High-precision three-dimensional automatic modeling method for substation

Abstract

A high-precision three-dimensional automatic modeling method for a substation, comprising: collecting laser point cloud data of a substation and multi-view images of the substation; performing dense point cloud data registration, camera parameter reconstruction for the multi-view images, projection of the multi-view images, and adaptive point cloud density control; using three-dimensional Gaussian distribution functions to approximate the point cloud data, summing all the three-dimensional Gaussian distribution functions to obtain a Gaussian distribution geometric field of a three-dimensional Gaussian distribution point cloud representation, and obtaining a three-dimensional mesh model by means of three-dimensional mesh reconstruction and rendering; setting constraint conditions of devices on the basis of the three-dimensional mesh model, and using a volume segmentation algorithm and a three-dimensional labeling algorithm to generate a three-dimensional device model, device coordinates, and device category names; and converting the device model into a GIM file, generating a PMS file from the device coordinates and the device category names, and synchronizing the files to a GIM and PMS3.0. The present invention can effectively solve the problems of low efficiency and inconsistent quality in conventional manual modeling processes.

Description

A high-precision 3D automatic modeling method for substations

[0001] Cross-references to related applications

[0002] This application claims priority to Chinese Patent Application No. 2024119215085, filed on December 25, 2024, entitled "A High-Precision Three-Dimensional Automatic Modeling Method for Substations", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This invention relates to the fields of digital twins and digital power grid construction, specifically a high-precision three-dimensional automatic modeling method for substations. Background Technology

[0004] As a crucial infrastructure component of the power grid, the planning, design, construction, and operation and maintenance of substations are vital to the reliability and security of the power system. Currently, modeling substations based on GIM (Grid Information Modeling) plays a significant role in improving the efficiency and safety of substation construction. However, traditional GIM modeling methods primarily rely on manual data collection, drawing processing, and model creation, which is not only inefficient and prone to errors and omissions but also difficult to update dynamically in real time. With the increasing demands for information-based and digital management of the power grid, traditional manual modeling methods can no longer meet the requirements of modern substations for accuracy, real-time performance, and efficiency.

[0005] The Power Distribution Management System (PMS3.0), by integrating with GIM, enables real-time monitoring, diagnosis, and optimization of substation equipment. However, equipment data updates and maintenance in PMS3.0 still require a significant amount of manual work, resulting in a large workload and difficulty in ensuring data accuracy and timeliness.

[0006] Therefore, how to efficiently and automatically generate 3D models of substations and update equipment data in GIM and PMS3.0 systems in real time is a major challenge in current technology. Summary of the Invention

[0007] This invention provides an artificial intelligence-based high-precision automatic 3D substation modeling method, which aims to generate accurate 3D substation models through automation technology and seamlessly integrate them with the GIM platform and PMS3.0 system, thereby effectively solving the problems of low efficiency and inconsistent quality in the traditional manual modeling process.

[0008] This invention provides a high-precision three-dimensional automatic modeling method for substations, comprising the following steps:

[0009] Collect laser point cloud data and multi-view images of the substation;

[0010] Dense point cloud data registration is performed on the laser point cloud data of the substation. Camera parameters are reconstructed from the multi-view images of the substation to obtain the camera intrinsic and extrinsic parameters. The multi-view stereo MVS method is used to project the multi-view images onto the corresponding laser point cloud plane according to the camera intrinsic and extrinsic parameters. The density of the point cloud data is adjusted according to the resolution and sharpness of the projected images.

[0011] The three-dimensional Gaussian distribution point cloud representation algorithm is used to approximate the output point cloud data with a three-dimensional Gaussian distribution function. The summation of all three-dimensional Gaussian distribution functions is used to obtain the Gaussian distribution geometric field of the three-dimensional Gaussian distribution point cloud representation. The obtained Gaussian distribution geometric field is then fused with the NeRF method to realize three-dimensional mesh reconstruction and rendering, resulting in a three-dimensional mesh model.

[0012] Based on the generated 3D mesh model, set the constraints of the equipment, and use volume segmentation algorithm and 3D labeling algorithm to generate 3D equipment model, equipment coordinates and equipment category name;

[0013] The generated equipment model is converted into a GIM model file, and the equipment coordinates and equipment category names are generated into a PMS file, which is then synchronized to the GIM and PMS 3.0 systems.

[0014] Furthermore, the laser point cloud data is collected by a laser scanner to record the three-dimensional spatial coordinate information of the substation; the multi-view images are two-dimensional image information captured by a camera at different positions.

[0015] Furthermore, when collecting laser point cloud data from substations, distance constraints are applied to control the accuracy of the collected data. The distance constraints are as follows: d ≤ d max (1)

[0016] Where d is the distance from the target to the laser scanner, d max The target distance is the maximum distance from the laser scanner;

[0017] When acquiring multi-view images of a substation, image parameter constraints are used to ensure the quality of the acquired images. These image parameter constraints are as follows:

[0018] Where C represents contrast, L represents brightness, and C max C is the maximum contrast threshold. min L is the minimum contrast threshold. max L is the maximum brightness threshold. min This is the minimum brightness threshold.

[0019] Furthermore, when performing dense point cloud data registration on the laser point cloud data of substations, the Iterative Closest Point (ICP) method, the Normal Distribution Transform (NDT) method, the 3D Shape Context (3DSC) method, or the Fast Point Feature Histogram (FPFH) method are used for processing.

[0020] Furthermore, when reconstructing camera parameters from multi-view images of substations, the NeRF-method (Ne-Frequency Field-Search) is used to set the camera's intrinsic and extrinsic parameters as learnable parameters, and the intrinsic parameters are obtained by training with a multilayer perceptron (MLP). Alternatively, simulation calculations can be performed using the COLMAP software tool pipeline with motion structures and multi-view stereo methods, and the position and orientation of the camera in three-dimensional space can be obtained by using the motion structure (SfM) method, thereby estimating the intrinsic and extrinsic parameters.

[0021] Furthermore, the three-dimensional Gaussian distribution point cloud representation algorithm approximates the point cloud data output in the second step with a three-dimensional Gaussian distribution function. Summing all three-dimensional Gaussian distribution functions yields the Gaussian distribution geometric field representing the three-dimensional Gaussian distribution point cloud. The obtained Gaussian distribution geometric field is then fused with the NeRF method to achieve three-dimensional mesh reconstruction and rendering, resulting in a three-dimensional mesh model. Specifically, this includes:

[0022] The registered point cloud data is read, and a 3D Gaussian calculation is performed on each point cloud data. Some perturbation points are randomly generated around each point to increase the density and diversity of the point cloud data.

[0023] A three-dimensional Gaussian distribution function is used to fit each point cloud data point and its perturbation points, obtaining the mean vector, covariance matrix, and weight coefficients for each point. This constitutes a three-dimensional Gaussian distribution point cloud representation, describing the position, orientation, and shape of the point cloud data. For a given i-th point cloud data point, its three-dimensional Gaussian distribution function is:

[0024] Where x is a three-dimensional vector representing a point in space; w i It is a weighting coefficient, representing the importance of the point cloud data; μ i It is the mean vector, representing the center location of the point cloud data; Σ i It is the covariance matrix, representing the orientation and shape of the point cloud data;

[0025] The three-dimensional Gaussian distribution functions of all point cloud data are summed to obtain a mixture Gaussian distribution function, which is represented by Equation (4) as the Gaussian distribution geometry field SUGAR of the entire point cloud dataset:

[0026] The camera's intrinsic and extrinsic parameters, obtained from the output of camera parameter reconstruction, are used as inputs; NeRF algorithm for neural radiation fields.

[0027] The NeRF algorithm for neural radiation field is fused with the Gaussian distributed geometric field SUGAR characterized by equation (4) to finally render a three-dimensional mesh model Mesh.

[0028] Furthermore, the NeRF algorithm for neural radiation field is fused with the Gaussian distributed geometric field SUGAR characterized by equation (4), and finally rendered to obtain a three-dimensional mesh model, Mesh, which specifically includes:

[0029] In the NeRF algorithm for neural radiation fields, x represents a point in space, and P represents the intrinsic and extrinsic parameters of the camera. Each is input into a deep learning network, and each outputs the density value ρ(x) = MLP of the point. density (x, P) and the color value of the point c(x) = MLP color (x,P), the volume density function σ(x) is obtained through optimization, and the density field of the entire substation is described by equation (5): σ(x)=[ρ(x),c(x)] (5)

[0030] Integrating the Gaussian distribution geometric field SUGAR and density field:

[0031] By weighted fusion of the geometric field f(x) in equation (4) and the density field σ(x) in equation (5), the fused geometric density field σ'(x) is obtained: σ'(x)=α·f(x)+(1-α)·σ(x) (6)

[0032] Texture generation and mapping utilize the NeRF method to generate a color field c(x,d), as shown in the following equation: c(x,d) = MLP c (x,d)(7)

[0033] The obtained color field c(x,d) is mapped onto the Gaussian distributed geometric field SUGAR. On the surface of the initial 3D mesh InitialMesh generated by SUGAR, the color field generated by NeRF is mapped onto the vertices or faces of the initial 3D mesh InitialMesh to achieve high-quality texture mapping. Specifically, a color value is assigned to each vertex of the initial 3D mesh InitialMesh through vertex shading, and continuous texture is obtained through interpolation. In addition, the color field c(x,d) generated by NeRF is projected onto the surface of the initial 3D mesh InitialMesh through texture projection to obtain a high-resolution texture mapping model.

[0034] The Marching Cubes algorithm is used to extract the mesh surface from the geometric density field σ'(x), and the smoothness and detail of the initial 3D mesh InitialMesh surface are optimized using the uncertainty information of SUGAR. The optimization function E is expressed by Equation (8). total E total =E geometry +λ·E texture (8)

[0035] Among them, E geometryE represents geometric error, i.e., minimizing the error between the point cloud and the mesh; texture This represents texture error, minimizing the error between the color generated by NeRF and the mesh, where N is the number of points in the point cloud data;

[0036] Finally, the optimized 3D mesh model of the substation was obtained.

[0037] Furthermore, the constraints of the device include geometric constraint parameters and texture constraint parameters. The geometric constraint parameters include the area, size, spacing, and position of the device, and the texture constraint parameters include texture coordinates, filtering, and sampling.

[0038] Furthermore, the volume segmentation algorithm divides the obtained 3D mesh model into different regions by using geometric constraint parameters, with each region representing a device or component; the 3D labeling algorithm assigns a texture map to each region according to texture constraint parameters and gives each region a classification name according to predefined rules.

[0039] The present invention has the following advantages and effects:

[0040] 1. Automated process: The entire process is automated and requires no manual intervention, which greatly improves modeling efficiency.

[0041] 2. High precision: Combining point cloud data and multi-view images ensures high precision in 3D modeling.

[0042] 3. Real-time updates: Through seamless integration with GIM and PMS3.0 systems, substation models and equipment data can be updated in real time.

[0043] 4. Reduce human error: Automated modeling and data processing effectively avoids errors and omissions in manual modeling. Attached Figure Description

[0044] Figure 1 is a flowchart illustrating the high-precision three-dimensional automatic modeling method for substations according to the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Figure 1 shows a flowchart of the high-precision three-dimensional automatic modeling method for substations according to the present invention. The method includes the following steps:

[0047] The first step is to collect laser point cloud data and multi-view images of the substation.

[0048] The laser point cloud data is collected by a laser scanner (LiDAR) to record the three-dimensional spatial coordinate information of the substation; the multi-view images are two-dimensional image information taken by a camera at different positions (including aerial photography).

[0049] When collecting laser point cloud data from substations, a distance constraint is applied to control the accuracy of the collected data. The distance constraint is as follows: d≤d max (1)

[0050] Where d is the distance from the target to the laser scanner, d max The target distance is the maximum distance from the laser scanner;

[0051] When acquiring multi-view images of a substation, image parameter constraints are used to ensure the quality of the acquired images. These image parameter constraints are as follows:

[0052] Where C represents contrast, L represents brightness, and C max C is the maximum contrast threshold. min L is the minimum contrast threshold. max L is the maximum brightness threshold. min This is the minimum brightness threshold.

[0053] The second step involves dense point cloud registration of the substation's laser point cloud data. Laser point cloud data from different positions and angles are aligned to the same coordinate system using rigid body transformation and least squares method to improve the accuracy and completeness of the point cloud data. Camera parameters are reconstructed from multi-view images of the substation to obtain camera intrinsic and extrinsic parameters. The multi-view stereo MVS method is then used to project the multi-view images onto the corresponding laser point cloud plane based on the camera intrinsic and extrinsic parameters. Simultaneously, the density of the point cloud data is adjusted according to the resolution and sharpness of the projected images to ensure the quality and efficiency of the point cloud data.

[0054] When performing dense point cloud data registration on laser point cloud data from substations, considering the requirements of large scenes and high accuracy in substation scenarios, ICP (Iterative Closest Point) methods can be used. For other scenes, such as small scenes requiring fast registration, NDT (Normal Distribution Transform) can be used; for non-rigid objects (moving targets or people), 3DSC (3D Shape Context) methods can be used; and for scenes with minimal environmental interference, FPFH (Fast Point Feature Histogram) methods can be used. All of the above methods can achieve point cloud data registration.

[0055] When reconstructing camera parameters from multi-view images of substations, the NeRF (Self-calibrated Neural Radiation Field) method can be used. This involves setting the camera's intrinsic parameters (focal length, focus, height) and extrinsic parameters (position, rotation) as learnable parameters and training them with an MLP (Multilayer Perceptron). Alternatively, simulation calculations can be performed using the COLMAP (Structure of Motion and Multi-view Stereo Method) software tool. By employing the SfM (Structure of Motion) method, the camera's position and orientation in three-dimensional space can be obtained, and the intrinsic parameters (focal length, focus, and height) and extrinsic parameters (position, rotation) can be estimated.

[0056] In adaptive point cloud density control, Open3D (a 3D data algorithm library that can process point clouds, meshes, and depth maps) or PCL (a point cloud library) can be used in general scenarios to downsample the number of points. For scenarios with higher accuracy requirements, specific algorithms can be used, such as sampling methods based on CCVT (Constrained Capacity Voronoi Segmentation) and optimal transport theory. All of the above methods can achieve adaptive control of point cloud data density.

[0057] The third step involves using a 3D Gaussian distribution point cloud representation algorithm to approximate the point cloud data output from the second step with a 3D Gaussian distribution function. This approximates the position, orientation, and shape of the point cloud data. Then, all 3D Gaussian distribution functions are summed to obtain the Gaussian geometric field (SUGAR) representing the 3D Gaussian distribution point cloud. This SUGAR is then fused with the NeRF method to achieve 3D mesh reconstruction and rendering, ultimately resulting in a 3D mesh model. The specific implementation steps of the third step are as follows:

[0058] 1. Read the registered point cloud data output from the second step, perform 3D Gaussian calculation on each point cloud data, and randomly generate some perturbation points around each point to increase the density and diversity of the point cloud data;

[0059] 2. A three-dimensional Gaussian distribution function is used to fit each point cloud data point and its perturbation points, obtaining the mean vector, covariance matrix, and weight coefficients for each point. This constitutes a three-dimensional Gaussian distribution point cloud representation to describe the position, orientation, and shape of the point cloud data. For a given i-th point cloud data point, its three-dimensional Gaussian distribution function is:

[0060] Where x is a three-dimensional vector representing a point in space; w i It is a weighting coefficient, representing the importance of the point cloud data; μ i It is the mean vector, representing the center location of the point cloud data; Σ i It is the covariance matrix, representing the direction and shape of the point cloud data.

[0061] 3. Sum the three-dimensional Gaussian distribution functions of all point cloud data to obtain a mixture Gaussian distribution function, which is expressed by equation (4) as the Gaussian geometric field (SUGAR) of the entire point cloud dataset:

[0062] 4. Read the camera's intrinsic parameters (focal length, focus, and height) and extrinsic parameters (position and rotation) from the output of the second step camera parameter reconstruction, and use them as one of the inputs to NeRF.

[0063] 5. The NeRF (Neural Radiation Field) algorithm is used to fuse the Gaussian distributed geometric field (SUGAR) represented by formula (4) and finally render the three-dimensional mesh model (Mesh).

[0064] The NeRF (Neural Radiation Field) algorithm is implemented as follows: x represents a point in space, P represents the camera's intrinsic parameters (focal length, focus, and height) and extrinsic parameters (position and rotation), which are respectively input into a deep learning network (such as a multilayer perceptron MLP) and output the density value of the point ρ(x) = MLP. density (x, P) and the color value of the point c(x) = MLP color (x, P), the volume density function σ(x) is obtained through optimization. Thus, the density field of the entire substation can be described by equation (5): σ(x)=[ρ(x),c(x)] (5)

[0065] Integrating the Gaussian distribution geometric field (SUGAR) and density field:

[0066] By weighted fusion of the geometric field f(x) in equation (4) and the density field σ(x) in equation (5), the fused geometric density field σ'(x) is obtained: σ'(x)=α·f(x)+(1-α)·σ(x) (6)

[0067] Texture generation and mapping utilize the NeRF method to generate a color field c(x,d), as shown in the following equation: c(x,d) = MLP c (x,d)(7)

[0068] The obtained color field c(x,d) is mapped onto a Gaussian distributed geometric field (SUGAR). On the surface of the initial 3D mesh generated by SUGAR, the color field generated by NeRF is mapped onto the vertices or faces of the initial 3D mesh to achieve high-quality texture mapping. Specifically, a vertex shading method is used to assign a color value to each vertex of the initial 3D mesh, and continuous textures are obtained through interpolation. Furthermore, a texture projection method is used to project the color field c(x,d) generated by NeRF onto the surface of the initial 3D mesh to obtain a high-resolution texture mapping model.

[0069] The MarchingCubes algorithm is used to extract the mesh surface from the geometric density field σ'(x), and the smoothness and detail of the initial 3D mesh surface are optimized using the uncertainty information of SUGAR. Here, Equation (8) is used to represent the optimization function E. total E total =E geometry +λ·E texture (8)

[0070] Among them, E geometry E represents geometric error, i.e., minimizing the error between the point cloud and the mesh; texture This represents texture error, minimizing the error between the color generated by NeRF and the mesh. N is the number of points in the point cloud data.

[0071] Finally, the optimized three-dimensional mesh model of the substation is obtained.

[0072] Step 4: Based on the generated 3D mesh model, set the constraints of the device, and use the volume segmentation algorithm and 3D labeling algorithm to generate the 3D device model, device coordinates and device category name. The constraints of the device include geometric constraint parameters and texture constraint parameters. The geometric constraint parameters include the area, size, spacing and position of the device. The texture constraint parameters include texture coordinates, filtering and sampling.

[0073] The geometric constraint parameters of the equipment are used to constrain the layout of the building or equipment, defining the equipment area A. device Total area A workspace Length L device Width W device Total length L workspace Total width W workspace Spacing D between devices device Safety distance D safety Equipment location P device Other equipment locations P others Let them be expressed by equations (9), (10), (11), and (12) respectively: Area constraint: A device ≤A workspace (9)

[0074] Equipment area A required device It cannot exceed the total area A of the workspace. workspace ;

[0075] Size constraints:

[0076] Required equipment length L device and width Wdevice Plus safety distance D safety It cannot exceed the length L of the workspace. workspace and width W workspace Spacing constraint: D device ≥D safety (11)

[0077] The distance between devices must be greater than or equal to the safe distance; Position constraint: P device1 ≠P others (12)

[0078] Required location P of the equipment device1 Location P cannot be compared with other devices others overlapping.

[0079] The texture constraint parameters constrain the texture coordinate range, filtering, and sampling of the building or device, defining texture coordinates T and texture magnification T. enlarge Texture reduction T shrink Texture sampling T sample Texture interpolation T interpolation The constraints are expressed by equations (13), (14), and (15), respectively:

[0080] Texture coordinate range constraint: 0≤T≤1(13)

[0081] The texture coordinates must be between 0 and 1.

[0082] Texture filtering constraints:

[0083] The requirements are: to enlarge the texture of a single pixel to cover multiple pixels, the value must be greater than 1; to map the texture of multiple pixels to a single pixel, the value must be less than 1.

[0084] Texture sampling constraint: T sample =T interpolation (T enlarge ,T shrink (15)

[0085] The requirement is to impose texture sampling constraints, i.e., to perform interpolation within the range of texture magnification and texture reduction constraints.

[0086] The volume segmentation algorithm divides the resulting 3D mesh model into different regions using geometric features such as device area, size, and spacing. Each region represents a device or component. The specific implementation steps are as follows:

[0087] 1. Read the constraint parameters such as equipment area, size, and spacing.

[0088] 2. Using a volume segmentation algorithm, the 3D mesh model is divided into different regions according to the constraint parameters, and the boundary and center point of each region are recorded.

[0089] The 3D labeling algorithm assigns a texture map to each region based on image features such as texture coordinates, filtering, and sampling, and then assigns a category name to each region according to predefined rules, such as transformer, switch, or wire. The specific implementation steps are as follows:

[0090] 1. Read the constraint parameters such as texture coordinates, filtering, and sampling, and initialize the texture map.

[0091] 2. Using a 3D labeling algorithm, assign a texture map to each region and give each region a category name according to predefined rules.

[0092] 3. Output the 3D device model, coordinates, and classification name.

[0093] Step 5: Convert the generated equipment model into a GIM model file, and generate a PMS file containing the equipment coordinates and equipment category names, then synchronize it to the GIM and PMS 3.0 systems.

[0094] The GIM model conversion method, specifically, is implemented as follows:

[0095] 1. Input the generated 3D device model;

[0096] 2. Automatically check model integrity;

[0097] 3. Repair the normals;

[0098] 4. Reduce the number of polygons;

[0099] 5. Remove redundant vertices and faces;

[0100] 6. Adjust UV coordinates;

[0101] 7. Adjust the model's scale and position;

[0102] 8. Convert to GIM format and write to a GIM file.

[0103] The specific algorithm implementation for generating PMS files is as follows:

[0104] 1. Obtain equipment data and equipment location data from the PMS system, including the category name, relative coordinates, and asset code for each device;

[0105] 2. Calculate the distance, match the device to the nearest location, iterate through each device, calculate its distance to each location, find the nearest location, and add the matching result of the device and the nearest location to the intermediate result list;

[0106] 3. Write the intermediate results generated in the previous step, including the equipment asset code, category name, equipment coordinates, matching location ID, matching coordinates and distance, into the PMS file.

[0107] Taking the 3D reconstruction function in the digital twin system of a 110kV substation as an example:

[0108] The first step is to collect laser point cloud data and multi-view images of the substation.

[0109] The second step involves dense point cloud data registration, camera parameter reconstruction of multi-view images, multi-view image projection, and adaptive point cloud density control.

[0110] Among them, dense point cloud data registration is performed on the laser point cloud data of the substation. Laser point cloud data at different positions and angles are aligned to the same coordinate system through rigid body transformation and least squares method to improve the accuracy and integrity of the point cloud data.

[0111] To achieve automated 3D reconstruction of substations, preprocessing of the input laser point cloud data and multi-view images is required, including denoising, filtering, and downsampling, to improve data quality and reduce computational load. Point cloud data preprocessing includes motion distortion compensation, denoising, filtering, elevation normalization, and point cloud rasterization. Multi-view image preprocessing includes camera parameter estimation, camera parameter transformation, and camera parameter optimization.

[0112] Camera parameter reconstruction for multi-view images involves matching feature points from the images to calculate the camera's intrinsic and extrinsic parameters for each image, including position, orientation, and focal length. The NeRF method is employed, where the camera's intrinsic parameters (focal length, focus, height) and extrinsic parameters (position, rotation) are set as learnable parameters, and the intrinsic parameters are obtained by training an MLP (Multilayer Perceptron).

[0113] Multi-view image projection refers to projecting each image onto the corresponding laser point cloud plane according to camera parameters to increase the color information of the point cloud data.

[0114] Adaptive point cloud density control refers to adjusting the density of point cloud data based on the resolution and sharpness of the projected image to ensure the quality and efficiency of the point cloud data.

[0115] The third step involves using a 3D Gaussian distribution point cloud representation algorithm to approximate the point cloud data output in the second step with a 3D Gaussian distribution function. This function can describe the position, orientation, and shape of the point cloud data. Subsequently, all 3D Gaussian distribution functions are summed to obtain the Gaussian distribution geometric field (SUGAR) representing the 3D Gaussian distribution point cloud. The obtained Gaussian distribution geometric field (SUGAR) is then fused with the NeRF method to achieve 3D mesh reconstruction and rendering, ultimately resulting in a 3D mesh model.

[0116] The fourth step involves setting constraint parameters for equipment area, size, spacing, texture coordinates, filtering, and sampling, based on the characteristics of a 110kV substation. Volume segmentation and 3D labeling algorithms are then used to generate 3D equipment models, coordinates, and classification names. The volume segmentation algorithm divides the 3D mesh model into different regions based on geometric features such as equipment area, size, and spacing, according to the constraint parameters. Each region represents a piece of equipment or component. The 3D labeling algorithm assigns a texture map to each region based on image features such as texture coordinates, filtering, and sampling, and assigns a classification name to each region according to predefined rules, such as transformer, switch, or conductor. The specific implementation steps are as follows:

[0117] 1. Read the constraint parameters such as equipment area, size, and spacing.

[0118] 2. Using a volume segmentation algorithm, the system is divided into different regions, and the boundaries and center points of each region are recorded.

[0119] 3. Read the constraint parameters such as texture coordinates, filtering, and sampling, and initialize the texture map.

[0120] 4. Using a 3D labeling algorithm, assign a texture map to each region and give each region a category name according to predefined rules.

[0121] 5. Output the 3D device model, coordinates, and classification name.

[0122] Step 5: Use the GIM model conversion method to convert the 3D equipment model generated in step 4 into a GIM model file, and use the PMS file generation method to generate a PMS file from the coordinates and category names generated in step 4. The method then concludes. The GIM model conversion method refers to converting the 3D equipment model into a universal 3D modeling format according to the format requirements of the GIM model file, so that it can be used on different platforms and software. The PMS file generation method refers to converting the coordinates and category names into a universal attribute management format according to the format requirements of the PMS file, so that it can be used on different platforms and software. The specific implementation steps are as follows:

[0123] 1. Read the 3D device model file generated in step 4, parse its structure and attributes, and obtain information such as the device's geometry, position, orientation, and material.

[0124] 2. Create an empty GIM model file according to the format requirements of the GIM model file, and define its metadata, coordinate system, units and other information.

[0125] 3. Traverse all devices in the 3D device model, generate corresponding GIM device objects based on their type and attributes, and add them to the GIM model file. A GIM device object includes attributes such as the device's unique identifier, name, type, location, orientation, shape, and material.

[0126] 4. Save the GIM model file to complete the conversion from the 3D equipment model to the GIM model file.

[0127] 5. Read the coordinate and category name file generated in step 4, parse its contents, and obtain the device's coordinate and category name information.

[0128] 6. Create an empty PMS file according to the PMS file format requirements, and define its metadata, version, encoding and other information.

[0129] 7. Traverse all records in the coordinates and category names files, generate corresponding PMS device objects based on their contents, and add them to the PMS file. A PMS device object includes attributes such as the device's unique identifier, coordinates, and category name.

[0130] 8. Save the PMS file to complete the conversion of coordinates and category names to the PMS file.

[0131] Examples of the output files and output data for each step in this embodiment of the invention are shown in Table 1:

[0132] Table 1

[0133] Table 2 shows an example of the output table showing the device location matching results between the device and the PMS system:

[0134] Table 2

[0135] Table field descriptions:

[0136] Asset Code: The asset of the device, from the PMS system.

[0137] CategoryName: The category name of the device, obtained from the PMS system.

[0138] Device Coordinates: The coordinates of the device, in the format (X,Y,Z), are derived from the results of this method.

[0139] Matched Location ID: The location identifier in the matched PMS system, derived from the results of this method.

[0140] Matched Coordinates: The matched position coordinates, in the format (X,Y,Z), are the results of this method.

[0141] Distance: The Euclidean distance between the device coordinates and the matching position coordinates, derived from the results of this method.

[0142] In this way, the location of the device and the PMS system can be clearly linked, which facilitates subsequent management and maintenance.

[0143] This invention has the following characteristics:

[0144] (1) Unlike manual modeling and conventional point cloud or depth image-based methods, this invention uses both point cloud and image as input, combining the advantages of traditional single point cloud-based or visible light (e.g., oblique photography) modeling, and employs artificial intelligence methods (NeRF) and Gaussian distributed geometric field (SUGAR) to improve the accuracy and speed of modeling.

[0145] (2) By proposing equipment geometric constraints (area, size, spacing) parameters and texture constraints (coordinates, filtering, sampling), this invention can achieve high-precision modeling of specific scenarios (such as substation scenarios) in three dimensions and objects.

[0146] (3) This invention proposes to use volume segmentation algorithm and three-dimensional marking algorithm to realize entity segmentation of the three-dimensional scene model representation and output a general GIM file (.GIM), and to use three-dimensional marking algorithm to detect the three-dimensional scene model representation and generate a compatible file of the power grid PMS3.0 system containing equipment coordinates and type names, so as to achieve seamless integration of three-dimensional automated modeling with the technical platform and business application system of power production.

[0147] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A high-precision substation three-dimensional automatic modeling method, characterized in that, Includes the following steps: Collect laser point cloud data and multi-view images of the substation; Dense point cloud data registration is performed on the laser point cloud data of the substation. Camera parameters are reconstructed from the multi-view images of the substation to obtain the camera intrinsic and extrinsic parameters. The multi-view stereo MVS method is used to project the multi-view images onto the corresponding laser point cloud plane according to the camera intrinsic and extrinsic parameters. The density of the point cloud data is adjusted according to the resolution and sharpness of the projected images. The three-dimensional Gaussian distribution point cloud representation algorithm is used to approximate the output point cloud data with a three-dimensional Gaussian distribution function. The summation of all three-dimensional Gaussian distribution functions is used to obtain the Gaussian distribution geometric field of the three-dimensional Gaussian distribution point cloud representation. The obtained Gaussian distribution geometric field is then fused with the NeRF method to realize three-dimensional mesh reconstruction and rendering, resulting in a three-dimensional mesh model. Based on the generated 3D mesh model, set the constraints of the equipment, and use volume segmentation algorithm and 3D labeling algorithm to generate 3D equipment model, equipment coordinates and equipment category name; The generated equipment model is converted into a GIM model file, and the equipment coordinates and equipment category names are generated into a PMS file, which is then synchronized to the GIM and PMS 3.0 systems.

2. The high precision substation three-dimensional automatic modeling method of claim 1, wherein: The laser point cloud data is collected by a laser scanner, recording the three-dimensional spatial coordinate information of the substation; the multi-view images are two-dimensional image information captured by a camera at different positions.

3. The high precision substation three-dimensional automatic modeling method of claim 1, wherein: When collecting laser point cloud data of a substation, a distance constraint condition is adopted to control the accuracy of the collected data, and the distance constraint condition is as follows: d≤d max (1) where d is the distance of the target distance laser scanner, d max is the maximum distance of the target distance laser scanner; When collecting multi-view pictures of a transformer substation, picture parameter constraint conditions are adopted to ensure the quality of the collected pictures, and the picture parameter constraint conditions are as follows: where C is the contrast, L is the luminance, C max is the maximum threshold for contrast, C min is the minimum threshold for contrast, L max is the maximum threshold for luminance, L min is the minimum threshold for luminance.

4. The high precision substation three-dimensional automatic modeling method of claim 1, wherein: When performing dense point cloud data registration on laser point cloud data of substations, the Iterative Closest Point (ICP) method, the Normal Distribution Transform (NDT) method, the 3D Shape Context (3DSC) method, or the Fast Point Feature Histogram (FPFH) method are used for processing.

5. The high precision substation three-dimensional automatic modeling method of claim 1, wherein: When reconstructing camera parameters from multi-view images of a substation, the NeRF-self-calibrated neural radiation field method is used to set the camera's intrinsic and extrinsic parameters as learnable parameters, and the intrinsic parameters are obtained by training with a multilayer perceptron (MLP). Alternatively, simulation calculations can be performed using the COLMAP software tool pipeline with motion structures and multi-view stereo methods. The position and orientation of the camera in three-dimensional space can be obtained by using the motion structure SfM method, and then the intrinsic and extrinsic parameters can be estimated.

6. The high precision substation three-dimensional automatic modeling method of claim 1, wherein: The algorithm employs a 3D Gaussian distribution point cloud representation to approximate the point cloud data output from the second step using a 3D Gaussian distribution function. Summing all 3D Gaussian distribution functions yields the Gaussian distribution geometric field representing the 3D Gaussian point cloud. This Gaussian geometric field is then fused with the NeRF method to achieve 3D mesh reconstruction and rendering, resulting in a 3D mesh model. Specifically, this includes: The registered point cloud data is read, and a 3D Gaussian calculation is performed on each point cloud data. Some perturbation points are randomly generated around each point to increase the density and diversity of the point cloud data. A three-dimensional Gaussian distribution function is used to fit each point cloud data and its perturbed points to obtain the mean vector, covariance matrix and weight coefficient of each point, which constitutes a three-dimensional Gaussian distribution point cloud representation to describe the position, direction and shape of the point cloud data. For a given i-th point cloud data, the three-dimensional Gaussian distribution function is: where x is a three-dimensional vector representing a point in space; w i is a weight coefficient representing the importance of the point cloud data; μ i is a mean vector representing the central position of the point cloud data; Σ i is a covariance matrix representing the direction and shape of the point cloud data. The three-dimensional Gaussian distribution functions of all point cloud data are added to obtain a mixed Gaussian distribution function, and the Gaussian distribution geometry field SUGAR of the entire point cloud data set is expressed by formula (4): The camera's intrinsic and extrinsic parameters, obtained from the output of camera parameter reconstruction, are used as inputs; NeRF algorithm for neural radiation fields. The NeRF algorithm for neural radiation field is fused with the Gaussian distributed geometric field SUGAR characterized by equation (4) to finally render a three-dimensional mesh model Mesh.

7. The high-precision substation three-dimensional automatic modeling method of claim 6, wherein: The neural radiation field NeRF algorithm is fused with the Gaussian distributed geometric field SUGAR characterized by equation (4) to finally render a three-dimensional mesh model, which specifically includes: In the neural radiance field NeRF algorithm, x represents a point in space, P represents the internal and external parameters of a camera, is input into a deep learning network respectively, and outputs the density value of the point and the color value of the point respectively, that is, ρ(x) = MLP density (x, P) and c(x) = MLP color (x, P), and the volume density function σ(x) is obtained by optimization, and the density field of the entire substation is described by formula (5): σ(x)=[ρ(x),c(x)] (5) Fusion of Gaussian distribution geometry field SUGAR and density field: By weighting fusion of the geometry field f(x) of formula (4) and the density field σ(x) of formula (5), the fused geometry density field σ'(x) is obtained: σ'(x)=α·f(x)+(1-α)·σ(x) (6) Texture generation and mapping, the color field c(x,d) is generated by using the NeRF method, as shown in the following formula: c(x,d) = MLP c (x,d) (7) The obtained color field c(x,d) is mapped to the Gaussian distribution geometry field SUGAR, and the color field generated by NeRF is mapped to the vertices or patches of the initial three-dimensional mesh Initial Mesh generated by SUGAR, realizing high-quality texture mapping, wherein a color value is assigned to each vertex of the initial three-dimensional mesh Initial Mesh by the vertex shading method, and a continuous texture is obtained by interpolation; in addition, the color field c(x,d) generated by NeRF is projected onto the surface of the initial three-dimensional mesh Initial Mesh by the texture projection method, and a high-resolution texture map model is obtained; The initial mesh surface is extracted from the geometric density field σ'(x) using the marching cubes algorithm, and the smoothness and details of the initial mesh surface are optimized using the uncertainty information of SUGAR. The optimization function E is expressed by formula (8) total : E total = E geometry + λ · E texture (8) where E geometry denotes the geometry error, i.e., minimizing the error between the point cloud and the mesh; E texture denotes the texture error, minimizing the error between the color generated by NeRF and the mesh, N is the number of points in the point cloud data; Finally, the three-dimensional mesh model Mesh of the substation is obtained after optimization.

8. The high precision substation three-dimensional automatic modeling method of claim 1, wherein: The constraint conditions of the device include geometry constraint parameters and texture constraint parameters, the geometry constraint parameters include the area, size, spacing and position of the device, and the texture constraint parameters include texture coordinates, filtering and sampling.

9. The high-precision substation three-dimensional automatic modeling method of claim 8, wherein: The volume segmentation algorithm is to divide the obtained three-dimensional mesh model Mesh into different regions by using the geometry constraint parameters, and each region represents a device or component; the three-dimensional labeling algorithm is to assign a texture map to each region according to the texture constraint parameters, and give each region a classification name according to the predefined rule.

Images