Transformer substation 3D Gaussian high-precision modeling method and device based on RTK real-time fusion and multi-scale anti-aliasing

Abstract

The invention provides a transformer substation 3D Gaussian high-precision modeling method and device based on RTK real-time fusion and multi-scale anti-aliasing, and belongs to the technical field of transformer substation digital twinning. According to the method, through multi-source data synchronous acquisition, RTK-vision-IMU real-time pose fusion, structure perception Gaussian density distribution, multi-scale anti-aliasing optimization and GIM format output, the problems of modeling dislocation, edge aliasing, storage redundancy, poor compatibility and the like in the prior art are solved. According to the method, the modeling precision reaches 0.5 cm, the edge SSIM is improved by 43%, the storage capacity is reduced by 65%, the method can be directly connected with an electric power PMS, 0.2 mm-level defect recognition and intelligent operation and maintenance are supported, and the method has remarkable engineering practical value.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology for substations, specifically a method and device for 3D Gaussian high-precision modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing. Background Technology

[0002] With the deepening of smart grid construction, digital twins of substations have become a core support for lean operation and maintenance, and high-precision 3D modeling is key. Existing technologies have gone through three stages: manual modeling (poor accuracy and low efficiency), laser point cloud modeling (large storage and weak texture), and neural radiation field / 3D Gaussian modeling (improved real-time performance but with engineering bottlenecks).

[0003] The existing technical solutions are described below:

[0004] Similar Solution 1: A method for surface modeling of porcelain bushing substation equipment based on UAVs and lightweight 3D Gaussian modeling (CN119832154A). This solution, proposed by State Grid Hunan Electric Power Company, represents a typical technical approach for current 3D Gaussian modeling of substations. The core implementation steps are as follows:

[0005] 1) Data acquisition: A drone equipped with a high-definition camera and LiDAR is used to acquire multi-view images and point cloud data of the ceramic equipment;

[0006] 2) Pose acquisition: The camera pose is estimated from the image using the SfM (Structure of Motion) algorithm and then registered with the point cloud data.

[0007] 3) Model building: Initialize a 3D Gaussian model based on the registered point cloud, and optimize the Gaussian parameters through a composite loss function;

[0008] 4) Model compression: Light Gaussian compression technology is used to reduce storage requirements, and the final output is a model of the device surface.

[0009] The core advantage of this solution is that it achieves lightweight modeling of ceramic sleeve equipment, reducing storage requirements by 80%. However, it does not solve problems such as pose drift and edge jaggedness. The modeling accuracy is only 1cm level, which is only suitable for displaying the appearance of the equipment and cannot support defect identification.

[0010] Similar Solution 2: AI-driven Gaussian 3D substation modeling optimization method and system (CN120355853B)

[0011] This solution, proposed and authorized by Jiangsu Haohan Information Technology Co., Ltd., is the first power modeling solution to combine AI and 3DGS. Its core implementation steps are as follows:

[0012] 1) Multidimensional data acquisition: Simultaneously acquire substation images, point clouds, and real-time operation data to construct a multidimensional feature set;

[0013] 2) AI-optimized modeling: The Gaussian model rendering and optimization are driven by a multi-task learning framework to improve the model's realism;

[0014] 3) Model update: The model is updated adaptively by combining real-time equipment operation data to support full lifecycle management.

[0015] The core advantage of this solution is that it enables the linkage between the model and operation and maintenance data. However, it still relies on traditional SfM pose estimation, has weak resistance to electromagnetic interference, and does not design optimization strategies for the edge characteristics of substation equipment, so the edge quality has not been fundamentally improved.

[0016] Based on the most similar solutions mentioned above and other relevant patents (such as CN119919578A and CN119885861A), the shortcomings of the prior art can be summarized into the following four categories, and these shortcomings directly restrict the engineering implementation of the technology:

[0017] 1. The pose acquisition method has fundamental flaws.

[0018] Existing technologies all employ a pose acquisition method of "image SfM estimation + post-processing point cloud registration," which has two major problems: First, the SfM algorithm is affected by strong electromagnetic interference in substations, increasing the feature point matching error rate by 40%, resulting in large pose estimation errors; second, post-processing registration cannot correct real-time drift during the acquisition process, with registration misalignment reaching 0.5m in densely populated equipment areas (such as busbar joints), causing "ghosting" phenomena in the model. For example, in CN119832154A, when modeling the main transformer area, the relative position error between the porcelain bushing group and the oil tank reaches 2-3cm, failing to reflect the actual equipment structure.

[0019] 2. Lack of edge optimization strategies and insufficient defect identification capabilities.

[0020] Existing technologies only use simple Gaussian or mean filtering to process model edges, without designing differentiated solutions for the edge characteristics of substation equipment (such as the curved edges of sheds and the sharp contacts of disconnectors). Actual measurements show that the SSIM of the insulator edges modeled by CN120355853B is only 0.65, and the PSNR is less than 25dB. Obvious serrations are visible to the naked eye, and cracks smaller than 0.5mm are completely masked by these serrations, making them undetectable by automatic detection systems.

[0021] 3. Gaussian density distribution machinery cannot balance precision and efficiency.

[0022] Existing technologies employ allocation strategies of "globally uniform density" or "fixed density according to equipment type," lacking awareness of the complexity of local equipment structures. For example, CN119919578A uses the same Gaussian density (1000 points / ㎡) for porcelain bushings (large curvature, rich texture) and structural beams (small curvature, simple texture), resulting in the loss of details in the gaps between the porcelain bushing skirts, while the Gaussian points in the structural beam area are redundant, increasing the storage volume by 300%. The storage of a single 220kV substation model reaches 5-8GB, making it impossible to load on mobile maintenance devices.

[0023] 4. Poor standard compatibility and high cost of engineering application.

[0024] The existing PLY / OBJ format models output by the technology lack semantic information of power equipment (such as asset number and model), which needs to be manually re-annotated in GIM modeling software. Annotation of a single station takes 1-2 weeks. At the same time, geometric accuracy is lost by 15% during the secondary format conversion process. For example, after the conversion of the CN119885861A model, the equipment positioning error increased from 1cm to 1.8cm, which cannot meet the positioning requirements of the PMS system. Summary of the Invention

[0025] In view of the shortcomings of the prior art, the core objective of this invention is to provide a method and apparatus for 3D Gaussian modeling of substations that takes into account high precision, high quality, high compatibility and engineering practicality.

[0026] A method for high-precision 3D Gaussian modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing includes the following steps:

[0027] Step S1. Multi-source data synchronous acquisition: Align RTK, IMU, LiDAR and HD camera data through a time synchronizer, use an improved A* algorithm to plan flight paths to avoid strong electromagnetic interference sources, and acquire image data and point cloud data with timestamps, RTK position and IMU attitude labels;

[0028] Step S2. Real-time RTK-Vision-IMU Fusion: The images and point cloud data acquired in Step S1 are linearly interpolated for time alignment. The high-precision RTK position and high-frequency vision-IMU pose are fused using an adaptive weighting algorithm to output device pose data.

[0029] Step S3. Data preprocessing: Using the device pose data from step S2, noise filtering, registration, and downsampling are performed on the point cloud data, and illumination equalization, distortion correction, and texture enhancement are performed on the image data to achieve cross-modal feature matching and obtain preprocessed point cloud and image data.

[0030] Step S4. Structure-aware Gaussian modeling: Based on the point cloud and image data preprocessed in step S3, calculate the structural complexity index C of the device, and adaptively allocate Gaussian density through the "split-merge-prune" strategy to construct a 3D Gaussian model;

[0031] Step S5. Multi-scale anti-aliasing optimization: Project the Gaussian model constructed in step S4 onto a 2D plane, detect and classify device edges, dynamically adjust the size of the filtering window according to the edge type, eliminate edge jagged artifacts, and obtain the optimized Gaussian model;

[0032] Step S6. GIM Format Conversion and System Integration: Convert the Gaussian model optimized in Step S5 into 3DTiles format, then map it to GIM format and embed semantic information to integrate with the power PMS system.

[0033] Furthermore, step S1 specifically includes:

[0034] A customized UAV data acquisition platform is used, with RTK modules, IMU units, LiDAR, and high-definition cameras achieving millisecond-level data alignment through a time synchronizer;

[0035] Data collection area is divided into three levels based on substation CAD drawings and equipment ledgers: Level 1: Main transformers, circuit breakers, disconnect switches, flight altitude 15-20m, flight path spacing 2m, ensuring unobstructed equipment surfaces; Level 2: Surge arresters, voltage transformers, flight altitude 20-25m, flight path spacing 3m; Level 3: Busbar structure, inspection passages, etc., flight altitude 30-40m, flight path spacing 5m.

[0036] Anti-interference route optimization: The improved A* algorithm is used to plan the route to avoid strong electromagnetic interference sources such as the main transformer and GIS equipment, while ensuring that the surface of each equipment is covered by at least 3 views to meet the multi-view constraints of 3D reconstruction.

[0037] Multi-device synchronous acquisition and control: The drone flight control system triggers the synchronous operation of each device. RTK and IMU data are uploaded to the ground station in real time. The lidar and camera collect data at 0.5-second intervals. Each frame of image and point cloud data is accompanied by a precise timestamp, RTK position and IMU attitude label.

[0038] Furthermore, step S2 specifically includes:

[0039] 2.1 Timestamp Alignment of Multi-Source Data

[0040] Time synchronization is achieved using a linear interpolation method to ensure that data from multiple sources at the same time correspond to the same device state.

[0041] 2.2 Adaptive Weight Fusion Model

[0042] Define the fused pose data as P fuse It includes position (x, y, z) and attitude (roll, pitch, yaw), with position in meters (m) and attitude in rads (rad), derived from RTK pose P. rtk and visual-IMU pose P vio The weighted average is obtained, and the weighting coefficient ω is dynamically adjusted based on the reliability of the two data sources, as shown in the following formula:

[0043] ;

[0044] ;

[0045] The meanings and calculation methods of the parameters in formulas (2) and (3):

[0046] 1)σ rtk RTK pose accuracy is the product of the position accuracy factor (HDOP) output by the RTK device and the fixed error, determined based on actual measurements at the substation: σ in static condition rtk =0.01m (1cm), σ during dynamic flight rtk =0.015m (1.5cm), σ when the signal is interrupted rtk Set to ∞ (weights are zero);

[0047] 2)σ vio Visual-IMU pose error is calculated by the reprojection error of feature points between two adjacent frames, using the formula σ. vio =√(Σ(d²) / n), where d is the reprojection error of the i-th feature point and n is the number of matching feature points. This is obtained by converting the error through "pixel error → angle error → position error" in conjunction with camera intrinsic parameters. When n < 50, σ vio Set to 0.05m;

[0048] 3) Weight adjustment logic: Under normal circumstances, the RTK pose accuracy σ rtk Much smaller than the visual-IMU pose error σ vio ω=0.8-0.9, prioritize trusting RTK data; when strong electromagnetic interference from the substation causes the RTK signal-to-noise ratio to be <20dB, σ rtk As the signal increases, ω automatically decreases to 0.1-0.2, prioritizing the trust of visual-IMU data; after signal recovery, ω smoothly recovers from 0.2 to 0.8 within 0.5s using exponential smoothing to avoid abrupt pose changes.

[0049] 2.3 Reliability Verification of Fusion Results

[0050] A dual verification mechanism of "outlier removal + fixed marker verification" is adopted to ensure pose reliability:

[0051] 1) RANSAC outlier removal: The RANSAC algorithm is used to remove outliers caused by sudden sensor errors for 10 consecutive frames of fused pose. The inlier rate must be ≥90%.

[0052] 2) Fixed marker verification: 10-15 optical markers are pre-positioned in the substation. Their actual coordinates are accurately measured using a Leica TS60 total station. The pixel coordinates of the markers in the image corresponding to the fused pose are converted into world coordinates through the camera projection matrix and compared with the measured coordinates. If the error is >3cm, the local area is re-acquired to ensure pose accuracy.

[0053] Furthermore, step S3 specifically includes:

[0054] 3.1 Laser point cloud preprocessing

[0055] 1) Noise filtering: A combination strategy of "pass-through filtering + statistical filtering" is adopted. The pass-through filtering is set to a Z-axis range of 0-50m, and the statistical filtering is set to a neighboring point count of 15 and a standard deviation threshold of 2. After filtering, the point cloud noise rate is ≤0.5%.

[0056] 2) Point cloud registration: Using the fused pose as the initial value, the NDT registration algorithm is used to fuse the point clouds from each viewpoint into a global point cloud, with a registration error ≤0.3cm;

[0057] 3) Downsampling optimization: Voxel grid downsampling is adopted to optimize the point cloud density to 500 points / ㎡ while preserving details, thereby reducing the amount of subsequent calculations;

[0058] 3.2 High-resolution image preprocessing

[0059] 1) Illumination equalization: The multi-scale Retinex algorithm is used to enhance the uniformity of image illumination and solve the texture distortion problem caused by backlighting and shadows in outdoor substation scenes;

[0060] 2) Distortion correction: The camera intrinsic parameters and distortion coefficients are obtained in advance using the Zhang Zhengyou calibration method. Radial and tangential distortion corrections are implemented using OpenCV. The pixel error after correction is ≤0.5 pixels.

[0061] 3) Texture Enhancement: The CLAHE algorithm is used to improve the contrast of the device's surface texture;

[0062] 3.3 Cross-modal feature matching

[0063] Achieve precise matching between point cloud geometric features and image texture features, thus endowing Gaussian models with realistic textures:

[0064] 1) Image feature extraction: The SIFT algorithm is used to extract key points and descriptors from the image;

[0065] 2) Point cloud feature extraction: The FPFH algorithm is used to extract local geometric features of the point cloud;

[0066] 3) Feature matching: Cross-modal feature matching is achieved based on the RANSAC algorithm, with matching error controlled within 2 pixels and matching success rate ≥95%.

[0067] Furthermore, step S4 specifically includes:

[0068] 4.1 Initial 3D Gaussian Model Construction

[0069] Taking the preprocessed global point cloud as input, a 3D Gaussian is initialized for each point cloud point. The Gaussian parameters include:

[0070] 1) Position parameters: Point cloud 3D coordinates (x, y, z), the coordinate system is unified to Beijing 54 coordinate system, which is obtained from WGS84 coordinates through seven-parameter transformation;

[0071] 2) Shape parameters: covariance matrix, initially set to 0.01I, controlling the size and orientation of the Gaussian ellipsoid, and subsequently optimized by gradient descent;

[0072] 3) Appearance parameters: color, opacity;

[0073] 4) Semantic parameters: Device type label, obtained from the device ledger via UUID association;

[0074] 5) The initial number of Gaussians is determined based on the point cloud size. The number of Gaussians for a single 220kV main transformer is controlled between 500,000 and 800,000 to ensure the integrity of the initial model.

[0075] 4.2 Evaluation of the structural complexity of substation equipment

[0076] A complexity index C, which integrates geometric and textural features, is constructed to quantify the modeling difficulty of different regions of the device and provide a basis for density allocation. This index is designed based on the engineering experience that "the more complex the geometry and the richer the texture, the higher the Gaussian density is required to preserve details." The complexity index is calculated as shown in formula (4):

[0077] ;

[0078] The meaning and calculation method of each parameter in the formula:

[0079] 1) K: Point cloud curvature, calculated using principal component analysis. For each point cloud point, a neighborhood point with a radius of 3cm is taken, and the eigenvalues ​​λ1≥λ2≥λ3 are solved through the covariance matrix. λ1 is the largest eigenvalue, and λ3 is the smallest eigenvalue. Curvature K=(λ2+λ3) / (λ1+λ2+λ3), K∈[0,1]. The larger the value of K, the more complex the surface.

[0080] 2) E: Image texture entropy, calculated using the gray-level co-occurrence matrix, with a window size of 5×5 and 256 gray levels. P(i,j) is the joint probability of gray values ​​i and j in GLCM. The larger the E value, the richer the texture.

[0081] 3) α: Weighting coefficient, determined based on actual measurements during substation modeling;

[0082] 4) K_max / K_min and E_max / E_min: These are the maximum / minimum curvature and maximum / minimum texture entropy of the entire device's point cloud, respectively, used for normalization calculations to ensure that C∈[0,1], thus ensuring that the complexity metrics of different devices can be compared.

[0083] Based on the C value, the device area is divided into three levels, corresponding to different density allocation strategies:

[0084] 1) High-complexity regions, C≥0.7: Geometrically complex and rich in texture, requiring high-density modeling;

[0085] 2) Medium complexity region, 0.3≤C<0.7: The geometry is relatively regular but has local details, requiring medium density modeling;

[0086] 3) Low complexity regions, C<0.3: geometrically regular and with simple textures, requiring low-density modeling;

[0087] 4.3 Dynamic Allocation and Optimization of Gaussian Density

[0088] Differentiated Gaussian optimization strategies based on "split-merge-prune" are employed for regions of varying complexity to achieve adaptive density adjustment.

[0089] 1) In high-complexity regions, C≥0.7: Gaussian splitting algorithm is adopted. When the gradient error of Gaussian is >0.05, it is split into two new Gaussian points along the direction of maximum gradient. The spacing between Gaussian points after splitting is controlled at 0.2-0.3cm to ensure the integrity of the umbrella skirt gap and the details of the joint thread.

[0090] 2) In the medium complexity region, 0.3≤C<0.7: the Gaussian merging algorithm is adopted. When the Euclidean distance between two adjacent Gaussians is <0.1cm and the RGB color difference is <5, they are merged into one Gaussian. After merging, the spacing between Gaussian points is controlled at 0.5-0.8cm to maintain the integrity of the structure while reducing redundancy.

[0091] 3) In the low complexity region, C<0.3: Gaussian pruning algorithm is adopted. When the opacity of Gaussian points is less than 0.01 after iterative optimization, they are removed from the model. After pruning, the spacing between Gaussian points is controlled at 1-2cm, which greatly reduces the storage requirements.

[0092] Furthermore, step S5 specifically includes:

[0093] 5.1 Precise edge detection of equipment

[0094] 1) 2D edge extraction: The 3D Gaussian model is projected onto the 2D image plane through camera intrinsic parameters, and the Canny edge detection algorithm is used to extract the device edges, ensuring an edge detection rate of ≥98%;

[0095] 2) Edge repair and classification: Broken edges are connected by morphological dilation, and continuous edges with a single pixel width are obtained by skeleton extraction; based on the edge gradient direction change rate, the edges are divided into three categories: curved edges, sharp edges, and smooth edges.

[0096] 5.2 Scale-Adaptive Filtering Optimization

[0097] Dynamically adjust the filter window size and parameters based on edge type and gradient value G to achieve differentiated anti-aliasing optimization:

[0098] Filter window size selection: The gradient value G reflects the sharpness of the edge. The edge gradient is calculated by the Sobel operator (G=√(G²+G²), where G and G are the gradients in the x and y directions, respectively). The larger G is, the sharper the edge, and the smaller the window needs to be to retain details. The selection rule for the window size S is as shown in formula (5). This rule is designed based on the balance between edge smoothing requirements and detail retention.

[0099] (5);

[0100] Bilateral filtering optimization: Bilateral filtering is applied to the Gaussian parameters of the edge region. The weight of the filter kernel is determined by the spatial distance weight W and the color distance weight W, and the formula is W=W×W.

[0101] The spatial distance weight W is calculated using a Gaussian function: W(x,x')=exp(-||x-x'||² / (2σ²)), where x and x' are the spatial coordinates of the Gaussian point, and σ is half the window size (e.g., σ=1.5 for a 3×3 window). The color distance weight W is calculated based on the Euclidean distance in the RGB color space: W(c,c')=exp(-||c-c'||² / (2σ²)), where c and c' are the RGB color values ​​of the Gaussian point, and σ is determined based on actual measurements of color differences on the device surface to ensure clear and jagged edges after filtering.

[0102] Furthermore, step S6 specifically includes:

[0103] 6.1 Embedding of Device Semantic Information

[0104] The entire lifecycle information of substation equipment is used as semantic tags and embedded into corresponding Gaussian clusters. The tag information follows the "GB / T 39554-2020 Unified Specification for Power Grid Information Model", specifically including:

[0105] 1) Basic information: Equipment asset number, model, manufacturer, and installation date;

[0106] 2) Operation and maintenance information: last maintenance time, defect record, and operating status;

[0107] 3) Geometric information: Equipment positioning coordinates and dimensional parameters;

[0108] Semantic tags are uniquely associated with Gaussian clusters via UUIDs, ensuring the accuracy of information retrieval.

[0109] 6.2 Mapping and Conversion from Gaussian Model to GIM Format

[0110] Design a two-step conversion process: "Gaussian model → 3DTiles → GIM" to ensure no loss of accuracy during the conversion:

[0111] 1) Gaussian model to 3DTiles: The optimized Gaussian model is converted to 3DTiles format using the open-source library CesiumGS. This format supports streaming loading and real-time rendering. The conversion process uses bilinear interpolation to maintain geometric accuracy. Specifically, when projecting the Gaussian ellipsoid onto the tile grid, bilinear interpolation is used to calculate the color value of the edge pixels, with an error ≤0.1cm.

[0112] 2) 3DTiles to GIM format conversion: Develop a dedicated format conversion tool to map the geometric data in 3DTiles to the components of the GIM model and the semantic tags to the GIM attributes. The converted GIM model conforms to the access specifications of the State Grid PMS 3.0 system (Q / GDW 1929.3-2023).

[0113] A high-precision 3D Gaussian modeling device for substations based on RTK real-time fusion and multi-scale anti-aliasing includes:

[0114] Multi-source data acquisition module: includes RTK module, IMU unit, lidar, high-definition camera and time synchronizer. The time synchronizer is connected to RTK unit, IMU unit, lidar and high-definition camera respectively. With the improved A* algorithm, the route is planned to avoid strong electromagnetic interference sources and acquire image data and point cloud data with timestamp, RTK position and IMU attitude labels.

[0115] Real-time pose fusion module: used to perform linear interpolation time alignment on the acquired image and point cloud data, and fuse RTK high-precision position and visual-IMU high-frequency pose through an adaptive weight algorithm to output device pose data;

[0116] Data preprocessing module: Used to perform noise filtering, registration and downsampling on point cloud data using device pose data, and to perform illumination equalization, distortion correction and texture enhancement on image data, and to achieve cross-modal feature matching to obtain preprocessed point cloud and image data;

[0117] Structure-aware modeling module: Based on preprocessed point cloud and image data, it calculates the structural complexity index C of the device and adaptively allocates Gaussian density through a "split-merge-prune" strategy to construct a 3D Gaussian model;

[0118] Multi-scale anti-aliasing optimization module: This module projects the constructed Gaussian model onto a 2D plane, detects and classifies device edges, dynamically adjusts the size of the filtering window according to the edge type, eliminates edge jagged artifacts, and obtains the optimized Gaussian model.

[0119] Standard format output module: used to convert the optimized Gaussian model into 3DTiles format, then map it to GIM format and embed semantic information to interface with the power PMS system.

[0120] A 3D Gaussian high-precision modeling system for substations based on RTK real-time fusion and multi-scale anti-aliasing includes: a computer-readable storage medium and a processor;

[0121] The computer-readable storage medium is used to store executable instructions;

[0122] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the RTK-based real-time fusion and multi-scale anti-aliasing substation 3D Gaussian high-precision modeling method.

[0123] A non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned RTK-based real-time fusion and multi-scale anti-aliasing method for 3D Gaussian high-precision modeling of substations.

[0124] The present invention has the following beneficial effects:

[0125] 1) Achieve centimeter-level accurate acquisition of device pose: Through real-time fusion of multi-source data, the device pose error is reduced to below 0.1° (attitude) and 1cm (position), eliminating modeling misalignment and ensuring that the geometric accuracy of the model reaches the 0.5cm level;

[0126] 2) Eliminate jagged artifacts at equipment edges: Design a multi-scale anti-aliasing algorithm adapted to substation equipment to improve the SSIM of model edges to above 0.92 and the PSNR to above 35dB, meeting the identification requirements of minute defects at the 0.2mm level;

[0127] 3) Achieve adaptive optimization of Gaussian density: Dynamically allocate Gaussian density based on the complexity of equipment structure, and reduce the storage of a single station model by more than 60% while ensuring the integrity of details, and control the storage of a single main transformer model to within 1GB;

[0128] 4) Achieve seamless integration with power standard systems: Directly output GIM format models with semantic information without secondary conversion; the model-PMS system integration error is ≤0.5cm, reducing engineering application costs. Attached Figure Description

[0129] Figure 1 This is a flowchart of the 3D Gaussian high-precision modeling method for substations based on RTK real-time fusion and multi-scale anti-aliasing, which is based on the present invention.

[0130] Figure 2 This is a flowchart of step two of the present invention: RTK-vision-inertial multi-source real-time fusion.

[0131] Figure 3 This is a flowchart of step five of the present invention: multi-scale anti-aliasing optimization for substation equipment adaptation. Detailed Implementation

[0132] 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.

[0133] The first aspect of this invention provides a 3D Gaussian high-precision modeling method for substations based on RTK real-time fusion and multi-scale anti-aliasing. The core link is "accurate acquisition of multi-source data - real-time pose fusion - structure-aware modeling - multi-scale optimization - standard format output," and the method consists of six steps. Each step works collaboratively to achieve high-precision modeling. The overall flowchart of the technical solution is shown below. Figure 1 As shown. The method includes the following steps:

[0134] Step 1: Construction and Data Acquisition of a Multi-Source Data Acquisition System

[0135] The core of this step is to build a synchronized, high-precision data acquisition hardware platform and design an acquisition strategy adapted to the substation scenario to provide high-quality raw data for subsequent modeling.

[0136] 1.1 Hardware System Configuration

[0137] A customized UAV data acquisition platform was adopted, and each device achieved millisecond-level data alignment through a time synchronizer. The hardware configuration is shown in Table 1.

[0138] Table 1

[0139]

[0140] 1.2 Data Acquisition Strategy for Substation Adaptation

[0141] To address the challenges of uneven equipment distribution and complex electromagnetic environments in substations, a data acquisition strategy of "zonal planning - path optimization - synchronous control" is adopted. The specific steps are as follows:

[0142] 1) Data collection area classification: Based on the substation CAD drawings and equipment ledger, the data collection area is divided into three levels. Level 1 area (core equipment area): main transformer, circuit breaker, disconnector, etc., flight altitude 15-20m, flight path interval 2m to ensure that the equipment surface is unobstructed.

[0143] 2) Secondary area (auxiliary equipment area): surge arresters, voltage transformers, etc., flight altitude 20-25m, flight path spacing 3m;

[0144] 3) Level 3 area (structure and passageway area): busbar structure, inspection passageway, etc., flight altitude 30-40m, line spacing 5m;

[0145] 4) Anti-interference route optimization: The improved A* algorithm is used to plan the route to avoid strong electromagnetic interference sources such as main transformers and GIS equipment (distance ≥ 5m), while ensuring that each equipment surface is covered by at least 3 viewpoints to meet the multi-view constraints of 3D reconstruction.

[0146] 5) Multi-device synchronous acquisition and control: The drone flight control system (DJI SDK) triggers the synchronous operation of each device. RTK and IMU data are uploaded to the ground station in real time (latency ≤50ms). The LiDAR and camera collect data at 0.5s intervals. Each frame of image and point cloud data is accompanied by a precise timestamp, RTK position and IMU attitude label.

[0147] Step 2: Real-time fusion of RTK-vision-inertial multi-source systems

[0148] This step is crucial for solving the pose drift problem. By designing an adaptive weight fusion algorithm, the high-precision position information from the RTK is fused with the high-frequency pose information from the vision-IMU in real time, outputting accurate and stable device pose data. The sub-flowchart is shown below. Figure 2 As shown.

[0149] 2.1 Timestamp Alignment of Multi-Source Data

[0150] Due to the different data update rates of each sensor (RTK 10Hz, IMU 100Hz, camera 2Hz, LiDAR 2Hz), a linear interpolation method is used to achieve time synchronization, ensuring that multi-source data at the same moment correspond to the same device state. This method is based on the engineering assumption that "the sensor's motion state is approximately uniform over a short period of time," and is suitable for substation scenarios where UAVs fly at low speeds (≤5m / s). The interpolation error can be controlled within the sub-centimeter level.

[0151] Taking the alignment of RTK data and camera images as an example, let the times of two consecutive adjacent frames of RTK data be t. i t i+1 (i is the RTK data index, and t) i+1 > t i The time of a single frame image from the camera is t. img If t i < t img <t i+1 (i.e., the image time is between two adjacent frames of RTK data time), then t img RTK position P at time img The calculation is as shown in formula (1):

[0152] (1)

[0153] In formula (1), For t img RTK 3D position coordinates at time (WGS84 coordinate system, unit: m); P i P i+1 t i t i+1 RTK 3D position coordinates at time t; denominator t i+1 -t i The time interval between adjacent RTK data (fixed at 0.1s because the RTK update rate is 10Hz); numerator t img -t i This represents the time difference (in seconds) between the image time and the time of the previous frame of RTK data. This linear interpolation method converts the time deviation of asynchronous data into positional deviation, with a measured interpolation error ≤0.1cm, meeting the modeling accuracy requirements.

[0154] 2.2 Adaptive Weight Fusion Model

[0155] Define the fused pose as P fuse (Includes position (x, y, z) and attitude (roll, pitch, yaw), position unit: m, attitude unit: rad), by RTK pose P rtk and visual-IMU pose P vioThe weighted average is obtained. The weighting coefficient ω is dynamically adjusted based on the reliability of the two data sources, based on the principle of "reciprocal weighting of variance"—the smaller the variance of the data source error, the higher the reliability, and the greater the weight. The formula is as follows:

[0156]

[0157]

[0158] The meanings and calculation methods of the parameters in formulas (2) and (3):

[0159] 1)σ rtk RTK pose accuracy (unit: m) is the product of the position accuracy factor (HDOP) output by the RTK device and the fixed error, determined based on actual measurements at the substation: σ in static condition rtk =0.01m (1cm), σ during dynamic flight rtk =0.015m (1.5cm), σ when signal interruption (signal-to-noise ratio <15dB) rtk Set to ∞ (weights are zero);

[0160] 2)σ vio Visual-IMU pose error (unit: m) is calculated by the reprojection error of feature points between two adjacent frames, using the formula σ. vio =√(Σ(d²) / n), where d is the reprojection error of the i-th feature point (unit: pixels), and n is the number of matching feature points (n≥50). Combined with camera intrinsic parameters (focal length f=35mm, pixel size 1.5μm), it is obtained through the conversion of "pixel error → angle error → position error". When n<50 (insufficient feature point matching, low reliability), σ vio Set to 0.05m (5cm);

[0161] 3) Weight adjustment logic: Under normal circumstances (RTK signal-to-noise ratio ≥ 25dB), the RTK pose accuracy σ rtk Much smaller than the visual-IMU pose error σ vio ω=0.8-0.9, RTK data is preferred; when strong electromagnetic interference in the substation (such as in the main transformer area) causes the RTK signal-to-noise ratio to be <20dB, σ rtk As the signal increases, ω automatically decreases to 0.1-0.2, prioritizing the trust of visual-IMU data; after signal recovery, ω smoothly recovers from 0.2 to 0.8 within 0.5s using exponential smoothing to avoid abrupt changes in pose (abrupt change amplitude ≤0.1cm).

[0162] 2.3 Reliability Verification of Fusion Results

[0163] A dual verification mechanism of "outlier removal + fixed marker verification" is adopted to ensure pose reliability:

[0164] 1) RANSAC outlier removal: The RANSAC algorithm (100 iterations, 2cm inlier threshold) is used to remove outliers caused by sudden sensor errors (such as camera shake, LiDAR noise) for 10 consecutive frames of fused pose. The inlier rate must be ≥90%.

[0165] 2) Fixed marker verification: 10-15 optical markers (10cm×10cm, checkerboard pattern) are pre-positioned within the substation. Their actual coordinates are accurately measured using a Leica TS60 total station (error ≤0.1cm). The pixel coordinates of the markers in the image corresponding to the fused pose are converted to world coordinates using the camera projection matrix and compared with the measured coordinates. If the error is >3cm, local area re-acquisition is triggered to ensure pose accuracy.

[0166] Step 3: Data Preprocessing and Feature Extraction

[0167] The collected point cloud and image data are preprocessed in a targeted manner to eliminate noise and distortion, and to extract accurate geometric and texture features, providing high-quality input data for 3D Gaussian modeling.

[0168] 3.1 Laser point cloud preprocessing

[0169] 1) Noise filtering: A combination strategy of "pass-through filtering + statistical filtering" is adopted. The pass-through filtering is set to a Z-axis range of 0-50m (to remove ground and high-altitude noise), and the statistical filtering is set to a neighboring point count of 15 and a standard deviation threshold of 2 (to remove isolated noise points). The point cloud noise rate after filtering is ≤0.5%.

[0170] 2) Point cloud registration: Using the fused pose as the initial value, the NDT registration algorithm (voxel size 0.2m, number of iterations 50) is used to fuse the point clouds from each viewpoint into a global point cloud, with a registration error ≤0.3cm;

[0171] 3) Downsampling optimization: Voxel grid downsampling (voxel size 0.1cm) is adopted to optimize the point cloud density to 500 points / ㎡ while preserving details, thereby reducing the amount of subsequent calculations.

[0172] 3.2 High-resolution image preprocessing

[0173] 1) Illumination equalization: The multi-scale Retinex algorithm is used to enhance the uniformity of image illumination, solving the texture distortion problem caused by backlighting and shadows in outdoor substation scenes. After processing, the image brightness variance is reduced by 60%.

[0174] 2) Distortion correction: The camera intrinsic parameters and distortion coefficients are obtained in advance using the Zhang Zhengyou calibration method (checkerboard size 5cm×5cm). Radial and tangential distortion corrections are implemented through OpenCV. The pixel error after correction is ≤0.5 pixels.

[0175] 3) Texture enhancement: The CLAHE algorithm (contrast limit threshold 2.0, grid size 8×8) is used to improve the texture contrast of the device surface, especially the texture clarity of defective areas such as ceramic sleeve cracks and rust.

[0176] 3.3 Cross-modal feature matching

[0177] Achieve precise matching between point cloud geometric features and image texture features, thus endowing Gaussian models with realistic textures:

[0178] 1) Image feature extraction: SIFT algorithm (6 layers of Gaussian difference pyramid, key point threshold 0.03) is used to extract key points and descriptors of the image;

[0179] 2) Point cloud feature extraction: The FPFH algorithm (neighborhood radius 5cm) is used to extract local geometric features of the point cloud;

[0180] 3) Feature matching: Cross-modal feature matching is achieved based on the RANSAC algorithm, with matching error controlled within 2 pixels and matching success rate ≥95%.

[0181] Step 4: Structure-Aware 3D Gaussian Modeling

[0182] This step is the core of modeling. By constructing an evaluation system for the structural complexity of the equipment, it enables the dynamic allocation of Gaussian density, reducing storage requirements while ensuring detail accuracy, and solving the mechanical problem of density allocation in existing technologies.

[0183] 4.1 Initial 3D Gaussian Model Construction

[0184] Taking the preprocessed global point cloud as input, a 3D Gaussian is initialized for each point cloud point. The Gaussian parameters include:

[0185] 1) Position parameters: Point cloud 3D coordinates (x, y, z), the coordinate system is unified to Beijing 54 coordinate system, which is obtained from WGS84 coordinates through seven-parameter transformation;

[0186] 2) Shape parameters: Covariance matrix, initially set to 0.01I (I is a 3×3 identity matrix), controlling the size and orientation of the Gaussian ellipsoid, and subsequently optimized by gradient descent;

[0187] 3) Appearance parameters: Color (RGB values ​​extracted from matched image pixels, range 0-255), Opacity (initial value 0.8, used to control the visibility of Gaussian points);

[0188] 4) Semantic parameters: Equipment type labels (such as "main transformer porcelain bushing" and "disconnecting switch") are obtained from the equipment ledger through UUID association.

[0189] 5) The initial number of Gaussians is determined based on the point cloud size. The number of Gaussians for a single 220kV main transformer is controlled between 500,000 and 800,000 to ensure the integrity of the initial model.

[0190] 4.2 Evaluation of the structural complexity of substation equipment

[0191] A complexity index C, which integrates geometric and textural features, is constructed to quantify the modeling difficulty of different regions of the device and provide a basis for density allocation. This index is designed based on the engineering experience that "the more complex the geometry and the richer the texture, the higher the Gaussian density is needed to preserve details." The complexity index is calculated as shown in formula (4):

[0192]

[0193] The meaning and calculation method of each parameter in the formula:

[0194] 1) K: Point cloud curvature, calculated using principal component analysis (PCA). For each point cloud point, a neighborhood point with a radius of 3 cm is taken, and the eigenvalues ​​λ1≥λ2≥λ3 (λ1 is the largest eigenvalue and λ3 is the smallest eigenvalue) are solved through the covariance matrix. Curvature K=(λ2+λ3) / (λ1+λ2+λ3), K∈[0,1]. The larger the value of K, the more complex the surface.

[0195] 2) E: Image texture entropy, calculated using the gray-level co-occurrence matrix (GLCM), with a window size of 5×5 and 256 gray levels. (The base of the logarithm is 2, which conforms to the standard definition of information entropy). P(i,j) is the joint probability of gray values ​​i and j in GLCM. The larger the E value, the richer the texture.

[0196] 3) α: Weighting coefficient, taken as 0.6, determined based on actual measurements of substation modeling—the impact of geometric complexity on model accuracy (weight 0.6) is greater than that of texture complexity (weight 0.4).

[0197] 4) K_max / K_min and E_max / E_min: These are the maximum / minimum curvature and maximum / minimum texture entropy of the point cloud of the entire device, respectively. They are used for normalization calculations to make C∈[0,1], ensuring that the complexity indicators of different devices can be compared.

[0198] Based on the C value, the device area is divided into three levels, corresponding to different density allocation strategies:

[0199] 1) High-complexity areas (C≥0.7): such as porcelain umbrella skirts and busbar joints, which are geometrically complex and rich in texture, requiring high-density modeling;

[0200] 2) Medium complexity region (0.3≤C<0.7): such as main transformer tank and circuit breaker casing, which have relatively regular geometry but local details and require medium density modeling;

[0201] 3) Low complexity areas (C<0.3): such as structural beams and equipment brackets, which have regular geometry and simple texture, require low-density modeling.

[0202] 4.3 Dynamic Allocation and Optimization of Gaussian Density

[0203] Differentiated Gaussian optimization strategies based on "split-merge-prune" are employed for regions of varying complexity to achieve adaptive density adjustment.

[0204] 1) High complexity region (C≥0.7): Gaussian splitting algorithm is adopted. When the gradient error of Gaussian (the pixel difference between the rendered image and the original image) > 0.05, it is split into 2 new Gaussian parts along the direction of maximum gradient. The spacing between Gaussian points after splitting is controlled at 0.2-0.3cm to ensure the integrity of details such as umbrella skirt gap and joint thread.

[0205] 2) Medium complexity region (0.3≤C<0.7): Gaussian merging algorithm is adopted. When the Euclidean distance between two adjacent Gaussians is <0.1cm and the RGB color difference is <5 (ΔR+ΔG+ΔB<5), they are merged into one Gaussian. After merging, the spacing between Gaussian points is controlled at 0.5-0.8cm to maintain the integrity of the structure while reducing redundancy.

[0206] 3) Low complexity region (C<0.3): Gaussian pruning algorithm is adopted. When the opacity of Gaussian points is less than 0.01 after iterative optimization, they are removed from the model. After pruning, the spacing between Gaussian points is controlled at 1-2cm, which greatly reduces storage requirements.

[0207] After optimization, the number of Gaussians in a single 220kV main transformer is controlled to within 1.2 million, which is 65% less than the fixed density scheme, while the integrity of key details is maintained at 99%.

[0208] Step 5: Multi-scale anti-aliasing optimization for substation equipment adaptation

[0209] This step designs a scale-adaptive anti-aliasing algorithm based on the edge features of substation equipment to eliminate edge artifacts in traditional 3DGS, thereby improving the model's visual quality and defect recognition capabilities. The sub-flowchart is shown below. Figure 3 As shown.

[0210] 5.1 Precise edge detection of equipment

[0211] 1) 2D Edge Extraction: The 3D Gaussian model is projected onto the 2D image plane through camera intrinsic parameters. The Canny edge detection algorithm (high threshold 150, low threshold 50, determined based on actual measurement of grayscale features of substation equipment) is used to extract the equipment edges, ensuring an edge detection rate ≥98%.

[0212] 2) Edge repair and classification: The broken edges are connected by morphological dilation (3×3 rectangular convolution kernel), and continuous edges with a single pixel width are obtained by skeleton extraction. Based on the edge gradient direction change rate (curved edge change rate <10° / pixel, sharp edge change rate >45° / pixel, and smooth edge in between), the edges are divided into three categories: curved edges (such as ceramic umbrella skirt), sharp edges (such as knife switch contacts), and smooth edges (such as equipment shell).

[0213] 5.2 Scale-Adaptive Filtering Optimization

[0214] Dynamically adjust the filter window size and parameters based on edge type and gradient value G to achieve differentiated anti-aliasing optimization:

[0215] Filter window size selection: The gradient value G reflects the sharpness of the edge. The edge gradient is calculated by the Sobel operator (G=√(G²+G²), where G and G are the gradients in the x and y directions, respectively). The larger G is, the sharper the edge, and the smaller the window needs to be to retain details. The selection rule for the window size S is as shown in formula (5). This rule is designed based on the balance between edge smoothing requirements and detail retention.

[0216] (5)

[0217] Bilateral filtering optimization: Bilateral filtering is applied to the Gaussian parameters of the edge region. The filter kernel weights are jointly determined by the spatial distance weight W and the color distance weight W, with the formula W = W × W. This method is based on the principle that "spatially adjacent and color-similar Gaussian points should have similar parameters," which can smooth jagged edges without blurring them.

[0218] The spatial distance weight W is calculated using a Gaussian function: W(x,x')=exp(-||x-x'||² / (2σ²)), where x and x' are the spatial coordinates of the Gaussian point, and σ is half the window size (e.g., σ=1.5 for a 3×3 window). The color distance weight W is calculated based on the Euclidean distance in the RGB color space: W(c,c')=exp(-||c-c'||² / (2σ²)), where c and c' are the RGB color values ​​of the Gaussian point, and σ is 20 (determined based on actual measurements of color differences on the device surface) to ensure clear and jagged edges after filtering.

[0219] 5.3 Quantitative Verification of Optimization Effect

[0220] The anti-aliasing effect is quantified using two metrics: PSNR and SSIM.

[0221] 1) Peak Signal-to-Noise Ratio (PSNR): Calculated using the formula PSNR=10log((2-1)² / MSE), where MSE is the mean square error between the optimized model rendering and the original high-definition image. The optimized edge region PSNR is improved to over 35dB, a 40% improvement over existing technologies (25dB), significantly reducing image noise.

[0222] 2) Structural Similarity (SSIM): Calculated using the formula SSIM=(2μμ+C)(2σ+C) / [(μ²+μ²+C)(σ²+σ²+C)], where μ is the mean, σ is the variance, σ is the covariance, and C and C are constants. After optimization, the SSIM of the edge region is improved to over 0.92, a 41.5% improvement compared to the existing technology (0.65), significantly enhancing the integrity of the edge structure.

[0223] Visual inspection shows that the optimized model has smooth edges, no jagged edges or floating artifacts, and artificial crack defects with a width of 0.2 mm can be clearly identified.

[0224] Step Six: Output GIM in Standard Format and Integrate with the System

[0225] This step enables the direct conversion of 3D Gaussian models to the standard format of the power industry, embedding device semantic information to ensure seamless integration with the PMS system and solve the problem of poor compatibility with existing technologies.

[0226] 6.1 Embedding of Device Semantic Information

[0227] The entire lifecycle information of substation equipment is used as semantic tags and embedded into corresponding Gaussian clusters (Gaussian points of the same equipment form a cluster). The tag information follows the "GB / T 39554-2020 Unified Specification for Power Grid Information Model", specifically including:

[0228] 1) Basic information: Equipment asset number (e.g., "220kV Wuchang Station #1 Main Transformer-001"), model (e.g., "SFZ11-240000 / 220"), manufacturer (e.g., "TBEA"), installation date (accurate to the day, e.g., "2023-05-10");

[0229] 2) Operation and maintenance information: last maintenance time (e.g., "2024-01-15"), defect record (associated with PMS defect ID, e.g., "Q20240115003"), and operation status (e.g., "normal", "alert", "defect").

[0230] 3) Geometric information: Equipment positioning coordinates (Beijing 54 coordinate system, such as "X:4235678.23, Y:356789.45, Z:23.67") and dimensional parameters (such as "height 1200cm, diameter 80cm").

[0231] Semantic tags are uniquely associated with Gaussian clusters via UUIDs, ensuring the accuracy of information retrieval.

[0232] 6.2 Mapping and Conversion from Gaussian Model to GIM Format

[0233] Design a two-step conversion process: "Gaussian model → 3DTiles → GIM" to ensure no loss of accuracy during the conversion:

[0234] 1) Gaussian Model to 3DTiles: The optimized Gaussian model is converted to 3DTiles format using the open-source library CesiumGS (which supports the KHR_gaussian_splatting extension). This format supports streaming loading and real-time rendering. The conversion process uses bilinear interpolation to maintain geometric accuracy. Specifically, when projecting the Gaussian ellipsoid onto the tile grid, bilinear interpolation is used to calculate the color value of edge pixels, with an error ≤0.1cm.

[0235] 2) 3DTiles to GIM format conversion: Develop a dedicated format conversion tool (based on GISBox SDK V3.0) to map the geometric data (vertices, faces) in 3DTiles to the components of the GIM model (such as "main transformer body" and "ceramic sleeve"), and map the semantic tags to GIM attributes (corresponding to "basic equipment attributes" and "operation and maintenance attributes" in "GB / T 39554-2020"). The converted GIM model conforms to the access specifications of the State Grid PMS 3.0 system (Q / GDW 1929.3-2023).

[0236] 6.3 Verification of Integration with PMS System

[0237] The converted GIM model was imported into the State Grid PMS 3.0 system for verification of three core indicators:

[0238] 1) Equipment positioning accuracy: By comparing the equipment coordinates in the model with the total station measured coordinates using the coordinate calibration tool built into the PMS system, the error is ≤0.5cm, which meets the positioning requirements of operation and maintenance (such as the navigation accuracy requirement of ≤1cm for drone inspection).

[0239] 2) Attribute query response: Clicking on the model device in the PMS system triggers the semantic information query interface. Optimized by Redis caching, the response time is ≤1s, meeting the real-time requirements of operation and maintenance.

[0240] 3) Model loading speed: The entire 220kV substation model (including 200 devices and approximately 15 million Gaussian points) is optimized through LOD (Level of Detail), with a loading time of ≤30s on mobile devices (such as Huawei Mate 60 Pro), supporting smooth browsing.

[0241] This invention achieves the following beneficial effects through a multi-module collaborative scheme:

[0242] 1. Significantly improved pose accuracy: The RTK-Vision-IMU real-time fusion module (adaptive weighting algorithm) reduces pose error to below 0.1° (attitude) and 1cm (position), with an error of ≤0.3cm under electromagnetic interference, representing an 85% improvement over existing technologies;

[0243] 2. Significantly improved edge quality: The multi-scale anti-aliasing module (differential filtering) achieves an edge SSIM of 0.93 (compared to 0.65) and PSNR ≥ 35dB, enabling the identification of minute defects as small as 0.2mm.

[0244] 3. Balance between storage efficiency and accuracy: The storage of a single main variable model in the structure-aware modeling module (complexity-driven density allocation) has been reduced to 0.8GB (from 2.3GB), a reduction of 65%, while maintaining 99% detail integrity;

[0245] 4. Seamless integration with power systems: The standard format output module directly generates GIM format, with an integration error of ≤0.3cm with PMS, eliminating the need for secondary conversion. The modeling cycle for a single station is shortened from 3 weeks to 5 days, and labor costs are reduced by 70%.

[0246] 5. Wide applicability: Supports 110kV-1000kV substations and can be extended to converter stations and other scenarios. The model is lightweight (≤5GB per station) and adaptable to multi-terminal deployment.

[0247] This invention also provides a 3D Gaussian high-precision modeling device for substations based on RTK real-time fusion and multi-scale anti-aliasing, comprising:

[0248] Multi-source data acquisition module: includes RTK module, IMU unit, lidar, high-definition camera and time synchronizer. The time synchronizer is connected to RTK unit, IMU unit, lidar and high-definition camera respectively. With the improved A* algorithm, the route is planned to avoid strong electromagnetic interference sources and acquire image data and point cloud data with timestamp, RTK position and IMU attitude labels.

[0249] Real-time pose fusion module: used to perform linear interpolation time alignment on the acquired image and point cloud data, and fuse RTK high-precision position and visual-IMU high-frequency pose through an adaptive weight algorithm to output device pose data;

[0250] Data preprocessing module: Used to perform noise filtering, registration and downsampling on point cloud data using device pose data, and to perform illumination equalization, distortion correction and texture enhancement on image data, and to achieve cross-modal feature matching to obtain preprocessed point cloud and image data;

[0251] Structure-aware modeling module: Based on preprocessed point cloud and image data, it calculates the structural complexity index C of the device and adaptively allocates Gaussian density through a "split-merge-prune" strategy to construct a 3D Gaussian model;

[0252] Multi-scale anti-aliasing optimization module: This module projects the constructed Gaussian model onto a 2D plane, detects and classifies device edges, dynamically adjusts the size of the filtering window according to the edge type, eliminates edge jagged artifacts, and obtains the optimized Gaussian model.

[0253] Standard format output module: used to convert the optimized Gaussian model into 3DTiles format, then map it to GIM format and embed semantic information to interface with the power PMS system.

[0254] Another aspect of the present invention provides a 3D Gaussian high-precision modeling system for substations based on RTK real-time fusion and multi-scale anti-aliasing, comprising: a computer-readable storage medium and a processor;

[0255] The computer-readable storage medium is used to store executable instructions;

[0256] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the substation 3D Gaussian high-precision modeling method based on RTK real-time fusion and multi-scale anti-aliasing as described in the first aspect.

[0257] In another aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the RTK-based real-time fusion and multi-scale anti-aliasing method for 3D Gaussian high-precision modeling of substations as described in the first aspect.

[0258] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0259] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0260] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0261] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0262] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for high-precision 3D Gaussian modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing, characterized in that, Includes the following steps: Step S1. Multi-source data synchronous acquisition: Align RTK, IMU, LiDAR and HD camera data through a time synchronizer, use an improved A* algorithm to plan flight paths to avoid strong electromagnetic interference sources, and acquire image data and point cloud data with timestamps, RTK position and IMU attitude labels; Step S2. Real-time RTK-Vision-IMU Fusion: The images and point cloud data acquired in Step S1 are linearly interpolated for time alignment. The high-precision RTK position and high-frequency vision-IMU pose are fused using an adaptive weighting algorithm to output device pose data. Step S3. Data preprocessing: Using the device pose data from step S2, noise filtering, registration, and downsampling are performed on the point cloud data, and illumination equalization, distortion correction, and texture enhancement are performed on the image data to achieve cross-modal feature matching and obtain preprocessed point cloud and image data. Step S4. Structure-aware Gaussian modeling: Based on the point cloud and image data preprocessed in step S3, calculate the structural complexity index C of the device, and adaptively allocate Gaussian density through the "split-merge-prune" strategy to construct a 3D Gaussian model; Step S5. Multi-scale anti-aliasing optimization: Project the Gaussian model constructed in step S4 onto a 2D plane, detect and classify device edges, dynamically adjust the size of the filtering window according to the edge type, eliminate edge jagged artifacts, and obtain the optimized Gaussian model; Step S6. GIM Format Conversion and System Integration: Convert the Gaussian model optimized in Step S5 into 3DTiles format, then map it to GIM format and embed semantic information to integrate with the power PMS system.

2. The method for high-precision 3D Gaussian modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing as described in claim 1, characterized in that: Step S1 specifically includes: A customized UAV data acquisition platform is used, with RTK modules, IMU units, LiDAR, and high-definition cameras achieving millisecond-level data alignment through a time synchronizer; Data collection area is divided into three levels based on substation CAD drawings and equipment ledgers: Level 1: Main transformers, circuit breakers, disconnect switches, flight altitude 15-20m, flight path spacing 2m, ensuring unobstructed equipment surfaces; Level 2: Surge arresters, voltage transformers, flight altitude 20-25m, flight path spacing 3m; Level 3: Busbar structure, inspection passages, etc., flight altitude 30-40m, flight path spacing 5m. Anti-interference route optimization: The improved A* algorithm is used to plan the route to avoid strong electromagnetic interference sources such as the main transformer and GIS equipment, while ensuring that the surface of each equipment is covered by at least 3 views to meet the multi-view constraints of 3D reconstruction. Multi-device synchronous acquisition and control: The drone flight control system triggers the synchronous operation of each device. RTK and IMU data are uploaded to the ground station in real time. The lidar and camera collect data at 0.5-second intervals. Each frame of image and point cloud data is accompanied by a precise timestamp, RTK position and IMU attitude label.

3. The method for high-precision 3D Gaussian modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing as described in claim 1, characterized in that: Step S2 specifically includes: 2.1 Timestamp Alignment of Multi-Source Data Time synchronization is achieved using a linear interpolation method to ensure that data from multiple sources at the same time correspond to the same device state. 2.2 Adaptive Weight Fusion Model Define the fused pose data as P fuse It includes position (x, y, z) and attitude (roll, pitch, yaw), with position in meters (m) and attitude in rads (rad), derived from RTK pose P. rtk and visual-IMU pose P vio The weighted average is obtained, and the weighting coefficient ω is dynamically adjusted based on the reliability of the two data sources, as shown in the following formula: ； ； The meanings and calculation methods of the parameters in formulas (2) and (3): 1)σ rtk RTK pose accuracy is the product of the position accuracy factor (HDOP) output by the RTK device and the fixed error, determined based on actual measurements at the substation: σ in static condition rtk =0.01m (1cm), σ during dynamic flight rtk =0.015m (1.5cm), σ when the signal is interrupted rtk Set to ∞ (weights are zero); 2)σ vio Visual-IMU pose error is calculated by the reprojection error of feature points between two adjacent frames, using the formula σ. vio =√(Σ(d²) / n), where d is the reprojection error of the i-th feature point and n is the number of matching feature points. This is obtained by converting the error through "pixel error → angle error → position error" in conjunction with camera intrinsic parameters. When n < 50, σ vio Set to 0.05m; 3) Weight adjustment logic: Under normal circumstances, the RTK pose accuracy σ rtk Much smaller than the visual-IMU pose error σ vio ω=0.8-0.9, prioritize trusting RTK data; when strong electromagnetic interference from the substation causes the RTK signal-to-noise ratio to be <20dB, σ rtk As the signal increases, ω automatically decreases to 0.1-0.2, prioritizing the trust of visual-IMU data; after signal recovery, ω smoothly recovers from 0.2 to 0.8 within 0.5s using exponential smoothing to avoid abrupt pose changes. 2.3 Reliability Verification of Fusion Results A dual verification mechanism of "outlier removal + fixed marker verification" is adopted to ensure pose reliability: 1) RANSAC outlier removal: The RANSAC algorithm is used to remove outliers caused by sudden sensor errors for 10 consecutive frames of fused pose. The inlier rate must be ≥90%. 2) Fixed marker verification: 10-15 optical markers are pre-positioned in the substation. Their actual coordinates are accurately measured using a Leica TS60 total station. The pixel coordinates of the markers in the image corresponding to the fused pose are converted into world coordinates through the camera projection matrix and compared with the measured coordinates. If the error is >3cm, the local area is re-acquired to ensure pose accuracy.

4. The method for 3D Gaussian high-precision modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing as described in claim 1, characterized in that: Step S3 specifically includes: 3.1 Laser point cloud preprocessing 1) Noise filtering: A combination strategy of "pass-through filtering + statistical filtering" is adopted. The pass-through filtering is set to a Z-axis range of 0-50m, and the statistical filtering is set to a neighboring point count of 15 and a standard deviation threshold of 2. After filtering, the point cloud noise rate is ≤0.5%. 2) Point cloud registration: Using the fused pose as the initial value, the NDT registration algorithm is used to fuse the point clouds from each viewpoint into a global point cloud, with a registration error ≤0.3cm; 3) Downsampling optimization: Voxel grid downsampling is adopted to optimize the point cloud density to 500 points / ㎡ while preserving details, thereby reducing the amount of subsequent calculations; 3.2 High-resolution image preprocessing 1) Illumination equalization: The multi-scale Retinex algorithm is used to enhance the uniformity of image illumination and solve the texture distortion problem caused by backlighting and shadows in outdoor substation scenes; 2) Distortion correction: The camera intrinsic parameters and distortion coefficients are obtained in advance using the Zhang Zhengyou calibration method. Radial and tangential distortion corrections are implemented using OpenCV. The pixel error after correction is ≤0.5 pixels. 3) Texture Enhancement: The CLAHE algorithm is used to improve the contrast of the device's surface texture; 3.3 Cross-modal feature matching Achieve precise matching between point cloud geometric features and image texture features, thus endowing Gaussian models with realistic textures: 1) Image feature extraction: The SIFT algorithm is used to extract key points and descriptors from the image; 2) Point cloud feature extraction: The FPFH algorithm is used to extract local geometric features of the point cloud; 3) Feature matching: Cross-modal feature matching is achieved based on the RANSAC algorithm, with matching error controlled within 2 pixels and matching success rate ≥95%.

5. The method for high-precision 3D Gaussian modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing as described in claim 1, characterized in that: Step S4 specifically includes: 4.1 Initial 3D Gaussian Model Construction Taking the preprocessed global point cloud as input, a 3D Gaussian is initialized for each point cloud point. The Gaussian parameters include: 1) Position parameters: Point cloud 3D coordinates (x, y, z), the coordinate system is unified to Beijing 54 coordinate system, which is obtained from WGS84 coordinates through seven-parameter transformation; 2) Shape parameters: covariance matrix, initially set to 0.01I, controlling the size and orientation of the Gaussian ellipsoid, and subsequently optimized by gradient descent; 3) Appearance parameters: color, opacity; 4) Semantic parameters: Device type label, obtained from the device ledger via UUID association; 5) The initial number of Gaussians is determined based on the point cloud size. The number of Gaussians for a single 220kV main transformer is controlled between 500,000 and 800,000 to ensure the integrity of the initial model. 4.2 Evaluation of the structural complexity of substation equipment A complexity index C, which integrates geometric and textural features, is constructed to quantify the modeling difficulty of different regions of the device and provide a basis for density allocation. This index is designed based on the engineering experience that "the more complex the geometry and the richer the texture, the higher the Gaussian density is required to preserve details." The complexity index is calculated as shown in formula (4): ； The meaning and calculation method of each parameter in the formula: 1) K: Point cloud curvature, calculated using principal component analysis. For each point cloud point, a neighborhood point with a radius of 3cm is taken, and the eigenvalues ​​λ1≥λ2≥λ3 are solved through the covariance matrix. λ1 is the largest eigenvalue, and λ3 is the smallest eigenvalue. Curvature K=(λ2+λ3) / (λ1+λ2+λ3), K∈[0,1]. The larger the value of K, the more complex the surface. 2) E: Image texture entropy, calculated using the gray-level co-occurrence matrix, with a window size of 5×5 and 256 gray levels. P(i,j) is the joint probability of gray values ​​i and j in GLCM. The larger the E value, the richer the texture. 3) α: Weighting coefficient, determined based on actual measurements during substation modeling; 4) K_max / K_min and E_max / E_min: These are the maximum / minimum curvature and maximum / minimum texture entropy of the entire device's point cloud, respectively, used for normalization calculations to ensure that C∈[0,1], thus ensuring that the complexity metrics of different devices can be compared. Based on the C value, the device area is divided into three levels, corresponding to different density allocation strategies: 1) High-complexity regions, C≥0.7: Geometrically complex and rich in texture, requiring high-density modeling; 2) Medium complexity region, 0.3≤C<0.7: The geometry is relatively regular but has local details, requiring medium density modeling; 3) Low complexity regions, C<0.3: geometrically regular and with simple textures, requiring low-density modeling; 4.3 Dynamic Allocation and Optimization of Gaussian Density Differentiated Gaussian optimization strategies, employing "split-merge-prune," are used to adapt density adjustment for regions of varying complexity. 1) In high-complexity regions, C≥0.7: Gaussian splitting algorithm is adopted. When the gradient error of Gaussian is >0.05, it is split into two new Gaussian points along the direction of maximum gradient. The spacing between Gaussian points after splitting is controlled at 0.2-0.3cm to ensure the integrity of the umbrella skirt gap and the details of the joint thread. 2) In the medium complexity region, 0.3≤C<0.7: the Gaussian merging algorithm is adopted. When the Euclidean distance between two adjacent Gaussians is <0.1cm and the RGB color difference is <5, they are merged into one Gaussian. After merging, the spacing between Gaussian points is controlled at 0.5-0.8cm to maintain the integrity of the structure while reducing redundancy. 3) In the low complexity region, C<0.3: Gaussian pruning algorithm is adopted. When the opacity of Gaussian points is less than 0.01 after iterative optimization, they are removed from the model. After pruning, the spacing between Gaussian points is controlled at 1-2cm, which greatly reduces the storage requirements.

6. The method for high-precision 3D Gaussian modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing as described in claim 1, characterized in that: Step S5 specifically includes: 5.1 Precise edge detection of equipment 1) 2D edge extraction: The 3D Gaussian model is projected onto the 2D image plane through camera intrinsic parameters, and the Canny edge detection algorithm is used to extract the device edges, ensuring an edge detection rate of ≥98%; 2) Edge repair and classification: Broken edges are connected by morphological dilation, and continuous edges with a single pixel width are obtained by skeleton extraction; based on the edge gradient direction change rate, the edges are divided into three categories: curved edges, sharp edges, and smooth edges. 5.2 Scale-Adaptive Filtering Optimization Dynamically adjust the filter window size and parameters based on edge type and gradient value G to achieve differentiated anti-aliasing optimization: Filter window size selection: The gradient value G reflects the sharpness of the edge. The edge gradient is calculated by the Sobel operator (G=√(G²+G²), where G and G are the gradients in the x and y directions, respectively). The larger G is, the sharper the edge, and the smaller the window needs to be to retain details. The selection rule for the window size S is as shown in formula (5). This rule is designed based on the balance between edge smoothing requirements and detail retention. （5）； Bilateral filtering optimization: Bilateral filtering is applied to the Gaussian parameters of the edge region. The weight of the filter kernel is determined by the spatial distance weight W and the color distance weight W, and the formula is W=W×W. The spatial distance weight W is calculated using a Gaussian function: W(x,x')=exp(-||x-x'||² / (2σ²)), where x and x' are the spatial coordinates of the Gaussian point, and σ is half the window size (e.g., σ=1.5 for a 3×3 window). The color distance weight W is calculated based on the Euclidean distance in the RGB color space: W(c,c')=exp(-||c-c'||² / (2σ²)), where c and c' are the RGB color values ​​of the Gaussian point, and σ is determined based on actual measurements of color differences on the device surface to ensure clear and jagged edges after filtering.

7. The method for 3D Gaussian high-precision modeling of substations based on RTK real-time fusion and multi-scale anti-aliasing as described in claim 1, characterized in that: Step S6 specifically includes: 6.1 Embedding of Device Semantic Information The entire lifecycle information of substation equipment is used as semantic tags and embedded into corresponding Gaussian clusters. The tag information follows the "GB / T 39554-2020 Unified Specification for Power Grid Information Model", specifically including: 1) Basic information: Equipment asset number, model, manufacturer, and installation date; 2) Operation and maintenance information: last maintenance time, defect record, and operating status; 3) Geometric information: Equipment positioning coordinates and dimensional parameters; Semantic tags are uniquely associated with Gaussian clusters via UUIDs, ensuring the accuracy of information retrieval; 6.2 Mapping and Conversion from Gaussian Model to GIM Format Design a two-step conversion process: "Gaussian model → 3DTiles → GIM" to ensure no loss of accuracy during the conversion: 1) Gaussian model to 3DTiles: The optimized Gaussian model is converted to 3DTiles format using the open-source library CesiumGS. This format supports streaming loading and real-time rendering. The conversion process uses bilinear interpolation to maintain geometric accuracy. Specifically, when projecting the Gaussian ellipsoid onto the tile grid, bilinear interpolation is used to calculate the color value of the edge pixels, with an error ≤0.1cm. 2) 3DTiles to GIM format conversion: Develop a dedicated format conversion tool to map the geometric data in 3DTiles to the components of the GIM model and the semantic tags to the GIM attributes. The converted GIM model conforms to the access specifications of the State Grid PMS 3.0 system (Q / GDW 1929.3-2023).

8. A 3D Gaussian high-precision modeling device for substations based on RTK real-time fusion and multi-scale anti-aliasing, characterized in that, include: Multi-source data acquisition module: includes RTK module, IMU unit, lidar, high-definition camera and time synchronizer. The time synchronizer is connected to RTK unit, IMU unit, lidar and high-definition camera respectively. With the improved A* algorithm, the route is planned to avoid strong electromagnetic interference sources and acquire image data and point cloud data with timestamp, RTK position and IMU attitude labels. Real-time pose fusion module: used to perform linear interpolation time alignment on the acquired image and point cloud data, and fuse RTK high-precision position and visual-IMU high-frequency pose through an adaptive weight algorithm to output device pose data; Data preprocessing module: Used to perform noise filtering, registration and downsampling on point cloud data using device pose data, and to perform illumination equalization, distortion correction and texture enhancement on image data, and to achieve cross-modal feature matching to obtain preprocessed point cloud and image data; Structure-aware modeling module: Based on preprocessed point cloud and image data, it calculates the structural complexity index C of the device and adaptively allocates Gaussian density through a "split-merge-prune" strategy to construct a 3D Gaussian model; Multi-scale anti-aliasing optimization module: This module projects the constructed Gaussian model onto a 2D plane, detects and classifies device edges, dynamically adjusts the size of the filtering window according to the edge type, eliminates edge jagged artifacts, and obtains the optimized Gaussian model. Standard format output module: used to convert the optimized Gaussian model into 3DTiles format, then map it to GIM format and embed semantic information to interface with the power PMS system.

9. A 3D Gaussian high-precision modeling system for substations based on RTK real-time fusion and multi-scale anti-aliasing, comprising: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the substation 3D Gaussian high-precision modeling method based on RTK real-time fusion and multi-scale anti-aliasing as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the RTK-based real-time fusion and multi-scale anti-aliasing method for 3D Gaussian high-precision modeling of substations according to any one of claims 1-7.

Images