A power transmission line defect three-dimensional quantification method based on RGB-D point cloud and PointNet++

Abstract

This invention discloses a three-dimensional quantification method for transmission line defects based on RGB-D point clouds and PointNet++, aiming to solve the technical problem that traditional two-dimensional visual inspection can only qualitatively identify defects and cannot achieve accurate quantification of physical parameters. The method includes constructing a dual-modal data acquisition framework of RGB texture and dense point cloud, achieving pixel-level spatiotemporal alignment through an online calibration algorithm for heterogeneous sensors; building a semantic segmentation model based on PointNet++ to accurately segment key components of the transmission line; and establishing a differential analysis mechanism between the standard equipment 3D model and the measured point cloud model to automatically calculate key parameters such as icing thickness, hardware corrosion volume, and insulator damage area. This invention breaks through the limitations of traditional two-dimensional visual qualitative identification, realizing a leap from visual inspection to physical quantification of defects, filling the gap in three-dimensional quantitative assessment in the industry, and providing accurate data support for the digital operation and maintenance of transmission lines.

Description

Technical Field

[0001] This invention belongs to the field of power transmission line inspection technology, and in particular relates to a three-dimensional quantification method for power transmission line defects based on RGB-D point cloud and PointNet++. Background Technology

[0002] With the advancement of new power system construction, transmission lines, as the core carriers of energy transmission, are directly related to the safe and stable operation of the power grid due to their operation and maintenance quality. Currently, the inspection of high-voltage transmission lines and remote energy facilities faces a serious pain point of lacking data dimensions: existing inspections mostly rely on ordinary visible light cameras to take two-dimensional photos, lacking depth information, and can only achieve qualitative identification of defects, making it difficult to accurately quantify key physical parameters such as the volume of insulator damage, the degree of hardware deformation, and the thickness of icing.

[0003] Traditional two-dimensional visual inspection has the following shortcomings: First, it cannot obtain three-dimensional spatial information, which leads to defect assessment relying on human experience, resulting in strong subjectivity and large errors; second, the transmission of massive image data consumes a lot of bandwidth, and manual image interpretation is inefficient, making it difficult to meet the needs of digital twin modeling in the digital transformation of the power grid; third, for defects that require quantitative assessment, such as icing and corrosion, there is a lack of effective automated calculation methods, which cannot provide accurate data support for operation and maintenance decisions.

[0004] While existing point cloud-based inspection technologies can acquire 3D information, they suffer from several bottlenecks: firstly, the heterogeneous sensor data from RGB and depth cameras exhibit spatiotemporal misalignment, leading to poor fusion results; secondly, there is a lack of precise semantic segmentation methods for transmission line components, resulting in low utilization of point cloud data; furthermore, an effective differential analysis mechanism between standard and measured models has not yet been established, hindering the automated calculation of defect physical parameters. Therefore, a technical method capable of accurate 3D defect quantization is urgently needed to fill this industry gap. Thus, this paper proposes a 3D defect quantization method and system for transmission lines based on RGB-D point clouds and PointNet++ to address these issues. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a three-dimensional quantization method for transmission line defects based on RGB-D point cloud and PointNet++, which solves the technical problems that traditional two-dimensional visual inspection cannot quantify the physical parameters of defects, and that existing point cloud technology has data alignment differences, low segmentation accuracy and lack of quantization logic, so as to realize the leap from visual inspection to physical quantization of defects.

[0006] To achieve the above technical solution, the technical solution adopted by the present invention is as follows: A method for 3D quantization of transmission line defects based on RGB-D point cloud and PointNet++, comprising: S1, Dual-modal data acquisition and pixel-level spatiotemporal alignment: A dual-modal data acquisition framework for RGB textures and dense point clouds is constructed. RGB images and dense point cloud data of power transmission line components are simultaneously acquired by a drone equipped with an RGB camera and a depth camera. Based on an online calibration algorithm for heterogeneous sensors, the intrinsic parameter matrix, distortion coefficient, and extrinsic parameter matrix between the RGB camera and the depth camera are solved to achieve pixel-level spatiotemporal alignment of the dual-modal data, resulting in spatiotemporally aligned dual-modal fused data. S2, semantic segmentation based on improved PointNet++: An improved PointNet++ semantic segmentation model is constructed. The improvements include introducing RGB texture feature weight factors in the grouping layer to fuse point cloud geometric features and RGB texture features. The dual-modal fusion data is input into the improved PointNet++ semantic segmentation model to output the semantic segmentation results of key components of the transmission line and extract the point cloud subset of the target component. S3, Defect region extraction based on standard model difference analysis: Construct a standard 3D point cloud model of key components of the transmission line; perform high-precision registration between the point cloud subset of the target component and the standard 3D point cloud model; perform difference operation on the registered measured point cloud model and the standard 3D point cloud model, and extract the point cloud with a distance difference greater than the preset distance threshold as the point cloud subset of the defect area; S4. Defect 3D Parameter Calculation: Based on a subset of the point cloud of the defect region, the physical parameters of the defect are automatically calculated. The physical parameters include at least one of the following: ice thickness, metal fitting corrosion volume, or insulator damage area.

[0007] Preferably, the heterogeneous sensor online calibration algorithm in step S1 specifically includes: A checkerboard calibration board was selected, and multiple sets of calibration board images in different poses were captured simultaneously by an RGB camera and a depth camera mounted on a drone to construct a calibration dataset. The intrinsic parameter matrix and distortion coefficients of the RGB camera are solved based on Zhang Zhengyou's calibration method, and its intrinsic parameters are obtained through the calibration tool built into the depth camera. An optimization function is constructed with the goal of minimizing pixel coordinate error. The extrinsic parameter matrix between the RGB camera and the depth camera is iteratively optimized so that the corresponding pixel position deviation between the depth point cloud and the RGB image is ≤1 pixel.

[0008] Preferably, the training process of the improved PointNet++ semantic segmentation model in step S2 includes: Collect sample data of transmission line components covering different defect types and environmental conditions, and label the samples with component categories and defect areas; In the sampling layer of the model, a strategy combining farthest point sampling and random sampling is adopted, and in the grouping layer, a texture feature weight factor that is adaptively adjusted based on the variance of pixel gray values ​​in RGB images is introduced. The total loss function is a weighted sum of the cross-entropy loss function and the Dice loss function. The model is trained until the semantic segmentation accuracy meets the preset threshold.

[0009] Preferably, the high-precision registration of the point cloud in step S3 includes: First, the ICP algorithm is used to perform coarse registration between the target component point cloud subset and the standard 3D point cloud model. Based on the improved Trimmed ICP algorithm, outliers containing background clutter and measurement noise are removed, and the registration error is optimized through iterative optimization to meet the preset accuracy requirements.

[0010] Preferably, the method for calculating the icing thickness in step S4 is as follows: In the defective point cloud subset of the icing area and the corresponding area of ​​the standard model, search the spatial distance of each pair of nearest points, remove outliers, and take the average value as the icing thickness.

[0011] Preferably, the method for calculating the corrosion volume of the hardware in step S4 is as follows: A voxelization method was used to perform 3D modeling of the point cloud subset of the defect area in the rust region, and the model volume was calculated as the rust volume by statistically counting the number of effective voxels.

[0012] Preferably, the method for calculating the damaged area of ​​the insulator in step S4 is as follows: Project a subset of the point cloud of the defective region into the XOY plane, fit the projection boundary using the polygon approximation method, and calculate the area of ​​the irregular region within the boundary as the damaged area.

[0013] Preferably, after step S4, the method further includes: Based on the preset defect level threshold, the calculated physical parameters are rated, and a structured inspection report containing defect location, type, level, and three-dimensional parameters is automatically generated.

[0014] Preferably, a computer device includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to implement the aforementioned method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++.

[0015] Preferably, a computer-readable storage medium stores computer instructions that cause a computer to execute the aforementioned method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++.

[0016] The beneficial effects of this invention are as follows: 1. This invention breaks through the limitations of traditional two-dimensional vision. By constructing a dual-modal acquisition framework of RGB texture and dense point cloud, and combining it with a heterogeneous sensor online calibration algorithm to achieve pixel-level spatiotemporal alignment, it effectively solves the problem of insufficient data dimension of a single sensor and provides a high-quality data foundation for the three-dimensional quantization of defects.

[0017] 2. The present invention has high segmentation accuracy and strong robustness. The improved PointNet++ semantic segmentation model integrates point cloud geometric features and RGB texture features, achieving a segmentation accuracy of ≥98% for power transmission line components, and can adapt to diverse measurement needs in complex field environments.

[0018] 3. This invention combines the advantages of quantitative accuracy and automation. By establishing a point cloud differential analysis mechanism and precise parameter calculation logic, it achieves automated quantification of key parameters such as icing thickness, corrosion volume, and damaged area, with a measurement error of ≤2mm, filling the gap in the industry's three-dimensional quantitative assessment. In addition, this method is highly adaptable to the needs of digital operation and maintenance. The calculation results and structured inspection reports can directly support the digital twin modeling of transmission lines, providing accurate data support for operation and maintenance decisions. This can significantly reduce manual intervention and greatly improve operation and maintenance efficiency. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a schematic diagram of the system structure for dual-modal data acquisition of power transmission lines using an RGB camera and a depth camera mounted on a drone according to the present invention. Figure 3 This is a schematic diagram illustrating the online calibration principle of heterogeneous sensors according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the improved PointNet++ semantic segmentation model structure according to an embodiment of the present invention; Figure 5 This is a flowchart of point cloud registration and differential analysis according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the logic for solving the three-dimensional parameters of defects in an embodiment of the present invention; Figure 7 This is a quantitative effect diagram of corrosion defects in hardware according to an embodiment of the present invention; Figure 8 This is a quantitative effect diagram of insulator damage defects according to an embodiment of the present invention; Figure 9 This is a quantitative effect diagram of conductor icing defects in an embodiment of the present invention. Detailed Implementation

[0020] Example 1: like Figure 1As shown, a three-dimensional quantization method for transmission line defects based on RGB-D point cloud and PointNet++ includes: S1, Dual-modal data acquisition and pixel-level spatiotemporal alignment: A dual-modal data acquisition framework for RGB textures and dense point clouds is constructed. RGB images and dense point cloud data of power transmission line components are simultaneously acquired by a drone equipped with an RGB camera and a depth camera. Based on an online calibration algorithm for heterogeneous sensors, the intrinsic parameter matrix, distortion coefficient, and extrinsic parameter matrix between the RGB camera and the depth camera are solved to achieve pixel-level spatiotemporal alignment of the dual-modal data, resulting in spatiotemporally aligned dual-modal fused data. S2, semantic segmentation based on improved PointNet++: An improved PointNet++ semantic segmentation model is constructed. The improvements include introducing RGB texture feature weight factors in the grouping layer to fuse point cloud geometric features and RGB texture features. The dual-modal fusion data is input into the improved PointNet++ semantic segmentation model to output the semantic segmentation results of key components of the transmission line and extract the point cloud subset of the target component. S3, Defect region extraction based on standard model difference analysis: Construct a standard 3D point cloud model of key components of the transmission line; perform high-precision registration between the point cloud subset of the target component and the standard 3D point cloud model; perform difference operation on the registered measured point cloud model and the standard 3D point cloud model, and extract the point cloud with a distance difference greater than the preset distance threshold as the point cloud subset of the defect area; S4. Defect 3D Parameter Calculation: Based on a subset of the point cloud of the defect region, the physical parameters of the defect are automatically calculated. The physical parameters include at least one of the following: ice thickness, metal fitting corrosion volume, or insulator damage area.

[0021] Preferably, the heterogeneous sensor online calibration algorithm in step S1 specifically includes: A checkerboard calibration board was selected, and multiple sets of calibration board images in different poses were captured simultaneously by an RGB camera and a depth camera mounted on a drone to construct a calibration dataset. The intrinsic parameter matrix and distortion coefficients of the RGB camera are solved based on Zhang Zhengyou's calibration method, and its intrinsic parameters are obtained through the calibration tool built into the depth camera. An optimization function is constructed with the goal of minimizing pixel coordinate error. The extrinsic parameter matrix between the RGB camera and the depth camera is iteratively optimized so that the corresponding pixel position deviation between the depth point cloud and the RGB image is ≤1 pixel.

[0022] Preferably, the training process of the improved PointNet++ semantic segmentation model in step S2 includes: Collect sample data of transmission line components covering different defect types and environmental conditions, and label the samples with component categories and defect areas; In the sampling layer of the model, a strategy combining farthest point sampling and random sampling is adopted, and in the grouping layer, a texture feature weight factor that is adaptively adjusted based on the variance of pixel gray values ​​in RGB images is introduced. The total loss function is a weighted sum of the cross-entropy loss function and the Dice loss function. The model is trained until the semantic segmentation accuracy meets the preset threshold.

[0023] Preferably, the high-precision registration of the point cloud in step S3 includes: First, the ICP algorithm is used to perform coarse registration between the target component point cloud subset and the standard 3D point cloud model. Based on the improved Trimmed ICP algorithm, outliers containing background clutter and measurement noise are removed, and the registration error is optimized through iterative optimization to meet the preset accuracy requirements.

[0024] Preferably, the method for calculating the icing thickness in step S4 is as follows: In the defective point cloud subset of the icing area and the corresponding area of ​​the standard model, search the spatial distance of each pair of nearest points, remove outliers, and take the average value as the icing thickness.

[0025] Preferably, the method for calculating the corrosion volume of the hardware in step S4 is as follows: A voxelization method was used to perform 3D modeling of the point cloud subset of the defect area in the rust region, and the model volume was calculated as the rust volume by statistically counting the number of effective voxels.

[0026] Preferably, the method for calculating the damaged area of ​​the insulator in step S4 is as follows: Project a subset of the point cloud of the defective region into the XOY plane, fit the projection boundary using the polygon approximation method, and calculate the area of ​​the irregular region within the boundary as the damaged area.

[0027] Preferably, after step S4, the method further includes: Based on the preset defect level threshold, the calculated physical parameters are rated, and a structured inspection report containing defect location, type, level, and three-dimensional parameters is automatically generated.

[0028] Preferably, a computer device includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to implement the aforementioned method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++.

[0029] Preferably, a computer-readable storage medium stores computer instructions that cause a computer to execute the aforementioned method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++.

[0030] Example 2: This embodiment provides a three-dimensional quantization method for transmission line defects based on RGB-D point cloud and PointNet++, including the following steps: The dual-modal data acquisition utilizes a framework that combines RGB textures and dense point clouds. A drone equipped with both an RGB camera and a depth camera simultaneously acquires data on critical components of power transmission lines, such as insulators, fittings, and conductors. During acquisition, the drone maintains stable flight attitude using RTK positioning technology, ensuring that the RGB and depth cameras maintain a consistent sampling frequency of 30fps. This allows the acquisition of RGB images containing texture information of the target components and dense point cloud data containing spatial three-dimensional coordinate information.

[0031] Due to the differences in hardware parameters between the RGB camera and the depth camera, the original acquired data is misaligned in time and space. Pixel-level alignment needs to be achieved through an online calibration algorithm using heterogeneous sensors. First, a checkerboard calibration board is selected as the calibration target, with a corner spacing of 50mm. The dual sensors mounted on the drone simultaneously capture 10-20 sets of calibration board images in different postures to construct a calibration dataset. The different postures include rotation angles of 0°-90° and shooting distances of 5-20m.

[0032] Secondly, based on Zhang Zhengyou's calibration method, the intrinsic parameter matrix composed of the focal length and principal point coordinates of the RGB camera and the distortion coefficients composed of radial and tangential distortions are solved. The initial intrinsic parameters are obtained through the calibration tool built into the depth camera. Finally, an optimization function is constructed with the goal of minimizing pixel coordinate error. The extrinsic parameter matrices of the RGB camera and the depth camera, including the rotation matrix R and the translation vector T, are iteratively optimized to ensure that the corresponding pixel position deviation between the depth point cloud and the RGB image is ≤1 pixel, thereby achieving pixel-level spatiotemporal alignment of the dual-modal data.

[0033] To accurately extract point cloud data of key components of transmission lines, the PointNet++ model was improved and used for semantic segmentation. First, sample data of different defect types (icing, corrosion, damage, deformation) and different environmental conditions (strong light, weak light, rain, fog) were collected, covering transmission line components of different voltage levels such as 500kV and 220kV, totaling over 10,000 samples. Component category and defect region labels were then applied to the samples. Second, a strategy combining farthest-point sampling (FPS) and random sampling was introduced at the sampling layer to improve sample representativeness. Finally, an RGB texture feature weighting factor was introduced at the grouping layer. This factor is adaptively adjusted based on the variance of RGB image pixel grayscale values ​​to enhance texture. To assess the fusion effect of geometric and textural features, a dual-modal feature extraction network was constructed, combining point cloud geometric features and RGB texture features. Then, a weighted sum of cross-entropy and Dice loss functions (7:3 weight ratio) was used as the loss function. The Adam optimizer was employed with an initial learning rate of 0.001. Through a learning rate decay strategy, the learning rate was reduced to 0.9 every 10 epochs until the semantic segmentation accuracy on the validation set reached ≥98%, resulting in a pre-trained PointNet++ semantic segmentation model. Finally, the spatiotemporally aligned dual-modal fused data was input into the pre-trained model to output the semantic segmentation results of key components of the transmission line, extracting a subset of the point cloud data of the target components.

[0034] Based on the design drawings and parameters of transmission line components, standard 3D models of key components such as insulators and fittings were constructed using SolidWorks software, and exported as point cloud files in .ply format as a defect comparison benchmark. In the point cloud registration stage, the target component point cloud subset was first coarsely registered with the standard point cloud model using the ICP algorithm to initially eliminate initial positional deviations. Then, the improved TrimmedICP algorithm was used to remove defects. Outliers such as background clutter and measurement noise are removed, and the registration parameters are iteratively optimized to make the registration error ≤0.5mm, resulting in a high-precision registration result. Finally, differential operation is performed to calculate the spatial distance difference between the registered measured point cloud model and the standard equipment 3D point cloud model. 0.5mm is set as the distance threshold, and point clouds with a distance difference greater than this threshold are identified as a subset of the defect area point cloud.

[0035] The 3D parameter calculation of defects is based on the automatic calculation of key physical parameters using a subset of the point cloud of the defect region. The specific logic is as follows: For the ice thickness, the spatial distance between each pair of nearest points in the corresponding region of the point cloud subset of the defect region and the standard model is searched. After removing outliers greater than 3 times the standard deviation, the average value is taken as the ice thickness. For the corrosion volume of the fittings, a voxelization method with a voxel size of 0.1mm×0.1mm×0.1mm is used to perform 3D modeling of the point cloud subset of the defect region. The model volume is calculated as the corrosion volume by counting the number of effective voxels. For the damaged area of ​​the insulator, the point cloud subset of the defect region is projected onto the XOY plane. The projection boundary is fitted using a polygon approximation method, and the area of ​​the irregular region within the boundary is calculated as the damaged area. At the same time, the accuracy is verified by measuring the defect parameters with a laser rangefinder to ensure that the measurement error of the calculation result is ≤2mm.

[0036] Example 3: like Figures 2-9 As shown, in order to make the technical solution, implementation process and beneficial effects of the present invention clearer, the following describes in detail the three-dimensional quantification method of transmission line defects based on RGB-D point cloud and PointNet++ of the present invention with specific embodiments. This embodiment takes a 500kV transmission line as the application scenario and carries out quantification implementation for three typical defects: hardware corrosion, insulator damage and conductor icing.

[0037] 1. Three-dimensional quantification of corrosion defects in hardware: The DJI M300RTK drone was selected as the flight platform, equipped with a Sony RX1RII RGB camera with a resolution of 7952×5304 and a focal length of 35mm, and an Intel RealSense D455 depth camera with a ranging range of 0.1-10m and a depth accuracy of ±2%@2m; the calibration target used a 50mm×50mm checkerboard calibration board; the accuracy verification equipment used a Leica DISTOS 910 laser rangefinder with a measurement accuracy of ±0.1mm; and the data processing terminal used a Lenovo Legion Y9000P laptop.

[0038] The operating system is Windows 11 Professional, the point cloud processing uses the Open3D 0.17.0 library, the deep learning framework is PyTorch 2.10.1, the model training and inference are based on Python 3.9, and the standard 3D model construction uses SolidWorks 2025.

[0039] Z-10 type suspension clamp hardware was selected from a 500kV transmission line section in Y City. The hardware had obvious corrosion defects. The site was unobstructed and had good lighting conditions, with outdoor sunshine from 10:00 to 14:00.

[0040] The drone was maneuvered to a position 8 meters directly in front of the fitting. RTK positioning technology was used to lock the flight attitude, with a hovering error ≤0.1m. Both the RGB and depth cameras were set to a sampling frequency of 30fps, simultaneously acquiring 30 sets of data. This yielded an RGB image containing complete texture information of the fitting and dense point cloud data with a density of 1000 points / cm², in JPG and PLY formats respectively. During the acquisition process, the drone's flight speed was maintained at ≤3m / s to avoid data blurring due to drone vibration.

[0041] The checkerboard calibration plate was fixed on a bracket next to the power transmission line. The attitude of the calibration plate was adjusted so that the UAV could simultaneously take 15 sets of calibration images from different angles and distances to build a calibration dataset, including rotation angles of 0°, 30°, 45°, 60°, and 90°, and distances of 5m, 8m, 10m, 15m, and 20m.

[0042] Based on Zhang Zhengyou's calibration method, the RGB camera intrinsic parameter matrix is ​​solved using the OpenCV library: ; Radial distortion coefficient k1= 0.012, k2=0.006, tangential distortion coefficient p1= 0.0002, p2=0.0001; Initial intrinsic parameters of the depth camera are obtained through Intel RealSense SDK, focal length f. x =384.2、f y =384.5, principal point coordinates u0=320.1, v0=240.3.

[0043] Construct the optimization objective function: ; Where u i ,v i X represents the pixel coordinates of an RGB image. i Let R be the 3D coordinates of the depth point cloud, R be the rotation matrix, and T be the translation vector. The Levenberg-Marquardt algorithm is used to iteratively optimize the extrinsic parameters, and finally pixel-level alignment of the dual-modal data is achieved with an alignment error of 0.7 pixels.

[0044] The spatiotemporally aligned bimodal fusion data is input into the pre-trained model, which is trained based on 10,000+ sets of transmission line component samples. The samples cover 500kV and 220kV voltage levels and include four types of defects such as corrosion and damage, as well as four environmental conditions such as strong light and rain.

[0045] The model selects core points by combining farthest-point sampling FPS with random sampling, and calculates the variance of pixel grayscale values ​​in the RGB image at the grouping layer. Adaptive generation of RGB texture feature weight factors Feature extraction is performed by fusing point cloud geometric features and texture features.

[0046] Output semantic segmentation results for hardware components, extract a subset of hardware point clouds, and verify that the segmentation accuracy is 98.6%, successfully removing irrelevant point clouds such as wires and supports in the background.

[0047] Based on the design drawings of the Z-10 type suspension clamp, a standard 3D model was constructed using SolidWorks 2025 and exported as a PLY format point cloud file.

[0048] Coarse registration was performed using the ICP algorithm with 50 iterations and an initial registration error of 3.2 mm. Subsequently, an improved TrimmedICP algorithm was used to eliminate errors. , For outliers other than those in the model, after 100 iterations of optimization, the registration error was reduced to 0.3 mm.

[0049] The spatial distance difference between the registered point cloud and the standard point cloud was calculated. A distance threshold of 0.5 mm was set, and a subset of point clouds with a distance difference greater than the threshold was selected. These subsets are the point clouds of the metal fitting corrosion area, with a total of 86,200 point clouds.

[0050] A voxelization method was used to create a 3D model of the point cloud subset of the rusted area. The voxel size was set to 0.1mm × 0.1mm × 0.1mm, and the number of effective voxels was counted to be 8,620,000. The rust volume was calculated to be 8,620,000 × (0.1 × 0.1 × 0.1) = 86.2cm³. A Leica DISTOS 910 laser rangefinder was used to perform multi-point measurements on the rusted area. The actual rust volume was calculated to be 85.9cm³ through geometric calculation. The measurement error was 0.3cm³, and the corresponding spatial error was ≤0.2mm, which meets the requirement of error ≤2mm.

[0051] The system presets a threshold for the volume of corrosion on fittings: ≥100cm³ is classified as a Level 1 defect, 50-100cm³ as a Level 2 defect, and <50cm³ as a Level 3 defect. This corrosion defect is determined to be a Level 2 defect. The system automatically generates a structured inspection report, which includes information such as the defect location, type, level, 3D parameters, on-site RGB image, and point cloud model of the corrosion area, and synchronizes it to the power transmission line operation and maintenance cloud platform.

[0052] 2. Three-dimensional quantification of insulator damage defects: The hardware and software environment are the same as in Example 1. The experimental object is the XP-70 type insulator of the 500kV transmission line section, which has a damaged shed. The data collection environment is outdoor cloudy weather to avoid reflection interference caused by strong direct sunlight.

[0053] The drone flew to a distance of 6 meters to the side of the insulator and hovered, simultaneously collecting 25 sets of RGB images and dense point cloud data to ensure complete coverage of the damaged area of ​​the insulator skirt.

[0054] After semantic segmentation using the improved PointNet++ model, a subset of the insulator point cloud was extracted with a segmentation accuracy of 98.3%. The subset of the point cloud in the damaged area was then selected through differential analysis.

[0055] The point cloud subset of the damaged area was projected onto the XOY plane, and the projection boundary was fitted using the polygon approximation method with a fitting error ≤0.3mm. The area of ​​the irregular region within the boundary was calculated to be 3.6cm² using the shoelace formula. The actual measured damaged area by the laser rangefinder was 3.5cm², with a measurement error of 0.1cm², corresponding to a spatial error ≤0.15mm.

[0056] The preset threshold for the damaged area of ​​an insulator is as follows: ≥5cm² is a level 1 defect, 2-5cm² is a level 2 defect, and <2cm² is a level 3 defect. If the defect is determined to be a level 2 defect, an inspection report containing the three-dimensional coordinates and projected area map of the damaged area is generated and synchronized to the cloud.

[0057] 3. Three-dimensional quantification of conductor icing defects: The hardware and software environment are the same as in Example 1. The experimental object is the 500kV transmission line conductor. Artificial ice blocks are used to simulate natural icing and icing defects. The outdoor low temperature environment is -2℃ to ensure stable icing.

[0058] The drone flew to a position 10m directly above the conductor and hovered, simultaneously collecting 30 sets of dual-modal data, covering a 5m length of the icy section of the conductor.

[0059] Within the corresponding region of the point cloud subset in the icy area and the 3D model of the standard traverse, search for the spatial distance between 1000 closest points, discarding those with a distance greater than 3 times the standard deviation. After identifying 12 outliers, the average value was taken as 8.6 mm; the actual ice thickness measured by the laser rangefinder was 8.5 mm, with a measurement error of 0.1 mm.

[0060] The preset ice thickness threshold is ≥10mm for Level 1 defects, 5-10mm for Level 2 defects, and <5mm for Level 3 defects. If the defect is determined to be Level 2, an inspection report containing a heat map of ice thickness distribution and a point cloud comparison model is generated and synchronized to the cloud.

[0061] The implementation process of this invention allows for flexible parameter adjustment based on actual application scenarios: for transmission line components of different voltage levels, the drone's flight distance and sampling frequency can be adjusted; for complex environments, the number of calibration image groups can be increased to improve alignment accuracy; for different defect types, parameter thresholds can be customized to adapt to specific operation and maintenance needs. In this embodiment, the quantization error of all defects is ≤0.3mm, far lower than the preset error threshold of ≤2mm, verifying the high precision and high reliability of the method of this invention, which can meet the practical application requirements of digital operation and maintenance of transmission lines.

Claims

1. A three-dimensional quantization method for transmission line defects based on RGB-D point cloud and PointNet++, characterized in that, include: S1, Dual-modal data acquisition and pixel-level spatiotemporal alignment: A dual-modal data acquisition framework for RGB textures and dense point clouds is constructed. RGB images and dense point cloud data of power transmission line components are simultaneously acquired by a drone equipped with an RGB camera and a depth camera. Based on an online calibration algorithm for heterogeneous sensors, the intrinsic parameter matrix, distortion coefficient, and extrinsic parameter matrix between the RGB camera and the depth camera are solved to achieve pixel-level spatiotemporal alignment of the dual-modal data, resulting in spatiotemporally aligned dual-modal fused data. S2, semantic segmentation based on improved PointNet++: An improved PointNet++ semantic segmentation model is constructed. The improvements include introducing RGB texture feature weight factors in the grouping layer to fuse point cloud geometric features with RGB texture features. The improved PointNet++ semantic segmentation model is input into the dual-modal fusion data, and the semantic segmentation results of key components of the transmission line are output, and the point cloud subset of the target component is extracted. S3, Defect region extraction based on standard model difference analysis: Construct a standard 3D point cloud model of key components of the transmission line; perform high-precision registration between the point cloud subset of the target component and the standard 3D point cloud model; perform difference operation on the registered measured point cloud model and the standard 3D point cloud model, and extract the point cloud with a distance difference greater than the preset distance threshold as the point cloud subset of the defect area; S4. Defect 3D Parameter Calculation: Based on a subset of the point cloud of the defect region, the physical parameters of the defect are automatically calculated. The physical parameters include at least one of the following: ice thickness, metal fitting corrosion volume, or insulator damage area.

2. The method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++ according to claim 1, characterized in that, The heterogeneous sensor online calibration algorithm in step S1 specifically includes: A checkerboard calibration board was selected, and multiple sets of calibration board images in different poses were captured simultaneously by an RGB camera and a depth camera mounted on a drone to construct a calibration dataset. The intrinsic parameter matrix and distortion coefficients of the RGB camera are solved based on Zhang Zhengyou's calibration method, and its intrinsic parameters are obtained through the calibration tool built into the depth camera. An optimization function is constructed with the goal of minimizing pixel coordinate error. The extrinsic parameter matrix between the RGB camera and the depth camera is iteratively optimized so that the corresponding pixel position deviation between the depth point cloud and the RGB image is ≤1 pixel.

3. The method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++ according to claim 1, characterized in that, The training process of the improved PointNet++ semantic segmentation model in step S2 includes: Collect sample data of transmission line components covering different defect types and environmental conditions, and label the samples with component categories and defect areas; In the sampling layer of the model, a strategy combining farthest point sampling and random sampling is adopted, and in the grouping layer, a texture feature weight factor that is adaptively adjusted based on the variance of pixel gray values ​​in RGB images is introduced. The total loss function is a weighted sum of the cross-entropy loss function and the Dice loss function. The model is trained until the semantic segmentation accuracy meets the preset threshold.

4. The method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++ according to claim 1, characterized in that, The high-precision registration of point clouds in step S3 includes: First, the ICP algorithm is used to perform coarse registration between the target component point cloud subset and the standard 3D point cloud model. Based on the improved Trimmed ICP algorithm, outliers containing background clutter and measurement noise are removed, and the registration error is optimized through iterative optimization to meet the preset accuracy requirements.

5. The method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++ according to claim 1, characterized in that, The method for calculating the icing thickness in step S4 is as follows: In the defective point cloud subset of the icing area and the corresponding area of ​​the standard model, search the spatial distance of each pair of nearest points, remove outliers, and take the average value as the icing thickness.

6. The method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++ according to claim 1, characterized in that, The method for calculating the corrosion volume of hardware in step S4 is as follows: A voxelization method was used to perform 3D modeling of the point cloud subset of the defect area in the rust region, and the model volume was calculated as the rust volume by statistically counting the number of effective voxels.

7. The method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++ according to claim 6, characterized in that, The method for calculating the damaged area of ​​the insulator in step S4 is as follows: Project a subset of the point cloud of the defective region into the XOY plane, fit the projection boundary using the polygon approximation method, and calculate the area of ​​the irregular region within the boundary as the damaged area.

8. The method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++ according to claim 1, characterized in that, Step S4 is followed by: Based on the preset defect level threshold, the calculated physical parameters are rated, and a structured inspection report containing defect location, type, level, and three-dimensional parameters is automatically generated.

9. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the three-dimensional quantization method for transmission line defects based on RGB-D point cloud and PointNet++ as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute any one of claims 1 to 8, a method for three-dimensional quantization of transmission line defects based on RGB-D point cloud and PointNet++.

Images