Intelligent retrieval of substation GIM model and BIM universal format conversion method

Abstract

The application provides a substation GIM model intelligent retrieval and BIM universal format conversion method, relates to the technical field of substation data processing, and solves the problems that the existing substation GIM model is difficult to realize cross-modal accurate retrieval, high-fidelity BIM format conversion and dynamic specification knowledge fusion. The method first constructs a power field knowledge graph and extracts key features in user multi-modal query intention data, then a multi-modal retrieval model understands the key features and conducts joint reasoning in the power field knowledge graph in a manner of combining a graph neural network and a visual-linguistic pre-training model, and retrieves a target GIM model therein; the geometric topology of the target GIM model is reconstructed, the professional attributes are mapped, and a generative adversarial network is used to convert material coding into PBR material, so as to jointly generate a corresponding universal BIM model; after completing compliance verification and optimization on the universal BIM model, the model can be output to realize business application.

Description

Technical Field

[0001] This invention relates to the field of substation data processing technology, specifically to a method for intelligent retrieval of substation GIM models and conversion to BIM universal format. Background Technology

[0002] With the deep application of artificial intelligence technology in the digitalization process of the power industry, the State Grid Corporation of China has established a large-scale 3D design general model library covering substations of multiple voltage levels. These models adopt the professional GIM format and deeply integrate the characteristic parameters of power equipment with 3D geometric information, providing an important data foundation for substation design, construction, and operation and maintenance. To improve the utilization rate of massive model assets, the industry continues to explore efficient management and retrieval technologies, but existing solutions still have significant limitations in cross-modal association and format compatibility.

[0003] In the field of model retrieval, mainstream technical approaches face multiple bottlenecks. Traditional methods based on structured attribute queries rely on key-value pair matching mechanisms using preset identifiers such as equipment IDs and types. While capable of handling basic retrieval needs, their rigid relational architecture fails to resolve semantic relationships between models and unstructured documents (such as construction drawings and material lists), resulting in a lack of responsiveness to complex query commands (e.g., "locate construction drawings related to equipment under specific operating conditions"). Recent cross-modal retrieval technologies attempt to integrate multi-source data through common knowledge graphs and utilize graph convolutional networks for entity embedding learning, but they still have shortcomings in power industry scenarios: their graph construction relies excessively on geometric feature clustering of multi-view rendered images, i.e., "geometric word" recognition. However, substation equipment has highly specialized morphological structures and engineering attributes, and simple geometric feature extraction is insufficient to accurately distinguish between equipment types with similar functions. Furthermore, they lack deep coupling with the power industry's standard material system and equipment attributes. In addition, some emerging solutions focus on the search function of AI-generated model family libraries. Although they can dynamically call equipment models based on parameters, they fail to effectively connect with the State Grid's existing standardized 3D model library resources, limiting their practical application scope.

[0004] At the model conversion and generation level, AI-driven methods mostly focus on forward modeling optimization. For example, generating high-precision equipment models through multi-angle data acquisition, feature extraction, and rendering training. While these technologies improve modeling efficiency, they pay insufficient attention to the reverse conversion from existing GIM professional formats to BIM universal formats. Existing conversion tools lack deep learning capabilities to handle the complex semantic mapping between GIM-specific material coding systems and universal rendering pipelines (such as PBR material systems), resulting in converted models often losing key material representations and engineering attribute information, degenerating into geometric shells lacking engineering significance. Simultaneously, the lack of a compliance verification process makes it difficult for the conversion results to meet the requirements of power industry design and construction standards, further limiting the model's application value in cross-platform collaboration.

[0005] In summary, the current application of 3D models in substations faces three core challenges: retrieval technologies struggle to support precise semantic queries across modalities, resulting in the target model being "not found"; format conversion processes suffer from material loss and attribute fragmentation, leading to poor model conversion; and industry-standard knowledge is not dynamically integrated into the processing flow, causing data "incompatibility." There is an urgent need for a new method that can deeply integrate multi-source information, achieve high-fidelity conversion, and embed domain knowledge to connect the data chain of design, construction, and operation and maintenance, fully releasing the value of large-scale 3D model assets. Summary of the Invention

[0006] The purpose of this invention is to address the challenges of accurate cross-modal retrieval, high-fidelity BIM format conversion, and dynamic specification knowledge integration in existing substation GIM models. Therefore, this invention proposes an intelligent retrieval method for substation GIM models and a universal BIM format conversion method. This invention uses substation GIM models and AI search technology as its application platform. By integrating multimodal AI with a power knowledge graph, it improves retrieval efficiency and accuracy. Simultaneously, based on generative adversarial networks (GANs) for material transfer, it achieves high-fidelity and lossless cross-format conversion. Furthermore, based on a multimodal knowledge graph, it provides a single, reliable data source for the entire lifecycle management of design, construction, and operation and maintenance.

[0007] The present invention employs the following technical solutions to achieve its objective: A method for intelligent retrieval of GIM models and conversion to universal BIM formats for substations includes the following steps: S1. In the intelligent retrieval stage, a knowledge graph of the power field is constructed based on the general three-dimensional design model library and corresponding design specifications in the power field. S2. Receive the user's multimodal query intent data input, and use an NLP engine that integrates BERT and rule templates to parse and extract key features from the multimodal query intent data; S3. Construct a multimodal retrieval model; The multimodal retrieval model combines graph neural networks with a vision-language pre-trained model to understand key features and perform joint reasoning in the power field knowledge graph to retrieve the target GIM model and its associated files. S4. In the format conversion stage, based on the retrieved target GIM model, a geometric deep learning algorithm is used to parse and reconstruct the model's geometric topology. At the same time, the professional attributes of the target GIM model are mapped to the model's geometric topology. Then, a generative adversarial network is used to convert the material encoding of the target GIM model into PBR materials. Based on the model's geometric topology with mapped professional attributes, combined with the corresponding PBR materials, a general BIM model corresponding to the target GIM model is generated. S5. The generated general BIM model is verified in real time through the rule engine. At the same time, a reinforcement learning agent is introduced to adjust the equipment layout in the general BIM model through simulated trial and error, thereby autonomously optimizing the non-compliant items in the general BIM model.

[0008] Specifically, after step S5, the method further includes: S6. Output the verified and optimized general BIM model to external platforms and integrate it into external 3D design platforms and / or digital twin platforms for design applications; when the general BIM model is used in the platform, it presents its corresponding complete geometric information, electrical properties and PBR materials.

[0009] Specifically, in step S1, a named entity recognition model from natural language processing technology is used to extract multiple types of key entities from the corresponding design specifications; based on predefined rules in the power industry, semantic relationships between multiple types of key entities are established to construct a structured knowledge graph; subsequently, a graph neural network is used to perform deep representation learning on each model in the 3D design general model library, embedding the geometric features and topological connections of each model into the corresponding entity nodes of the structured knowledge graph, thus completing the construction of the power industry knowledge graph.

[0010] Specifically, in step S2, the multimodal query intent data is a fuzzy search request initiated by the user, which includes three structural types: natural language text, two-dimensional drawing images, and three-dimensional sample models. After integrating BERT and NLP engines, the input data is jointly encoded and parsed to extract the structured key features. The key features include device type, spatial location, connection relationship, and associated documents.

[0011] Specifically, in step S3, the multimodal retrieval model uses graph neural networks to analyze the model topology relationships in key features and uses a vision-language pre-trained model to understand the graph-text semantics in key features. This enables multi-hop reasoning and semantic similarity calculation on the knowledge graph of the power field, locating multiple target model nodes and their associated nodes that meet the conditions corresponding to the key features, and forming multiple candidate retrieval results.

[0012] Specifically, the multimodal retrieval model sorts each candidate retrieval result according to its relevance score, determines the most relevant candidate retrieval result as the target retrieval result, and uses the target GIM model and its associated files corresponding to the target retrieval result as the output of the intelligent retrieval stage.

[0013] Specifically, in step S4, a geometric deep learning algorithm is first used to parse the boundary representation and construct the entity geometric data of the target GIM model source file, and reconstruct it into a precise geometry and assembly hierarchy tree in the general BIM format to form the model geometric topology; then, based on the knowledge graph of the power industry, the professional attributes of the target GIM model are mapped to the corresponding fields of the model geometric topology in the general BIM format using entity linking.

[0014] Specifically, the generative adversarial network takes the original material encoding and rendering snapshot of the target GIM model as input, learns its visual features, and generates a complete set of material textures that conform to the physically based rendering workflow, which serve as the PBR material corresponding to the target GIM model. Finally, the PBR material is applied to the model's geometric topology to obtain the corresponding general BIM model.

[0015] Specifically, in step S5, the design specifications in the power sector are first encoded into machine-executable rules and integrated into the rule engine; the rule engine performs real-time scanning on the generated general BIM model to detect whether there are any parts of its geometric topology that violate the design specifications.

[0016] Specifically, if the rule engine detects that there are parts in the model's geometric topology that violate the design specifications, the reinforcement learning agent will adjust the equipment layout represented by the model's geometric topology, determine an optimization scheme that meets the design specifications and is close to the original general BIM model, and present the corresponding correction suggestions for the optimization scheme to the outside world or automatically execute the corresponding correction operation.

[0017] In summary, due to the adoption of this technical solution, the beneficial effects of this invention are as follows: This invention significantly improves the application efficiency of substation 3D models through the deep integration of multimodal artificial intelligence and knowledge in the power industry. In model retrieval, it completely changes the inefficient traditional model that relies on manual catalog traversal or single keyword matching, achieving a breakthrough improvement in retrieval efficiency and accuracy. Based on natural language understanding and cross-modal association analysis, the system can accurately respond to complex semantic query commands, quickly locate the target model and its associated multidimensional data assets such as construction drawings and material lists, and significantly shorten design retrieval time.

[0018] Regarding format conversion, this invention solves the bottleneck problem in the conversion process from GIM professional format to general BIM format. Through deep learning material semantic mapping mechanism and topology relationship reconstruction technology, it achieves high-fidelity lossless transfer of material properties, equipment parameters and engineering information for the first time, effectively avoiding the material loss and attribute fragmentation problems caused by traditional conversion, and ensuring that the converted model fully retains the original engineering semantics and visualization expressiveness.

[0019] Meanwhile, based on a dynamically constructed unified multimodal knowledge graph, the barriers between heterogeneous data such as 3D models, 2D drawings, and technical specifications are broken down. Any design change can be automatically synchronized to all related data entities, ensuring the consistency and uniqueness of data throughout the entire lifecycle. This provides a unified and reliable data source for design, construction, and operation and maintenance collaboration, fundamentally enhancing the efficiency of cross-disciplinary collaboration and the level of engineering quality management. Attached Figure Description

[0020] The present invention is further described in detail with reference to the following figures, which include 6 figures as follows: Figure 1 This is a schematic diagram illustrating the overall process of the method of the present invention; Figure 2 This is a schematic diagram of the process for constructing a knowledge graph in the power field in the method of this invention; Figure 3 This is a flowchart illustrating the process of processing multimodal query intent data in the method of the present invention; Figure 4 This is a schematic diagram of the process for retrieving the target GIM model in the method of the present invention; Figure 5 This is a schematic diagram of the conversion process from GIM to a general BIM model in the method of the present invention; Figure 6 This is a schematic diagram of the process for optimizing the compliance verification of a general BIM model in the method of this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] A method for intelligent retrieval of GIM models and conversion to universal BIM formats for substations. Figure 1 This document provides a brief overview of the overall process of the method, which can be viewed concurrently. The key steps of the method can be summarized as follows: S1. In the intelligent retrieval stage, a knowledge graph of the power field is constructed based on the general three-dimensional design model library and corresponding design specifications in the power field. S2. Receive the user's multimodal query intent data input, and use an NLP engine that integrates BERT and rule templates to parse and extract key features from the multimodal query intent data; S3. Construct a multimodal retrieval model; The multimodal retrieval model combines graph neural networks with a vision-language pre-trained model to understand key features and perform joint reasoning in the power field knowledge graph to retrieve the target GIM model and its associated files. S4. In the format conversion stage, based on the retrieved target GIM model, a geometric deep learning algorithm is used to parse and reconstruct the model's geometric topology. At the same time, the professional attributes of the target GIM model are mapped to the model's geometric topology. Then, a generative adversarial network is used to convert the material encoding of the target GIM model into PBR materials. Based on the model's geometric topology with mapped professional attributes, combined with the corresponding PBR materials, a general BIM model corresponding to the target GIM model is generated. S5. The generated general BIM model is verified in real time through the rule engine. At the same time, a reinforcement learning agent is introduced to adjust the equipment layout in the general BIM model through simulated trial and error, thereby autonomously optimizing the non-compliant items in the general BIM model.

[0024] This embodiment will describe the details and preferred methods of each step in the order described above.

[0025] Step S1: Construct a knowledge graph for the power sector. Based on the State Grid's 3D design general model library and related design specifications, a knowledge graph for the power sector is constructed. This graph links entities in GIM model components and unstructured documents to form a unified semantic network, providing underlying data support for intelligent retrieval. Among them, unstructured documents include construction drawings and material lists, and entities include equipment, parameters, and materials.

[0026] In this embodiment, such as Figure 2 As shown, step S1 specifically includes the following sub-steps: S1-1, Data Extraction and Entity Recognition: Using Natural Language Processing (NLP) technology, especially Named Entity Recognition (NER) models, key entities are automatically extracted from unstructured text such as State Grid design specifications, equipment nameplate parameter tables, and design specification documents; key entities include terms such as "750kV", "SF6 circuit breaker", and "tension insulator". S1-2, Relationship Extraction and Graph Construction: Based on the relationship extraction model and predefined rules in the power field, semantic relationships between entities are established to construct a structured power knowledge graph; semantic relationships include, for example, "is a type of", "connected to", "rated parameters are", etc. S1-3, Model Library Feature Embedding: Using graph neural networks, deep representation learning is performed on each model in the GIM model library, embedding its geometric features and topological connections into the corresponding entity nodes of the knowledge graph mentioned above, thereby achieving deep integration of models and knowledge.

[0027] Specifically, in constructing the knowledge graph for the power sector, multi-source heterogeneous data is first extracted from the State Grid's 3D design general model library and related design specifications. This data encompasses unstructured text such as equipment nameplate parameter tables, design specifications, construction drawings, and material lists, containing a large number of specialized entities unique to the power sector. By deploying a named entity recognition model in natural language processing, the system automatically identifies and extracts key entities with clear engineering semantics from the text, such as equipment models for specific voltage levels, insulator types, or circuit breaker specifications. These entities constitute the basic nodes of the knowledge graph, and the recognition process fully considers the complexity and ambiguity of power terminology, ensuring that core concepts are accurately captured.

[0028] Furthermore, based on predefined semantic rules and relation extraction models in the power sector, a structured network of relationships between entities is established. The semantic rules originate from the technical logic system of the State Grid design specifications, clearly defining the functional connections, parameter dependencies, and classification levels between equipment. The relation extraction model automatically constructs semantic relationships between entities, such as "belongs to a certain type of equipment," "connected to a certain system," and "has a certain rated parameter," by analyzing the textual context. This process integrates scattered entities into a knowledge network with engineering logic, forming a semantic framework that supports intelligent reasoning and effectively solving the problem of semantic fragmentation between unstructured data.

[0029] Specifically, this implementation method employs graph neural networks for deep feature learning of GIM model components in a general 3D design model library. The geometric structural features and topological connections of each GIM model are mapped to corresponding entity nodes in the knowledge graph using a graph embedding algorithm. For example, the spatial attributes of a circuit breaker model, such as its geometric dimensions and interface locations, are fused with the "circuit breaker" entity node represented in the knowledge graph. This deep fusion mechanism ensures that the knowledge graph not only contains semantic relationships at the textual level but also carries the geometric and topological information of the 3D model, providing a cross-dimensional data association foundation for subsequent multimodal retrieval. The resulting knowledge graph constitutes a unified semantic network for the power industry, making the implicit relationships between equipment parameters, design specifications, and 3D models explicit, providing underlying data support for multimodal joint reasoning in the intelligent retrieval stage.

[0030] Step S2: Multimodal query intent understanding based on deep learning. Receive retrieval requests initiated by users in natural language, sketches, or example models; use an NLP engine that integrates BERT and rule templates to parse the query intent and extract key features; key features include equipment type "main transformer," parameter "750kV," and spatial relationship "located on the north side," etc.

[0031] In this embodiment, such as Figure 3 As shown, step S2 specifically includes the following sub-steps: S2-1, Multimodal input reception; receiving fuzzy search requests initiated by users in the form of natural language text, uploaded two-dimensional drawing images, or three-dimensional reference models; S2-2, Intent Parsing and Structuring: A multimodal understanding model integrating pre-trained language models (such as BERT / ERNIE) and visual encoders (such as ResNet / ViT) is used to jointly encode and parse the input; fuzzy queries are parsed into structured, machine-understandable query conditions; for example, the original fuzzy query is "find the circuit breaker connected to the main transformer on the north side and its foundation diagram", and the parsed query conditions are: Equipment type: circuit breaker; Spatial location: north side; Connection relationship: main transformer; Associated document: foundation construction drawing.

[0032] Specifically, the system employing the method of this embodiment receives fuzzy search requests initiated by users in various forms, such as natural language descriptions, hand-drawn sketches, or uploaded reference models. These multimodal inputs may contain incomplete equipment names, vague spatial location descriptions, or non-standard engineering terminology, such as "the circuit breaker with a bracket near the main transformer." To accurately understand such complex intentions, this embodiment employs a multimodal parsing engine that integrates a pre-trained language model and a visual encoder. The language model captures core elements such as equipment type and technical parameters from the text through deep semantic analysis, while the visual encoder encodes the geometric features of the sketch or example model, recognizing the topological relationships and spatial constraints within them.

[0033] During the parsing process, the language model and visual encoder perform collaborative feature alignment. For example, when a user describes "circuit breaker connecting to the main transformer on the north side" along with a location sketch, spatial relation words such as "north side" and "connect" in the text are matched with the coordinate features of the sketch using vector matching. At the same time, entities such as "circuit breaker" and "main transformer" are mapped to the standardized equipment terminology library in the model database. This cross-modal feature fusion effectively eliminates ambiguity in natural language and transforms fragmented input into structured query conditions that can be processed by the machine.

[0034] The final output contains retrieval instructions with clear engineering semantics, and its structure is so sophisticated that it can be directly integrated with the reasoning engine of the knowledge graph. The original fuzzy request is transformed into a combination of multi-dimensional constraints such as device type, spatial location, connection relationship, and related documents, ensuring the completeness of the user's intent and the technical accuracy, thus laying the foundation for subsequent joint retrieval of the knowledge graph.

[0035] Step S3: Knowledge Graph-Driven Cross-Modal Semantic Retrieval. The structured features parsed in Step S2 are input into the multimodal retrieval model. This model integrates graph neural networks to analyze the topological relationships of the model and combines a vision-language pre-trained model to understand the semantics of the graph and text. Joint reasoning is performed in the knowledge graph to accurately retrieve the target GIM model and all its associated construction drawings, bills of materials, and technical documents.

[0036] In this embodiment, such as Figure 4 As shown, step S3 specifically includes the following sub-steps: S3-1, Graph Reasoning and Semantic Matching: The structured query conditions obtained in step S2 are used to perform multi-hop reasoning and semantic similarity calculation on the knowledge graph constructed in step S1 under the drive of graph neural network, to accurately locate multiple target model nodes that meet the conditions and their associated nodes, and to form multiple candidate search results. S3-2, Result Sorting and Presentation: The multimodal retrieval model sorts the candidate retrieval results according to their relevance scores, determines the most relevant candidate retrieval results as the target retrieval results, returns the target GIM model corresponding to the target retrieval results, and automatically associates and packages all related digital assets such as construction drawings and material lists as the output of the intelligent retrieval stage.

[0037] Specifically, in the knowledge graph-driven retrieval process, structured query conditions are input into a multimodal retrieval model for deep semantic reasoning. This model first utilizes a graph neural network to traverse entity nodes and relationship paths within the knowledge graph, capturing topological associations between devices through a multi-hop reasoning mechanism. For example, when the query involves "circuit breakers connected to the main transformer," the system automatically traces along the "connected to" relationship chain in the graph to the main transformer node, filters all circuit breaker models with this topological relationship, and simultaneously identifies associated digital assets such as foundation construction drawings. This graph-based reasoning capability effectively uncovers the implicit engineering logic chains between entities, avoiding the limitations of traditional keyword matching.

[0038] The multimodal retrieval model integrates the semantic understanding capabilities of a vision-language pre-trained model. For spatial descriptions in the query conditions, such as "north side," or geometric features, such as "with support," the visual encoder calculates the similarity between the extracted directional features and the geometric attributes of the model embedded in the knowledge graph. The language model performs semantic alignment of device parameters, ensuring that technical terms such as "750kV" accurately match the standardized parameters in the knowledge graph. This implementation uses a joint computation mechanism of text and image channels to enable the retrieval model to simultaneously consider topological connectivity accuracy and spatial layout rationality, forming multi-dimensional retrieval constraints.

[0039] Finally, a relevance fusion algorithm is used to intelligently rank the candidate results. The system comprehensively evaluates multiple indicators, including the topological matching degree output by the graph neural network, the geometric similarity calculated by the visual-language model, and parameter consistency, to generate a relevance score for each candidate target. The highest-ranked target model and its associated technical documents, bill of materials, and other digital assets are automatically packaged and output. This process not only returns a precisely retrieved 3D model but also provides complete associated documents required for construction, forming a digital asset package usable for the project, significantly improving the efficiency of retrieving and reusing design data.

[0040] Step S4: AI-powered high-fidelity conversion from GIM to general BIM format. The core of this step lies in using AI to solve the problem of information loss during format conversion. It uses geometric deep learning algorithms to losslessly analyze and reconstruct the model's geometric topology, achieves accurate mapping of professional attributes based on knowledge graph entity linking technology, and employs Generative Adversarial Networks (GANs) to automatically convert the GIM material encoding into PBR materials containing channels such as metallicity and roughness. Ultimately, it generates a general BIM model (such as IFC / FBX) with excellent visual effects and engineering information, achieving true high-fidelity conversion.

[0041] In this embodiment, such as Figure 5 As shown, step S4 specifically includes the following sub-steps: S4-1, Geometric and Topological Structure Analysis and Reconstruction: Using geometric deep learning algorithms to analyze the boundary representation of GIM source files and construct entity geometric data, and non-destructively reconstruct them into accurate geometry and assembly hierarchy trees in general BIM formats such as FBX. S4-2, Intelligent mapping of attributes and metadata: Based on the knowledge graph constructed in step S1, the professional attributes embedded in the GIM, such as device_id and rated voltage, are automatically and accurately mapped to the corresponding fields in the target BIM format through entity linking technology. S4-3. Material Transfer Based on Generative Adversarial Networks (GANs): Train a dedicated conditional generative adversarial network, such as Pix2PixHD / CycleGAN, whose generator takes the original material encoding and rendering snapshot of the GIM as input, learns its visual features, and generates a complete set of material textures that conform to the physically based rendering (PBR) workflow. These textures are then applied to the output FBX model to achieve photorealistic visual effects. The PBR workflow includes base color, metallicity, roughness, and normal maps.

[0042] Specifically, during the format conversion process, the system applying the method of this embodiment first performs deep analysis and reconstruction of the geometric structure of the target GIM model. Through a geometric deep learning algorithm, boundary representation data and structural entity geometric information are accurately extracted from the source file and converted into standardized geometries in the universal BIM format. This algorithm maintains the original assembly hierarchy while non-destructively reconstructing the topological connection structure of the 3D model, ensuring that key engineering features such as bolt hole locations and equipment interfaces are fully preserved after conversion. This deep learning-based reconstruction mechanism of this embodiment effectively avoids the geometric distortion problems caused by format differences in traditional conversion tools.

[0043] The transfer of professional attributes relies on the semantic association capabilities of knowledge graphs. The system automatically identifies professional metadata such as equipment numbers and electrical parameters embedded in the GIM model, and establishes a precise mapping with standardized fields in the target BIM format through entity linking technology of the knowledge graph. For example, the rated voltage parameter of a circuit breaker model is linked to the electrical attribute set of the BIM model through the "rated parameter is" relationship chain in the knowledge graph, achieving semantic-level lossless transfer. The mapping method based on knowledge reasoning ensures that the converted model not only retains its geometric shape but also fully embodies the technical characteristics of the circuit equipment.

[0044] As a preferred embodiment of this approach, to address the visual fidelity requirements of material conversion, a generative adversarial network (GAN) is employed to achieve intelligent transfer of material encoding. The generator network takes a visual snapshot and encoding parameters of the original material as input, learns the generation rules of physically rendered materials through adversarial training, and automatically outputs a complete set of PBR material parameters, including primary color maps, metallicity channels, and roughness textures. This process pays particular attention to the unique material characteristics of electrical equipment, such as the reflective effect of the glaze on porcelain insulators or the texture of the oxide layer on metal components, ensuring that the converted model presents a physical appearance consistent with the real object in the 3D view. The final general-purpose BIM model, integrating geometric structure, professional attributes, and highly realistic materials, can be directly applied to design collaboration platforms and construction management systems.

[0045] Step S5: Automatic Compliance Verification and Optimization Based on Rule Engine and Reinforcement Learning. This step aims to ensure that the output model complies with mandatory industry standards. It uses a rule engine to verify the model's safety clearance, installation space, and other aspects in real time, and introduces a reinforcement learning agent to autonomously optimize for any violations. The reinforcement learning agent learns how to fine-tune the equipment layout to meet all regulatory constraints through simulated trial and error, thereby automatically outputting compliant optimization solutions or correction suggestions, transforming traditional time-consuming manual verification into a highly efficient automated process.

[0046] In this embodiment, such as Figure 6 As shown, step S5 specifically includes the following sub-steps: S5-1, Real-time Standard Verification: Industry design standards are encoded into machine-executable rules and integrated into the rule engine; the converted BIM model is scanned in real time to detect whether it violates mandatory clauses such as safety clearance and installation space. S5-2, AI-assisted correction: For detected violations, a layout optimization agent is trained using reinforcement learning algorithms. This agent continuously tries to fine-tune the device position and orientation to find an optimization scheme that satisfies all specification constraints and is closest to the original design, and provides correction suggestions or automatically executes corrections.

[0047] Specifically, the converted general-purpose BIM model must undergo rigorous compliance verification. The system uses a rules engine to transform industry design specifications into executable logical constraints, such as safety clearance requirements for electrical equipment and operating space standards for equipment installation. This engine performs real-time 3D spatial scanning of the model to accurately detect any violations. When it identifies a device's spacing as not meeting the specified threshold, the engine automatically marks the violation area and generates a diagnostic report, forming a preliminary verification conclusion.

[0048] In response to identified violations, the reinforcement learning agent initiates an autonomous optimization mechanism. This agent explores layout adjustment strategies through simulated trial and error, fine-tuning the position and orientation of equipment while maintaining the original design intent. After each adjustment, the agent evaluates its compliance score based on real-time feedback from the rules engine, gradually learning the optimal solution that satisfies multiple constraints. For example, when the safety clearance between a circuit breaker and an adjacent structure is insufficient, the agent attempts to translate or rotate the equipment multiple times, ultimately finding a solution that both meets clearance specifications and minimizes design changes.

[0049] The optimization results can directly output a revised compliance model or generate operational suggestions for designers' reference. The entire process transforms the traditional spatial verification that relies on manual experience into an automated closed loop, significantly improving design efficiency and compliance assurance capabilities, and ensuring that the output model meets the mandatory technical requirements for power engineering construction.

[0050] After completing the above steps, the method of this embodiment further includes the following steps: Step S6: Result Output and System Integration. The final output is a general-purpose BIM model file containing complete geometric information, electrical properties, and high-fidelity PBR materials. The file format can be common and mature BIM file formats such as .ifc, .fbx, .glb, and .blend. At the same time, the file has the ability to seamlessly integrate with mainstream design software such as Revit, Bentley, and Bochao, mainstream 3D software such as Blender, and digital twin platforms.

Claims

1. A method for intelligent retrieval of GIM models and conversion to BIM universal format for substations, characterized in that, Includes the following steps: S1. In the intelligent retrieval stage, a knowledge graph of the power field is constructed based on the general three-dimensional design model library and corresponding design specifications in the power field. S2. Receive the user's multimodal query intent data input, and use an NLP engine that integrates BERT and rule templates to parse and extract key features from the multimodal query intent data; S3. Construct a multimodal retrieval model; The multimodal retrieval model combines graph neural networks with a vision-language pre-trained model to understand key features and perform joint reasoning in the power field knowledge graph to retrieve the target GIM model and its associated files. S4. In the format conversion stage, based on the retrieved target GIM model, a geometric deep learning algorithm is used to parse and reconstruct the model's geometric topology. At the same time, the professional attributes of the target GIM model are mapped to the model's geometric topology. Then, a generative adversarial network is used to convert the material encoding of the target GIM model into PBR materials. Based on the model's geometric topology with mapped professional attributes, combined with the corresponding PBR materials, a general BIM model corresponding to the target GIM model is generated. S5. The generated general BIM model is verified in real time through the rule engine. At the same time, a reinforcement learning agent is introduced to adjust the equipment layout in the general BIM model through simulated trial and error, thereby autonomously optimizing the non-compliant items in the general BIM model.

2. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 1, characterized in that, After step S5, the method further includes: S6. Output the verified and optimized general BIM model to external platforms and integrate it into external 3D design platforms and / or digital twin platforms for design applications; when the general BIM model is used in the platform, it presents its corresponding complete geometric information, electrical properties and PBR materials.

3. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 1, characterized in that: In step S1, the named entity recognition model in natural language processing technology is used to extract multiple types of key entities from the corresponding design specifications; according to predefined rules in the power field, semantic relationships between multiple types of key entities are established to construct a structured knowledge graph. Subsequently, graph neural networks were used to perform deep representation learning on each model in the general model library for 3D design, embedding the geometric features and topological connections of each model into the corresponding entity nodes of the structured knowledge graph, thus completing the construction of the knowledge graph in the power field.

4. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 1, characterized in that: In step S2, the multimodal query intent data is the fuzzy search request initiated by the user, which includes three structural types: natural language text, two-dimensional drawing images, and three-dimensional sample models. After fusing BERT and NLP engines, the input data is jointly encoded and parsed to extract the structured key features. Key features include device type, spatial location, connectivity, and associated documents.

5. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 1, characterized in that: In step S3, the multimodal retrieval model uses graph neural networks to analyze the model topology relationships in key features and uses a vision-language pre-trained model to understand the graph-text semantics in key features. This enables multi-hop reasoning and semantic similarity calculation on the knowledge graph of the power field, locating multiple target model nodes and their associated nodes that meet the conditions corresponding to the key features, and forming multiple candidate retrieval results.

6. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 5, characterized in that: The multimodal retrieval model sorts each candidate retrieval result according to its relevance score, determines the most relevant candidate retrieval result as the target retrieval result, and uses the target GIM model and its associated files corresponding to the target retrieval result as the output of the intelligent retrieval stage.

7. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 1, characterized in that: In step S4, a geometric deep learning algorithm is first used to parse the boundary representation and construct the entity geometry data of the target GIM model source file, and reconstruct it into a precise geometry and assembly hierarchy tree in the general BIM format to form the model geometric topology. Then, based on the knowledge graph of the power industry, the professional attributes of the target GIM model are mapped to the corresponding fields of the model geometric topology in the general BIM format using entity linking.

8. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 7, characterized in that: Generative Adversarial Networks (GANs) take the original material codes and rendering snapshots of the target GIM model as input, learn its visual features, and generate a complete set of material textures that conform to the physically based rendering workflow, which serve as the PBR materials corresponding to the target GIM model. Finally, the PBR materials are applied to the model's geometric topology to obtain the corresponding general BIM model.

9. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 1, characterized in that: In step S5, the design specifications for the power industry are first encoded into machine-executable rules and integrated into the rule engine. The rule engine then scans the generated general BIM model in real time to detect whether there are any parts of its geometric topology that violate the design specifications.

10. The method for intelligent retrieval of GIM models and conversion to BIM universal format for substations according to claim 9, characterized in that: If the rule engine detects that there are parts in the model's geometric topology that violate the design specifications, the reinforcement learning agent will adjust the equipment layout represented by the model's geometric topology, determine an optimization scheme that meets the design specifications and is close to the original general BIM model, and present the corresponding correction suggestions for the optimization scheme to the outside world or automatically execute the corresponding correction operation.

Images