Multi-modal engineering design intelligent generation method and system based on knowledge graph and RAG

Abstract

The invention discloses a knowledge graph and RAG-based multi-modal engineering design intelligent generation method and system. The method comprises the following steps of S100, multi-modal data analysis and structured processing; s200, constructing a domain knowledge graph; s210, constructing a standardized module library; s300, performing intelligent retrieval and scheme generation; and S400, compliance verification and feedback learning are carried out. Through deep collaboration of the domain knowledge graph and the standardized module library, spanning of engineering design knowledge from'retrieval 'to'understanding and driving' is achieved, the problems of knowledge fragmentation and low reuse rate are solved, and the design efficiency, quality and automation level are remarkably improved.

Description

Technical Field

[0001] This invention relates to a multimodal engineering design intelligent generation method and system based on knowledge graphs and RAG, belonging to the interdisciplinary field of artificial intelligence and computer-aided design technology. Background Technology

[0002] The engineering design field has accumulated a large amount of scattered, multi-format historical project data over a long period, including design documents, CAD drawings, BIM models, calculation sheets, and standards. This data is mostly unstructured or semi-structured, and comes in various forms (text, graphics, models), resulting in "data silos." This leads to low knowledge reuse rates, inefficient design processes, and a heavy reliance on the personal experience of designers.

[0003] Currently, traditional computer-aided design tools primarily focus on graphic drawing and calculation, lacking the ability to mine implicit knowledge and intelligently reuse it. Although some design assistance systems based on case-based reasoning or database retrieval exist, they typically rely on precise keyword matching, struggle to handle complex semantic queries and multimodal data, and are unable to generate innovative design solutions that fit the context. Furthermore, traditional design knowledge management largely depends on document-based archiving, lacking in-depth refinement and management of design knowledge in a granular, standardized, and modular manner. This makes it impossible to support the efficient and accurate identification, retrieval, and reorganization of design elements by artificial intelligence. In particular, existing technologies generally lack the ability to generate structured results composed of standardized and parametric components that can be directly embedded into the design process, limiting the direct usability and engineering value of solutions produced by intelligent design systems.

[0004] In recent years, Retrieval-Augmented Generation (RAG) technology has made significant progress in question-answering systems and business analytics, with existing patents (such as CN120386858A) demonstrating its application in tasks such as generating text reports. Simultaneously, some methods combining RAG with Large Language Models (LLM) for AI-assisted design have emerged (e.g., CN118350110B, CN119939863A). However, existing technologies are mostly focused on processing general text or less structured business data. Their knowledge organization methods, retrieval objects, and output results all revolve around natural language text (such as business reports or question-answering texts), making it difficult to directly transfer to highly specialized, multimodal, and strongly constrained fields such as engineering design. Specifically, current technologies have the following limitations:

[0005] Firstly, in terms of the objects being processed, engineering design requires in-depth processing of multimodal data (such as CAD drawings, BIM models, specification documents, calculation sheets, etc.) and the generation of directly usable structured design results (such as system diagrams, parametric models, equipment lists, etc.). However, existing RAG solutions are still mainly based on text retrieval and generation, lacking the ability to structure and integrate multimodal information.

[0006] Secondly, in terms of knowledge representation and utilization, existing solutions mostly rely on unstructured vector retrieval or dedicated small-scale domain models, resulting in problems such as knowledge fragmentation, weak semantic connections, and insufficient generalization ability. In particular, they lack explicit definitions and deep reasoning mechanisms for engineering design entities (such as equipment, functions, and parameters), their attributes, and complex relationships. They have not constructed domain knowledge graphs that support semantic reasoning, causing the generation process to heavily rely on the general capabilities of LLMs rather than domain knowledge logic, making it difficult to guarantee the professionalism, accuracy, and interpretability of the results.

[0007] In addition, existing methods generally lack automated compliance checks and multi-disciplinary conflict detection mechanisms (such as geometric conflicts, logical conflicts, and specification compliance) in terms of the reliability and usability of the generated results, making them difficult to apply to real-world engineering scenarios.

[0008] Finally, in terms of knowledge base construction and evolution, even with some attempts to introduce knowledge graphs, there are still obvious limitations in the construction methods: relying entirely on manual construction by experts is costly and has poor scalability; while the method of automatically extracting from data often results in high knowledge noise, low standardization, and poor logical consistency.

[0009] In addition to the RAG scheme based on pure vector retrieval, existing technologies also include technical approaches such as the Model Context Protocol (MCP). MCP primarily aims to provide standardized interfaces for general-purpose large language models to call external tools, thereby extending their ability to perform specific operations (such as computation and query APIs). However, for engineering design—a task highly dependent on structured domain knowledge and complex semantic reasoning—simple action execution capabilities cannot solve the core problems of knowledge understanding and integration.

[0010] Therefore, there is an urgent need for a next-generation intelligent design assistance technology that can deeply understand engineering semantics, deeply integrate multimodal domain knowledge, and intelligently generate structured design schemes based on a standardized knowledge system. It is necessary to establish a complete end-to-end intelligent design generation framework to solve the fundamental problems mentioned above in terms of cross-domain universality, knowledge interpretability, multimodal data processing capabilities, and generation determinism. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multimodal intelligent generation method and system for engineering design based on knowledge graphs and RAG (Research-Based Engineering). This method combines RAG models with domain-specific knowledge graphs to address the pain points of data fragmentation and low knowledge reuse rates in the engineering design field. It constructs a sustainably evolving intelligent design ecosystem, aiming to deeply integrate domain knowledge graphs and standardized module libraries. Compared to simple retrieval-enhanced generation technologies, it achieves a leap from "knowledge retrieval" to "knowledge understanding and driving," solving problems such as difficulty in reusing historical design knowledge, low automation of the design process, and poor universality and interpretability of existing generative AI solutions. Its core lies in the deep synergy and bidirectional correlation between domain knowledge graphs, standardized module libraries, and retrieval-enhanced generation, ensuring that the generated solution possesses both professional accuracy and direct usability.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] This invention provides a multimodal engineering design intelligent generation method based on knowledge graphs and RAG, comprising the following steps:

[0014] S100: Multimodal data parsing and structured processing: Collect multimodal data, use parsing technology to extract design entities, attributes and relationships, generate structured data with standardized labels and store it;

[0015] S200: Domain Knowledge Graph Construction: Based on the structured data, construct a domain knowledge graph with "design entity-function-parameter-constraint" as the core for semantic representation and reasoning, and establish its dynamic update mechanism;

[0016] S210: Construction of Standardized Module Library: Based on the domain knowledge graph and the structured data, the standardized module library is extracted and constructed; the standardized module library consists of at least one reusable standardized design module with parameter-driven capabilities; each standardized design module has a unique identifier, a clear interface definition, attribute specifications, and associated design files; the standardized design module establishes a bidirectional association mapping relationship with the design entities in the domain knowledge graph;

[0017] S300: Intelligent Retrieval and Solution Generation: Analyze the design requirements of the new project, and based on the bidirectional association mapping relationship, use hybrid retrieval technology and re-ranking technology to collaboratively retrieve the most relevant design knowledge fragments and standardized design modules from the domain knowledge graph and the standardized module library, and input them into the large language model that has been pre-trained and fine-tuned with engineering design domain text and code, to generate a draft of a structured design solution composed of standardized design module identifiers and parameterized instructions;

[0018] S400: Compliance Verification and Feedback Learning: The initial draft of the structured design scheme is subjected to engineering compliance checks and human-computer interaction verification, and the final approved scheme is fed back to the domain knowledge graph and the standardized module library to achieve continuous evolution.

[0019] Preferably, in S100, the multimodal data includes PDF documents, CAD drawings, BIM models, Office documents, and database tables; the parsing technology includes OCR parsing technology, NLP parsing technology, and CAD / BIM parsing technology; the design entities include physical equipment, logical modules, connecting structural components, materials, processes, or service interfaces; and the standardized labels are generated based on industry classification standards and internationally accepted coding systems.

[0020] Preferably, in S200, the domain knowledge graph construction adopts a hybrid construction method combining automated extraction and expert review. Instances, data, and relationships are automatically extracted from multimodal data using natural language processing and machine learning techniques to populate and update the domain knowledge graph.

[0021] Preferably, in S300, the hybrid retrieval technology is implemented by combining keyword-based sparse retrieval and vector embedding-based dense retrieval; the re-ranking technology is implemented through a cross-encoder model.

[0022] Preferably, in S300, the initial draft of the structured design scheme includes one or more of the following: system configuration diagram, parametric model, equipment selection list, scheme description, or calculation and analysis report; the generation of the initial draft of the structured design scheme is specifically as follows: the large language model selects and combines standardized design modules that meet the requirements from the standardized module library based on the retrieved context, and generates the initial draft of the structured design scheme composed of standardized design module identifiers and parametric instructions.

[0023] Preferably, the standardized design module includes a primary system wiring module, a secondary protection principle module, and an equipment layout module; the attribute specifications of the standardized design module include electrical parameters, physical dimensions, and applicable specification codes.

[0024] Preferably, in S400, the feedback learning includes: after the final approved scheme is reviewed, the designs with reusable value can be abstracted into new standardized design modules, assigned unique identifiers and added to the standardized module library, and the design entities, relationships and bidirectional association mapping relationships in the domain knowledge graph are updated synchronously.

[0025] The present invention also provides an intelligent engineering design generation system for implementing the intelligent engineering design generation method described above, comprising:

[0026] A multimodal data preprocessing module is used to perform S100;

[0027] The domain knowledge graph management module is used to execute S200;

[0028] A standardized module library management module is used to execute S210;

[0029] The intelligent generation engine module is used to execute S300;

[0030] The verification and feedback module is used to execute S400;

[0031] The joint knowledge base includes an engineering design database, a domain knowledge graph storage, and a standardized module library, used to store the structured data, the domain knowledge graph, and the standardized design modules; wherein, the domain knowledge graph management module and the standardized module library management module work together to maintain the bidirectional association mapping relationship.

[0032] Preferably, the intelligent generation engine module includes a requirement parsing unit, a hybrid retrieval unit, a reordering unit, and a generator; the generator is able to generate a structured design scheme containing design descriptions, configuration suggestions, and technical parameters based on the retrieved context.

[0033] Preferably, the intelligent engineering design generation system is deployed in the form of a software-as-a-service cloud platform, providing users with two access methods: a web interface and an API interface.

[0034] This invention deeply integrates AI's data processing, knowledge management, and generation capabilities into the traditional engineering design process, realizing the transformation from "data" to "knowledge" and then to "intelligent design." It ensures the professionalism, accuracy, and compliance of the generated content, representing an innovative practice of AI-enabled engineering design industry upgrading. Its beneficial effects include:

[0035] 1. Knowledge integration and efficient reuse: By deeply integrating fragmented multimodal engineering design knowledge through domain knowledge graphs, the problem of "data silos" is solved, and the reuse rate of historical experience is greatly improved.

[0036] 2. High standardization and usability of generated design results: By introducing a standardized module library that is bidirectionally linked to the knowledge graph, the final approved solution is composed of predefined and verified parametric modules (standardized design module identifiers and parametric instructions), ensuring that the results output by the intelligent engineering design generation system not only conform to the specifications, but also have high feasibility, consistency, compliance and maintainability. This greatly enhances the direct application value of the generated results, reduces the illusion risk of generative AI, and provides a reliable guarantee for serious industrial design.

[0037] 3. Accuracy and interpretability of intelligent design generation: Utilizing RAG technology, the generation process is based on accurately retrieved domain knowledge provided by the collaboration of domain knowledge graphs and standardized module libraries. This enables AI models to generate high-quality, structured design results with clear and traceable decision sources based on precise context, significantly improving design efficiency and quality, reducing over-reliance on designers' personal experience, and is fundamentally different from business analysis systems that generate text reports.

[0038] 4. Continuous self-evolution capability of the intelligent engineering design generation system: By introducing a feedback learning mechanism, the intelligent engineering design generation system can continuously absorb new knowledge, automatically optimize the domain knowledge graph, and supplement and transform standardized design modules during use, thereby achieving self-improvement and evolution and forming a virtuous cycle.

[0039] 5. Deep Knowledge Understanding and Reasoning: Through domain knowledge graphs with clear semantic relationships, the engineering design intelligent generation system can understand the complex relationships between design entities and perform logical reasoning, rather than simple text matching, thereby supporting more complex intelligent design scenarios.

[0040] 6. Domain versatility: The multimodal engineering design intelligent generation method and system based on "design entity" proposed in this invention has strong versatility and can be quickly adapted to multiple engineering design sub-domains such as electrical, chemical, and architectural engineering, breaking through the domain limitations of dedicated AI models.

[0041] 7. System Independence and Output Practicality: This invention constitutes a complete intelligent design generation solution, independent of any specific downstream execution system. The generated solutions possess high practical value, can be directly applied to design practice, or serve as high-quality input for various design processes, significantly improving design efficiency and quality.

[0042] 8. Compared with technologies such as MCP, this invention achieves semantic retrieval, deep understanding, and intelligent combination of design knowledge by constructing a deeply structured domain knowledge system (knowledge graph) and establishing a bidirectional association between parameterized standardized design modules, rather than simply calling tools. The RAG system constructed by this invention ensures that generative AI relies on verified, semantically related, and parameterized domain knowledge entities during the design process, thereby fundamentally guaranteeing the professionalism, accuracy, and reliability of the generated solutions.

[0043] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the overall process of the intelligent generation method for engineering design of the present invention.

[0046] Figure 2 This is a schematic diagram of the multimodal data parsing and structuring process of the present invention.

[0047] Figure 3 This is a schematic diagram illustrating an example of the core structure of the knowledge graph in this invention field.

[0048] Figure 4 This is a detailed flowchart of the intelligent retrieval and solution generation process of this invention.

[0049] Figure 5 This is an architectural block diagram of the intelligent generation system for engineering design of this invention.

[0050] Figure 6 This is a schematic diagram illustrating an example of the layout of the human-computer interaction interface of the present invention. Detailed Implementation

[0051] In this invention, the terms are defined as follows:

[0052] Design entity: This is a broad concept, referring to any core element in engineering design. Specifically, it manifests as physical equipment (such as switchgear, circuit breakers, pumps, beams and columns), logical modules (such as software function blocks and control loops), structural components, materials, processes, or service interfaces. Domain knowledge graphs describe these "design entities" and the functional, parameter, and constraint relationships between them, thus forming a general framework for representing engineering knowledge.

[0053] Attributes: These are the characteristics or parameters of a design entity used to describe its static or dynamic properties, such as electrical parameters (rated current), physical dimensions (length, width, height), material type, performance indicators, etc.

[0054] Relationship: refers to the semantic association between different design entities, or between a design entity and terms such as function, parameter, and constraint. It is used to express the interaction or dependency between entities, such as functional relationships (e.g., "circuit breaker control bus"), structural relationships (e.g., "contains", "connects"), and constraint relationships (e.g., "complies with specifications").

[0055] Function: refers to the purpose, function, or behavior of a designed entity, such as "the breaking function of a circuit breaker" or "the power distribution function of a busbar".

[0056] Parameters: refer to the quantitative characteristics or qualitative indicators of the performance, specifications or status of the design entity, such as "rated current (1250A)" or "dimensions (800mm×600mm)".

[0057] Constraints: These refer to the rules, specifications, or limitations that must be followed during the design process, such as "specification requirements (e.g., GB50016)" or "physical constraints (e.g., minimum safety clearance)".

[0058] Functional intent: refers to the design goals, core functional requirements, or tasks expressed in user needs, such as "automatic transfer switch (automatic switching on of backup power)".

[0059] Parametric-driven capability: refers to the ability of a standardized design module to automatically generate or modify its associated design files (such as CAD drawings, BIM models, or equipment lists) based on the input parameter values.

[0060] This invention provides complete intelligent design generation capabilities, and its output results have multiple applications: they can be directly used as a reference by designers, imported into traditional design software for further refinement, or used as an input source for a standardized module library system.

[0061] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0062] This invention provides a multimodal engineering design intelligent generation method and system based on knowledge graphs and RAG. Figure 1 This is a schematic diagram of the overall process of the intelligent engineering design generation method of the present invention. In one embodiment, the overall process of the multimodal intelligent engineering design generation method based on knowledge graph and RAG of the present invention is as follows: Figure 1 As shown. The following is combined with... Figure 1 Each step is described in detail. The multimodal engineering design intelligent generation method and system based on knowledge graphs and RAG of the present invention has wide applicability and can be widely applied to engineering fields such as power, chemical, and construction that require complex designs and have strict accuracy requirements. It can be especially applied to the field of electrical engineering design. The present invention will be described in detail using the intelligent design of substations as an example.

[0063] S100: Multimodal data parsing and structured processing.

[0064] Figure 2 This is a schematic diagram of the multimodal data parsing and structuring process of the present invention. Figure 2 This demonstrates the complete workflow of multimodal data parsing and structuring, including multimodal data input, parsing and extraction, and outputting structured data. For example... Figure 2As shown, in S100, multimodal data is collected, and using parsing techniques (including but not limited to OCR, NLP, and CAD / BIM parsing techniques), design entities, attributes (such as the rated current and dimensions of equipment) and relationships (such as the installation location and connection relationships of equipment) are extracted. Structured data with standardized labels is generated (for example, through data standardization processing (e.g., using NLP tools to identify entities and relationships and standardizing based on predefined domain ontology)), and stored in the engineering design database of the joint knowledge base.

[0065] Multimodal data (also known as multimodal engineering design knowledge) includes, for example, multi-source heterogeneous historical engineering design data, or distributed multimodal data from historical engineering projects. This data includes unstructured text (such as PDF documents and design specifications), semi-structured data (such as Excel bills of materials), and specialized engineering format files (such as CAD drawings and BIM models). More specifically, multi-source heterogeneous historical engineering design data includes, but is not limited to, PDF documents, CAD drawings, BIM models, Office documents, and database tables. Standardized labels can be generated based on industry classification standards and internationally accepted coding systems.

[0066] For example, taking the intelligent design of substations as an example, multimodal data includes, but is not limited to, primary system single-line diagrams (PDF / DWG format), secondary schematic diagrams, equipment lists (Excel), calculation sheets (Word), equipment technical manuals, relevant design specifications (GB and industry standards), as well as BIM models and IEC61850 SCL model files of intelligent electrical components (such as intelligent circuit breakers).

[0067] After collecting multimodal data, OCR, NLP, and CAD / BIM parsing technologies (such as CAD / BIM file parsing interfaces) are used to analyze the multimodal data, extracting design entities, attributes, and relationships. This generates structured data with standardized labels (e.g., entities are extracted using Named Entity Recognition (NER), relationships are identified using a relation extraction model, and standardized labels are assigned to entities and attributes based on a predefined domain ontology (especially its classification system and coding rules)). This data is then stored in the engineering design database of a joint knowledge base. Additionally, after collection, the multimodal data can be cleaned. Cleaning refers to quality verification and correction of the original data, including deduplication, missing value handling, format standardization, and error correction.

[0068] It should be added that this invention does not limit the data extraction technology, and those skilled in the art can also use other existing technologies to process multimodal data.

[0069] For example, taking the intelligent design of substations as an example, CAD parsing technology can be used to analyze DWG drawings and extract design entities such as circuit breakers, contactors, disconnectors, transformers, and busbars, as well as their connection relationships. NLP parsing technology can be used to analyze textual materials such as calculation sheets, manuals, lists, and specifications to extract key parameters, such as rated current, short-circuit current, thermal stability current, dynamic stability current, and protection settings. Dedicated parsers can be used to process SCL models or databases (e.g., XML parsers for IEC61850 SCL models, or SQL query interfaces for specific database formats) to extract the intelligent capabilities and communication parameters of the equipment. All extracted information is converted into structured data and stored in the engineering design database.

[0070] The dedicated parser can process multimodal data, including BIM models, CAD drawings, and substation configuration description language (SCL) files conforming to the IEC 61850 standard. Among these, the SCL file, as the core configuration data of intelligent substations, can be efficiently and accurately extracted by the dedicated parser from key design entities and attributes such as intelligent electronic devices, primary equipment logic models, functional definitions, and communication relationships between devices. This provides structured input for constructing a high-precision domain knowledge graph.

[0071] The joint knowledge base includes an engineering design database, a domain knowledge graph storage, and a standardized module library. These are used to store the structured data, domain knowledge graph, and standardized design modules, providing data support for the entire system. Specifically, the joint knowledge base is, for example, an integrated storage system comprising: an engineering design database storing structured data with standardized tags generated in S100; a domain knowledge graph storage typically using a graph database (such as Neo4j) to store a domain knowledge graph centered on "design entity-function-parameter-constraint"; and a standardized module library, for example, storing standardized design modules and their metadata, parameterized scripts, and associated files. These three components collectively provide data support for the system.

[0072] S200: Domain knowledge graph construction.

[0073] In S200, a domain knowledge graph is constructed based on structured data, centered on "design entity-function-parameter-constraint" for semantic representation and reasoning, and a dynamic update mechanism is established for it. The construction of the domain knowledge graph adopts a hybrid approach combining automated extraction and expert review to balance efficiency and accuracy.

[0074] The construction of a domain knowledge graph can utilize methods from existing technologies. For example, the hybrid construction method is as follows:

[0075] Automated extraction: Using natural language processing (NLP) tools, the structured data generated in step S100 is initially extracted using entity recognition and relation extraction techniques to extract "design entities" and their "functions", "parameters" and "constraints" relationships. For structured data sources such as databases, mapping tools (such as D2RQ) can be used for conversion.

[0076] Expert review: Domain experts review, correct, and enrich the results of automated extraction to ensure the accuracy and completeness of knowledge. Domain experts can manually add complex logical constraints and industry standards.

[0077] Graph Storage and Management: Finally, the finalized domain knowledge graph is stored using a graph database (such as Neo4j), and a dynamic update mechanism is established. For example, by setting rules, incremental updates of the graph are triggered when new structured data is added to the database.

[0078] Through the above steps, based on the structured data obtained in step S100, a (multi-dimensional engineering) domain knowledge graph with "design entity-function-parameter-constraint" as its core is constructed and maintained through the hybrid construction method. The domain knowledge graph forms an engineering knowledge representation framework that supports semantic representation and reasoning by describing the complex semantic relationships between design entities and their functions, parameters, and constraints. A dynamic update mechanism for the domain knowledge graph is also established, which continues to evolve and grow with the completion of new projects.

[0079] S210: Construction of standardized module libraries.

[0080] Based on domain knowledge graphs and structured data, a standardized module library is extracted and constructed. The standardized module library consists of at least one reusable engineering design unit (i.e., a standardized design module) with parameter-driven capabilities; each standardized design module has a unique identifier, a clear interface definition, attribute specifications, and associated design files; the standardized design module establishes a bidirectional association mapping relationship with the design entities in the domain knowledge graph.

[0081] This invention does not limit the construction method of the standardized module library. For example, in specific construction, it can be completed by analyzing historical project data and combining domain experience, such as by experts manually defining it or by using clustering algorithms (such as K-Means) to help identify reusable design patterns.

[0082] More specifically, it can be constructed in the following way:

[0083] Module identification and definition: By analyzing historical project data (such as historical engineering design results including drawings, models, documents, etc.), domain experts or clustering algorithms identify recurring and fully functional design scheme units, which are then defined as candidate modules.

[0084] Parametric interface definition: Define the functional interface, key parameters, and constraints for each standardized design module. For example, for an "outgoing line cabinet module", its key parameter set may include "rated current (In)", "short-circuit breaking current (Ics)" and "cabinet width (W)".

[0085] Model and File Association: Create or associate a parametric driving model (such as a script based on AutoLISP or Dynamo) and associated design files (such as DWG drawing blocks, Revit family files, and calculation sheet templates) for each standardized design module. A parametric driving model refers to a software unit within the standardized design module that can receive input parameters and automatically drive and generate or adjust specific design content through integrated business logic and algorithms. It is the core execution component of the standardized design module. A parametric driving model can be a script or program that automatically adjusts the design content based on input parameters.

[0086] Identifier allocation and storage: Assign a globally unique identifier to each standardized design module and store it in the standardized module library.

[0087] Through the above steps, based on the structured data in the domain knowledge graph and engineering design database, a highly standardized module library is extracted and constructed, consisting of at least one reusable engineering design unit (standardized design module) with parameter-driven capabilities. Each standardized design module corresponds to a reusable engineering design unit or solution template and has a unique identifier, clear interface definition, attribute specifications, and associated design files. The standardized module library establishes a bidirectional mapping relationship with the domain knowledge graph. Entity nodes in the domain knowledge graph (i.e., nodes representing design entities) can be associated with one or more specific standardized design modules. The metadata of the standardized design modules (such as functions, (performance) parameters, and (applicable) constraints) is also synchronously updated in the domain knowledge graph.

[0088] Figure 3 This is a schematic diagram illustrating an example of the core structure of the knowledge graph in this invention's field. For example... Figure 3As shown, it illustrates a multi-dimensional relationship network centered on "design entity-function-parameter-constraint" in the intelligent design of substations. A knowledge graph for the electrical design domain is constructed based on data. Entity types include: electrical equipment (design entity), model, electrical parameters (parameters), design specifications (constraints), and system schemes (functions, such as "single busbar segmented wiring"). Relationship types include: being a certain model, a module containing equipment, having parameters, conforming to specifications, and being a component of the scheme. A dynamic update mechanism for the domain knowledge graph ensures that new project data can be continuously incorporated. When the standardized module library is applied to the field of electrical engineering design, standardized design modules include, but are not limited to: primary system wiring modules, secondary protection principle modules, and equipment layout modules; the attribute specifications of these standardized design modules include electrical parameters, physical dimensions, and applicable specification codes.

[0089] Simultaneously, a standardized module library is constructed, incorporating validated typical design solutions into reusable modules, such as:

[0090] Primary system wiring modules: such as "10kV incoming line cabinet module", "PT cabinet module", "outgoing line cabinet module", and "bus tie cabinet module". Each module is a parameterized unit, containing the corresponding typical primary wiring diagram (DWG block), equipment configuration logic, and key electrical parameter ranges.

[0091] Secondary protection principle modules: such as "microcomputer protection configuration module", "metering circuit module", and "communication interface module". Each module is associated with a specific secondary schematic diagram and configuration instructions.

[0092] Equipment layout modules: such as the "KYN28A-12 switch cabinet single row layout module", which includes civil engineering requirements, cabinet dimensions, aisle width, etc.

[0093] Each module, as a parameterized unit, establishes a bidirectional association mapping with entities in the domain knowledge graph. For example, the "circuit breaker" design entity in the domain knowledge graph can be associated with all "outgoing cabinet modules" in the standardized module library that contain that type of circuit breaker.

[0094] S300: Intelligent search and solution generation.

[0095] The design requirements of the new project are analyzed, and the most relevant design knowledge fragments and standardized design modules are retrieved from the domain knowledge graph and standardized module library using hybrid retrieval and re-ranking techniques. These are then input into a large language model that has been pre-trained and fine-tuned with engineering design domain text and code to generate a preliminary draft of a structured design scheme consisting of standardized design module identifiers and parameterized instructions.

[0096] Hybrid retrieval technology combines keyword-based sparse retrieval with vector embedding-based dense retrieval. Re-ranking techniques, such as cross-encoder models, refine the scoring and ranking of the initial hybrid retrieval results (the most relevant design knowledge fragments and standardized design modules). Large language models, pre-trained and fine-tuned with engineering design text and code, are domain expert models capable of generating structured design descriptions, code snippets, or parametric modeling commands.

[0097] The generation of the initial draft of the structured design scheme involves the large language model selecting and combining standardized design modules that meet the requirements from the standardized module library based on the retrieved context. This generates an initial draft of the structured design scheme consisting of standardized design module identifiers and parameterized instructions. The initial draft of the structured design scheme may include, for example, one or more of the following: system configuration diagram, parameterized model, equipment selection list, scheme description, or calculation and analysis report.

[0098] Figure 4 This is a detailed flowchart of the intelligent retrieval and solution generation process of this invention. For example... Figure 4 As shown, the detailed process of intelligent retrieval and solution generation is as follows:

[0099] S301: Requirements Analysis and Contextual Retrieval: Receive natural language or structured requirements for new projects; analyze requirements using an NLP model to extract key design parameters, performance indicators, constraints, and functional intents; based on the analyzed new project design requirements, employ hybrid retrieval techniques to perform vector similarity matching and semantic retrieval in the domain knowledge graph and standardized module library. This hybrid retrieval technique combines keyword-based sparse retrieval with vector embedding-based dense retrieval; finally, use re-ranking techniques (models) to finely score and rank the retrieval results, selecting the Top-K design knowledge fragments and standardized design modules most relevant to the new project design requirements.

[0100] S302: Intelligent Generation of Design Scheme: The contextual design knowledge fragments retrieved in step S301 are combined with the new project design requirements to form a prompt, which is then input into the large language model. Based on the enhanced context, the large language model calls and combines standardized design modules from the standardized module library to generate a structured initial draft of a design scheme that conforms to engineering specifications and consists of standardized design module identifiers and parameterized instructions. The initial draft of the structured design scheme includes one or more of the following: a system configuration diagram of standardized design modules and parameterized instructions, a parameterized model, an equipment selection list, a scheme description, or a calculation and analysis report. This scheme can be directly used to guide the design or as input for downstream automated design systems.

[0101] Taking the intelligent design of substations as an example, in S300, the design requirements of new projects are obtained in the following way: User inputs natural language requirements: "Design a 10kV substation with two incoming lines with automatic transfer switch, one standby line, eight outgoing lines, a maximum short-circuit current of 25kA, and using KYN28 cabinets."

[0102] In S301, requirement parsing and context retrieval are, for example:

[0103] Analyze the requirements and extract key design parameters and functional intent (such as voltage 10kV, 2 incoming lines, 8 outgoing lines, Ics=25kA, cabinet type: switch cabinet KYN28-12, automatic transfer switch).

[0104] Based on the parsed new project design requirements, a hybrid retrieval technique (combining keyword and vector retrieval) is used to search the domain knowledge graph and standardized module library. For example, the most relevant "single busbar segmented with backup power automatic transfer function" scheme and its corresponding primary system wiring module, secondary protection principle module and equipment layout module combination are finally retrieved.

[0105] In S301, the intelligent generation of design schemes is as follows: Call module: [Incoming Line Cabinet Module-ML-IN-01], parameters: {Rated Current=1250A, Breaking Current=25kA}, quantity: 2; Call module: [Outgoing Line Cabinet Module-ML-FEED-02], parameters: {Rated Current=630A, Breaking Current=25kA}, quantity: 8; Generate a single-line diagram and layout diagram of the primary system.

[0106] S400: Compliance verification and feedback learning.

[0107] The initial draft of the structured design scheme undergoes engineering compliance checks and human-computer interaction verification. The final approved scheme is then fed back to the domain knowledge graph and standardized module library to enable continuous learning and optimization of the system. For example, engineering compliance checks can be performed by executing predefined specification rules based on a rule engine or by using logical reasoning with the domain knowledge graph; human-computer interaction verification is achieved by providing a graphical user interface.

[0108] Specifically, the generated initial draft of the structured design scheme will be automatically checked for compliance with the design rules and constraints in the domain knowledge graph (corresponding to the "constraint" nodes and relationships in the knowledge graph, and the rules inferred from them); engineering compliance checks can be implemented based on rule engines (such as Drools); the intelligent engineering design generation system provides a human-computer interaction interface for review, and human-computer interaction verification can be implemented through web front-end technologies (such as React / Vue); Figure 6 This is a schematic diagram illustrating an example of the layout of the human-computer interaction interface of the present invention. Figure 6An exemplary schematic diagram of a feasible interface layout is shown, which receives review, modification and confirmation feedback from designers and domain experts through a human-computer interaction interface (interface); the final adopted design scheme (final approved scheme) and data will be fed back to the engineering design database and domain knowledge graph as new knowledge samples; among them, new excellent design schemes (such as designs with reusable value) can be abstracted into new standardized design modules and incorporated into the standardized module library to realize the system's self-iterative optimization.

[0109] Taking intelligent design of substations as an example, in S400, engineering compliance checks include: verifying whether the electrical spacing, rated data, short-circuit parameters, etc., in the initial draft of the structured design scheme conform to the specifications in the domain knowledge graph, based on the specifications in the domain knowledge graph. Human-computer interaction verification includes: designers reviewing the generated scheme on the interactive interface, and adjusting or replacing certain modules.

[0110] Figure 6 This illustrates a feasible layout for a human-computer interaction interface, including, for example, a top navigation bar, a main content area, and a bottom operation area (verification and feedback area). The top navigation bar may include, for example, a menu, project name, and user information. The main content area may include, for example, a solution display area and a requirement input area. The solution display area may include, for example, a single-line diagram preview, a layout diagram preview, and an equipment list table; the requirement input area may include a natural language input box, a parsing button, and a list of key parameters. The bottom operation area may include, for example, input boxes for compliance check results and expert review comments, and operation buttons.

[0111] The adopted design scheme (final approved scheme) can be systematically learned and updated in the domain knowledge graph due to its novel modular combination method. Among them, the new design scheme that has been reviewed and confirmed to have reuse value can be abstracted into a new standardized design module, assigned a unique identifier and added to the standardized module library. The design entities, relationships and bidirectional association mapping relationships in the domain knowledge graph are updated synchronously to achieve feedback learning.

[0112] This invention also provides a multimodal engineering design intelligent generation system based on knowledge graphs and RAG. Figure 5 This is an architectural block diagram of the intelligent generation system for engineering design of this invention, as shown below. Figure 5 As shown, the multimodal engineering design intelligent generation system based on knowledge graphs and RAG of the present invention includes:

[0113] Multimodal data preprocessing module: Used to execute S100, for example, responsible for the collection, parsing, cleaning and storage of multi-source heterogeneous engineering data, and has the ability to process CAD / BIM specific format files.

[0114] Domain Knowledge Graph Management Module: Used to execute S200, such as responsible for the construction, storage, query (retrieval), visualization (e.g. through human-computer interaction interface) and dynamic updates of the domain knowledge graph.

[0115] Standardized module library management module: used to execute step S210, such as being responsible for the construction of standardized design modules, version management (dynamic updates), retrieval, and bidirectional association mapping with the domain knowledge graph.

[0116] Intelligent generation engine module: used to execute step S300.

[0117] The intelligent generation engine module includes, for example, a requirement parsing unit: used to parse the design requirements of new projects, such as user-input requirements. A hybrid retrieval unit: used to retrieve the most relevant design knowledge fragments and standardized design modules from the domain knowledge graph and standardized module library, i.e., recalling candidate knowledge from the knowledge base, such as integrating sparse and dense retrieval functions. A reordering unit: used to fine-tune and rank the retrieved most relevant design knowledge fragments and standardized design modules. A generator: a large language model pre-trained and fine-tuned with engineering design domain text and code generates a preliminary draft of a structured design scheme consisting of standardized design module identifiers and parameterized instructions based on the retrieved context.

[0118] Verification and Feedback Module: Used to execute step S400, such as being responsible for engineering compliance checks on the initial draft of the structured design scheme, providing a human-computer interaction interface for human-computer interaction verification, and managing the feedback learning data flow (system's self-iterative optimization).

[0119] Joint Knowledge Base: Includes an engineering design database, domain knowledge graph storage, and standardized module library, used to store structured data, domain knowledge graphs, and standardized design modules, providing data support for intelligent engineering design generation systems.

[0120] To more clearly illustrate the technical effects of this invention, a comparison is made with the existing RAG scheme based on pure vector retrieval, taking the main wiring design of a substation as an example.

[0121] Existing technical solution: Inputting the requirement "10kV outgoing switchgear, short-circuit current 25kA", the system retrieves several text descriptions (such as design specification paragraphs and historical scheme descriptions) containing keywords like "10kV", "outgoing switchgear", and "short-circuit current" through vector similarity. LLM (Large Language Modeling) then generates a natural language description of the design scheme based on these text fragments, potentially including a recommended circuit breaker model. However, LLM cannot guarantee whether this model precisely matches the short-circuit current or is compatible with other equipment on the same busbar (such as busbars and instrument transformers) in terms of parameters and geometry. This requires extensive manual verification and calculation by the designer.

[0122] The solution of this invention is as follows: Input the requirement of "10kV outgoing line cabinet, short-circuit current 25kA". The intelligent engineering design generation system first performs semantic retrieval in the domain knowledge graph: 1) Locates the "outgoing line cabinet" functional entity; 2) Associates it with the "circuit breaker" entity through the "module contains equipment" relationship; 3) Filters out a list of circuit breaker models with a "breaking current (Ics)" attribute ≥25kA through the "having parameters" relationship; 4) Confirms that the selected model conforms to relevant standards through the "complying with specifications" relationship. Subsequently, the intelligent engineering design generation system calls the parameterized driven model of the "outgoing line cabinet module", takes 25kA as the input parameter, and executes internally predefined business rules (e.g., IF Ics > 25kA THEN Circuit breaker model = VD4-12 / 1250-31.5), accurately outputs the specific model, and automatically generates a list of equipment containing that model, schematic diagram segments, and cabinet dimension diagram. The outgoing line cabinet module also automatically calls the conflict detection function to verify whether the bus rated current meets the requirements of the newly added outgoing line cabinet. This invention ensures the accuracy, compliance, and usability of the generated results by using semantic understanding of domain knowledge graphs and precise execution of parameterized driven models, fundamentally overcoming the shortcomings of the comparative schemes.

[0123] The following is a brief, layered description of the input and output processes of the intelligent generation method for multimodal engineering design provided by this invention, using specific examples:

[0124] 1. Data Layer (Historical Data Structure):

[0125] Input: Dispersed, multi-format historical project data (multi-source heterogeneous historical engineering design data) (such as PDF drawings, Word reports, Excel lists, CAD files, etc.).

[0126] Process: Using technologies such as OCR, NLP entity recognition, and relation extraction, unstructured text is converted into structured data. For example, design entities and parameters (attributes) such as "transformer," "model: SCB13," "capacity: 2000kVA," and "turns ratio: 10 / 0.4kV" are automatically extracted from a calculation report and labeled with standardized tags such as "equipment" and "performance parameters."

[0127] Output: A rich, standardized tagged "engineering design database" that provides raw materials for domain knowledge graphs.

[0128] 2. Application Layer (Intelligent Generation and Verification):

[0129] Input: New project design requirements (e.g., “Design a substation capable of supplying power to a 450kW Class I low-voltage load”).

[0130] process:

[0131] Retrieval: Parse the requirements, convert them into vectors, and perform high-speed retrieval in the domain knowledge graph and engineering design database to find the most relevant historical cases, design rules, and equipment parameters (e.g., retrieve three substation design documents of similar size and their equipment selection lists).

[0132] Generation: The retrieved (essential) information is used as context and combined with the current design requirements of the new project, then sent to the LLM (Large Language Model). The LLM then creates a new, standardized, structured design draft based on this information (e.g., it can output one or more of the following: power supply single-line diagram, main equipment list, layout diagram, design description, etc.).

[0133] Verification and Feedback: The intelligent engineering design generation system checks whether the parameters in the initial draft of the structured design scheme match and comply with specifications. Designers or domain experts review, modify, and finally approve the scheme on the interface. The adopted final approved scheme is automatically fed back into the database and domain knowledge graph, making the intelligent engineering design generation system increasingly sophisticated.

[0134] Output: A high-quality, compliant final approved solution, significantly improving design efficiency and quality throughout the process.

[0135] This invention provides complete intelligent design generation capabilities, and its output results have multiple applications: they can be directly used as a reference by designers, imported into traditional design software for further refinement, or used as an input source for a standardized module library system.

[0136] In another embodiment, the intelligent engineering design generation system of the present invention can be integrated with traditional software and deployed as a Software as a Service (SaaS) cloud platform. Users can access the service through a web browser or API interface, integrating it into existing PLM or CAD design processes. Mainstream CAD / BIM design software (such as AutoCAD and Revit) can obtain knowledge services in real time during the designer's drawing process by calling their provided RESTful API interfaces. For example, when placing a switch cabinet in Revit, the system can verify in real time whether its distance from surrounding components is compliant, realizing "design and verification simultaneously" and fundamentally eliminating design errors.

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multimodal engineering design intelligent generation method based on knowledge graph and RAG, characterized in that, Includes the following steps: S100: Multimodal data parsing and structured processing: Collect multimodal data, use parsing technology to extract design entities, attributes and relationships, generate structured data with standardized labels and store it; S200: Domain Knowledge Graph Construction: Based on the structured data, construct a domain knowledge graph with "design entity-function-parameter-constraint" as the core for semantic representation and reasoning, and establish its dynamic update mechanism; S210: Construction of Standardized Module Library: Based on the domain knowledge graph and the structured data, the standardized module library is extracted and constructed; the standardized module library consists of at least one reusable standardized design module with parameter-driven capabilities; each standardized design module has a unique identifier, a clear interface definition, attribute specifications, and associated design files; the standardized design module establishes a bidirectional association mapping relationship with the design entities in the domain knowledge graph; S300: Intelligent Retrieval and Solution Generation: Analyze the design requirements of the new project, and based on the bidirectional association mapping relationship, use hybrid retrieval technology and re-ranking technology to collaboratively retrieve the most relevant design knowledge fragments and standardized design modules from the domain knowledge graph and the standardized module library, and input them into the large language model that has been pre-trained and fine-tuned with engineering design domain text and code, to generate a draft of a structured design solution composed of standardized design module identifiers and parameterized instructions; S400: Compliance Verification and Feedback Learning: The initial draft of the structured design scheme is subjected to engineering compliance checks and human-computer interaction verification, and the final approved scheme is fed back to the domain knowledge graph and the standardized module library to achieve continuous evolution.

2. The intelligent generation method for engineering design according to claim 1, characterized in that, In S100, the multimodal data includes PDF documents, CAD drawings, BIM models, Office documents, and database tables; the parsing technologies include OCR parsing technology, NLP parsing technology, and CAD / BIM parsing technology; the design entities include physical equipment, logical modules, connecting structural components, materials, processes, or service interfaces; and the standardized labels are generated based on industry classification standards and internationally recognized coding systems.

3. The intelligent generation method for engineering design according to claim 1, characterized in that, In S200, the domain knowledge graph is constructed using a hybrid approach that combines automated extraction with expert review.

4. The intelligent generation method for engineering design according to claim 1, characterized in that, In S300, the hybrid retrieval technology is achieved by combining keyword-based sparse retrieval and vector embedding-based dense retrieval; the re-ranking technology is achieved through a cross-encoder model.

5. The intelligent generation method for engineering design according to claim 1, characterized in that, In S300, the initial draft of the structured design scheme includes one or more of the following: system configuration diagram, parametric model, equipment selection list, scheme description or calculation analysis report; the generation of the initial draft of the structured design scheme is specifically as follows: the large language model selects and combines standardized design modules that meet the requirements from the standardized module library based on the retrieved context, and generates the initial draft of the structured design scheme composed of standardized design module identifiers and parametric instructions.

6. The intelligent generation method for engineering design according to claim 5, characterized in that, The standardized design module includes a primary system wiring module, a secondary protection principle module, and an equipment layout module; the attribute specifications of the standardized design module include electrical parameters, physical dimensions, and applicable specification codes.

7. The intelligent generation method for engineering design according to claim 1, characterized in that, In S400, the feedback learning includes: after the final approved scheme is reviewed, the designs with reusable value are abstracted into new standardized design modules, assigned unique identifiers and added to the standardized module library, and the design entities, relationships and bidirectional association mapping relationships in the domain knowledge graph are updated synchronously.

8. An intelligent engineering design generation system for implementing the intelligent engineering design generation method as described in any one of claims 1-7, characterized in that, include: A multimodal data preprocessing module is used to perform S100; The domain knowledge graph management module is used to execute S200; A standardized module library management module is used to execute S210; The intelligent generation engine module is used to execute S300; The verification and feedback module is used to execute S400; A joint knowledge base, including an engineering design database, a domain knowledge graph storage, and a standardized module library, is used to store the structured data, the domain knowledge graph, and the standardized design modules; The domain knowledge graph management module and the standardized module library management module work together to maintain the bidirectional association mapping relationship.

9. The intelligent engineering design generation system according to claim 8, characterized in that, The intelligent generation engine module includes a requirement parsing unit, a hybrid retrieval unit, a reordering unit, and a generator; the generator generates a structured design scheme containing design descriptions, configuration suggestions, and technical parameters based on the retrieved context.

10. The intelligent engineering design generation system according to claim 8, characterized in that, The intelligent engineering design generation system is deployed in the form of a software-as-a-service cloud platform, providing users with two access methods: web interface and API interface.

Images