An Automatic Semantic Annotation Method for 3D Models of Power Engineering

Abstract

This invention discloses an automatic semantic annotation method for 3D models of power engineering, relating to the field of semantic annotation technology, including the following steps: Step 1: Multi-dimensional feature extraction. A power-specific multi-dimensional feature extraction engine is constructed to simultaneously extract geometric and power-related features from various components in the 3D model of the power engineering, ultimately generating a unique multi-dimensional feature vector for each component, providing data support for subsequent differentiation of similar components; Step 2: Knowledge graph matching and similar component differentiation. A power engineering knowledge graph containing nodes and relationships such as equipment type, specifications, and functional attributes is constructed; a power semantic embedding model is trained using a graph neural network (GNN) to spatially map the multi-dimensional feature vectors of the components generated in Step 1 with the semantic vectors in the knowledge graph, accurately distinguishing similar components and avoiding mislabeling caused by human visual confusion from the source.

Description

Technical Field

[0001] This invention relates to the field of semantic annotation technology, specifically to an automatic semantic annotation method for three-dimensional models of power engineering. Background Technology

[0002] Semantic annotation methods refer to the automatic identification of feature information of target objects (such as components of a 3D model of a power engineering project) through technical means, combined with relevant professional knowledge and rules, to assign semantic tags with clear meanings and conforming to standards (such as equipment type, specifications, functional attributes, etc.) to the target objects, realizing the transformation of target objects from the physical / geometric level to the semantic level, without the need for manual annotation of each object. According to the search, a Chinese patent with publication number 1 discloses a method for segmentation and semantic annotation of a geometric mesh scene model, which includes the following steps: establishing a 3D model training set, requiring that each 3D model in the training set is a single object; The automatic scene model segmentation method uses a training set to segment the scene model into multiple objects based on a hierarchical clustering algorithm. For the classification of the segmentation results, shape features are extracted from each segmented object, and a category label is determined based on a classification algorithm. Finally, the semantics of the scene model are summarized, a set of semantic labels for each object is obtained. Compared to existing technologies, this invention has the following advantages: the automatic scene model segmentation method uses existing shape knowledge from the training set to assist in decision-making during the segmentation process, thus solving the difficult problem of handling contact objects in scene segmentation; and the semantic annotation of the scene model is more consistent with human visual perception of scenes.

[0003] Currently, when semantic annotation is used in 3D power engineering models, it mostly relies on manual work, resulting in long annotation cycles, easy omissions and errors, and high costs for complex power models. Secondly, the data sources are relatively limited, relying solely on 3D models without integrating multi-source information such as drawings, equipment lists, and specification texts in other file formats. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic semantic annotation method for three-dimensional models of power engineering projects, so as to solve the problems mentioned in the background art.

[0005] In view of the above problems, the technical solution proposed by the present invention is as follows: An automatic semantic annotation method for 3D models of power engineering projects includes the following steps: Step 1: Multi-dimensional feature extraction. Construct a multi-dimensional feature extraction engine for power engineering. Simultaneously extract geometric features and power engineering features from various components in the 3D model of power engineering. Finally, generate a multi-dimensional feature vector for each component, providing data support for the subsequent differentiation of similar components. Step 2: Knowledge Graph Matching and Similar Component Differentiation. Construct a power engineering knowledge graph containing nodes and relationships such as equipment type, specifications, and functional attributes; train a power semantic embedding model through a graph neural network (GNN), and spatially map the multi-dimensional feature vectors of components generated in Step 1 with the semantic vectors in the knowledge graph to accurately distinguish similar components and avoid mislabeling caused by human visual confusion from the source. Step 3: Multi-view collaboration and hidden component annotation. CNN+multi-view collaboration technology is used to project the 3D model into multi-view 2D images, capture the local features of hidden components and complete semantic segmentation; the segmentation results are back-projected onto the 3D model to achieve accurate annotation of hidden components, and initially solve the problem of missing annotation of hidden components. Step 4: GUID Mapping and Omission Detection. Establish a globally unique GUID mapping relationship to achieve the correspondence between 3D components and drawing elements and equipment list; by comparing the component list in the equipment list with the current AI annotation results, automatically identify unannotated omission components and issue prompts to ensure that all components are annotated without omission. Step 5: Quality Control and Optimization of Labeling. Establish a unified four-level semantic labeling system for power engineering, clarify labeling standards, and rely on the power specification verification rule library based on national and industry standards to automatically verify the labeling results, correct inconsistencies in labeling, generate labeling confidence scores, and only prompt low-confidence labeling results for manual review to reduce manual workload. Through an incremental learning mechanism, the manually corrected results are supplemented into the model training set to continuously improve the labeling accuracy and achieve precise control of labeling quality.

[0006] As a preferred technical solution of the present invention, in step 1, the geometric features include shape, size, curvature, and orientation.

[0007] As a preferred technical solution of the present invention, in step 1, the power professional features include the insulation distance corresponding to the voltage level and the equipment size parameters.

[0008] As a preferred technical solution of the present invention, in step 3, the concealed component is set as a cable in the cable trench or an underground foundation component, wherein the underground foundation component is any one or any combination of a transformer, circuit breaker, disconnector, and instrument transformer.

[0009] As a preferred technical solution of the present invention, in step 5, the labeling system includes equipment category, equipment type, specifications, functions and standards.

[0010] On the other hand, the present invention further refines the content of step 4 above, including the following steps: Step 4.1: Construct a globally unique GUID mapping relationship. Assign a globally unique GUID (Globally Unique Identifier) ​​to each component in the 3D model of the power engineering, each element in the CAD drawing, and each piece of equipment in the equipment list, and establish a one-to-one correspondence mapping relationship between 3D component, GUID, CAD element, and equipment in the equipment list. Step 4.2: Comparison of Component List and AI Annotation Results Based on the GUID mapping relationship established in Step 4.1, extract the complete component list (including all regular components and hidden components that need to be annotated) from the equipment register, and compare it with the AI ​​annotation results completed in Step 3 (including annotations of regular components and hidden components) one by one. Filter out the components that are recorded in the equipment register but not annotated in the 3D model to form a preliminary list of missing components. Step 4.3: Identification and Supplementation of Missing Components. The system performs a second check on the list of missing components. After confirming that there are no misjudgments or duplicates, the system automatically issues a supplementation prompt, clearly marking the GUID of the missing component, the corresponding equipment information, the location of the CAD element, and the approximate coordinates in three-dimensional space, guiding the staff to complete the supplementation operation of the missing components.

[0011] Compared with existing technologies, the beneficial effects of this invention are as follows: This method can effectively solve the problems of mislabeling caused by visual confusion of similar components, omission of hidden components due to perspective occlusion, and inconsistent labeling standards and results in the semantic annotation of 3D power engineering models. Through the application of multi-dimensional feature extraction, knowledge graph matching, multi-view collaborative annotation, GUID mapping verification, and a unified semantic system and standard verification mechanism, it accurately distinguishes similar components, comprehensively captures hidden components, and standardizes annotation logic. At the same time, it filters low-confidence annotation results through confidence scoring for targeted manual review, and continuously optimizes model performance by combining incremental learning, which greatly reduces the cost and workload of manual review, significantly improves the accuracy, completeness, and standardization of semantic annotation, effectively avoids problems such as human visual misjudgment, omission, and mislabeling, and efficiently realizes the transformation of 3D power engineering models from the geometric level to the professional semantic level, providing reliable support for subsequent model applications. Attached Figure Description

[0012] Figure 1 A flowchart of an automatic semantic annotation method for a three-dimensional model of a power engineering project provided by the present invention; Figure 2 The present invention provides a module information block diagram of an automatic semantic annotation method for a three-dimensional model of a power engineering project. Detailed Implementation

[0013] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Please see Figures 1-2 This invention provides a technical solution: an automatic semantic annotation method for a three-dimensional model of a power engineering project, comprising the following steps: Step 1: Multi-dimensional feature extraction. Construct a multi-dimensional feature extraction engine for power engineering. Simultaneously extract geometric features and power engineering features from various components in the 3D model of power engineering. Finally, generate a multi-dimensional feature vector for each component, providing data support for the subsequent differentiation of similar components. Step 2: Knowledge Graph Matching and Similar Component Differentiation. Construct a power engineering knowledge graph containing nodes and relationships such as equipment type, specifications, and functional attributes; train a power semantic embedding model through a graph neural network (GNN), and spatially map the multi-dimensional feature vectors of components generated in Step 1 with the semantic vectors in the knowledge graph to accurately distinguish similar components and avoid mislabeling caused by human visual confusion from the source. Step 3: Multi-view collaboration and hidden component annotation. CNN+multi-view collaboration technology is used to project the 3D model into multi-view 2D images, capture the local features of hidden components and complete semantic segmentation; the segmentation results are back-projected onto the 3D model to achieve accurate annotation of hidden components, and initially solve the problem of missing annotation of hidden components. Step 4: GUID Mapping and Omission Detection. Establish a globally unique GUID mapping relationship to achieve the correspondence between 3D components and drawing elements and equipment list; by comparing the component list in the equipment list with the current AI annotation results, automatically identify unannotated omission components and issue prompts to ensure that all components are annotated without omission. Step 5: Quality Control and Optimization of Labeling. Establish a unified four-level semantic labeling system for power engineering, clarify labeling standards, and rely on the power specification verification rule library based on national and industry standards to automatically verify the labeling results, correct inconsistencies in labeling, generate labeling confidence scores, and only prompt low-confidence labeling results for manual review to reduce manual workload. Through an incremental learning mechanism, the manually corrected results are supplemented into the model training set to continuously improve the labeling accuracy and achieve precise control of labeling quality.

[0015] Example 1: Automatic Semantic Annotation of 3D Models of Substations This embodiment applies to the automatic semantic annotation of a 3D model of a 110kV outdoor substation. The specific implementation steps are as follows: Step 1: Multi-dimensional Feature Extraction. A dedicated multi-dimensional feature extraction engine for the power industry is constructed to simultaneously extract geometric and power-related features from various components (including transformers, circuit breakers, disconnectors, instrument transformers, cables in cable trenches, transformer underground foundations, circuit breaker underground foundations, etc.) in the 3D model of the 110kV substation. Geometric features include the component's shape, size, curvature, and orientation. For example, the transformer component is extracted as a cuboid with dimensions of 5m x 3m x 4m, a surface curvature approaching 0, and a north-south orientation. Power-related features include insulation distances corresponding to voltage levels (insulation distance for 110kV equipment is not less than 1.5m) and equipment size parameters (the body size of a 110kV circuit breaker is 1.2m x 0.8m x 2.5m). Finally, a unique multi-dimensional feature vector is generated for each component, providing data support for distinguishing similar components in the future.

[0016] Step 2: Knowledge Graph Matching and Similar Component Differentiation. A power engineering knowledge graph is constructed, containing nodes and relationships related to substation equipment types, specifications, functional attributes, etc. Nodes include 110kV transformers, 35kV circuit breakers, 110kV circuit breakers, supporting insulators, and wall bushings. The relationships are: 110kV circuit breaker - belongs to - high-voltage substation equipment; 110kV circuit breaker - voltage level - 110kV supporting insulator - function - supporting conductive component. A power semantic embedding model is trained using a graph neural network (GNN). The multi-dimensional feature vectors of the components generated in Step 1 are spatially mapped to the semantic vectors in the knowledge graph, accurately distinguishing between 110kV and 35kV circuit breakers (through differences in feature vectors of insulation distance and equipment size parameters) and supporting insulators and wall bushings (through differences in feature vectors of shape and orientation), thus avoiding mislabeling caused by human visual confusion at the source.

[0017] Step 3: Multi-view Collaboration and Hidden Component Annotation. CNN+multi-view collaboration technology is used to project the substation's 3D model into multiple 2D images from top, bottom, side, and specialized perspectives such as inside the cable trench and directly above the underground foundation. CNN is used to capture local features of hidden components (cables in the cable trench, transformer underground foundations, and circuit breaker underground foundations) and perform semantic segmentation. The underground foundation components are combinations of transformers and circuit breakers. The segmentation results are then back-projected onto the 3D model to achieve accurate annotation of hidden components, initially solving the problem of missing annotations of hidden components.

[0018] Step 4: GUID mapping and missing label verification specifically includes the following sub-steps: Step 4.1: Construct a globally unique GUID mapping relationship. Assign a globally unique GUID to each component in the 3D model of the substation, each element in the CAD drawing, and each piece of equipment in the equipment list. Establish a one-to-one correspondence mapping relationship between 3D component, GUID, CAD element, and equipment in the equipment list. For example, the GUID of the 3D component of the 110kV transformer is GUID-001, the corresponding GUID of the CAD element is CAD-GUID-001, and the corresponding GUID of the 110kV main transformer (capacity 50MVA) in the equipment list is EQ-GUID-001.

[0019] Step 4.2: Comparison of Component List and AI Annotation Results. Based on the GUID mapping relationship established in Step 4.1, extract all component lists from the equipment inventory (including conventional and concealed components such as 110kV transformers, circuit breakers, disconnect switches, instrument transformers, cables in cable trenches, and underground foundations). Compare all components with the AI ​​annotation results completed in Step 3 to filter out components that are recorded in the equipment inventory but not annotated in the 3D model (such as the underground foundation of a disconnect switch), forming a preliminary list of missing components.

[0020] Step 4.3: Identification and Supplementation Prompt for Missing Components. The list of missing components selected is verified a second time. After confirming that there are no misjudgments or duplicates, the system automatically issues a supplementation prompt, specifying the GUID (GUID-056) of the underground foundation of the disconnector switch, the corresponding equipment information (underground foundation of the disconnector switch, dimensions 1.0m 1.0m 0.8m), the location of the CAD element (C zone of the high-voltage bay in the substation), and the approximate three-dimensional spatial coordinates (X=120.5m, Y=85.3m, Z=-1.2m), guiding the staff to complete the supplementation operation for the missing components.

[0021] Step 5: Quality Control and Optimization of Labeling. Establish a unified four-level semantic labeling system for power engineering. This system includes equipment categories (transformer equipment), equipment types (transformers, circuit breakers, disconnectors, etc.), specifications (voltage level 110kV, capacity 50MVA, etc.), and functions and specifications (high-voltage side of main transformer, meeting GB50229-2019 specifications), clearly defining labeling standards. Based on the national standard (GB50229-2019 "Code for Fire Protection Design of Thermal Power Plants and Substations") and industry standards' power specification verification rule library, automatically verify the labeling results, correcting inconsistencies (such as mislabeling a 35kV circuit breaker as a 110kV circuit breaker). Generate labeling confidence scores, setting a confidence threshold of 90%. Only labeling results with a confidence score <90% (such as the labeling of a small insulator) are prompted for manual review, reducing manual workload. Through an incremental learning mechanism, manually corrected results are added to the model training set to continuously improve labeling accuracy and achieve precise control of labeling quality.

[0022] After this embodiment was completed, all components of the 3D model of the 110kV substation were accurately semantically labeled, with no errors or omissions. The labeling accuracy rate was over 96%, and the workload of manual review was reduced by 80% compared to traditional manual labeling.

[0023] The data table is as follows: Annotation type Component categories Number of components (pieces) Please indicate the correct quantity (pieces). Labeling accuracy Remark Automatic semantic annotation Conventional components (transformers, circuit breakers, etc.) 120 118 98.3% No mislabeling, mainly for large, easily identifiable components. Concealed components (cables in cable trenches, underground foundations, etc.) 80 75 93.8% After GUID verification and relabeling, no missing labels were found. Total (all components) 200 193 96.5% The overall accuracy rate is 96%, consistent with the effect described in the example. Traditional manual annotation Full components 200 182 91.0% For comparison, the workload of manual verification is 5 times that of automatic annotation. Automatic annotation + manual review (low confidence level) Low-confidence labeled components (confidence level < 90%) 12 12 100% By only reviewing low-confidence components, manual workload is reduced by 80%. Example 2: Automatic Semantic Annotation of 3D Models of Power Distribution Rooms This embodiment is applied to the automatic semantic annotation of a 3D model of a 10kV indoor power distribution room. The specific implementation steps are as follows: Step 1: Multi-dimensional Feature Extraction. A dedicated multi-dimensional feature extraction engine for the power industry is constructed to simultaneously extract geometric and power-related features from various components in the 3D model of the power distribution room (including distribution cabinets, transformers, cables in cable trenches, transformer underground foundations, supports, etc.). Geometric features include the component's shape, size, curvature, and orientation. For example, the transformer underground foundation is extracted as a cube with dimensions of 0.8m x 0.8m x 0.6m, a surface curvature of 0, and an east-west orientation. Power-related features include insulation distances corresponding to voltage levels (insulation distance for 10kV equipment is not less than 0.2m) and equipment size parameters (the body size of a 10kV transformer is 0.5m x 0.4m x 0.6m). Finally, a unique multi-dimensional feature vector is generated for each component, providing data support for distinguishing similar components in the future.

[0024] Step 2: Knowledge Graph Matching and Similar Component Differentiation. A power engineering knowledge graph is constructed, containing nodes and relationships related to equipment types, specifications, and functional attributes in the power distribution room. Nodes include 10kV distribution cabinets, 10kV transformers, cables, and supports. The relationships are: 10kV transformer - belongs to - power distribution equipment; 10kV transformer - voltage level - 10kV cable - function - power transmission. A power semantic embedding model is trained using a graph neural network (GNN). The multi-dimensional feature vectors of the components generated in Step 1 are spatially mapped to the semantic vectors in the knowledge graph to accurately distinguish different models of 10kV transformers (through differences in feature vectors of equipment size parameters and curvature), avoiding mislabeling problems caused by human visual confusion.

[0025] Step 3: Multi-view collaboration and hidden component annotation. Using CNN+multi-view collaboration technology, the 3D model of the power distribution room is projected into multiple 2D images from various indoor angles, inside the cable trench, and directly above the underground foundation. The local features of the hidden components (cables in the cable trench, transformer underground foundation) are captured by CNN and semantic segmentation is completed. The underground foundation component is the transformer. The segmentation results are back-projected onto the 3D model to achieve accurate annotation of the hidden components, initially solving the problem of missing annotation of hidden components.

[0026] Step 4: GUID mapping and missing label verification specifically includes the following sub-steps: Step 4.1: Construct a globally unique GUID mapping relationship. Assign a globally unique GUID to each component in the 3D model of the power distribution room, each element in the CAD drawing, and each piece of equipment in the equipment list. Establish a one-to-one correspondence mapping relationship between 3D component, GUID, CAD element, and equipment in the equipment list. For example, the GUID of the 3D component of the cable in the cable trench is GUID-102, the corresponding GUID of the CAD element is CAD-GUID-102, and the GUID of the YJV22-8.7 / 15kV-3120mm cable in the equipment list is EQ-GUID-102.

[0027] Step 4.2: Comparison of Component List and AI Annotation Results. Based on the GUID mapping relationship established in Step 4.1, extract all component lists from the equipment inventory (including distribution cabinets, transformers, cables in cable trenches, underground foundations of transformers, etc.) and compare them one by one with the AI ​​annotation results completed in Step 3. Filter out components that are recorded in the equipment inventory but not annotated in the 3D model (such as a section of cable in a cable trench) to form a preliminary list of missing components.

[0028] Step 4.3: Identification and Supplementation of Missing Components. The list of missing components selected is verified a second time. After confirming that there are no misjudgments or duplicates, the system automatically issues a supplementation prompt, specifying the GUID (GUID-108) of the cable section, the corresponding equipment information (YJV22-8.7 / 15kV-3120mm cable, length 15m), the CAD element location (cable trench on the west side of the power distribution room), and the approximate three-dimensional spatial coordinates (X=35.2m, Y=18.6m, Z=-0.5m), guiding the staff to complete the supplementation operation of the missing components.

[0029] Step 5: Quality Control and Optimization of Labeling. Establish a unified four-level semantic labeling system for power engineering. This system includes equipment categories (power distribution equipment), equipment types (distribution cabinets, transformers, cables, etc.), specifications (voltage level 10kV, cable type YJV22-8.7 / 15kV-3120mm, etc.), and functions and specifications (power transmission, meeting DL / T5210.1-2018 specifications), clearly defining labeling standards. Based on a power specification verification rule library based on national and industry standards, automatically verify the labeling results and correct inconsistencies (such as mislabeling cables as conductors). Generate labeling confidence scores, prompting manual review only for labels with confidence scores <90%, reducing manual workload. Through an incremental learning mechanism, manually corrected results are added to the model training set to continuously improve labeling accuracy and achieve precise quality control of labeling.

[0030] After this embodiment is completed, the semantic annotation accuracy of the 10kV substation 3D model reaches over 95%, and the annotation of hidden components is complete, effectively meeting the semantic requirements of substation design and operation and maintenance.

[0031] The data table is as follows: Annotation type Component categories Number of components (pieces) Please indicate the correct quantity (pieces). Labeling accuracy Remark Automatic semantic annotation Standard components (distribution cabinets, instrument transformers, brackets, etc.) 90 88 97.8% No mislabeling; primarily features easily identifiable indoor components, suitable for indoor power distribution room scenarios. Concealed components (cables in cable trenches, underground foundations for instrument transformers, etc.) 70 65 92.9% After GUID verification and supplementary labeling, no omissions were found, with a focus on covering cables in cable trenches and underground foundations of instrument transformers. Total (all components) 160 153 95.6% The overall accuracy rate is 95%, which matches the description in the example where the accuracy rate is over 95%. Traditional manual annotation Full components 160 145 90.6% For comparison, manual annotation is inefficient and prone to errors due to the obstruction of indoor spaces. Automatic annotation + manual review (low confidence level) Low-confidence labeled components (confidence level < 90%) 10 10 100% By only reviewing low-confidence components, the amount of manual work is reduced by more than 75% compared to traditional manual annotation. Example 3: Automatic Semantic Annotation of 3D Models of Power Distribution Rooms This embodiment is applied to the automatic semantic annotation of a 3D model of a 10kV indoor power distribution room. The specific implementation steps are as follows: Step 1: Multi-dimensional feature extraction. Construct a power-specific multi-dimensional feature extraction engine to simultaneously extract geometric features (shape, size, curvature, orientation) and power-related features (10kV equipment insulation distance, equipment size parameters) for various components of the three-dimensional model of the power distribution room, generating a unique multi-dimensional feature vector for each component. The core calculation formula is as follows: (1) Calculation formula for multi-dimensional feature vector of component: Where: is the multi-dimensional feature vector of the i-th component; is the geometric feature vector of the i-th component (composed of normalized shape, size, curvature, and orientation); is the power-related feature vector of the i-th component (composed of normalized 10kV equipment insulation distance and equipment size parameters); and are feature weight coefficients, and satisfy (in this embodiment, the power-related features are emphasized). (2) Calculation of component curvature (for geometric feature extraction): Where: is the curvature of a point on the surface of the component; is the radius of curvature of that point (curvature of planar components, such as the underground foundation of the transformer).

[0032] Step 2: Knowledge Graph Matching and Similar Component Differentiation. A power engineering knowledge graph containing substation equipment nodes and their relationships is constructed. A semantic embedding model is trained using GNN to map feature vectors to semantic vectors in the knowledge graph, accurately distinguishing different models of 10kV transformers and avoiding mislabeling.

[0033] Step 3: Multi-view collaboration and hidden component annotation. Using CNN+ multi-view collaboration technology, the 3D model is projected into multi-view 2D images to capture and segment the features of hidden components such as cables and underground foundations of transformers in the cable trench, and then back-projected onto the 3D model to complete the annotation.

[0034] Step 4: GUID mapping and missing label verification specifically includes the following sub-steps: Step 4.1: Construct a globally unique GUID mapping relationship. Assign unique GUIDs to 3D components, CAD elements, and equipment in the equipment list, and establish a one-to-one mapping relationship.

[0035] Step 4.2: Comparison of Component List and AI Annotation Results. Based on the GUID mapping relationship established in Step 4.1, extract all component lists from the equipment inventory (including distribution cabinets, transformers, cables in cable trenches, underground foundations of transformers, etc.) and compare them one by one with the AI ​​annotation results completed in Step 3. Filter out components that are recorded in the equipment inventory but not annotated in the 3D model (such as a section of cable in a cable trench) to form a preliminary list of missing components.

[0036] Step 4.3: Identification and Supplementation of Missing Components. The list of missing components selected is verified a second time. After confirming that there are no misjudgments or duplicates, the system automatically issues a supplementation prompt, specifying the GUID (GUID-108) of the cable section, the corresponding equipment information (YJV22-8.7 / 15kV-3120mm cable, length 15m), the CAD element location (cable trench on the west side of the power distribution room), and the approximate three-dimensional spatial coordinates (X=35.2m, Y=18.6m, Z=-0.5m), guiding the staff to complete the supplementation operation of the missing components.

[0037] Step 5: Quality Control and Optimization of Labeling. Establish a unified four-level semantic labeling system for power engineering. This system includes equipment categories (power distribution equipment), equipment types (distribution cabinets, transformers, cables, etc.), specifications (voltage level 10kV, cable type YJV22-8.7 / 15kV-3120mm, etc.), and functions and specifications (power transmission, meeting DL / T5210.1-2018 specifications), clearly defining labeling standards. Based on a power specification verification rule library based on national and industry standards, automatically verify the labeling results and correct inconsistencies (such as mislabeling cables as conductors). Generate labeling confidence scores, prompting manual review only for labels with confidence scores <90%, reducing manual workload. Through an incremental learning mechanism, manually corrected results are added to the model training set to continuously improve labeling accuracy and achieve precise quality control of labeling.

[0038] After this embodiment is completed, the accuracy of the 3D model annotation of the 10kV substation reaches over 95%, with no hidden components missing from the annotation, meeting the semantic requirements of design and operation and maintenance.

Claims

1. An automatic semantic annotation method for a three-dimensional model of a power engineering project, characterized in that, Includes the following steps: Step 1: Multi-dimensional feature extraction. Construct a multi-dimensional feature extraction engine for power engineering. Simultaneously extract geometric features and power engineering features from various components in the 3D model of power engineering. Finally, generate a multi-dimensional feature vector for each component, providing data support for the subsequent differentiation of similar components. Step 2: Knowledge Graph Matching and Similar Component Differentiation. Construct a power engineering knowledge graph containing nodes and relationships such as equipment type, specifications, and functional attributes; train a power semantic embedding model through a graph neural network (GNN), and spatially map the multi-dimensional feature vectors of components generated in Step 1 with the semantic vectors in the knowledge graph to accurately distinguish similar components and avoid mislabeling caused by human visual confusion from the source. Step 3: Multi-view collaboration and hidden component annotation. CNN+multi-view collaboration technology is used to project the 3D model into multi-view 2D images, capture the local features of hidden components and complete semantic segmentation; the segmentation results are back-projected onto the 3D model to achieve accurate annotation of hidden components, and initially solve the problem of missing annotation of hidden components. Step 4: GUID Mapping and Omission Detection. Establish a globally unique GUID mapping relationship to achieve the correspondence between 3D components and drawing elements and equipment list; by comparing the component list in the equipment list with the current AI annotation results, automatically identify unannotated omission components and issue prompts to ensure that all components are annotated without omission. Step 5: Quality Control and Optimization of Labeling. Establish a unified four-level semantic labeling system for power engineering, clarify labeling standards, and rely on the power specification verification rule library based on national and industry standards to automatically verify the labeling results, correct inconsistencies in labeling, generate labeling confidence scores, and only prompt low-confidence labeling results for manual review to reduce manual workload. Through an incremental learning mechanism, the manually corrected results are supplemented into the model training set to continuously improve the labeling accuracy and achieve precise control of labeling quality.

2. The automatic semantic annotation method for a three-dimensional model of a power engineering project according to claim 1, characterized in that, In step 1, geometric features include shape, size, curvature, and orientation.

3. The automatic semantic annotation method for a three-dimensional model of a power engineering project according to claim 1, characterized in that, In step 1, the electrical engineering features include the insulation distance corresponding to the voltage level and the equipment size parameters.

4. The automatic semantic annotation method for a three-dimensional model of a power engineering project according to claim 1, characterized in that, In step 3, the concealed components are either cables in the cable trench or underground foundation components, wherein the underground foundation components are any one or any combination of transformers, circuit breakers, disconnect switches, and instrument transformers.

5. The automatic semantic annotation method for a three-dimensional model of a power engineering project according to claim 1, characterized in that, Step 4 further includes: Step 4.1: Construct a globally unique GUID mapping relationship. Assign a globally unique GUID (Globally Unique Identifier) ​​to each component in the 3D model of the power engineering, each element in the CAD drawing, and each piece of equipment in the equipment list, and establish a one-to-one correspondence mapping relationship between 3D component, GUID, CAD element, and equipment in the equipment list. Step 4.2: Comparison of Component List and AI Annotation Results Based on the GUID mapping relationship established in Step 4.1, extract the complete component list (including all regular components and hidden components that need to be annotated) from the equipment register, and compare it with the AI ​​annotation results completed in Step 3 (including annotations of regular components and hidden components) one by one. Filter out the components that are recorded in the equipment register but not annotated in the 3D model to form a preliminary list of missing components. Step 4.3: Identification and Supplementation of Missing Components. The system performs a second check on the list of missing components. After confirming that there are no misjudgments or duplicates, the system automatically issues a supplementation prompt, clearly marking the GUID of the missing component, the corresponding equipment information, the location of the CAD element, and the approximate coordinates in three-dimensional space, guiding the staff to complete the supplementation operation of the missing components.

6. The automatic semantic annotation method for a three-dimensional model of a power engineering project according to claim 1, characterized in that, In step 5, the labeling system includes equipment category, equipment type, specifications, functions and standards.

Images