A large model-based building intelligent design auxiliary system and method

Abstract

This invention discloses a building intelligent design assistance system and method based on a large model, relating to the field of building intelligent design technology. The system adopts a hybrid cloud architecture, including a local subsystem deployed on the user's local machine and a cloud-based intelligent service platform deployed in the cloud. This invention possesses a design data collaboration mechanism under the hybrid cloud architecture, a specific system architecture and method for dynamic, on-demand, and context-driven collaborative scheduling between local private design data and cloud public data resources through a secure gateway. In the traditional model, designers need to actively memorize or frequently switch software / web pages to query specifications and data, often resulting in workflow interruptions. This invention, through real-time context awareness and domain-driven intent understanding based on a large model, proactively pushes specifications, material parameters, or design suggestions precisely related to the current operation, enabling designers to focus on core creative ideas and significantly improving design efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent building design technology, and in particular to an intelligent building design assistance system and method based on a large model. Background Technology

[0002] Currently, the construction industry is transforming towards digitalization and intelligentization. The widespread adoption of BIM (Building Information Modeling) technology has improved the information integration in design, but it hasn't fundamentally solved the problems of intelligent decision-making and automated knowledge application. On the other hand, artificial intelligence, especially Natural Language Processing (NLP) and Computer Vision (CV) technologies, has made significant progress in image recognition and text generation, resulting in powerful, general-purpose pre-trained large models such as DeepSeek. Cloud computing technology provides the foundation for the storage, computation, and sharing of massive amounts of data. However, how to deeply integrate general AI capabilities, cloud resource sharing, and local professional design environments to build a real-time, accurate, and professionally-savvy intelligent design assistant remains a pressing challenge for the industry.

[0003] Currently, there are some relatively similar solutions on the market, which can be divided into two categories: The first category consists of cloud-based design resource libraries and plugins. For example, some cloud platforms offer online specification queries and component library downloads, which can be quickly accessed through lightweight plugins (such as Revit plugins). However, these plugins have limited functionality, restricted to "querying" and "inserting," operating on a "person finds information" model. They cannot provide proactive intelligent assistance by "information finding people," and lack deep semantic connection with the design context (i.e., the drawing or model being edited).

[0004] The second category is design inspection software based on rules or traditional machine learning. Some software can perform compliance checks on BIM models based on predefined rules (such as fire separation distances and evacuation widths). However, their rule bases are complex to maintain, lack flexibility, struggle to handle complex regulatory clauses that require semantic understanding, and are usually limited to "post-event checks," failing to provide "real-time guidance" during the designer's drawing process.

[0005] In the current architectural design process, local design files (such as CAD drawings and BIM models) are isolated from external public data resources (such as national standard databases, new material parameter databases, and standard drawing sets). Designers need to manually switch, query, and verify between different software and platforms, resulting in data fragmentation, low information acquisition efficiency, and difficulty in ensuring the timeliness and accuracy of the data used.

[0006] Mainstream architectural design software (such as AutoCAD and Revit) are essentially efficient tools for graphic creation and data management, but they lack deep semantic understanding and intelligent assistance capabilities. They cannot understand design intent, automatically associate relevant code clauses, or perform consistency checks and logical reasoning based on multi-source information.

[0007] The architectural design process generates various modalities of data, including texts (design specifications, standards), drawings (plans, elevations, sections), and tables (bills of quantities, material lists). Current technologies lack effective automated means to perform deep semantic alignment and correlation of this heterogeneous data, resulting in information fragmentation. Modifying one part may cause inconsistencies in many other parts, leading to high costs for collaboration and verification.

[0008] General-purpose AI models (such as various pre-trained language models) lack expertise in the architectural field and cannot be directly used to handle specialized design problems. On the other hand, training a domain-specific model from scratch is extremely costly, time-consuming, and requires massive amounts of labeled data, which most design firms cannot afford. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by proposing a building intelligent design assistance system and method based on a large model.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A building intelligent design assistance system based on a large model, the system adopts a hybrid cloud architecture, including a local subsystem deployed on the user's local machine and a cloud intelligent service platform deployed in the cloud; The local subsystem includes: This design software integration plugin is used to deeply embed into mainstream architectural design software, capture the user's design operation context information in real time, and send this context information to the cloud; at the same time, it receives and parses the intelligent assistance results returned from the cloud and presents them in a graphical way in the original design software interface; Local data proxy and security gateway are used to manage on-demand, secure access to local private design data, extracting and encrypting only the minimum subset of data necessary for intelligent analysis, and ensuring secure authentication for communication with the cloud. The cloud-based intelligent service platform includes: The domain-specific large model engine is based on a general pre-trained large model and is obtained by efficiently fine-tuning the parameters using a knowledge dataset in the architectural domain. It possesses professional knowledge and reasoning capabilities in the architectural domain. The multimodal alignment and fusion module is used to map heterogeneous design data such as text, graphics, and tables to a unified semantic space through a deep learning network, thereby achieving cross-modal semantic association. The dynamic design knowledge graph management module is used to automatically build and update a knowledge graph reflecting the entities and relationships of the design project based on multimodal alignment results. The intelligent auxiliary decision engine is used to coordinate and call the above modules, perform intent understanding and reasoning verification based on the received context information, and generate specific intelligent auxiliary instructions.

[0011] Preferably, the design software integration plugin captures the design operation context by hooking the application programming interface events of the design software. The context information includes the currently active view type, the identifier of the selected primitive object, the geometric and attribute parameters of the primitive, and the user operation type.

[0012] Preferably, the data upload strategy executed by the local data agent and security gateway is incremental upload, which only extracts and uploads data fragments related to the current analysis context when a specific change in the design data is detected or in response to a cloud query request.

[0013] Preferably, the domain-specific large model engine employs one of LoRA, Adapter Tuning, or Prefix Tuning as its parameter-efficient fine-tuning technology, and the architectural domain knowledge dataset used for training includes architectural design codes, standard atlases, material library parameters, and annotated design case texts.

[0014] Preferably, the multimodal alignment and fusion module specifically includes: A text encoder is used to extract feature vectors from text data; The graphics encoder uses a graph convolutional network for vector graphics and a convolutional neural network for raster images to extract feature vectors from the graphics data. The table encoder, using the Transformer architecture, is used to extract feature vectors from tabular data. A shared projection network is used to project the feature vectors output by each of the encoders onto a unified semantic space; Additionally, the contrastive learning loss function used during the training phase is used to bring the feature distance between semantically positive sample pairs closer and push the feature distance between semantically negative sample pairs further apart.

[0015] Preferably, the dynamic design knowledge graph management module uses a graph database as the storage backend. Its knowledge graph is initially constructed based on the industry standard model of building information model and is dynamically updated during the design process according to the incremental information of design changes received.

[0016] Preferably, the intelligent auxiliary decision-making engine generates intelligent auxiliary instructions in the following types: automatically prompting for specification clauses relevant to the current design context, performing compliance pre-checks on the current design and identifying potential conflicts, and providing multi-view support when the design is modified.Figure 1 Consistent synchronization suggestions.

[0017] Preferred: A building intelligent design assistance method based on the system, characterized in that the method includes the following steps: Step S101: Capture the user's operational context in the design software in real time through the design software integration plugin of the local subsystem; Step S102: The context information and the necessary minimum data subset are encrypted and uploaded to the cloud intelligent service platform through a local data proxy and security gateway; Step S201: The intelligent auxiliary decision engine of the cloud-based intelligent service platform receives contextual information and drives the domain big model engine to understand the design intent; Step S202: Invoke the multimodal alignment and fusion module to associate the specification text, drawing information and table data related to the current context; Step S203: Query the dynamic design knowledge graph to perform design consistency checks and change impact analysis; Step S204: The intelligent auxiliary decision engine integrates information from all parties and generates intelligent auxiliary suggestions; Step S301: Return the generated intelligent assistance suggestions to the local subsystem through a secure channel; Step S302: The design software integration plugin of the local subsystem receives and parses the suggestion, and integrates it into the design software interface in a non-intrusive graphical manner to prompt the user; Step S401: When the user adopts the suggestion or makes design modifications, the plugin will upload the change information, driving the cloud-based dynamic design knowledge graph management module to update the knowledge graph.

[0018] Preferably, in step S202, the association specifically involves: calculating the similarity between the graphic element features and attribute text features involved in the current design context and the multimodal data in the cloud specification library and standard atlas library, and retrieving the entry with the highest semantic relevance. In step S203, the design consistency check includes checking whether the spatial relationships between entities in the knowledge graph meet the logical constraints of the relevant specification clauses; the change impact analysis includes traversing other related entities in the knowledge graph when an entity attribute changes, and determining whether it may cause new specification conflicts or design contradictions.

[0019] Preferably, in step S302, the non-intrusive graphical method includes highlighting relevant graphic elements in the design view, popping up prompt information boxes on the side or corner of the graphical interface, and providing a preview of modification schemes that can be adopted with one click.

[0020] The beneficial effects of this invention are as follows: This invention presents a design data collaboration mechanism under a hybrid cloud architecture. It describes a specific system architecture and method for dynamic, on-demand, and context-driven collaborative scheduling between local private design data and cloud-based public data resources through a secure gateway. In the traditional model, designers need to actively memorize or frequently switch software / web pages to query specifications and data, often resulting in workflow interruptions. This invention, through real-time context awareness and domain-driven intent understanding, proactively pushes specifications, material parameters, or design suggestions that are precisely relevant to the current operation, enabling designers to focus on core creative ideas and significantly improving design efficiency.

[0021] This invention is based on a general pre-trained large model, uses a specially constructed knowledge dataset in the field of architecture, and employs efficient parameter fine-tuning techniques for transfer learning, thereby obtaining a complete training process and technical solution for intelligent models in the field of architecture.

[0022] This invention jointly encodes text, drawings, and tables, and implements a specific algorithm for alignment in a unified semantic space through a specially designed loss function and network structure. Combining the real-time reasoning capabilities of dynamic knowledge graphs, this invention can perform compliance pre-checks and conflict warnings during the design process, effectively avoiding major design rework caused by oversights and significantly improving design quality.

[0023] This invention achieves a complete interactive process and system implementation by deeply embedding a plugin into mainstream design software, capturing the design operation context in real time, triggering backend intelligent engine calculations, and feeding the results back to the design interface in a non-intrusive, graphical manner. It seamlessly integrates intelligent assistance into the existing design process without changing the designer's original work habits, resulting in a smooth and natural user experience and greatly improving the tool's ease of use and acceptance.

[0024] This invention utilizes multimodal alignment results to automatically construct and update a knowledge graph reflecting the entities and relationships of a design project, and uses this graph to perform design consistency checks, change impact analysis, and intelligent recommendations. Attached Figure Description

[0025] Figure 1 This is a flowchart of a building intelligent design assistance method based on a large model proposed in this invention. Detailed Implementation

[0026] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0027] Example 1: A Building Intelligent Design Assistance System Based on a Large Model This system includes a local subsystem and a cloud-based intelligent service platform, as detailed below: 1. Local Subsystem 1.1 Design software integration plug-ins It is loaded directly into the process space of the main design software (such as Revit, AutoCAD) as a dynamic link library. This allows it to access the design software's internal object model with the highest privileges and lowest latency.

[0028] For Revit: Plugins are developed using the Revit.NET API. Integration with Revit is achieved by registering the IExternalCommand and IExternalApplication interfaces. The core is registering event listeners, for example: DocumentChangedEvent: Listens for any addition, deletion, or modification operation on any graphic element in the document.

[0029] ElementSelectionEvent: Listens for changes in the element selected by the user.

[0030] UIApplication.ViewActivated: Listens for user view switching (such as switching from a 2D view to a 3D view).

[0031] For AutoCAD: The plugin is developed using the AutoCAD.NET API. It is initialized by overriding the IExtensionApplication interface and listens for events such as DocumentCollection.DocumentActivated (document activation) and Editor.SelectionAdded (selection set change).

[0032] When the above event is triggered, the plugin immediately creates a "context snapshot" object. This object is structured JSON data and contains: eventType: "ElementMove" (Event Type) documentId:"ProjectABC.rvt" (Document ID) activeView: "Level1FloorPlan" (Active View) selectedElements:[{id:"12345",category:"Door",parameters:{Width:900,Height:2100,...}}] (Selects the list of graphic elements and their key parameters) additionalData:{previousPosition:[x1,y1],currentPosition:[x2,y2]} (Event-related additional data) After receiving the JSON-formatted auxiliary instructions returned from the cloud, the plugin calls the design software's graphics API for visualization rendering. For example, for highlighting prompts, it calls NewTransaction().SetElementColor(ElementId,Color.Yellow); for information boxes, it calls TaskDialog.Show("Prompt","Specification conflict detected...").

[0033] 1.2 Local Data Proxy and Security Gateway As a standalone Windows service (or Linux daemon), it communicates with plugin processes via a local loopback network (localhost) (such as gRPC or WebSocket). It is responsible for all communication with the external cloud and implements strict data security policies.

[0034] When the context information sent by the plugin is insufficient for in-depth analysis (e.g., the cloud needs to analyze the entire geometry of the current view), the gateway will initiate a query to the design software API on behalf of the plugin. However, it follows the "minimum necessary" principle; for example, it does not upload the entire BIM model, but only extracts and uploads a snapshot of the 2D vector graphics of the current view (generated using ExportToSVG ​​or a similar API), or the geometric bounding box information of specific elements.

[0035] Upon startup, the gateway uses a pre-installed digital certificate or client credentials obtained from a cloud authentication server to authenticate with cloud services.

[0036] All uploaded and downloaded data is end-to-end encrypted using the TLS 1.3 protocol.

[0037] Based on configurable rules of the gateway, sensitive information (such as project name and internal number) in the design data is automatically filtered out and replaced with anonymous identifiers.

[0038] 2. Cloud-based intelligent service platform 2.1 Domain-Specific Large Model Engine Its goal is to transform general language understanding and generation capabilities into the ability to solve professional problems in the field of architecture.

[0039] Choose a large open-source language model, such as DeepSeek-Coder-V2 or Qwen1.5-7B.

[0040] The data sources are design specifications (PDF text), standard drawing sets (such as drawing set numbers and explanatory text), and building material libraries (parameter tables in CSV / JSON format).

[0041] The aforementioned unstructured data is converted into a (Instruction, Input, Output) fine-tuning format. Using the base model itself, a large amount of dialogue data simulating designer questions and answers is generated through prompts, followed by review and correction.

[0042] We employ the LoRA technique from parameter-efficient fine-tuning. Specifically, we inject a low-rank adapter next to the Query, Key, Value, and Output projection matrices of each attention layer in the original Transformer model. During fine-tuning, all parameters of the original model are frozen, and only these newly added adapters with a very small number of parameters are trained. This reduces computational and storage costs.

[0043] 2.2 Multimodal Alignment and Fusion Module It includes: Text encoder: A text encoder that uses the large domain model itself.

[0044] Vector Graphics Encoder: The primitives (such as Line, Circle, BlockReference) in CAD / DWG drawings have hierarchical and attribute structures, which can be regarded as a graph network. Therefore, Graph Convolutional Network (GCN) or Graph Attention Network (GAT) is used to process them, treating each primitive as a node and its geometric relationships (such as connection and containment) as edges.

[0045] Raster image encoder: For images converted from view screenshots or PDF drawings, a pre-trained VisionTransformer (ViT) or ResNet-50 is used as an encoder to extract visual features.

[0046] Table Encoder: Treats each row of a table as a sequence and processes it using a lightweight Transformer encoder.

[0047] Alignment training: Construct a training dataset containing the correspondence between "text-image-graphical element-table". For example, a data sample contains: the standard clause "the clear width of the evacuation door shall not be less than 0.9 meters", a door graphic element marked with a rectangle on a floor plan, and the width parameter "900" corresponding to the door in the material list.

[0048] Data from different modalities are input into their respective encoders to obtain feature vectors Vtext, Vimage, Vvector, and Vtable. All feature vectors are then projected onto the same d-dimensional unified semantic space through a shared projection layer (typically a multilayer perceptron, MLP).

[0049] Training was performed using contrastive learning.

[0050] 2.3 Dynamically Designed Knowledge Graph Management Module This transforms structured BIM information and unstructured design context into a knowledge graph capable of logical reasoning. It defines the node types and relationships of the ontology based on the industry standard IFC. For example: Node types: IfcWall, IfcDoor, IfcSpace, RegulationClause.

[0051] Relationship types: containedIn (door containedIn room), hasProperty (wall hasProperty fire resistance rating), isGovernedBy (room isGovernedBy code clause).

[0052] When a new project is imported into the system, the BIM model (such as an IFC file or a Revit model) is parsed, and nodes and relationships are automatically created.

[0053] Design change information uploaded by the plugin is converted into graph transaction operations. For example, when a user moves a door, the corresponding operations are DELETErelation(Door_D123containedInRoom_R101) and CREATErelation(Door_D123containedInRoom_R102). These operations are executed in a graph database (such as Neo4j) using Cypher (Graph Database Query Language) statements.

[0054] The system can perform complex queries based on the graph. For example, it can check whether the opening direction of the evacuation doors in all "crowded rooms" points to the "evacuation direction". This can be achieved using the Cypher statement: MATCH(room:IfcSpace)-[:hasFunction]->(func:Function{name:"Crowded"}),(door:IfcDoor)-[:containedIn]->(room),(door)-[:hasProperty]->(prop:Property{name:"Opening Direction"})WHEREprop.value<>"Facing Evacuation Direction"RETURNdoor. This query can quickly identify all doors that do not meet the requirements.

[0055] 2.4 Intelligent Assisted Decision Engine The engine connects the capabilities of the previous modules into a complete workflow. It is itself a state-based rules engine or a lightweight workflow framework. It triggers predefined processing pipelines based on the received context event types (such as ElementMove, NewElementAdded).

[0056] The specific steps are as follows: Intent understanding: The engine packages the contextual information into a natural language question and sends it to the domain-wide model engine; Multimodal retrieval: Based on the "firewall" keyword returned by the large model, the engine calls the multimodal alignment module to perform a semantic search in the specification library, find all relevant terms, and sort them by semantic relevance.

[0057] Knowledge Graph Verification: The engine feeds the geometric information of the new wall and the retrieved regulatory clauses into the knowledge graph module. The knowledge graph module performs a spatial query to verify whether it meets the regulatory requirements.

[0058] Decision-making and generation: The engine integrates all results: if fully compliant, it may not prompt or simply confirm; if there are risks or conflicts, it generates specific auxiliary suggestions and cites the original text of the specification.

[0059] Example 2: Method Flow The implementation process of this method is illustrated by taking the example of a user changing the location of a room's doorway in Revit: Context capture and upload: When a user selects a doorway element and begins dragging, the Revit integration plugin immediately captures this action, including context information such as the element ID being manipulated (doorway), action type (move), current location coordinates, and room ID. The plugin then encrypts this context information, along with brief geometric attributes of the doorway and room, and uploads it to the cloud via a local security gateway.

[0060] Cloud-based intelligent analysis and reasoning: The cloud-based intelligent decision-making engine receives information. The engine first invokes a domain-wide model to understand the intent to "move the door." Based on this knowledge, the model outputs key specifications that this operation may involve, such as "evacuation width requirements" and "passage turning radius." Next, the engine uses a multimodal alignment module to semantically match the geometric information of the current door with the numerical requirements in the specification text. Simultaneously, the engine queries a dynamic knowledge graph to analyze the impact of the door's movement on evacuation paths and adjacent spaces.

[0061] Feedback and Interaction: The engine detected that the width of the moved doorway might not meet the minimum evacuation width requirement. Therefore, it generated a warning suggestion: "Note: The effective width of the moved doorway is 850mm, lower than the standard requirement of 900mm. Adjustment is recommended." This suggestion was returned to the Revit plugin. The plugin then highlights the doorway in the Revit interface and displays a non-modal tooltip to inform the user of this information in real time.

[0062] Knowledge graph dynamic updates: After the user adjusts the doorway width according to the suggestions and confirms the move, the plugin uploads the final new doorway position and parameters. The cloud-based knowledge graph management module automatically updates the corresponding node attributes and relationships with other entities in the graph.

[0063] Through the above methods, this invention achieves a closed loop from data perception and intelligent analysis to interactive feedback, providing full-process, real-time intelligent assistance for architectural design.

[0064] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A building intelligent design assistance system based on a large model, characterized in that, The system adopts a hybrid cloud architecture, including a local subsystem deployed on the user's local machine and a cloud-based intelligent service platform deployed in the cloud. The local subsystem includes: This design software integration plugin is used to deeply embed into mainstream architectural design software, capture the user's design operation context information in real time, and send this context information to the cloud; at the same time, it receives and parses the intelligent assistance results returned from the cloud and presents them in a graphical way in the original design software interface; Local data proxy and security gateway are used to manage on-demand, secure access to local private design data, extracting and encrypting only the minimum subset of data necessary for intelligent analysis, and ensuring secure authentication for communication with the cloud. The cloud-based intelligent service platform includes: The domain-specific large model engine is based on a general pre-trained large model and is obtained by efficiently fine-tuning the parameters using a knowledge dataset in the architectural domain. It possesses professional knowledge and reasoning capabilities in the architectural domain. The multimodal alignment and fusion module is used to map heterogeneous design data such as text, graphics, and tables to a unified semantic space through a deep learning network, thereby achieving cross-modal semantic association. The dynamic design knowledge graph management module is used to automatically build and update a knowledge graph reflecting the entities and relationships of the design project based on multimodal alignment results. The intelligent auxiliary decision engine is used to coordinate and call the above modules, perform intent understanding and reasoning verification based on the received context information, and generate specific intelligent auxiliary instructions.

2. The building intelligent design assistance system based on a large model according to claim 1, characterized in that, The design software integration plugin captures the design operation context by hooking the application programming interface events of the design software. The context information includes the currently active view type, the identifier of the selected primitive object, the geometric and attribute parameters of the primitive, and the user operation type.

3. The building intelligent design assistance system based on a large model according to claim 1, characterized in that, The local data agent and security gateway implement an incremental upload strategy, extracting and uploading data fragments related to the current analysis context only when a specific change in the design data is detected or in response to a cloud query request.

4. The building intelligent design assistance system based on a large model according to claim 1, characterized in that, The domain-specific large model engine employs one of the following parameter-efficient fine-tuning techniques: LoRA, Adapter Tuning, or Prefix Tuning. The architectural domain knowledge dataset used for training includes architectural design codes, standard atlases, material library parameters, and annotated design case texts.

5. The building intelligent design assistance system based on a large model according to claim 1, characterized in that, The multimodal alignment and fusion module specifically includes: A text encoder is used to extract feature vectors from text data; The graphics encoder uses a graph convolutional network for vector graphics and a convolutional neural network for raster images to extract feature vectors from the graphics data. The table encoder, using the Transformer architecture, is used to extract feature vectors from tabular data. A shared projection network is used to project the feature vectors output by each of the encoders onto a unified semantic space; Additionally, the contrastive learning loss function used during the training phase is used to bring the feature distance between semantically positive sample pairs closer and push the feature distance between semantically negative sample pairs further apart.

6. The building intelligent design assistance system based on a large model according to claim 1, characterized in that, The dynamic design knowledge graph management module uses a graph database as its storage backend. Its knowledge graph is initially built based on the industry standard model of building information model and is dynamically updated during the design process according to the incremental information of design changes received.

7. The building intelligent design assistance system based on a large model according to claim 1, characterized in that, The intelligent auxiliary decision engine generates intelligent auxiliary instructions in the following types: automatically prompting for specification clauses related to the current design context, performing compliance pre-checks on the current design and identifying potential conflicts, and providing synchronous suggestions to maintain consistency across multiple views when the design is modified.

8. A building intelligent design assistance method based on the system according to any one of claims 1-7, characterized in that, The method includes the following steps: Step S101: Capture the user's operational context in the design software in real time through the design software integration plugin of the local subsystem; Step S102: The context information and the necessary minimum data subset are encrypted and uploaded to the cloud intelligent service platform through a local data proxy and security gateway; Step S201: The intelligent auxiliary decision engine of the cloud-based intelligent service platform receives contextual information and drives the domain big model engine to understand the design intent; Step S202: Invoke the multimodal alignment and fusion module to associate the specification text, drawing information and table data related to the current context; Step S203: Query the dynamic design knowledge graph to perform design consistency checks and change impact analysis; Step S204: The intelligent auxiliary decision engine integrates information from all parties and generates intelligent auxiliary suggestions; Step S301: Return the generated intelligent assistance suggestions to the local subsystem through a secure channel; Step S302: The design software integration plugin of the local subsystem receives and parses the suggestion, and integrates it into the design software interface in a non-intrusive graphical manner to prompt the user; Step S401: When the user adopts the suggestion or makes design modifications, the plugin will upload the change information, driving the cloud-based dynamic design knowledge graph management module to update the knowledge graph.

9. The building intelligent design assistance method based on a large model according to claim 8, characterized in that, In step S202, the association specifically involves: calculating the similarity between the graphic element features and attribute text features involved in the current design context and the multimodal data in the cloud specification library and standard atlas library, and retrieving the entry with the highest semantic relevance. In step S203, the design consistency check includes checking whether the spatial relationships between entities in the knowledge graph meet the logical constraints of the relevant specification clauses; the change impact analysis includes traversing other related entities in the knowledge graph when an entity attribute changes, and determining whether it may cause new specification conflicts or design contradictions.

10. The building intelligent design assistance method based on a large model according to claim 8, characterized in that, In step S302, the non-intrusive graphical method includes highlighting relevant elements in the design view, popping up prompt information boxes on the side or corner of the graphical interface, and providing a preview of modification schemes that can be adopted with one click.

Images