High-fidelity three-dimensional reconstruction method and system based on multi-source data fusion

Abstract

The invention relates to a high-fidelity three-dimensional reconstruction method and system based on multi-source data fusion, and the method comprises the steps: obtaining multi-source heterogeneous data of a power grid target region, carrying out the time-space alignment and standardization preprocessing, and generating a standardized data set; performing multi-modal feature extraction and fusion coding on the standardized data set to generate uniform multi-dimensional feature representation; performing entity matching and semantic enhancement on the multi-dimensional feature representation and a pre-constructed power grid semantic knowledge graph to generate a semantic enhanced three-dimensional scene graph; carrying out three-dimensional geometric reconstruction and dynamic detail enhancement on the three-dimensional scene graph based on semantic enhancement, and constructing a high-fidelity power grid model; performing multi-level-of-detail optimization processing and format packaging on the high-fidelity power grid model, and outputting a high-fidelity three-dimensional power grid model suitable for the digital twin platform; according to the scheme, multi-source heterogeneous data can be fully fused, a high-precision and high-fidelity three-dimensional power grid model is constructed, and the requirements of fine management and real-time monitoring of a power grid are met.

Description

Technical Field

[0001] This application relates to the fields of power grid and information technology, and in particular to a high-fidelity 3D reconstruction method and system based on multi-source data fusion. Background Technology

[0002] In the field of power grid technology, accurate 3D modeling of the target area of ​​the power grid is a crucial foundation for realizing digital management and intelligent operation and maintenance of the power grid. Traditional 3D modeling methods for power grids mainly rely on a single data source, such as Geographic Information System (GIS) data or point cloud data. GIS data can provide geospatial distribution information of the power grid, such as the location of substations and transmission lines on a map; point cloud data can present the 3D geometry of power equipment.

[0003] However, single data sources have significant limitations. The information contained in a single data source is limited, making it difficult to comprehensively and accurately reflect the true state of the power grid. To overcome this problem, some methods attempt to fuse multiple data sources. For example, geospatial data is simply overlaid with equipment point cloud data to obtain richer information. However, this simple fusion method lacks effective data processing and analysis, failing to fully explore the inherent relationships between data points. This results in reconstructed 3D models with missing semantic information and low accuracy, making it difficult to meet the needs of refined power grid management and real-time monitoring. Summary of the Invention

[0004] The main objective of this application is to provide a high-fidelity 3D reconstruction method and system based on multi-source data fusion for power grids. This method can fully integrate multi-source heterogeneous data from the target area of ​​the power grid to construct a high-fidelity 3D power grid model suitable for digital twin platforms, thereby improving the accuracy and semantic information integrity of the 3D power grid model and meeting the needs of refined management and real-time monitoring of the power grid.

[0005] To achieve the above objectives, embodiments of the present invention provide a high-fidelity 3D reconstruction method based on multi-source data fusion, the method comprising the following steps: Acquire multi-source heterogeneous data of the target area of ​​the power grid, including geospatial data, equipment point cloud data, design parameter data and real-time monitoring data; and perform spatiotemporal alignment and standardization preprocessing on the multi-source heterogeneous data to generate a standardized dataset; Multimodal feature extraction is performed on the standardized dataset to generate feature vectors containing geometric feature vectors, texture feature vectors, and semantic feature vectors; and the feature vectors are fused and encoded to generate a unified multidimensional feature representation. The multi-dimensional feature representation is matched with a pre-constructed power grid semantic knowledge graph for entity matching, and the matching results are subjected to relational reasoning and semantic enhancement through the semantic constraints of the power grid semantic knowledge graph to generate a semantically enhanced 3D scene graph. Based on the semantically enhanced 3D scene graph, 3D geometry reconstruction is performed to generate a preliminary 3D model; based on real-time monitoring data, dynamic detail enhancement is performed on the preliminary 3D model to construct a high-fidelity power grid model. The high-fidelity power grid model is optimized at multiple levels of detail to generate an optimized model that supports real-time rendering; the optimized model is then encapsulated in a specific format and output as a high-fidelity 3D power grid model suitable for a digital twin platform.

[0006] Accordingly, this application also provides a high-fidelity 3D reconstruction system based on multi-source data fusion, the system comprising: The acquisition module is used to acquire multi-source heterogeneous data of the target area of ​​the power grid, including geospatial data, equipment point cloud data, design parameter data and real-time monitoring data; and to perform spatiotemporal alignment and standardization preprocessing on the multi-source heterogeneous data to generate a standardized dataset. The feature extraction module is used to perform multimodal feature extraction on the standardized dataset, generate feature vectors containing geometric feature vectors, texture feature vectors, and semantic feature vectors; and fuse and encode the feature vectors to generate a unified multidimensional feature representation. The semantic enhancement processing module is used to perform entity matching between the multi-dimensional feature representation and the pre-constructed power grid semantic knowledge graph, and to perform relational reasoning and semantic enhancement on the matching results through the semantic constraints of the power grid semantic knowledge graph to generate a semantically enhanced three-dimensional scene graph. The model reconstruction module is used to perform three-dimensional geometric reconstruction based on the semantically enhanced three-dimensional scene map to generate a preliminary three-dimensional model; and to perform dynamic detail enhancement on the preliminary three-dimensional model based on real-time monitoring data to construct a high-fidelity power grid model. The optimization processing module is used to perform multi-level optimization processing on the high-fidelity power grid model to generate an optimized model that supports real-time rendering; the optimized model is encapsulated in a format and output as a high-fidelity three-dimensional power grid model suitable for a digital twin platform.

[0007] In summary, the technical solution of this application acquires multi-source heterogeneous data from the target area of ​​the power grid and performs spatiotemporal alignment and standardization preprocessing to generate a standardized dataset, providing a high-quality data foundation for subsequent feature extraction and model construction. Multimodal feature extraction and fusion encoding are performed on the standardized dataset to generate a unified multi-dimensional feature representation, which can fully explore the inherent features of the data. Entity matching and semantic enhancement are performed between the multi-dimensional feature representation and a pre-constructed power grid semantic knowledge graph to generate a semantically enhanced 3D scene graph, improving the semantic information integrity of the model. Based on the semantically enhanced 3D scene graph, 3D geometric reconstruction and dynamic detail enhancement are performed to construct a high-fidelity power grid model. Multi-level detail optimization and format encapsulation are then performed, ultimately outputting a high-fidelity 3D power grid model suitable for digital twin platforms. This improves the accuracy and real-time rendering performance of the power grid 3D model, meeting the needs of refined power grid management and real-time monitoring. Attached Figure Description

[0008] Figure 1 This is a scene illustration of the high-fidelity 3D reconstruction method based on multi-source data fusion in the embodiments of this application; Figure 2 A flowchart of a high-fidelity 3D reconstruction method based on multi-source data fusion is provided for embodiments of this application; Figure 3 A schematic diagram of the multidimensional feature extraction process provided in the embodiments of this application; Figure 4 A schematic diagram of the semantic feature extraction process provided in the embodiments of this application; Figure 5 This is a schematic diagram of the semantic enhancement processing provided in the embodiments of this application; Figure 6 A schematic diagram illustrating the process of reconstructing a three-dimensional power grid model provided in an embodiment of this application; Figure 7 A schematic diagram illustrating the process of generating a high-fidelity 3D power grid model provided in this application embodiment; Figure 8 A schematic diagram of the structure of a high-fidelity 3D reconstruction system based on multi-source data fusion provided in an embodiment of this application; Figure 9 A schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0009] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0010] This application provides a high-fidelity 3D reconstruction method and system based on multi-source data fusion, which will be described in detail below.

[0011] In this embodiment, the high-fidelity 3D reconstruction method based on multi-source data fusion is a comprehensive 3D power grid modeling method. This method involves processing and fusing multi-source heterogeneous data of the target area of ​​the power grid, aiming to construct a high-precision, high-fidelity 3D power grid model to meet the needs of digital management and intelligent operation and maintenance of the power grid. Specifically, it acquires multi-source heterogeneous data such as geospatial data, equipment point cloud data, design parameter data, and real-time monitoring data of the target area of ​​the power grid. These data undergo spatiotemporal alignment and standardization preprocessing to generate a standardized dataset. Multimodal feature extraction and fusion encoding are performed on the standardized dataset to generate a unified multi-dimensional feature representation. The multi-dimensional feature representation is then matched with a pre-constructed power grid semantic knowledge graph for entity matching and semantic enhancement, generating a semantically enhanced 3D scene graph. Based on the semantically enhanced 3D scene graph, 3D geometric reconstruction and dynamic detail enhancement are performed to construct a high-fidelity power grid model. Finally, the high-fidelity power grid model undergoes multi-level detail optimization and format encapsulation to output a high-fidelity 3D power grid model suitable for a digital twin platform. The entire process involves multi-source data fusion, feature extraction, semantic enhancement, model building, and optimization, which can effectively improve the accuracy, semantic information integrity, and real-time rendering performance of the power grid 3D model.

[0012] As shown in Figure 1, a high-fidelity 3D reconstruction method based on multi-source data fusion for power grids is provided. The power grid 3D reconstruction scenario can include an UAV oblique photography system, a ground laser scanning device, a power grid asset management system, a sensor network, and a control platform. The UAV oblique photography system, the ground laser scanning device, the power grid asset management system, the sensor network, and the control platform are connected through a wireless network.

[0013] Taking a large urban power grid scenario as an example, this scenario contains a large number of power facilities such as substations, transmission lines, and power distribution equipment. It is necessary to perform high-precision 3D modeling of the entire power grid target area in order to achieve refined management and real-time monitoring of the power grid.

[0014] Among them, the UAV oblique photography system is used to collect geospatial data of the power grid target area. The UAV can carry multiple cameras at different angles to photograph the target area from various perspectives. For example, when the UAV flies over the target area, the cameras can capture information such as the appearance of substations and the route of transmission lines. By performing aerial triangulation processing on the collected image data, digital surface model data is generated, accurately presenting the topography and features of the power grid target area.

[0015] Ground-based laser scanning equipment can be placed at suitable locations within the target area of ​​the power grid to collect point cloud data of electrical equipment in that area. The laser scanning equipment scans the surface of the electrical equipment by emitting a laser beam, acquiring three-dimensional coordinate point cloud data of the equipment surface. For example, when scanning a transformer in a substation, the laser scanning equipment can accurately capture information such as the transformer's shape and dimensions. By performing noise reduction and filtering on the acquired equipment point cloud data, clean point cloud data is generated, removing noise points and useless information, thus improving the quality of the point cloud data.

[0016] The power grid asset management system stores equipment design parameter data for circuit equipment in the target area of ​​the power grid. Equipment design parameter data, such as the rated power of transformers and the conductor type of transmission lines, is exported from this system. The exported design parameter data is then parsed and formatted to generate structured parameter data, facilitating subsequent processing and analysis.

[0017] Sensor networks are distributed across various locations within the target area of ​​the power grid to collect real-time monitoring data. Sensors can collect information such as temperature and vibration from power equipment. For example, temperature and vibration sensors are installed on transmission lines to monitor temperature changes and vibration levels in real time. The collected real-time monitoring data undergoes timestamp alignment and missing value interpolation to generate time-series monitoring data, ensuring data accuracy and completeness.

[0018] After acquiring data from UAV oblique photography systems, ground-based laser scanning equipment, power grid asset management systems, and sensor networks, the control platform performs spatiotemporal alignment and standardization preprocessing on this multi-source heterogeneous data to generate a standardized dataset. Then, multimodal feature extraction and fusion encoding are performed on the standardized dataset to generate a unified multi-dimensional feature representation. This multi-dimensional feature representation is then matched with a pre-constructed power grid semantic knowledge graph for entity matching and semantic enhancement, generating a semantically enhanced 3D scene graph. Based on the semantically enhanced 3D scene graph, 3D geometric reconstruction and dynamic detail enhancement are performed to construct a high-fidelity power grid model. Finally, the high-fidelity power grid model undergoes multi-level detail optimization and format encapsulation to output a high-fidelity 3D power grid model suitable for digital twin platforms. Through this model, power grid managers can monitor the power grid's operating status in real time, promptly identify potential faults and safety hazards, and improve the power grid's operational efficiency and reliability.

[0019] refer to Figure 2 , Figure 2 This is a flowchart illustrating a high-fidelity 3D reconstruction method based on multi-source data fusion provided in this application embodiment. The execution subject of this method can be computer equipment (such as a control platform), such as a server. The high-fidelity 3D reconstruction method based on multi-source data fusion provided in this application embodiment specifically includes:

[0020] S10: Acquire multi-source heterogeneous data of the target area of ​​the power grid, including geospatial data, equipment point cloud data, design parameter data and real-time monitoring data; and perform spatiotemporal alignment and standardization preprocessing on the multi-source heterogeneous data to generate a standardized dataset.

[0021] In this embodiment, the target area of ​​the power grid refers to a specific geographical area designated when carrying out power grid management, monitoring, or modeling. Its scope can be defined according to specific needs; it can be a part of a city, such as a downtown business district, or an entire city or even a larger administrative region, such as a county-level city or prefecture-level city. This area can include various power facilities, including power generation equipment (such as urban distributed solar photovoltaic arrays and wind turbines), transmission equipment (high-voltage transmission lines and transmission towers), transformer equipment (substation transformers), and distribution equipment (distribution cabinets and distribution boxes). Simultaneously, its geographical environment also has a significant impact on the power grid. Geographical topography (mountains and plains) affects the laying of transmission lines and the selection of substation sites, while climatic conditions (heavy rain, strong winds, and snow) may lead to transmission line faults, affecting power grid stability. The target area of ​​the power grid is the basic scope for data collection, processing, and model construction in this application. Operations are carried out around its power grid information to achieve goals such as refined power grid management, efficient operation, and accurate 3D modeling.

[0022] In this embodiment, multi-source heterogeneous data refers to data from different data sources with different formats and characteristics. Geospatial data describes the geographic information of the target area of ​​the power grid, such as terrain, landforms, and land features; equipment point cloud data is three-dimensional coordinate point cloud data of the surface of power equipment obtained through technologies such as laser scanning; design parameter data is the design specifications and parameter information of power equipment; real-time monitoring data is the operating status data of power equipment collected in real time by sensors, such as temperature and vibration information. Spatiotemporal alignment refers to unifying data from different data sources in time and space to ensure data consistency and accuracy; standardization preprocessing refers to performing operations such as format conversion, noise removal, and missing value handling on the data to make the data meet the requirements of subsequent processing.

[0023] In one embodiment, the process of acquiring multi-source heterogeneous data can be implemented in various ways. For geospatial data, an unmanned aerial vehicle (UAV) oblique photogrammetry system can be used for acquisition; for equipment point cloud data, a ground-based laser scanning device can be used for acquisition; for design parameter data, it can be exported from a power grid asset management system; and for real-time monitoring data, it can be acquired through a sensor network. In one embodiment, step S10 can be implemented in the following manner, which will be described in detail below:

[0024] A1: Collect geospatial data of the target area of ​​the power grid using an unmanned aerial vehicle (UAV) oblique photography system, and perform aerial triangulation processing on the collected geospatial data to generate digital surface model data.

[0025] In this embodiment, the UAV oblique photography system is a UAV system equipped with multiple cameras at different angles, used to collect geospatial data from different perspectives. Aerial triangulation processing refers to processing the collected image data to determine the exterior orientation elements of the image and the three-dimensional coordinates of ground points, generating digital surface model data.

[0026] In one embodiment, during the geospatial data acquisition process of the UAV oblique photography system, the UAV flies along a preset flight path and altitude, and the camera captures images of the power grid target area from different angles. For example, the UAV can capture images at certain intervals over the power grid target area, acquiring image data from multiple perspectives. After acquiring the image data, aerial triangulation processing is performed. First, feature extraction and matching are performed on the image data to find corresponding points between images; then, adjustment calculations are performed using methods such as collinearity equations to determine the exterior orientation elements of the images; finally, digital surface model data is generated based on the exterior orientation elements and the image data. For example, during the geospatial data acquisition process of a large substation, the UAV acquires image data from multiple perspectives, and through aerial triangulation processing, generates digital surface model data of the substation, accurately presenting the substation's topography and feature information.

[0027] A2: Obtain point cloud data of power equipment in the target area of ​​the power grid through ground laser scanning equipment, and perform noise reduction and filtering on the obtained point cloud data to generate clean point cloud data.

[0028] In this embodiment, the ground laser scanning device is a device that scans the surface of an object by emitting a laser beam to obtain three-dimensional coordinate point cloud data of the object's surface. Noise denoising and filtering processes refer to removing noise points and useless information from the point cloud data to improve its quality.

[0029] In one embodiment, a ground-based laser scanning device is positioned at a suitable location within the target area of ​​the power grid to scan electrical equipment. For example, when scanning transmission line towers, the laser scanning device can emit a laser beam to scan the surface of the tower, acquiring three-dimensional coordinate point cloud data of the tower surface. After acquiring the point cloud data, the data undergoes denoising and filtering. Statistical filtering, radius filtering, and other methods can be used to remove noise points, and Gaussian filtering and other methods can be used to smooth the data, generating clean point cloud data. For example, after scanning multiple electrical devices within a substation, denoising and filtering processes remove noise points and useless information from the point cloud data, generating high-quality clean point cloud data.

[0030] A3: Export the equipment design parameter data of circuit equipment in the target area of ​​the power grid from the power grid asset management system, and parse and format the exported design parameter data to generate structured parameter data.

[0031] In this embodiment, the power grid asset management system is a system for storing power grid equipment design parameter data. Parsing and formatting refer to processing the exported design parameter data, converting it into a structured data format to facilitate subsequent processing and analysis.

[0032] In one embodiment, equipment design parameter data, such as the rated power of transformers and the conductor type of transmission lines, is exported from a power grid asset management system. The exported design parameter data is then parsed and formatted. First, the data is categorized and organized to determine its structure and fields. Then, the data is cleaned and transformed to remove useless information and erroneous data. Finally, the data is converted into structured parameter data. For example, after exporting design parameter data for multiple circuit devices in a target area of ​​a power grid, the data is converted into structured parameter data through parsing and formatting, such as storing it in a table format, containing fields such as equipment name, equipment type, and rated power.

[0033] A4: Real-time monitoring data is collected through a sensor network, and the collected real-time monitoring data is processed by timestamp alignment and missing value interpolation to generate time series monitoring data.

[0034] In this embodiment, the sensor network is a network composed of multiple sensors used to collect real-time operational data of power grid equipment. Timestamp alignment refers to unifying the data collected by different sensors in terms of time to ensure data temporal consistency; missing value imputation refers to supplementing missing values ​​in the collected data to ensure data integrity.

[0035] In one embodiment, a sensor network is distributed across key locations within the target area of ​​the power grid, such as transformers and transmission lines, to collect real-time data on temperature, vibration, and other parameters. For example, temperature and vibration sensors are installed at regular intervals along a transmission line to continuously collect data on the line's operational status. Due to factors such as sensor sampling frequency and transmission delay, the collected data may exhibit temporal inconsistencies and data gaps. Timestamp alignment of the collected real-time monitoring data can be achieved using methods such as linear interpolation to unify data from different sensors onto the same time scale. For missing value imputation, imputation can be performed based on historical statistical information such as the mean and median, or predictive imputation can be performed using machine learning-based methods, such as the K-nearest neighbor algorithm, to generate time-series monitoring data. Taking a transformer in a substation as an example, some temperature data collected by the sensor network is missing. By using mean imputation, the missing temperature values ​​are filled in, and the timestamps of the temperature data collected by different sensors are aligned, resulting in complete and consistent time-series monitoring data, which facilitates subsequent analysis and prediction of the transformer's operational status.

[0036] A5: Perform coordinate system unification processing on digital surface model data, clean point cloud data, structured parameter data, and time series monitoring data, and perform spatiotemporal alignment processing on the data in the unified coordinate system to generate a standardized dataset.

[0037] In this embodiment of the application, coordinate system one processing refers to converting data from different data sources to the same coordinate system to eliminate positional deviations caused by different coordinate systems; spatiotemporal alignment processing is based on coordinate system one, further ensuring the consistency of data in time and space.

[0038] In one embodiment, digital surface model data, cleanroom point cloud data, structured parameter data, and time-series monitoring data may originate from different acquisition devices and systems, and may use different coordinate systems. First, a unified reference coordinate system, such as the WGS-84 coordinate system, is selected based on the coordinate system information of each data source. For digital surface model data and cleanroom point cloud data, coordinate transformation formulas are used to transform them from their original coordinate systems to the unified coordinate system. For structured parameter data and time-series monitoring data, if they contain location information, corresponding coordinate transformations are also performed. After coordinate system unification, spatiotemporal alignment processing is performed on the data in the unified coordinate system. By combining the timestamps and spatial location information of the data, it is ensured that data from the same time and spatial location correspond to each other, generating a standardized dataset. For example, in a power grid target area containing multiple substations and transmission lines, geospatial data, equipment point cloud data, design parameter data, and real-time monitoring data undergo coordinate system unification and spatiotemporal alignment processing, enabling different types of data to be accurately correlated, providing a high-quality dataset for subsequent multimodal feature extraction.

[0039] S20: Perform multimodal feature extraction on the standardized dataset to generate feature vectors containing geometric feature vectors, texture feature vectors, and semantic feature vectors; and perform fusion encoding on the feature vectors to generate a unified multidimensional feature representation.

[0040] In this embodiment, multimodal feature extraction refers to extracting different types of features from different types of data. Geometric feature vectors are feature vectors that describe the geometric shape and structure of the data; texture feature vectors are feature vectors that describe the texture information of the data; and semantic feature vectors are feature vectors that describe the semantic information of the data. Fusion encoding refers to combining and encoding different types of feature vectors to generate a unified multidimensional feature representation in order to better represent the overall features of the data.

[0041] In one embodiment, the process of multimodal feature extraction on a standardized dataset can be performed separately for different types of data. For 3D point cloud data, normal vector and curvature calculations can be performed to extract local geometric features and generate geometric feature vectors. For texture image data, grayscale and filtering processes can be performed to extract texture statistical features and generate texture feature vectors. For device parameter data, parsing and entity extraction processes can be performed to extract semantic features and generate semantic feature vectors. After generating feature vectors, the geometric feature vectors, texture feature vectors, and semantic feature vectors are concatenated, and then the concatenated features are subjected to dimensionality reduction to generate a unified multidimensional feature representation. For example, methods such as principal component analysis (PCA) can be used for dimensionality reduction to reduce the dimensionality of the features and improve their expressive power.

[0042] S30: Perform entity matching between the multi-dimensional feature representation and the pre-constructed power grid semantic knowledge graph, and perform relational reasoning and semantic enhancement on the matching result through the semantic constraints of the power grid semantic knowledge graph to generate a semantically enhanced three-dimensional scene graph.

[0043] In this embodiment, the power grid semantic knowledge graph is a structured knowledge representation method used to describe entities, attributes, and relationships in the power grid domain. Entity matching refers to matching multi-dimensional feature representations with entities in the power grid semantic knowledge graph to find similar entities. Relationship reasoning refers to reasoning about the matching results based on the relation axioms in the power grid semantic knowledge graph to identify the topological connection relationships between devices. Semantic enhancement refers to semantically annotating the matching results using the identified topological connection relationships to enhance the semantic information of the matching results.

[0044] In one embodiment, the process of matching multi-dimensional feature representations with entities in a pre-built power grid semantic knowledge graph can be achieved through similarity calculation. First, the pre-built power grid semantic knowledge graph is loaded, and the graph query engine is initialized. Then, using the graph query engine, the similarity between the multi-dimensional feature representations and entities in the power grid semantic knowledge graph in the geometric and semantic feature spaces is calculated to generate initial matching results. The initial matching results are verified based on attribute constraint rules in the power grid semantic knowledge graph, and matching pairs that meet the constraints are selected. Relationship reasoning is performed on the selected matching pairs according to the relation axioms defined in the power grid semantic knowledge graph to identify the topological connection relationships between devices. The identified topological connection relationships are used to semantically enhance the matching pairs, generating a semantically enhanced 3D scene graph. For example, if the power grid semantic knowledge graph defines the connection relationship between transformers and transmission lines, relationship reasoning can identify the connection relationship between transformers and transmission lines in the matching pairs, and semantic annotation can be performed on the matching pairs to enhance the semantic information of the 3D scene graph.

[0045] S40: Perform 3D geometric reconstruction based on the semantically enhanced 3D scene graph to generate a preliminary 3D model; perform dynamic detail enhancement on the preliminary 3D model based on real-time monitoring data to construct a high-fidelity power grid model.

[0046] In this embodiment, 3D geometric reconstruction refers to constructing the geometry of a 3D model based on information from a semantically enhanced 3D scene graph. The preliminary 3D model is a basic model generated through 3D geometric reconstruction. Dynamic detail enhancement refers to supplementing and dynamically updating the preliminary 3D model with details using real-time monitoring data, making the model more realistic and accurate in reflecting the actual operating status of the power grid.

[0047] In one embodiment, the process of 3D geometric reconstruction based on semantically enhanced 3D scene graphs can be achieved through point cloud surface reconstruction processing. For example, the Poisson reconstruction algorithm is used to reconstruct the point cloud data in the semantically enhanced 3D scene graph to generate a preliminary 3D mesh model. Texture mapping and lighting processing are applied to the preliminary 3D mesh model to generate an enhanced 3D mesh model, improving the model's visual effect. A heat map texture is generated based on temperature data from real-time monitoring data, and the heat map texture is mapped onto the surface of the enhanced 3D mesh model to achieve temperature visualization; based on vibration data from real-time monitoring data, an approximate dynamic deformation effect is added to the enhanced 3D mesh model using a pre-calculated deformation parameter lookup table. The equipment model, after texture mapping and deformation enhancement, is assembled according to electrical connection relationships to construct a high-fidelity power grid model. For example, if the temperature rise of a transformer is detected in the real-time monitoring data, a heat map texture is generated and mapped onto the surface of the transformer model to intuitively display the temperature change of the transformer; when the vibration of a transmission line is monitored, a dynamic deformation effect is added to the transmission line model using a deformation parameter lookup table to more realistically reflect the operating status of the transmission line.

[0048] S50: Perform multi-level detail optimization on the high-fidelity power grid model to generate an optimized model that supports real-time rendering; encapsulate the optimized model in a format and output it as a high-fidelity 3D power grid model suitable for a digital twin platform.

[0049] In this embodiment, multi-level detail optimization refers to simplifying and optimizing the high-fidelity power grid model to generate models with different levels of detail to support real-time rendering. The optimized model is the model after multi-level detail optimization. Format encapsulation refers to converting the optimized model into a standard format suitable for the digital twin platform, facilitating its display and application on the digital twin platform.

[0050] In one embodiment, the process of optimizing a high-fidelity power grid model at multiple levels of detail can be achieved through methods such as mesh simplification, texture compression, and material optimization. For example, the Quadratic Error Metric (QEM) algorithm is used to simplify the mesh of the high-fidelity power grid model, generating a preliminary optimized model with multiple levels of detail. Based on the preliminary optimized model, texture compression and material optimization are performed to reduce the model's data volume, improve rendering efficiency, and generate an optimized model that supports real-time rendering. The optimized model undergoes coordinate system transformation and format standardization to generate a standardized model compatible with the digital twin platform. The standardized model is then encapsulated and metadata is added to generate an encapsulated model in a standard 3D file format. The encapsulated model undergoes quality verification and performance testing to ensure that its quality and performance meet requirements, outputting a high-fidelity 3D power grid model suitable for the digital twin platform. For example, the high-fidelity 3D power grid model can be displayed in real time on the digital twin platform, allowing power grid managers to monitor and analyze the model in real time and promptly identify potential problems in the power grid.

[0051] In one embodiment, reference Figure 3 Step S20 may include steps S21-S24, which will be described in detail below: S21: Perform normal vector and curvature calculations on the 3D point cloud in the standardized dataset to extract local geometric features and generate geometric feature vectors.

[0052] In this embodiment, the normal vector is a vector perpendicular to a point on the surface of a 3D point cloud, reflecting the directional information of the point cloud surface at that point; curvature is a parameter describing the degree of bending of the point cloud surface. By calculating the normal vector and curvature, local geometric features of the 3D point cloud can be extracted, thereby generating a geometric feature vector.

[0053] In one embodiment, for a 3D point cloud in a standardized dataset, a point is first selected as the center, and its neighboring points are determined. For this point and its neighboring points, a plane is fitted using the least squares method, and the normal vector of this plane is the normal vector of the point. For example, in the 3D point cloud data of a power pole, a point on the surface of the pole is selected, and neighboring points within a certain radius are selected with this point as the center. The normal vector of this point is obtained by fitting a plane using the least squares method.

[0054] Curvature can be calculated based on changes in the normal vector, such as calculating the rate of change of the normal vector at neighboring points. The calculated normal vector and local geometric features such as curvature are combined to generate a geometric feature vector. Taking the 3D point cloud data of multiple devices in a substation as an example, the normal vector and curvature of the point cloud data for each device are calculated to generate the corresponding geometric feature vector.

[0055] S22: Perform grayscale and filtering on the texture images in the standardized dataset to extract texture statistical features and generate texture feature vectors.

[0056] In this embodiment, grayscale processing converts a color texture image into a grayscale image, reducing the dimensionality of the data; filtering processing removes noise from the texture image, smoothing the image. Texture statistical features refer to features obtained through statistical analysis of the grayscale values ​​of the texture image, such as mean, variance, and contrast.

[0057] In one embodiment, for texture images in a standardized dataset, grayscale processing is first performed. A weighted average method can be used, where the pixel values ​​of the RGB channels of the color image are summed according to certain weights to obtain the grayscale value. For example, grayscale value = 0.299×R + 0.587×G + 0.114×B. After obtaining the grayscale image, filtering processing is performed, such as using Gaussian filtering, and convolution operations are used to smooth the image and remove noise. Then, texture statistical features are extracted from the filtered grayscale image. Statistical features such as the mean, variance, and contrast of the grayscale image are calculated, and these features are combined into a texture feature vector. Taking the texture image of a power transmission line as an example, after grayscale and filtering processing, its texture statistical features are extracted to generate a texture feature vector. This vector can reflect the texture information of the power transmission line surface, such as surface roughness and uniformity, which helps to distinguish different types of power transmission lines.

[0058] S23: Parse and extract entities from the device parameters in the standardized dataset to extract semantic features and generate semantic feature vectors.

[0059] In this embodiment, the parsing process involves performing structured analysis on the equipment parameter data to extract key information; the entity extraction process identifies relevant entities from the equipment parameters, such as equipment type and specifications. Semantic features refer to features that reflect the semantic information of the equipment, and semantic feature vectors are generated by extracting semantic features.

[0060] In one embodiment, equipment parameters in a standardized dataset are parsed. For example, for transformer equipment parameters, key information such as rated power, voltage level, and number of winding turns is parsed. Based on this parsed information, entity extraction is performed to determine the equipment type as a transformer and its specific specifications and model. This semantic information is then encoded to generate semantic feature vectors. Taking the parameter data of multiple devices in a power grid as an example, the parameters of each device are parsed and entity extracted to generate corresponding semantic feature vectors.

[0061] In one embodiment, reference Figure 4 Step S23 may further include steps S231-S233, as follows: S231: Perform structured parsing processing on the equipment parameters in the standardized dataset to extract equipment attribute information including equipment type, specification parameters and electrical attribute information.

[0062] In this embodiment, structured parsing processing involves decomposing and organizing equipment parameter data according to certain rules to extract useful equipment attribute information. Equipment type refers to the category to which the equipment belongs, such as transformer, circuit breaker, etc.; specification parameters are information describing the specific specifications of the equipment, such as capacity, size, etc.; electrical attribute information is information related to the electrical performance of the equipment, such as voltage, current, etc.

[0063] In one embodiment, for equipment parameters in a standardized dataset, such as parameter records of various equipment in a substation, these parameters are first identified and classified according to their format. For example, the text-based parameter data is parsed according to a preset template to extract the equipment type field and determine whether it is a transformer, switchgear, or other equipment. For specification parameters, information such as capacity and rated current are extracted from the parameter data. Regarding electrical attribute information, information such as voltage level and power factor are extracted. Taking the parameter data of a transformer as an example, through structured parsing, equipment attribute information such as equipment type as transformer, specification parameter as rated capacity of 1000kVA, and electrical attribute information as voltage level of 10kV is extracted.

[0064] S232: Based on the extracted device attribute information, perform graph structure construction processing to form a device attribute map.

[0065] In this embodiment, the graph structure construction process involves building a graph structure to represent the relationships between device attribute information. A device attribute graph is a knowledge representation method that uses a graph to represent device attributes and their relationships.

[0066] In one embodiment, a device association graph structure is constructed using extracted device attribute information as nodes and relationships between device attributes as edges. For example, for transformers and circuit breakers, if there is an electrical connection between them, an edge connects these two device nodes in the graph. Graph feature learning is then performed on the constructed device association graph, using methods such as Graph Convolutional Networks (GCNs), to learn the graph's topology and node features, generating a vectorized representation of the device attribute map. Taking multiple devices in a power grid as an example, by constructing a device association graph and performing graph feature learning, a device attribute map is formed. This map clearly displays the relationships and attribute information between devices, providing a powerful tool for semantic analysis and fault diagnosis of the power grid.

[0067] In one embodiment, step S232 can be implemented as follows: B1: Define the rules for functional dependencies between devices based on device type and electrical parameters.

[0068] In this embodiment, the functional dependency rules are rules that describe the functional relationships between different devices based on their type and electrical parameters. By defining these rules, the interactions and dependencies between devices can be clearly defined.

[0069] In one embodiment, the functional relationships between devices are analyzed based on device type and electrical parameters. For example, for transformers and load devices, electrical parameters such as the transformer's capacity and output voltage determine the power and voltage support it can provide to the load device. Based on this, a rule can be defined: when the power demand of the load device exceeds the transformer's rated capacity, the transformer may be unable to supply power to the load device normally. Taking a power grid system containing multiple transformers and different types of loads as an example, functional dependency rules between devices are defined based on electrical parameters such as transformer capacity, voltage level, and load device power demand. These rules help to understand the operating logic and interrelationships of devices in the power grid.

[0070] B2: Construct a device association graph structure based on the aforementioned functional dependency rules.

[0071] In this embodiment of the application, the device association graph structure is a structure that represents the relationship between devices in the form of a graph, where nodes represent devices and edges represent the functional dependencies between devices.

[0072] In one embodiment, based on defined functional dependency rules, devices in the power grid are treated as nodes, and the functional dependencies between devices are treated as edges to construct a device association graph. For example, in a substation, there are functional dependencies between transformers and switchgear, and between switchgear and electrical equipment; these dependencies are connected to the corresponding nodes in the graph via edges. In this way, a complete device association graph structure is constructed, which can intuitively display the connection relationships and functional dependencies of devices in the power grid.

[0073] B3: Perform graph feature learning on the device association graph to generate a vectorized representation of the device attribute graph.

[0074] In this embodiment, graph feature learning refers to extracting graph features by learning the structure and node information of the device association graph. Vectorization of the device attribute graph involves converting the device attribute graph into vector form, which facilitates subsequent calculations and analysis.

[0075] In one embodiment, graph neural networks (GNNs) and other methods are used to learn graph features from the device association graph. For example, graph convolutional networks (GCNs) can learn the feature representation of nodes by aggregating neighborhood information. Multiple layers of graph convolution operations are performed on the device association graph to update and propagate the feature information of each node, ultimately generating a vectorized representation of the device attribute map. Taking the device association graph of a large power grid as an example, graph feature learning transforms the complex device association graph into a vector representation. This vector effectively captures the relationships and attribute information between devices, providing an effective data format for semantic analysis and fault prediction of the power grid.

[0076] S233: Perform graph embedding processing on the device attribute map to generate semantic feature vectors.

[0077] In this embodiment of the application, graph embedding processing converts the device attribute graph into a low-dimensional vector space representation, thereby generating semantic feature vectors so that semantic information can be stored and processed in vector form.

[0078] In one embodiment, a graph embedding algorithm, such as the Node2Vec algorithm, is used for the device attribute graph. This algorithm samples node sequences in the graph through random walks and then uses a deep learning model, such as the Skip-Gram model, to map the nodes to a low-dimensional vector space. For example, in a device attribute graph containing multiple power devices and their relationships, the Node2Vec algorithm performs random walks to sample node sequences, and then the Skip-Gram model is used to map the nodes into vectors, ultimately generating semantic feature vectors. These semantic feature vectors can preserve the semantic information in the device attribute graph, such as device type and relationships between devices, providing an effective semantic representation for subsequent semantic matching and model construction.

[0079] S24: Concatenate the geometric feature vector, texture feature vector, and semantic feature vector, and perform dimensionality reduction on the concatenated features to generate a unified multi-dimensional feature representation.

[0080] In this embodiment, the concatenation process involves sequentially linking the geometric feature vector, texture feature vector, and semantic feature vector together to form a longer feature vector. Dimensionality reduction reduces the dimensionality of the concatenated feature vector, removes redundant information, and generates a unified multi-dimensional feature representation.

[0081] In one embodiment, the generated geometric feature vector, texture feature vector, and semantic feature vector are concatenated. For example, if the geometric feature vector has a dimension of m, the texture feature vector has a dimension of n, and the semantic feature vector has a dimension of p, then the concatenated feature vector has a dimension of m + n + p. Dimensionality reduction of the concatenated feature vector can be performed using Principal Component Analysis (PCA). The covariance matrix of the concatenated feature vector is calculated, its principal components are identified, and the top k principal components with the largest variances are selected. The feature vector is then projected onto the subspace formed by these k principal components to obtain the dimensionality-reduced feature vector, i.e., a unified multi-dimensional feature representation. Taking the data of multiple devices in a power grid as an example, the geometric, texture, and semantic feature vectors of each device are concatenated and dimensionality-reduced to generate a unified multi-dimensional feature representation. This representation can comprehensively reflect the geometric, texture, and semantic information of the device, providing a more comprehensive and effective feature input for subsequent entity matching and model construction.

[0082] In one embodiment, reference Figure 5 Step S30 may further include steps S31-S35, as follows: S31: Load the pre-built power grid semantic knowledge graph and initialize the graph query engine.

[0083] In this embodiment, the pre-constructed power grid semantic knowledge graph is a pre-built graph used to represent knowledge in the power grid domain, containing information such as power grid equipment, equipment attributes, and relationships between equipment. The graph query engine is a tool used for querying and retrieving information within the power grid semantic knowledge graph.

[0084] In one embodiment, a pre-built power grid semantic knowledge graph is stored in a database. When entity matching is required, the knowledge graph is loaded into memory. Simultaneously, a graph query engine is initialized, and its parameters, such as query syntax rules and index structure, are configured. For example, a semantic knowledge graph of a large power grid, containing information on numerous substations, transmission lines, equipment, and their relationships, is loaded from the database into memory, and a graph database-based query engine is initialized. This engine can quickly and accurately perform query operations within the knowledge graph, providing a foundation for subsequent entity matching.

[0085] S32: Using the graph query engine, perform similarity calculation between the multi-dimensional feature representation and entities in the power grid semantic knowledge graph to generate initial matching results.

[0086] In this embodiment of the application, similarity calculation measures the degree of similarity between the multi-dimensional feature representation and entities in the power grid semantic knowledge graph. By calculating the similarity, the entity most similar to the multi-dimensional feature representation is found, and an initial matching result is generated.

[0087] In one embodiment, a graph query engine is used to calculate the similarity between multi-dimensional feature representations and entities in the power grid semantic knowledge graph. This can be achieved through the following steps:

[0088] C1: Using the graph query engine, calculate the Euclidean distance between the multi-dimensional feature representation and the entities in the power grid semantic knowledge graph in the geometric feature space to obtain the geometric similarity.

[0089] In this embodiment of the application, Euclidean distance is a method for measuring the distance between two vectors in geometric feature space. Geometric similarity is obtained by calculating the Euclidean distance between the multi-dimensional feature representation and the entity in the power grid semantic knowledge graph in geometric feature space.

[0090] In one embodiment, the Euclidean distance between the geometric feature portion of the multi-dimensional feature representation and the geometric feature vector of the entity in the power grid semantic knowledge graph is calculated. For example, let the geometric feature vector of the multi-dimensional feature representation be A = (a1, a2, ..., an), and the geometric feature vector of the entity in the power grid semantic knowledge graph be B = (b1, b2, ..., bn), then the Euclidean distance d = √((a1 - b1)^2 + (a2 - b2)^2 + ... + (an - bn)^2). The smaller the distance, the higher the geometric similarity. Taking the multi-dimensional feature representation of a power device and entities of similar devices in the knowledge graph as an example, the Euclidean distance between them in the geometric feature space is calculated to obtain the geometric similarity, which reflects the degree of similarity between the devices in terms of geometric shape.

[0091] C2: Based on the geometric similarity, calculate the cosine similarity of the semantic feature space to obtain the semantic similarity.

[0092] In this embodiment of the application, cosine similarity is a measure of the cosine value of the angle between two vectors in the semantic feature space. By calculating the cosine similarity in the semantic feature space, the semantic similarity is obtained, which can reflect the degree of similarity between devices in semantic information.

[0093] In one embodiment, the cosine similarity is calculated for the semantic feature portion of the multi-dimensional feature representation and the semantic feature vector of the power grid semantic knowledge graph entity. Let the semantic feature vector of the multi-dimensional feature representation be C = (c1, c2, ..., cm), and the semantic feature vector of the power grid semantic knowledge graph entity be D = (d1, d2, ..., dm). ,in It is the dot product of vectors, where ||C|| and ||D|| are the magnitudes of the vectors, respectively. Taking the multi-dimensional feature representation of a transformer in a power grid and the transformer entity in a knowledge graph as examples, we calculate their cosine similarity in the semantic feature space to obtain semantic similarity. This similarity can reflect the degree of similarity of transformers in semantic information such as function and type.

[0094] C3: Using the geometric similarity and semantic similarity, a weighted fusion process is performed to generate a comprehensive matching score.

[0095] In this embodiment, the weighted fusion process involves summing the geometric similarity and semantic similarity according to certain weights to generate a comprehensive matching score, which takes into account both geometric and semantic information.

[0096] In one embodiment, let the geometric similarity be g, the semantic similarity be s, and the weights be w1 and w2 respectively (w1 + w2 = 1). Then, the comprehensive matching score S = w1×g + w2×s. For example, for matching a power device, if more emphasis is placed on geometric shape matching, w1 can be set to 0.6 and w2 to 0.4. The calculated comprehensive matching score can more comprehensively reflect the similarity between the multi-dimensional feature representation and the entities in the power grid semantic knowledge graph.

[0097] C4: Based on the comprehensive matching score, perform adaptive threshold filtering to retain high-quality matching pairs.

[0098] In this embodiment, the adaptive threshold filtering process automatically determines a suitable threshold based on the comprehensive matching score, filters out matching pairs with scores higher than the threshold, and retains high-quality matching pairs.

[0099] In one embodiment, the distribution of the overall matching scores for all matching pairs is first calculated, for example, by statistically analyzing parameters such as the mean and variance of the scores. Then, a threshold is adaptively determined based on this statistical information. For example, the mean plus a certain multiple of the standard deviation can be used as the threshold. Matching pairs with scores higher than this threshold are considered high-quality matching pairs and are retained. Taking the matching results of multiple devices in a power grid as an example, if the mean of the overall matching scores is 0.6 and the standard deviation is 0.1, setting the threshold to 0.7 filters out matching pairs with scores higher than 0.7. These matching pairs have a high degree of similarity, both geometrically and semantically, to entities in the power grid semantic knowledge graph.

[0100] C5: Perform topological consistency verification on the screening results to generate the final initial matching results.

[0101] In this embodiment, the topology consistency verification process checks whether the selected matching pairs conform to the power grid topology, i.e., whether the connection relationships and layout between devices are reasonable. Through the verification process, a final initial matching result is generated, ensuring the reasonableness and accuracy of the matching result.

[0102] In one embodiment, the selected matching pairs are verified based on the topological relationships between devices defined in the power grid semantic knowledge graph. For example, the knowledge graph specifies a particular electrical connection between transformers and switchgear. If the connection between the transformer and switchgear in a selected matching pair does not conform to this topological relationship, the matching pair is considered to be topologically inconsistent. All screening results are verified one by one, and matching pairs that do not conform to topological consistency are removed to generate the final initial matching result. Taking the device matching of a substation as an example, after topological consistency verification, some matching pairs that do not conform to the actual topological relationships are removed, resulting in an accurate and reliable final initial matching result.

[0103] S33: Verify the initial matching results based on the attribute constraint rules in the power grid semantic knowledge graph, and filter out matching pairs that meet the constraints.

[0104] In this embodiment, the attribute constraint rules are restrictions on equipment attributes defined in the power grid semantic knowledge graph, such as the rated power range and voltage level of the equipment. These rules are used to verify the initial matching results, filtering out matching pairs that meet the constraints and improving the accuracy of the matching results.

[0105] In one embodiment, the power grid semantic knowledge graph specifies that the rated power of a certain type of transformer should be within a certain range. For transformer matching pairs in the initial matching results, it is checked whether their rated power attribute conforms to this range. If not, the matching pair is considered to violate the attribute constraint rules and is discarded. Taking multiple transformer matching pairs in a power grid as an example, by verifying the attribute constraint rules, matching pairs that meet attribute constraints such as the rated power range are selected. These matching pairs are consistent with the definitions in the knowledge graph in terms of attributes, providing more accurate matching information for subsequent relational reasoning and semantic enhancement.

[0106] S34: Perform relational reasoning on the filtered matching pairs based on the relational axioms defined in the power grid semantic knowledge graph to identify the topological connection relationships between devices.

[0107] In this embodiment, the relation axioms are logical rules defining the relationships between devices in the power grid semantic knowledge graph, such as series and parallel connections. These axioms are used to perform relational reasoning on the filtered matching pairs, identifying the topological connections between devices and further refining the semantic information of the matching results.

[0108] In one embodiment, the power grid semantic knowledge graph defines axioms regarding the connection relationships between transformers and transmission lines, such as transformers being connected to other devices via transmission lines. For selected transformer-to-transmission-line pairs, reasoning is performed based on these axioms to determine their specific connection methods and topological relationships. Taking multiple device pairs in a power grid as an example, the topological connection relationships between devices are identified through relational reasoning, such as which transformers are connected in series with which transmission lines, and which are connected in parallel, providing topological information for constructing a semantically enhanced 3D scene graph.

[0109] S35: Use the identified topological connectivity to perform semantic annotation enhancement on the matching pairs, and generate a semantically enhanced 3D scene graph.

[0110] In this embodiment, semantic annotation enhancement refers to annotating the matched pairs using the identified topological connections to enrich their semantic information. Through semantic annotation enhancement, a semantically enhanced 3D scene graph is generated, enabling the scene graph to more accurately reflect the actual situation of the power grid.

[0111] In one embodiment, the identified topological connections between transformers and transmission lines, such as series connections, are annotated in the matching pairs. The annotated matching pairs are added to the 3D scene graph, simultaneously updating the connection relationships and semantic information between devices in the scene graph. Taking a 3D scene graph of a power grid as an example, through semantic annotation enhancement, the topological connections and semantic information between devices are clearly displayed in the scene graph, such as the connection methods between transformers and transmission lines, and the functions of the devices, providing richer and more accurate semantic information for subsequent 3D geometric reconstruction and model construction.

[0112] In one embodiment, reference Figure 6 Step S40 may further include steps S41-S45, as follows: S41: Perform point cloud surface reconstruction processing based on semantically enhanced 3D scene graphs to generate a preliminary 3D mesh model.

[0113] In this embodiment, the point cloud surface reconstruction process constructs a surface model of a 3D object based on point cloud data from a semantically enhanced 3D scene graph. This process generates a preliminary 3D mesh model, providing a foundation for subsequent model optimization and enhancement.

[0114] In one embodiment, the Poisson reconstruction algorithm is used to process point cloud data in a semantically enhanced 3D scene map. The algorithm first constructs an implicit function based on the normal vector information of the point cloud, and then obtains the surface model of the object by solving the Poisson equation. For example, in a semantically enhanced 3D scene map of a substation, point cloud data of various power equipment is included. Using the Poisson reconstruction algorithm, this point cloud data is reconstructed into a preliminary 3D mesh model, which roughly represents the shape and structure of the equipment. Taking the point cloud data of a transmission line tower as an example, after point cloud surface reconstruction processing, a preliminary 3D mesh model of the tower is generated, providing a basic geometric model for subsequent texture mapping and lighting processing.

[0115] S42: Perform texture mapping and lighting processing on the initial 3D mesh model to generate an enhanced 3D mesh model.

[0116] In this embodiment, texture mapping maps texture images onto the surface of a preliminary 3D mesh model, giving the model a more realistic appearance; lighting processing simulates different lighting conditions to enhance the model's three-dimensionality and realism. These two processing steps generate an enhanced 3D mesh model.

[0117] In one embodiment, for a preliminary 3D mesh model, texture mapping is first performed. Based on the vertex coordinates of the model surface and the coordinates of the texture image, the texture image is mapped onto the model surface. For example, in a preliminary 3D mesh model of substation equipment, the acquired texture image of the equipment surface is mapped onto the model, giving the model surface a realistic texture effect. Then, lighting processing is performed using a lighting model, such as the Phong lighting model, to simulate ambient light, diffuse light, and specular light. Based on the normal vector of the model surface and the lighting direction, the lighting intensity of each vertex is calculated, thereby enhancing the model's three-dimensionality and realism. Taking a preliminary 3D mesh model of a transformer as an example, after texture mapping and lighting processing, an enhanced 3D mesh model is generated. This model is more realistic in appearance and can better reflect the actual situation of the transformer.

[0118] S43: Generate a heat map texture based on temperature data in real-time monitoring data, and map the heat map texture onto the surface of the enhanced three-dimensional mesh model to achieve temperature visualization.

[0119] In this embodiment, the heat map texture is a texture generated based on temperature data from real-time monitoring data. By mapping it onto the surface of an enhanced three-dimensional mesh model, the temperature distribution of the device can be displayed intuitively, thus achieving temperature visualization.

[0120] In one embodiment, temperature data from real-time monitoring is first normalized, mapping the temperature values ​​to a specific color range, such as from blue (low temperature) to red (high temperature). A heatmap texture is then generated based on the normalized temperature data. For example, in real-time monitoring data of a transformer in a power grid, temperature values ​​at different locations on the transformer are acquired, normalized, and then a heatmap texture is generated. This heatmap texture is mapped onto the surface of an enhanced 3D mesh model, visually displaying the temperature distribution of the transformer through color variations. Taking multiple devices within a substation as an example, by generating heatmap textures and mapping them onto the surface of the device model, maintenance personnel can quickly understand the temperature status of the equipment and promptly identify potential faults.

[0121] S44: Based on vibration data from real-time monitoring, an approximate dynamic deformation effect is added to the enhanced 3D mesh model through a pre-calculated deformation parameter lookup table.

[0122] In this embodiment, the pre-calculated deformation parameter lookup table is a table pre-calculated based on vibration data and the physical characteristics of the model, recording the deformation parameters of the model under different vibration conditions. This lookup table adds an approximate dynamic deformation effect to the enhanced 3D mesh model, enabling the model to simulate the vibration state of the equipment.

[0123] In one embodiment, for vibration data in real-time monitoring data, frequency domain analysis is first performed to extract the main vibration frequency components. Based on these frequency components, modal parameter identification is performed to determine the vibration characteristic parameters of the equipment. A deformation parameter lookup table is pre-calculated based on these parameters. For example, for vibration data of a transmission line, the main vibration frequencies are obtained through frequency domain analysis, modal parameter identification is performed to determine the vibration characteristics of the line, and a deformation parameter lookup table for the line under different vibration conditions is pre-calculated. After acquiring real-time vibration data, the corresponding deformation parameters are retrieved from the lookup table and applied to an enhanced 3D mesh model to add an approximate dynamic deformation effect. Taking a transformer in a substation as an example, when vibration data of the transformer is monitored, a dynamic deformation effect is added to the transformer model through the deformation parameter lookup table, which more realistically simulates the operating state of the transformer and provides more intuitive information for real-time monitoring and fault diagnosis of the power grid.

[0124] In one embodiment, step S44 may further include the following sub-steps: D1: Perform frequency domain analysis on the vibration signal in the real-time monitoring data to extract the main vibration frequency components.

[0125] In this embodiment, frequency domain analysis processing converts the vibration signal in the time domain to the frequency domain, and extracts the main vibration frequency components by analyzing the frequency domain signal. These frequency components can reflect the vibration characteristics of the equipment.

[0126] In one embodiment, the vibration signal in the real-time monitoring data is transformed from the time domain to the frequency domain using a Fast Fourier Transform (FFT). For example, an FFT is performed on the vibration signal of a transmission line to obtain a frequency domain signal. In the frequency domain signal, the frequency components with larger amplitudes are identified as the main vibration frequency components. Taking the vibration signal of power equipment in a substation as an example, the main vibration frequency components are extracted through frequency domain analysis. These components can help determine the operating status of the equipment, such as whether abnormal vibrations exist.

[0127] D2: Based on the frequency components, perform modal parameter identification processing to generate equipment vibration characteristic parameters.

[0128] In this embodiment, the modal parameter identification process determines the modal parameters of the equipment, such as natural frequency and damping ratio, based on the extracted main vibration frequency components, and generates equipment vibration characteristic parameters. These parameters can accurately describe the vibration characteristics of the equipment.

[0129] In one embodiment, a frequency domain modal parameter identification method, such as peak picking, is employed to determine the natural frequencies of the equipment based on the main vibration frequency components. The damping ratio of the equipment is determined by analyzing the attenuation characteristics of the frequency domain signal. For example, for the vibration frequency components of a transformer, the peak picking method is used to determine its natural frequencies, and the damping ratio is determined by analyzing signal attenuation, generating the transformer's vibration characteristic parameters. Taking multiple devices in a power grid as an example, the vibration characteristic parameters of each device are obtained through modal parameter identification processing, providing an accurate basis for subsequent deformation parameter calculations and dynamic simulations.

[0130] D3: Using the vibration characteristic parameters, generate a deformation parameter lookup table through pre-calculation.

[0131] In this embodiment, the deformation parameter lookup table is prepared by pre-calculating the deformation parameters of the equipment under different vibration conditions based on the equipment's vibration characteristic parameters and storing them in a table. This lookup table allows for the rapid acquisition of deformation parameters under different vibration states.

[0132] In one embodiment, for the vibration characteristic parameters of the equipment, such as natural frequency and damping ratio, the deformation of the equipment under different vibration frequencies and amplitudes is calculated using methods such as finite element analysis, combined with the physical model of the equipment. The calculated deformation parameters are stored in a deformation parameter lookup table. For example, for a transmission line, based on its vibration characteristic parameters, the deformation of the line under different vibration conditions is calculated through finite element analysis, and a deformation parameter lookup table is generated. Taking multiple devices in a substation as an example, the pre-calculation generation of the deformation parameter lookup table provides a fast and effective method for real-time simulation of equipment vibration deformation.

[0133] D4: Based on the real-time monitoring data, retrieve the corresponding deformation parameter from the deformation parameter lookup table.

[0134] In this embodiment of the application, based on the vibration data monitored in real time, the corresponding deformation parameter is searched from the pre-calculated deformation parameter lookup table so as to add dynamic deformation effect to the equipment model.

[0135] In one embodiment, for the vibration data monitored in real time, the vibration frequency and amplitude information are extracted. Based on this information, the corresponding deformation parameter is looked up in a deformation parameter lookup table. For example, if the vibration frequency of a transformer is monitored to be 52Hz and the amplitude is 0.5, the record with the closest frequency and amplitude is found in the corresponding deformation parameter lookup table, and the corresponding deformation parameter is obtained. Taking multiple devices in a substation as an example, the deformation parameters can be quickly retrieved from the deformation parameter lookup table through real-time monitoring data, realizing the dynamic updating of the device model.

[0136] D5: Apply the retrieved deformation parameters to the 3D model to achieve dynamic deformation effects.

[0137] In this embodiment, the deformation parameters retrieved from the deformation parameter lookup table are applied to the enhanced 3D mesh model. By adjusting the vertex coordinates of the model, a dynamic deformation effect is achieved, which more realistically simulates the vibration state of the equipment.

[0138] In one embodiment, for the retrieved deformation parameters, the vertex coordinates of the model can be adjusted accordingly based on the physical structure of the equipment and the vertex information of the model. For example, when a deformation parameter of a transmission line is retrieved, the vertex coordinates of the transmission line model are offset according to the parameter to achieve a dynamic deformation effect of the line. Taking multiple equipment models within a substation as an example, by applying the deformation parameters to the models, dynamic deformation of the equipment models is achieved, providing a more intuitive and realistic model display for real-time monitoring and fault diagnosis of the power grid.

[0139] S45: Assemble the device model after texture mapping and deformation enhancement according to the electrical connection relationship to construct the high-fidelity power grid model.

[0140] In this embodiment, electrical connection refers to the electrical connection method between devices in the power grid, such as series or parallel connection. By assembling the device models after texture mapping and deformation enhancement according to the electrical connection relationship, a high-fidelity power grid model is constructed. This model can accurately reflect the actual structure and operating status of the power grid.

[0141] In one embodiment, the equipment models, after texture mapping and deformation enhancement, are assembled according to the electrical connection relationships defined in the power grid semantic knowledge graph. For example, equipment models such as transformers, switchgear, and transmission lines are connected and assembled according to their electrical connections. During the assembly process, it is ensured that the position and orientation of the equipment models conform to the actual situation. Taking a substation equipment model as an example, a high-fidelity power grid model is constructed by assembling it according to electrical connection relationships. This model is closer to the actual power grid in appearance, structure, and operating status.

[0142] In one embodiment, reference Figure 7 Step S50 may further include steps S51-S55, as detailed below: S51: Perform mesh simplification processing on the high-fidelity power grid model to generate a preliminary optimized model with multiple levels of detail.

[0143] In this embodiment, mesh simplification reduces the number of meshes in the high-fidelity power grid model while preserving the model's geometry and details as much as possible. This process generates a preliminary optimized model with multiple levels of detail, improving the model's rendering efficiency.

[0144] In one embodiment, the Quadratic Error Metric (QEM) algorithm is used to simplify the mesh of a high-fidelity power grid model. This algorithm calculates the quadratic error of each triangular facet, selects faces with smaller errors for merging, and gradually reduces the number of meshes. For example, a high-fidelity model of a large power grid contains a large number of triangular meshes. By simplifying the model using the QEM algorithm, the number of meshes is reduced while maintaining the basic shape of the model, generating a preliminary optimized model with multiple levels of detail. Taking a substation model as an example, after mesh simplification, the model's complexity is reduced, rendering speed is improved, and models with different levels of detail can be displayed according to actual needs, meeting the requirements of different scenarios.

[0145] S52: Based on the preliminary optimization model, perform texture compression and material optimization to generate an optimized model that supports real-time rendering.

[0146] In this embodiment, texture compression reduces the storage space of texture data and improves texture loading speed; material optimization adjusts the material properties of the model to improve its rendering effect. Through these two processing steps, an optimized model supporting real-time rendering is generated.

[0147] In one embodiment, for the texture data of the initial optimized model, a texture compression algorithm, such as the ETC (Ericsson Texture Compression) algorithm, is used to compress the texture data. This algorithm can significantly reduce the storage space of texture data while maintaining a certain texture quality. Simultaneously, the model's material properties are optimized, such as adjusting parameters like reflectivity and transparency, to improve the model's rendering effect. For example, in the initial optimized model of a power grid device, its texture is compressed, and its material properties are optimized to generate an optimized model that supports real-time rendering. Taking a large power grid model as an example, after texture compression and material optimization, the model's texture loading speed is faster, the rendering effect is more realistic, and it can meet the requirements of real-time rendering.

[0148] S53: Perform coordinate system transformation and format standardization on the optimized model to generate a standardized model compatible with the digital twin platform.

[0149] In this embodiment, coordinate system transformation converts the coordinate system of the optimized model to the coordinate system used by the digital twin platform, ensuring the model is displayed correctly on the platform; format standardization converts the file format of the optimized model to a standard format supported by the digital twin platform. Through these two processing steps, a standardized model compatible with the digital twin platform is generated.

[0150] In one embodiment, for the optimized model, a coordinate system transformation is first performed. According to the coordinate system definition of the digital twin platform, the vertex coordinates of the model are transformed accordingly. For example, if the digital twin platform uses the WGS-84 coordinate system, while the optimized model uses the local coordinate system, the vertex coordinates of the model are converted from the local coordinate system to the WGS-84 coordinate system. Then, format standardization processing is performed, converting the model's file format to a standard format supported by the digital twin platform, such as FBX format. Taking a power grid optimization model as an example, after coordinate system transformation and format standardization, a standardized model compatible with the digital twin platform is generated, which can be displayed and used normally on the digital twin platform.

[0151] S54: Perform data encapsulation and metadata addition processing on the standardized model to generate an encapsulated model in a standard 3D file format.

[0152] In this embodiment, data encapsulation involves packaging and storing the data of the standardized model, while metadata addition processing involves adding extra descriptive information to the model file, such as the model name, creation time, and device type. These two processing steps generate an encapsulated model in a standard 3D file format.

[0153] In one embodiment, for a standardized model, its vertex data, texture data, material data, etc., are packaged and stored in a single file. Simultaneously, metadata is added, such as the model name being "a substation power grid model," the creation date being "January 1, 2024," and the equipment type being "substation," etc. For example, in a standardized power grid model, data encapsulation and metadata addition processing are performed to generate an encapsulated model in a standard 3D file format. Taking a large power grid model as an example, the generated encapsulated model contains complete model data and detailed metadata, facilitating management and use on a digital twin platform. This metadata helps users quickly identify and understand relevant model information, improving data manageability and traceability.

[0154] S55: Perform quality verification and performance testing on the encapsulated model to output a high-fidelity 3D power grid model suitable for the digital twin platform.

[0155] In this embodiment, quality verification checks whether the geometry, texture, and materials of the encapsulated model meet the requirements, ensuring the model's accuracy and integrity. Performance testing evaluates the model's rendering speed, memory usage, and other performance metrics on the digital twin platform, ensuring the model can run smoothly on the platform. Through these two processing steps, a high-fidelity 3D power grid model suitable for the digital twin platform is output.

[0156] In one embodiment, for quality verification of the encapsulation model, a geometric inspection algorithm can be used to check for issues such as overlapping or cracks in the model's triangular faces; and to check the correctness of texture mapping and the rationality of material properties. For example, in an encapsulation model of a substation, geometric inspection revealed overlapping issues in some triangular faces, which were promptly repaired, ensuring the geometric quality of the model. For performance testing, the model can be rendered in real time on a digital twin platform, recording rendering time and memory usage. If the rendering time is too long or memory usage is too high, the model is further optimized, such as by further simplifying the mesh or compressing the texture. Taking an encapsulation model of a large power grid as an example, after multiple quality verifications and performance tests, the model is continuously optimized, ultimately outputting a high-fidelity 3D power grid model suitable for the digital twin platform. This model can be rendered and displayed quickly and accurately on the platform, meeting the needs of digital management and real-time monitoring of the power grid.

[0157] Accordingly, to better implement the above methods, this application also provides a high-fidelity 3D reconstruction system based on multi-source data fusion. For example... Figure 8 As shown, the high-fidelity 3D reconstruction system 80 based on multi-source data fusion includes:

[0158] The acquisition module 801 is used to acquire multi-source heterogeneous data of the target area of ​​the power grid, including geospatial data, equipment point cloud data, design parameter data and real-time monitoring data; and to perform spatiotemporal alignment and standardization preprocessing on the multi-source heterogeneous data to generate a standardized dataset. The feature extraction module 802 is used to perform multimodal feature extraction on the standardized dataset, generate feature vectors containing geometric feature vectors, texture feature vectors and semantic feature vectors; and perform fusion encoding on the feature vectors to generate a unified multidimensional feature representation; The semantic enhancement processing module 803 is used to perform entity matching between the multi-dimensional feature representation and the pre-constructed power grid semantic knowledge graph, and to perform relational reasoning and semantic enhancement on the matching result through the semantic constraints of the power grid semantic knowledge graph to generate a semantically enhanced three-dimensional scene graph. The model reconstruction module 804 is used to perform three-dimensional geometric reconstruction based on the semantically enhanced three-dimensional scene graph to generate a preliminary three-dimensional model; and to perform dynamic detail enhancement on the preliminary three-dimensional model based on real-time monitoring data to construct a high-fidelity power grid model. The optimization processing module 805 is used to perform multi-level optimization processing on the high-fidelity power grid model to generate an optimized model that supports real-time rendering; and to encapsulate the optimized model in a format and output it as a high-fidelity three-dimensional power grid model suitable for a digital twin platform.

[0159] The implementation details of each module are provided in the preceding method embodiments and will not be repeated here. The technical effects achieved by each module and device are described in the foregoing method embodiments.

[0160] like Figure 9 As shown, this application embodiment also provides a computer device 90, which includes a processor 901 and a memory 902, wherein the memory 902 stores a computer program, and when the computer program is executed by the processor 901, the processor 901 performs the steps of any of the methods described above.

[0161] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application are still within the scope of this application.

Claims

1. A high-fidelity 3D reconstruction method based on multi-source data fusion, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data of the target area of ​​the power grid, including geospatial data, equipment point cloud data, design parameter data and real-time monitoring data; and perform spatiotemporal alignment and standardization preprocessing on the multi-source heterogeneous data to generate a standardized dataset; Multimodal feature extraction is performed on the standardized dataset to generate feature vectors containing geometric feature vectors, texture feature vectors, and semantic feature vectors; and the feature vectors are fused and encoded to generate a unified multidimensional feature representation. The multi-dimensional feature representation is matched with a pre-constructed power grid semantic knowledge graph for entity matching, and the matching results are subjected to relational reasoning and semantic enhancement through the semantic constraints of the power grid semantic knowledge graph to generate a semantically enhanced 3D scene graph. Based on the semantically enhanced 3D scene graph, 3D geometry reconstruction is performed to generate a preliminary 3D model; based on real-time monitoring data, dynamic detail enhancement is performed on the preliminary 3D model to construct a high-fidelity power grid model. The high-fidelity power grid model is optimized at multiple levels of detail to generate an optimized model that supports real-time rendering; the optimized model is then encapsulated in a specific format and output as a high-fidelity 3D power grid model suitable for a digital twin platform.

2. The method according to claim 1, characterized in that, The process of extracting multimodal features from the standardized dataset to generate feature vectors containing geometric, texture, and semantic feature vectors, and then fusing and encoding these feature vectors to generate a unified multidimensional feature representation, includes the following steps: Normal vectors and curvature calculations are performed on 3D point clouds in a standardized dataset to extract local geometric features and generate geometric feature vectors. The texture images in the standardized dataset are grayscaled and filtered to extract texture statistical features and generate texture feature vectors. The device parameters in the standardized dataset are parsed and entity extraction is performed to extract semantic features and generate semantic feature vectors. Geometric feature vectors, texture feature vectors, and semantic feature vectors are concatenated, and the concatenated features are then subjected to dimensionality reduction to generate a unified multi-dimensional feature representation.

3. The method according to claim 1, characterized in that, The step of matching the multi-dimensional feature representation with a pre-constructed power grid semantic knowledge graph, and performing relational reasoning and semantic enhancement on the matching results through the semantic constraints of the power grid semantic knowledge graph to generate a semantically enhanced 3D scene graph includes the following steps: Load the pre-built power grid semantic knowledge graph and initialize the graph query engine. Using the graph query engine, the multi-dimensional feature representation is compared with the entities in the power grid semantic knowledge graph to calculate the similarity and generate an initial matching result. The initial matching results are verified based on the attribute constraint rules in the power grid semantic knowledge graph, and matching pairs that meet the constraints are selected. Based on the relational axioms defined in the power grid semantic knowledge graph, relational reasoning is performed on the filtered matching pairs to identify the topological connection relationships between devices; The identified topological connections are used to semantically enhance the matching pairs, generating a semantically enhanced 3D scene graph.

4. The method according to claim 1, characterized in that, The process of performing 3D geometric reconstruction based on the semantically enhanced 3D scene graph to generate a preliminary 3D model, and then performing dynamic detail enhancement on the preliminary 3D model based on real-time monitoring data to construct a high-fidelity power grid model includes the following steps: Point cloud surface reconstruction is performed on semantically enhanced 3D scene graphs to generate a preliminary 3D mesh model. Texture mapping and lighting processing are performed on the initial 3D mesh model to generate an enhanced 3D mesh model. A heatmap texture is generated based on temperature data from real-time monitoring data, and then mapped onto the surface of the enhanced 3D mesh model to achieve temperature visualization. Based on vibration data from real-time monitoring, an approximate dynamic deformation effect is added to the enhanced 3D mesh model using a pre-calculated deformation parameter lookup table. The device model, after texture mapping and deformation enhancement, is assembled according to electrical connection relationships to construct the high-fidelity power grid model.

5. The method according to claim 1, characterized in that, The process of performing multi-level detail optimization on the high-fidelity power grid model to generate an optimized model that supports real-time rendering, and then encapsulating the optimized model into a high-fidelity 3D power grid model suitable for a digital twin platform, includes the following steps: The high-fidelity power grid model is simplified by meshing to generate a preliminary optimized model with multiple levels of detail. Based on the preliminary optimization model, texture compression and material optimization are performed to generate an optimized model that supports real-time rendering. The optimized model is subjected to coordinate system transformation and format standardization to generate a standardized model compatible with the digital twin platform; The standardized model is encapsulated and metadata is added to generate an encapsulated model in a standard 3D file format; The encapsulated model is subjected to quality verification and performance testing to output a high-fidelity 3D power grid model suitable for the digital twin platform.

6. The method according to claim 2, characterized in that, The process of parsing and extracting entities from device parameters in the standardized dataset to generate semantic feature vectors includes the following steps: The equipment parameters in the standardized dataset are subjected to structured parsing processing to extract equipment attribute information, including equipment type, specification parameters, and electrical attribute information; Based on the extracted device attribute information, graph structure construction is performed to form a device attribute map. The device attribute map is subjected to graph embedding processing to generate semantic feature vectors.

7. The method according to claim 3, characterized in that, The step of using the graph query engine to calculate the similarity between the multi-dimensional feature representation and entities in the power grid semantic knowledge graph to generate initial matching results includes the following steps: Using the graph query engine, the Euclidean distance between the multi-dimensional feature representation and the entities in the power grid semantic knowledge graph in the geometric feature space is calculated to obtain the geometric similarity. Based on the geometric similarity, the cosine similarity in the semantic feature space is calculated to obtain the semantic similarity. By utilizing the geometric and semantic similarities, a weighted fusion process is performed to generate a comprehensive matching score. Based on the comprehensive matching score, an adaptive threshold filtering process is performed to retain high-quality matching pairs. The filtering results are subjected to topological consistency verification to generate the final initial matching results.

8. The method according to claim 4, characterized in that, The process of adding an approximate dynamic deformation effect to the enhanced 3D mesh model based on vibration data from real-time monitoring data and using a pre-calculated deformation parameter lookup table includes the following steps: The vibration signals in the real-time monitoring data are subjected to frequency domain analysis to extract the main vibration frequency components; Based on the frequency components, modal parameter identification processing is performed to generate equipment vibration characteristic parameters; Using the aforementioned vibration characteristic parameters, a deformation parameter lookup table is generated through pre-calculation; Based on real-time monitoring data, the corresponding deformation parameter is retrieved from the deformation parameter lookup table; The retrieved deformation parameters are applied to the 3D model to achieve dynamic deformation effects.

9. The method according to claim 6, characterized in that, The process of constructing a graph structure based on the extracted device attribute information to form a device attribute graph includes the following steps: Define rules for functional dependencies between devices based on device type and electrical parameters; Based on the functional dependency rules, construct the device association graph structure; Graph feature learning is performed on the device association graph to generate a vectorized representation of the device attribute graph.

10. A high-fidelity 3D reconstruction system based on multi-source data fusion, characterized in that, The system includes: The acquisition module is used to acquire multi-source heterogeneous data of the target area of ​​the power grid, including geospatial data, equipment point cloud data, design parameter data and real-time monitoring data; and to perform spatiotemporal alignment and standardization preprocessing on the multi-source heterogeneous data to generate a standardized dataset. The feature extraction module is used to perform multimodal feature extraction on the standardized dataset, generate feature vectors containing geometric feature vectors, texture feature vectors, and semantic feature vectors; and fuse and encode the feature vectors to generate a unified multidimensional feature representation. The semantic enhancement processing module is used to perform entity matching between the multi-dimensional feature representation and the pre-constructed power grid semantic knowledge graph, and to perform relational reasoning and semantic enhancement on the matching results through the semantic constraints of the power grid semantic knowledge graph to generate a semantically enhanced three-dimensional scene graph. The model reconstruction module is used to perform three-dimensional geometric reconstruction based on the semantically enhanced three-dimensional scene map to generate a preliminary three-dimensional model; and to perform dynamic detail enhancement on the preliminary three-dimensional model based on real-time monitoring data to construct a high-fidelity power grid model. The optimization processing module is used to perform multi-level optimization processing on the high-fidelity power grid model to generate an optimized model that supports real-time rendering; the optimized model is encapsulated in a format and output as a high-fidelity three-dimensional power grid model suitable for a digital twin platform.

Images