analyze the image to identify the component type and estimate the number of faces.

[0037] When a component is loaded into the BIM environment, the system invokes a lightweight graphics engine to perform multi-view rendering, generating a set of orthogonal projection images (front view, side view, top view, and isometric view), with each image resolution controlled within 256×256. These images are fed as input into a pre-trained convolutional neural network model, which is fine-tuned on a large dataset of labeled BIM component images. The fine-tuned model is capable of distinguishing between standard components and complex freeform surfaces, where standard components include walls, columns, and beams, and complex freeform surfaces include irregular curtain walls and sculptural components. The model output includes the component category confidence distribution and regression predictions of surface complexity based on visual texture density.

[0038] In one specific implementation, for a hyperboloid roof with high curvature, the model can initially determine that it belongs to the freeform surface category and estimate that its number of triangular facets is about 480,000, which exceeds the recommended threshold of most platforms. The recommended threshold of most platforms is usually below 100,000 facets, thus marking the number of triangular facets as a potential performance bottleneck.

[0039] S1022 represents components and their sub-components as nodes in a graph, and reference relationships as edges, using a graph neural network to detect reference cycles and nesting depth.

[0040] For nested family structures that may exist within a component, in a specific implementation, such as a device family containing multiple sub-components, material references, shared parameter links, etc., the system abstracts it as a directed graph G=(V,E), where each sub-component or parameter combination is a node v∈V, and logical relationships such as parent-child affiliation, parameter-driven, geometric reference, or shared resource invocation are edges e∈E.

[0041] In one specific implementation, the topology graph is automatically generated by parsing the family hierarchy exposed by the Revit software's API, preserving the semantic labels of reference paths. Subsequently, a graph neural network model is employed—performing message passing and aggregation operations on the graph, updating node representation vectors layer by layer—ultimately achieving an awareness of the overall structural pattern. The graph neural network model is used to identify two typical problems: first, reference cycles, where A references B, and B indirectly references A, leading to unresolvable issues; second, excessive nesting, defined as nesting levels exceeding a preset threshold, which can be 5 or more levels. Based on the anomaly score output by the graph neural network model and the maximum nesting depth, a structural complexity index is calculated, ranging from 0 to 1, with higher values ​​indicating greater maintenance difficulty and higher compatibility risks.

[0042] After completing the above two analyses, the geometric feature vectors from S1021 and the topological structure indicators from S1022 are integrated. The geometric feature vectors include component type, main contour size, estimated number of facets, visual complexity score, etc., while the topological structure indicators include nesting depth, reference cycle flag, number of sub-components, etc. Combined with the original parameter list extracted from the component metadata interface, which includes material, weight, manufacturer information, IFC classification code, etc., a structured component digital twin archive is constructed.

[0043] Based on this, a weighted fusion model is introduced to comprehensively evaluate various risk factors. In one specific implementation, the weight for geometric excess is set to 0.4, the weight for structural anomaly is 0.35, and the parameter missingness is 0.25, ultimately outputting a risk score between 0 and 10. In one implementation, a certain electromechanical equipment family, due to having 600,000 faces, nested loops, and missing key parameters, has a risk score of 9.2 and is identified by the system as a high-risk upload object requiring priority intervention. This digital twin profile is then pushed to the next stage for matching and comparison with the compliance knowledge graph generated in S101.

[0044] S103 Matches and compares the digital twin file of the component with the compliance knowledge graph to determine the potential compliance conflict points of the component on the target platform, and dynamically generates a sequence of processing strategies for the component based on the comparison results and preset decision logic.

[0045] This process is a comprehensive judgment step in intelligent compliance processing, initiated immediately after the component's digital twin file is constructed by S102. The system uses a unified semantic interface to perform multi-dimensional correlation queries between fields such as geometric attributes, parameter lists, and structural complexity indicators in the file and the compliance knowledge graph generated in S101, identifying deviations of the current component from the target platform's specification requirements. These deviations are categorized into four main conflict types: missing or non-standard parameter naming, excessive geometric complexity, nested structure violations, and missing IFC semantic mapping. Each conflict is labeled with its source clause number, severity level, and scope of impact.

[0046] After completing the matching and comparison, the system invokes the decision engine to dynamically orchestrate the processing strategy sequence. This engine has a built-in pre-defined decision logic rule library. In one specific implementation, the rules take the form: if a specific conflict combination is detected, several processing modules are scheduled according to priority. When the number of faces exceeds 100,000 and a reference cycle exists, the decision logic determines that structural repair must be performed before geometric simplification; if only parameter naming issues exist, the data encapsulation module is directly scheduled for semantic remapping. The strategy sequence is output in the form of an executable workflow, or it can be encoded to determine the scenario and then execute the corresponding processing code. The entire process requires no manual intervention, ensuring end-to-end automation from risk identification to solution generation.

[0047] In response to the component upload operation, the system records the corresponding processing strategy sequence and upload success or failure status for each component, and uploads the encrypted processing log to the cloud training platform with user authorization. Each time a component is processed and attempted to be uploaded to the target BIM platform, the client plugin listens for the API return status code to determine if the operation was successful. Regardless of success or failure, the system generates a structured processing log locally, including the original component ID, the processing strategy sequence used, the time consumed at each stage, resource usage, and the final upload result. Provided the user has enabled the anonymous learning and sharing option beforehand, this log is encrypted and any identifiable project information is removed before being uploaded to the cloud training platform, forming the data foundation for continuous learning.

[0048] The logs of failed uploads are used as new training samples in the joint training process of the natural language processing model and the graph neural network. For cases marked as failed uploads, the system automatically extracts the corresponding knowledge graph comparison results, digital twin profile feature vectors, and executed processing paths to construct a problem-response-result triplet sample. These failed samples are injected into the incremental training process of the natural language processing model and the graph neural network model to enhance the model's ability to identify compliant samples that were not successfully identified.

[0049] In one specific implementation, if multiple similar curtain wall families fail consecutively due to the lack of a specific set of custom attributes, the model will strengthen the judgment logic that such components should be checked in advance during the retraining process, thereby improving the prediction accuracy in similar scenarios in the future.

[0050] The system adjusts the call weights of each module in the processing strategy sequence based on the successfully uploaded component processing logs. It performs cluster analysis on the logs of all successful cases to evaluate the effectiveness of different processing paths for specific component types.

[0051] In one specific implementation, statistics showed that for window families with high polygon counts and standard naming, the success rate of using the geometry lightweighting → intelligent IFC mapping path reached 98%, while adding redundant parameter completion steps actually increased processing latency. Therefore, the system updates the strategy weight table in the background to reduce the trigger probability of unnecessary modules and optimize overall processing efficiency.

[0052] Upon receiving a new version of the standard document or accumulating a certain number of failed samples, an incremental update of the compliance knowledge graph and decision logic is triggered. When the cloud detects that the target platform has released a new version of the specification document, or when a certain type of failure event reaches a preset threshold, and in a specific implementation, the same type of component fails to upload five times consecutively, the knowledge base update process is automatically activated.

[0053] The knowledge base update process first uses a text model to compare the differences between old and new documents to identify the changed clauses. Then, it updates the node states and dependencies in the compliance knowledge graph and simultaneously adjusts the rule condition branches in the decision engine. After the update is complete, the new version of the knowledge graph and the optimized strategy model are packaged and pushed to the edge plugin, achieving a silent upgrade. This forms a complete closed loop of diagnosis, processing, feedback, and update, enabling the system not only to meet current compliance requirements but also to continuously update as the external environment changes.

[0054] S104 According to the processing strategy sequence, the corresponding geometry optimization module and data encapsulation module perform adaptive processing.

[0055] The above steps constitute the final execution stage of the entire compliance processing flow. After the processing strategy sequence generated in S103 is parsed into an executable instruction stream, the system activates the corresponding functional modules in sequence to achieve precise intervention on the components. The entire scheduling process is driven by a small model coprocessor, which dynamically coordinates the calling order and parameter configuration of the geometry optimization module and the data encapsulation module based on the operation priorities and dependencies defined in the strategy sequence, ensuring seamless connection between each processing stage.

[0056] In one specific implementation, the components are generated by the local Revit software, and all operations run in a non-blocking manner within the local Revit plugin environment, allowing users to monitor the processing progress in real time and intervene to confirm at key nodes.

[0057] See Figure 4 The specific processing procedure is as follows: The geometry optimization module described in S1041 uses a deep learning-driven mesh simplification algorithm to perform non-uniform simplification of high-face-count components by preserving key areas, and calls a three-dimensional geometry repair tool to correct zero-thickness or self-intersecting defects.

[0058] When processing policy sequences involving geometric lightweighting tasks, the system first loads the triangular mesh of the component to export an intermediate format. This intermediate format is then fed into a pre-deployed deep learning model for feature analysis. This model is a network jointly trained using graph convolutional networks and attention mechanisms. It has been trained on a large amount of engineering component data and can automatically identify and protect key geometric regions of engineering significance. See details... Figure 5 .

[0059] S10411 identifies key area features through a trained neural network model, preserving the mesh details corresponding to rounded corners, bolt holes, and load-bearing edges.

[0060] The model takes vertex coordinates, normal vectors, and edge connectivity of the component mesh as input and outputs an importance score map for each local region. In one specific implementation, high-fidelity regions and simplifiable regions are divided based on a score threshold. High-fidelity regions, such as equipment mounting holes, structural connection faces, and areas of abrupt curvature changes, maintain their vertex density during simplification. Simplifiable regions allow for a greater degree of topology compression. In one specific implementation, for a family of steel structure nodes, the model significantly preserves the original accuracy of its weld bevels and bolt hole areas, while drastically simplifying its large planar web areas.

[0061] S10412 performs vertex clustering and face merging operations on the mesh corresponding to the flat region to reduce the number of faces. After the key region is locked, the system calls the efficient algorithm in the underlying 3D processing library to perform non-uniform simplification on the remaining region: clustering is used to spatially group the vertices of the low curvature region, each group is merged into a single representative vertex, and the neighboring triangular faces are reconstructed; at the same time, an edge collapse strategy is applied to recursively remove redundant edges until the total number of faces meets the requirements of the target platform. In this embodiment, the number of faces needs to be reduced to less than 100,000 faces.

[0062] After simplification, the geometric errors of the previous and current versions are compared. In a specific implementation, the magnitude of the geometric error is determined by the Hausdorff distance to ensure that the maximum deviation is controlled within the allowable tolerance range.

[0063] S10413 uses a trained small discriminative model to detect zero-thickness surfaces and self-intersecting geometries, and triggers the 3D geometry repair tool to call the target interface to correct defects.

[0064] Before or after geometric processing, a lightweight hybrid discriminant model is run to identify common modeling errors. In one specific implementation, if a degenerate triangle or self-intersecting polygon entity with an area close to zero is detected, a repair instruction package is automatically generated, invoking the Boolean operation engine, offset reconstruction tool, or topology reorganization function in the interface for automatic repair.

[0065] In one specific implementation, for self-intersecting shells, the system attempts to decompose them into independent closed bodies and then re-merge them; for invalid stretched bodies, it regenerates the contour path and performs a safe stretching operation. The repair results are fed back to the digital twin archive in real time, updating its geometric health indicators.

[0066] S1042 The data encapsulation module uses a semantic mapping model to intelligently match the custom parameters of the component to the standard attribute set of the target platform and bind the corresponding entity and attribute set.

[0067] When a strategy sequence indicates the need for parameter standardization or IFC encapsulation, the data encapsulation module is activated. This module is responsible for transforming the component's proprietary parameter system into a semantic structure that conforms to the target platform's specifications. The core of the data encapsulation module relies on a finely tuned natural language processing model. This model possesses the ability to understand engineering semantics and establish accurate mappings between different naming conventions.

[0068] See Figure 6 The specific matching process is as follows: S10421 uses a natural language processing model to parse the semantic meaning of custom parameters of components and maps non-standard parameter names to the standardized attribute fields corresponding to the target platform.

[0069] In one specific implementation, field names are extracted from the component parameter list. In this implementation, these field names include width, Fire_Rating_Value, and material fire resistance rating. After text normalization, they are input into a semantic matching model. This model, based on the Sentence-BERT architecture, is trained on a terminology alignment corpus in the Building Information Modeling (BIM) domain and outputs the most likely target attribute candidates and their confidence scores. In this implementation, fire resistance time is mapped with high confidence to the IFC standard attribute value IfcFireRating.PeriodOfResistance, while the manufacturer model is associated with the attribute value Pset_Manufacturer.Common.ModelNumber. For ambiguous or polysemous parameters, the front end provides multiple candidate options for the user to choose from and records the selection behavior for subsequent model optimization.

[0070] S10422 matches suitable container family templates from a preset template library based on the complexity index of component type and nested hierarchy.

[0071] Based on the component classification information and structural complexity score extracted in S102, the optimal encapsulation scheme is retrieved from a container family template library, either locally or in the cloud. In one specific implementation, the template library contains various standardized shell structures, such as simple geometric containers, composite containers with nesting support, and MEP-specific smart families. The matching algorithm uses weighted similarity calculation, prioritizing templates with high compatibility and low maintenance costs. Once selected, the original component is automatically embedded into the template, inheriting its IFC classification and attribute set framework.

[0072] S10423 binds the component to the selected container family template and associates it with the corresponding entity type and attribute set.

[0073] In the Revit software implementation, during the binding process, the system creates a new FamilyInstance via the Revit API, sets its category to the target IFC entity type, and batch-writes the mapped standard parameter values. Simultaneously, necessary attributes are automatically attached to ensure that data integrity checks on the target platform are met.

[0074] The final generated components are not only geometrically compliant and structurally clear, but also possess a complete semantic tagging system, allowing them to be directly uploaded to the BIM collaboration platform and correctly recognized and referenced by other professional models. After the entire adaptive processing workflow is completed, the system generates a processing report, recording various operation logs, before-and-after comparison data, and changes in risk scores, for user review and archiving.

[0075] Please see Figure 7 , Figure 7 This is a schematic diagram of a building information model compliance processing device based on artificial intelligence, provided as an embodiment of this application. Figure 7 As shown, the processing device 200 includes: The standard document parsing module 210 is used to receive standard documents published by the target platform, perform semantic parsing on the unstructured text in the standard documents using a natural language processing model, and generate a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints, and data mapping relationships. The standard document parsing module 210 is further configured to extract text blocks and their hierarchical structure from the BIM standard documents, preserving the contextual relationships between chapters, clauses, and tables; apply a fine-tuned multimodal natural language processing model to perform joint semantic parsing on the extracted text, identifying the subject objects, constraint types, and target attributes in the normative statements; construct a heterogeneous graph containing standard clauses, component types, parameter names, entity and constraint condition nodes, connect each node through semantic relationship edges, and store them in a graph database in the form of triples; when a new version of the standard document is received, a comparison process is triggered to identify newly added, modified, or obsolete clauses and update the graph node status and dependency paths.

[0076] The component file generation module 220 is used to respond to the loading operation of components in the building information model, to perform image representation of the three-dimensional geometry of the components and extract geometric feature vectors, and to model the nested hierarchical structure of the components as a topological graph composed of nodes and edges, to identify abnormal reference relationships and structural complexity indicators, and to generate a digital twin file of the components containing geometric attributes, parameter lists and risk scores; wherein, the component file generation module 220 is further configured to render the components as lightweight three-dimensional images, use convolutional neural networks to analyze the images to identify the component type and estimate the number of faces; represent the components and their sub-components as nodes in the graph, and the reference relationships as edges, and use graph neural networks to detect reference cycles and nesting depth.

[0077] The compliance comparison and decision module 230 is used to match and compare the digital twin file of the component with the compliance knowledge graph, determine the potential compliance conflict points of the component on the target platform, and dynamically generate a sequence of processing strategies for the component based on the comparison results and preset decision logic.

[0078] The adaptive processing execution module 240 is used to schedule the corresponding geometry optimization module and data encapsulation module to perform adaptive processing according to the processing strategy sequence. The geometry optimization module employs a deep learning-driven mesh simplification algorithm to perform non-uniform simplification of high-face-count components while preserving key areas, and calls a 3D geometry repair tool to correct zero-thickness or self-intersecting defects. Specifically, it identifies key area features through a trained neural network model, preserving mesh details corresponding to fillets, bolt holes, and load-bearing edges; it performs vertex clustering and face merging operations on the mesh corresponding to flat areas to reduce the face count; and it uses a trained small discriminant model to detect zero-thickness defects. The thickness surface and self-intersecting geometry are defined, and a 3D geometry repair tool is triggered to call the target interface for defect correction. The data encapsulation module uses a semantic mapping model to intelligently match the component's custom parameters to the standard attribute set of the target platform and bind the corresponding entities and attribute sets. Specifically, the semantic meaning of the component's custom parameters is parsed through a natural language processing model, and non-standard parameter names are mapped to the standardized attribute fields corresponding to the target platform. Based on the complexity index of the component type and nested hierarchical structure, a suitable container family template is matched from the preset template library. The component is bound to the selected container family template and associated with the corresponding entity type and attribute set.

[0079] Furthermore, the processing device 200 also includes a log feedback and model update module 250, which is used to respond to the execution result of the component upload operation, record the processing strategy sequence corresponding to the component and the success or failure status of the upload, and upload the encrypted processing log to the cloud training platform under the user authorization condition; input the processing log of the failed upload component as a new training sample into the joint training process of natural language processing model and graph neural network; adjust the call weight of each module in the processing strategy sequence based on the successfully uploaded component processing log; and trigger the incremental update of compliance knowledge graph and decision logic after receiving a new version of standard document or accumulating a certain number of failure samples.

[0080] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 300 includes a processor 310, a memory 320, and a bus 330. The memory 320 stores machine-readable instructions executable by the processor 310. When the electronic device 300 is running, the processor 310 and the memory 320 communicate via the bus 330. When the machine-readable instructions are executed by the processor 310, they are used to implement the steps of the AI-based building information model compliance processing method, specifically including: receiving a standard document published by the target platform; using a natural language processing model to perform semantic parsing on the unstructured text in the standard document and generating a structured compliance knowledge graph; generating a digital twin file of the component in response to a component loading operation; executing a compliance comparison and decision generation processing strategy sequence; scheduling the geometry optimization module and the data encapsulation module to complete adaptive processing; and providing feedback logs based on the uploaded results to support model iterative updates.

[0081] This application also provides a computer-readable storage medium storing a computer program. When the computer program is run by a processor, it performs steps of an AI-based building information model compliance processing method, including but not limited to: semantic parsing and knowledge graph construction of standard documents, generation of component digital twin archives, compliance conflict identification, dynamic generation of processing strategies, collaborative execution of geometric optimization and data encapsulation, and log collection and incremental model update process based on actual upload results.

[0082] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0083] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0084] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0085] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0086] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A compliance processing method for building information modeling based on artificial intelligence, characterized in that, Receive standard documents published by the target platform, use a natural language processing model to perform semantic parsing on the unstructured text in the standard documents, and generate a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints, and data mapping relationships; In response to the loading operation of components in the Building Information Model, the three-dimensional geometry of the components is represented graphically and geometric feature vectors are extracted. At the same time, the nested hierarchical structure of the components is modeled as a topological graph composed of nodes and edges. Anomalies in reference relationships and structural complexity indicators are identified, and a digital twin profile of the components containing geometric attributes, parameter lists, and risk scores is generated. The components are rendered as lightweight three-dimensional images, and convolutional neural networks are used to analyze the images to identify the component type and estimate the number of faces. Components and their sub-components are represented as nodes in a graph, and reference relationships are represented as edges. A graph neural network is used to detect reference cycles and nesting depth. The component's digital twin file is matched and compared with the compliance knowledge graph to determine the component's potential compliance conflict points on the target platform. Based on the comparison results and a preset decision logic, a processing strategy sequence for the component is dynamically generated. The operation priorities and dependencies defined in the processing strategy sequence are used to dynamically coordinate the calling order and parameter configuration of the geometry optimization module and the data encapsulation module. According to the processing strategy sequence, the corresponding geometry optimization module and data encapsulation module are scheduled to perform adaptive processing. The geometry optimization module identifies key area features through a trained neural network model and retains the mesh details corresponding to rounded corners, bolt holes and load-bearing edges. Vertex clustering and patch merging operations are performed on the mesh corresponding to the flat area to reduce the number of faces; a trained small discriminative model is used to detect zero-thickness surfaces and self-intersecting geometry, and the 3D geometry repair tool is triggered to call the target interface to correct defects; The data encapsulation module parses the semantic meaning of the custom parameters of the components using a natural language processing model, maps non-standard parameter names to the standardized attribute fields corresponding to the target platform; matches suitable container family templates from the preset template library based on the complexity index of the component type and nested hierarchy structure; binds the component to the selected container family template, and associates the corresponding entity type and attribute set.

2. The compliance processing method for building information model based on artificial intelligence according to claim 1, characterized in that... The adaptive processing performed by the geometry optimization module and data encapsulation module according to the processing strategy sequence includes: The geometry optimization module uses a deep learning-driven mesh simplification algorithm to perform non-uniform simplification of high-face-count components, preserving key areas, and correcting zero-thickness or self-intersection defects. The data encapsulation module uses a semantic mapping model to intelligently match the custom parameters of the components to the standard attribute set of the target platform and bind the corresponding entities and attribute sets.

3. The compliance processing method for building information modeling based on artificial intelligence according to claim 1, characterized in that, The standard document received from the target platform is used to perform semantic parsing of the unstructured text in the standard document using a natural language processing model, and a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints, and data mapping relationships is generated, including: Extract text blocks and their hierarchical structure from the standard document, preserving the contextual relationships between chapters, clauses, and tables; A fine-tuned multimodal natural language processing model is applied to perform joint semantic parsing on the extracted text to identify the subject, constraint type, and target attribute in normative statements; The extracted results are constructed into a heterogeneous graph containing standard clauses, component types, parameter names, entity and constraint nodes. Each node is connected by semantic relation edges and stored in the graph database in the form of triples. When a new version of the standard document is received, a comparison process is triggered to identify newly added, modified, or obsolete clauses and update the graph node status and dependency paths.

4. The compliance processing method for building information modeling based on artificial intelligence according to claim 1, characterized in that, In response to the execution result of the component upload operation, the processing strategy sequence corresponding to the component and the upload success or failure status are recorded, and the encrypted processing log is uploaded to the cloud training platform under the user's authorization. The logs of failed upload component processing are used as new training samples to be input into the joint training process of natural language processing model and graph neural network; Adjust the call weight of each module in the processing strategy sequence based on the successfully uploaded component processing logs; Upon receiving a new version of the standard document or accumulating a certain number of failure samples, an incremental update of the compliance knowledge graph and decision-making logic is triggered.

5. A building information model compliance processing device based on artificial intelligence, characterized in that, include: The standard document parsing module is used to receive standard documents published by the target platform, use a natural language processing model to perform semantic parsing on the unstructured text in the standard documents, and generate a structured compliance knowledge graph containing mandatory parameter requirements, naming rules, geometric constraints and data mapping relationships. The component archive generation module is used to respond to the loading operation of components in the building information model, to perform image representation of the three-dimensional geometry of the component and extract geometric feature vectors, and to model the nested hierarchical structure of the component as a topological graph composed of nodes and edges. It identifies abnormal reference relationships and structural complexity indicators, and generates a digital twin archive of the component containing geometric attributes, parameter list and risk score. The component is rendered as a lightweight three-dimensional image, and a convolutional neural network is used to analyze the image to identify the component type and estimate the number of faces. Components and their sub-components are represented as nodes in a graph, and reference relationships are represented as edges. A graph neural network is used to detect reference cycles and nesting depth. The compliance comparison and decision module is used to match and compare the digital twin file of the component with the compliance knowledge graph to determine the potential compliance conflict points of the component on the target platform, and dynamically generate a processing strategy sequence for the component based on the comparison results and preset decision logic; the operation priority and dependency relationship defined in the processing strategy sequence dynamically coordinate the calling order and parameter configuration of the geometry optimization module and the data encapsulation module. An adaptive processing execution module is used to schedule the corresponding geometry optimization module and data encapsulation module to perform adaptive processing according to the processing strategy sequence; the geometry optimization module identifies key area features through a trained neural network model and retains the mesh details corresponding to rounded corners, bolt holes and load-bearing edges; Vertex clustering and patch merging operations are performed on the mesh corresponding to the flat area to reduce the number of faces; a trained small discriminative model is used to detect zero-thickness surfaces and self-intersecting geometry, and the 3D geometry repair tool is triggered to call the target interface to correct defects; The data encapsulation module parses the semantic meaning of the custom parameters of the components using a natural language processing model, maps non-standard parameter names to the standardized attribute fields corresponding to the target platform; matches suitable container family templates from the preset template library based on the complexity index of the component type and nested hierarchy structure; binds the component to the selected container family template, and associates the corresponding entity type and attribute set.

6. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. The machine-readable instructions are executed by the processor to perform the steps of the AI-based building information model compliance processing method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the AI-based building information model compliance processing method as described in any one of claims 1 to 4.

Images