Full-automatic electric power safety line inspection method and system based on laser point cloud

Abstract

This invention discloses a fully automated power line safety inspection method and system based on laser point clouds, relating to the field of power inspection. The method involves acquiring three-dimensional point cloud data of the power transmission and distribution system; processing the point cloud data to obtain preprocessed point cloud data; inputting the preprocessed point cloud data into an improved RandLA-Net model for point-by-point classification, outputting semantic segmentation results for conductors, insulators, power poles, and vegetation; constructing a complete scene model of the power transmission and distribution system based on the semantically segmented point cloud data; and completing the power line safety inspection based on this complete scene model. The technical solution provided by this invention enables the timely detection of potential safety hazards and the implementation of corresponding repair and prevention measures.

Description

Technical Field

[0001] This invention relates to the field of power line inspection, specifically a fully automated power line safety inspection method and system based on laser point clouds. Background Technology

[0002] Currently, optical remote sensing drones used for power line inspections (e.g., visible light and infrared light inspections) have limitations in analyzing acquired two-dimensional image data, such as a lack of depth information, limited field of view, and inability to directly create 3D models. However, 3D point cloud data can perfectly solve these limitations (accurately measuring parameters such as distance, height, and volume; acquiring omnidirectional 3D information in a single scan without multiple shots). Especially in power systems, such as long-distance inspections of overhead transmission lines, laser point cloud data can acquire global information, directly displayed on monitors in front of power workers, and provides high-precision, high-density 3D spatial information. This makes inspection work much more convenient and improves inspection efficiency. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a fully automated power line safety inspection method based on laser point clouds, comprising the following steps:

[0004] Step 1: Obtain 3D point cloud data of the power transmission and distribution system;

[0005] Step 2: Remove outliers from the point cloud data using statistical filtering, perform voxel downsampling at a preset resolution, segment the conductors and power poles using RANSAC fitting, remove environmental noise based on density clustering, and optimize the surface details of the equipment using morphological filtering to obtain the preprocessed point cloud data.

[0006] Step 3: Input the preprocessed point cloud data into the improved RandLA-Net model for point-by-point classification, and output the semantic segmentation results of conductors, insulators, power poles and towers and vegetation;

[0007] Step four involves performing planar segmentation and voxelization on the point cloud data of power poles and towers after semantic segmentation to generate a Level of Dimension (LOD) model; calculating the sag of the conductor point cloud data using RANSAC cylindrical fitting and B-spline shaping; performing surface defect detection on the insulator point cloud after loose wave reconstruction; repairing the broken parts of the broken conductor point cloud using RANSAC linear fitting; removing discrete points with a density below a threshold from the vegetation point cloud; smoothing the boundaries of residual noise on the power tower surface using an corrosion-expansion operation; and constructing a complete scene model of the power transmission and distribution system.

[0008] Step 5: Based on the complete scenario model of the power transmission and distribution system, calculate the minimum distance from the conductor to the vegetation, mark the tree obstacle risk where the distance exceeds the safety threshold; calculate the conductor sag and compare it with the design threshold to judge the risk and issue an early warning; construct the tower topology map to verify the connection integrity and complete the power line safety inspection.

[0009] Furthermore, the improved RandLA-Net model includes: optimizing the shallow sampling rate, employing multi-scale feature fusion for small targets, and improving the attention mechanism;

[0010] The shallow layer sampling rate is optimized: RandLA-Net's default layer-by-layer sampling rate Define the sampling attenuation amplitude as Δ, for the th l layer sampling rate Adjusted to:

[0011]

[0012] Multi-scale feature fusion is used for small targets: the original point features are combined. fraw Geometric features including curvature, normal vector, and density fgeo By concatenating the features, we obtain the enhanced feature vector. fenhanced :

[0013] fenhanced = Concat ( fraw , fgeo )

[0014] Attention mechanism improvement: Assign independent learnable query transformation matrices to conductors, towers, and vegetation categories respectively. Bond change matrix Through the transformation matrix Bond change matrix Mapping the point feature vectors to a class-specific subspace, the Softmax normalization yields the following probability distribution:

[0015]

[0016] in For the first Feature vectors of points, For the first Feature vectors of points, For feature dimensions.

[0017] Furthermore, the repair of the broken conductor point cloud using RANSAC linear fitting includes:

[0018] The equations for the straight lines of the conductors and towers are expressed in polar coordinates:

[0019]

[0020] In each iteration, two points are randomly selected to calculate the line parameters. , definition point The distance to the line is the error: like , If the preset threshold is used, then this point is an interior point;

[0021] Based on the interior point ratio w and confidence probability p according to Determine the number of iterations k For the multiple straight lines obtained by fitting, the endpoint spacing and direction consistency of adjacent line segments are calculated, and two line segments that meet the requirements of short endpoint spacing and small angle are merged into a straight line.

[0022] Furthermore, the removal of discrete points with a density below a threshold from the vegetation point cloud includes:

[0023] Suppose any point in the semantic point cloud :

[0024]

[0025] in For indicator functions, r Set the neighborhood radius; set the density threshold. ρth ,when ρ ( pi )< ρth Remove the point at that time.

[0026] Furthermore, the calculation of conductor sag is performed using the following formula:

[0027]

[0028] in L The distance between the highest points at both ends of the conductor. H For the horizontal tension of the conductor, w This refers to the weight per unit length of the conductor.

[0029] The fully automated power line safety inspection system based on laser point clouds, applying the aforementioned fully automated power line safety inspection method based on laser point clouds, includes: a data processing module, a data acquisition module, a preprocessing module, a semantic segmentation module, a scene restoration module, and a spatial analysis module; the data acquisition module, preprocessing module, semantic segmentation module, scene restoration module, and spatial analysis module are respectively connected to the data processing module.

[0030] The beneficial effects of this invention are: using drones to replace line patrol personnel for power transmission and distribution system inspections, acquiring 3D point cloud data, firstly preprocessing the point cloud data, then using deep learning methods for semantic segmentation, secondly reconstructing the scene, and finally analyzing the spatial relationships of various elements in the power scene to monitor the status of power lines and detect anomalies. This provides a basis for power planning and operation and maintenance decisions, timely identifies potential safety hazards, and allows for corresponding repair and prevention measures. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating a fully automated power line safety inspection method based on laser point clouds.

[0032] Figure 2 This is a schematic diagram illustrating the implementation process of a fully automated power line safety inspection method based on laser point clouds. Detailed Implementation

[0033] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0034] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0035] like Figure 1 As shown, the fully automated power line safety inspection method based on laser point clouds includes the following steps:

[0036] Step 1: Obtain 3D point cloud data of the power transmission and distribution system;

[0037] Step 2: Remove outliers from the point cloud data using statistical filtering, perform voxel downsampling at a preset resolution, segment the conductors and power poles using RANSAC fitting, remove environmental noise based on density clustering, and optimize the surface details of the equipment using morphological filtering to obtain the preprocessed point cloud data.

[0038] Step 3: Input the preprocessed point cloud data into the improved RandLA-Net model for point-by-point classification, and output the semantic segmentation results of conductors, insulators, power poles and towers and vegetation;

[0039] Step four involves performing planar segmentation and voxelization on the point cloud data of power poles and towers after semantic segmentation to generate a Level of Dimension (LOD) model; calculating the sag of the conductor point cloud data using RANSAC cylindrical fitting and B-spline shaping; performing surface defect detection on the insulator point cloud after loose wave reconstruction; repairing the broken parts of the broken conductor point cloud using RANSAC linear fitting; removing discrete points with a density below a threshold from the vegetation point cloud; smoothing the boundaries of residual noise on the power tower surface using an corrosion-expansion operation; and constructing a complete scene model of the power transmission and distribution system.

[0040] Step 5: Based on the complete scenario model of the power transmission and distribution system, calculate the minimum distance from the conductor to the vegetation, mark the tree obstacle risk where the distance exceeds the safety threshold; calculate the conductor sag and compare it with the design threshold to judge the risk and issue an early warning; construct the tower topology map to verify the connection integrity and complete the power line safety inspection.

[0041] The improved RandLA-Net model includes: optimizing the shallow sampling rate, using multi-scale feature fusion for small targets, and improving the attention mechanism;

[0042] The shallow layer sampling rate is optimized: RandLA-Net's default layer-by-layer sampling rate Define the sampling attenuation amplitude as Δ, for the th l layer sampling rate Adjusted to:

[0043]

[0044] Multi-scale feature fusion is used for small targets: the original point features are combined. fraw Geometric features including curvature, normal vector, and density fgeo By concatenating the features, we obtain the enhanced feature vector. fenhanced :

[0045] fenhanced = Concat ( fraw , fgeo )

[0046] Attention mechanism improvement: Assign independent learnable query transformation matrices to conductors, towers, and vegetation categories respectively. Bond change matrix Through the transformation matrix Bond change matrix Mapping the point feature vectors to a class-specific subspace, the Softmax normalization yields the following probability distribution:

[0047]

[0048] in For the first Feature vectors of points, For the first Feature vectors of points, For feature dimensions.

[0049] The method of repairing the broken conductor point cloud using RANSAC linear fitting includes:

[0050] The equations for the straight lines of the conductors and towers are expressed in polar coordinates:

[0051]

[0052] In each iteration, two points are randomly selected to calculate the line parameters. , definition point The distance to the line is the error: like , If the preset threshold is used, then this point is an interior point;

[0053] Based on the interior point ratio w and confidence probability p according to Determine the number of iterations k For the multiple straight lines obtained by fitting, the endpoint spacing and direction consistency of adjacent line segments are calculated, and two line segments that meet the requirements of short endpoint spacing and small angle are merged into a straight line.

[0054] The removal of discrete points with a density below a threshold from the vegetation point cloud includes:

[0055] Suppose any point in the semantic point cloud :

[0056]

[0057] in For indicator functions, r Set the neighborhood radius; set the density threshold. ρth ,when ρ ( pi )< ρth Remove the point at that time.

[0058] The calculation of conductor sag uses the following formula:

[0059]

[0060] in L The distance between the highest points at both ends of the conductor. H For the horizontal tension of the conductor, w This refers to the weight per unit length of the conductor.

[0061] The fully automated power line safety inspection system based on laser point clouds, applying the aforementioned fully automated power line safety inspection method based on laser point clouds, includes: a data processing module, a data acquisition module, a preprocessing module, a semantic segmentation module, a scene restoration module, and a spatial analysis module; the data acquisition module, preprocessing module, semantic segmentation module, scene restoration module, and spatial analysis module are respectively connected to the data processing module.

[0062] Specifically, this invention discloses a fully automated transmission line inspection method based on laser point cloud semantic space construction and analysis. It is primarily used for high-precision and comprehensive monitoring and analysis of power equipment, improving the safety and reliability of power systems. Power workers can directly view equipment status and information on a monitor, overcoming the limitations of traditional manual line inspections, such as requiring a large workforce and being difficult to conduct inspections in complex terrain. This invention relates to point cloud registration technology, point cloud semantic segmentation technology, and transmission line monitoring technology.

[0063] The system is applied in scenarios including power transmission line inspection (generating a 3D model of the power corridor using laser point clouds to detect conductor sag, insulator damage, conductor tension, etc.) and substation inspection (generating a 3D model of the substation using laser point clouds to analyze equipment spacing, structural deformation, etc.). First, a drone carrying a lidar acquires 3D point cloud data of the power transmission line in a single flight. Then, statistical filtering is used to remove obvious outliers, followed by voxel downsampling at a resolution of 0.1 meters. Random Access Scale (RANSAC) is then used to segment the conductors and towers, and density-based clustering is applied to the remaining point cloud data to remove environmental noise. Finally, morphological filtering is used on the equipment surfaces to optimize details. Next, the processed data is semantically segmented using the improved RandLA-Net model, and the type is predicted point by point. Conductors, insulators, poles, and vegetation are visualized using different colors. Scene reconstruction is then performed. The segmented pole point cloud data is further segmented into planes and voxels to generate a Level of Detail (LOD) model. RANSAC cylindrical fitting and B-spline shaping are used for conductor point cloud data to calculate sag. Loose wave reconstruction and surface defect detection are performed on insulator point clouds. Through semantic segmentation and scene reconstruction, the original point cloud data is transformed into quantifiable, operable, and decision-making data, facilitating subsequent analysis. Finally, the spatial relationships of the laser point cloud scene are constructed and analyzed. The minimum distance from conductors to vegetation is calculated point by point, and tree obstacle risks exceeding the safety threshold are marked. Conductor sag is calculated and compared with the design threshold to assess risk and provide early warnings. A pole topology map is constructed to verify connection integrity.

[0064] like Figure 2As shown, this invention aims to study fully automated laser point cloud line inspection. To obtain a system capable of precise application, accurate semantic segmentation, efficient geometric calculation algorithms, and a rule base tailored to business requirements are needed. By optimizing random sampling, improving local feature aggregation, and refining the attention mechanism in RandLA-Net, the model's ability to capture small targets such as slender conductors is enhanced through multi-scale feature fusion. Edge point weights are added to the loss function to improve the segmentation accuracy at conductor-insulator connections, resulting in accurate semantic segmentation results. For regular shapes such as conductors and towers, preset model parameters (conductor radius, etc.) are used to reduce the number of iterations. The FLANN library is used to accelerate KD-tree queries and improve distance calculation efficiency. Rules such as tree obstacle distance (≥5 meters) and conductor sag error (±10 cm) are set according to the power industry standard "Operation Regulations for Overhead Transmission Lines," defining tower-conductor connection rules. Differentiated rules are configured based on safety standards for different voltage levels. Through the synergistic advancement of data precision, algorithm efficiency, and business-oriented rules, a complete closed loop for laser point cloud processing can be achieved, enabling precise application of the entire system.

[0065] Based on domestic and international research progress, factors such as scene texture, number of targets, and structural complexity have a certain impact on the acquisition of 3D point cloud data for the system, and the accuracy of the system needs to be improved. This invention focuses on solving the problems of semantic segmentation, scene restoration, and spatial construction and analysis.

[0066] Semantic segmentation, due to the variability and diversity of power transmission scenarios, presents a significant challenge in providing the necessary feature information while simultaneously understanding the content of the scene, thereby ensuring the stability and accuracy of the entire system. Semantic segmentation methods can obtain the classification result of a specific point cloud point in the scene, and simultaneously segment the target point cloud point into its corresponding category, outputting the semantic segmentation result.

[0067] Adjusting the model structure and parameters of the preprocessed point cloud data:

[0068] Because power lines are sparse in power scenarios, RandLA-Net uses a default layer-by-layer sampling rate. =[4,4,4,4,2] would prematurely lose key points of the conductor, so we choose to reduce the shallow sampling rate, as expressed in mathematical expression (1). This represents the sampling attenuation amplitude.

[0069] (1)

[0070] If using =2 as an example, then =[2,2,4,4,2]( =[4-2,4-2,4,4,2]). In this way, the number of samples in the first and second layers increases from 25% to 50%, while the sampling rate of the third, fourth, and fifth layers remains unchanged.

[0071] For small objects such as conductors and insulators, 10 points are forcibly retained to achieve sampling optimization for power scenarios. At least k=10 key points are forcibly retained. Initially, 50% of the points are retained to reduce shallow sampling, and then for each key point region... If there are fewer than 10 sampling points, supplement them to 10 points after sorting by density features. Let the point cloud set be P, and the key object region set be:

[0072] For each region Sampling must meet the following requirements:

[0073] (2)

[0074] Point Whether to retain (1 for retain, 0 for discard).

[0075] By increasing the search radius of the slender wire, continuity over longer distances can be captured. The current search radius is... The current domain search radius is the same as the original search radius. Add to The length of this section of the conductor is times that of the previous section. .

[0076] (3)

[0077] When aggregating local features, use explicit addition of geometric features to help distinguish between guide lines and vegetation. This involves adding original point features. Geometric features including curvature, normal vector, and density The features are concatenated to obtain the enhanced feature vector. .

[0078] (4)

[0079] Independent attention weights are used for different categories such as conductors and power towers. For each category (conductors, towers, vegetation, etc.),... Assign independent learnable query transformation matrices Bond change matrix .pass and Will (No. (feature vectors of points) (No. The feature vectors of each point are mapped to a class-specific subspace. This is measured by the dot product. and In category The similarity is calculated based on the association strength under the following conditions, and finally divided by . To prevent gradient explosion, Softmax normalizes the data to a probability distribution:

[0080] (5)

[0081] In LocalSpatialEncoding (LSE), the axial direction encoding of the conductor is added to enhance the modeling of the conductor's orientation. Such improvements in local features and attention mechanisms can be better applied in power scenarios.

[0082] For scene restoration, we first separate the semantic point cloud data with labels by category. Then, we perform RANSAC straight line fitting on the conductors and power towers to repair the broken sections.

[0083] RANSAC straight line fitting process: The straight line equations for conductors and power towers are used in polar coordinates to avoid the problem of infinitely large slopes of vertical lines.

[0084] (6)

[0085] In each iteration, two points are randomly selected to calculate the line parameters. Definition point The distance to the line is the error: like If a point falls below a preset threshold, it is considered an interior point. Number of iterations. From the interior point ratio and confidence probability Decide: ,in This represents the estimated percentage of in-place points.

[0086] The break repair process involves running RANSAC multiple times on the semantic point cloud, removing interior points of fitted line segments until no significant line segments remain, resulting in multiple straight lines. For adjacent line segments, endpoint spacing and orientation consistency are calculated. Two line segments that meet the criteria of short endpoint spacing and small included angle are merged into a single straight line.

[0087] For the segmented vegetation noise, remove discrete points with a density below a threshold. Let any point in the semantic point cloud... Its local density radius The number of points in the neighborhood, of which This is an indicator function; it takes the value 1 if the distance condition is met, and 0 otherwise.

[0088] (7)

[0089] Set density threshold Directly remove low-density points, when Removed at that time.

[0090] Finally, the residual noise on the power tower surface is smoothed using an erosion-expansion operation to obtain a complete scene model that conforms to the power topology scenario.

[0091] Semantic space construction and analysis: After obtaining a complete model of the power topology scenario, the point cloud is registered and aligned with the designed CAD model using ICP. When calculating the safe distance between conductors and vegetation, the distance between the conductors and vegetation is calculated point-by-point. Using a KD-tree accelerates the query process and visualizes and marks risk points. Let a certain point in the vegetation be... and Calculate the distance according to formula (5).

[0092] (8)

[0093] When measuring a sag, the span is determined by measuring the distance between the highest points of the two ends of the conductor. The horizontal tension of the measuring conductor is , The weight per unit length of the conductor is given. The sag is measured according to formula (6).

[0094] (9)

[0095] The measured sag is compared with the designed sag, and an early warning is issued when the difference exceeds 10%.

Claims

1. An automatic power line safety inspection method based on laser point cloud semantic space construction and analysis, characterized in that, Includes the following steps: Step 1: Obtain 3D point cloud data of the power transmission and distribution system; Step 2: Remove outliers from the point cloud data using statistical filtering, perform voxel downsampling at a preset resolution, segment the conductors and power poles using RANSAC fitting, remove environmental noise based on density clustering, and optimize the surface details of the equipment using morphological filtering to obtain the preprocessed point cloud data. Step 3: Input the preprocessed point cloud data into the improved RandLA-Net model for point-by-point classification, and output the semantic segmentation results of conductors, insulators, power poles and towers and vegetation; Step four involves performing planar segmentation and voxelization on the point cloud data of power poles and towers after semantic segmentation to generate a Level of Dimension (LOD) model; calculating the sag of the conductor point cloud data using RANSAC cylindrical fitting and B-spline shaping; performing surface defect detection on the insulator point cloud after loose wave reconstruction; repairing the broken parts of the broken conductor point cloud using RANSAC linear fitting; removing discrete points with a density below a threshold from the vegetation point cloud; smoothing the boundaries of residual noise on the power tower surface using an corrosion-expansion operation; and constructing a complete scene model of the power transmission and distribution system. Step 5: Based on the complete scenario model of the power transmission and distribution system, calculate the minimum distance from the power line to the vegetation and mark the risk of tree obstacles where the distance exceeds the safety threshold; Calculate conductor sag and compare it with the design threshold to assess risk and issue early warnings; Construct a pole and tower topology diagram to verify connection integrity and complete power line safety inspection; The improved RandLA-Net model includes: optimizing the shallow sampling rate, using multi-scale feature fusion for small targets, and improving the attention mechanism; The shallow layer sampling rate is optimized: RandLA-Net's default layer-by-layer sampling rate Define the sampling attenuation amplitude as Δ, for the th l layer sampling rate Adjusted to: Multi-scale feature fusion is used for small targets: the original point features are combined. fraw Geometric features including curvature, normal vector, and density fgeo By concatenating the features, we obtain the enhanced feature vector. fenhanced : fenhanced = Concat ( fraw , fgeo ) Attention mechanism improvement: Assign independent learnable query transformation matrices to conductors, towers, and vegetation categories respectively. Bond change matrix Through the transformation matrix Bond change matrix Mapping the point feature vectors to a class-specific subspace, the Softmax normalization yields the following probability distribution: in For the first Feature vectors of points, For the first Feature vectors of points, For feature dimensions.

2. The automatic power line safety inspection method based on laser point cloud semantic space construction and analysis according to claim 1, characterized in that, The method of repairing the broken conductor point cloud using RANSAC linear fitting includes: The equations for the straight lines of the conductors and towers are expressed in polar coordinates: In each iteration, two points are randomly selected to calculate the line parameters. , definition point The distance to the line is the error: like , If the preset threshold is used, then this point is an interior point; Based on the interior point ratio w and confidence probability p according to Determine the number of iterations k For the multiple straight lines obtained by fitting, the endpoint spacing and direction consistency of adjacent line segments are calculated, and two line segments that meet the requirements of short endpoint spacing and small angle are merged into a straight line.

3. The automatic power line safety inspection method based on laser point cloud semantic space construction and analysis according to claim 1, characterized in that, The removal of discrete points with a density below a threshold from the vegetation point cloud includes: Suppose any point in the semantic point cloud : in For indicator functions, r Set the neighborhood radius; set the density threshold. ρth ,when ρi < ρth Remove the point at that time.

4. The automatic power line safety inspection method based on laser point cloud semantic space construction and analysis according to claim 1, characterized in that, The calculation of conductor sag uses the following formula: in L The distance between the highest points at both ends of the conductor. H For the horizontal tension of the conductor, w This refers to the weight per unit length of the conductor.

5. An automatic power line safety inspection system based on laser point cloud semantic space construction and analysis, characterized in that, The automatic power line safety inspection method based on laser point cloud semantic space construction and analysis according to any one of claims 1-4 includes: a data processing module, a data acquisition module, a preprocessing module, a semantic segmentation module, a scene restoration module, and a spatial analysis module; wherein the data acquisition module, preprocessing module, semantic segmentation module, scene restoration module, and spatial analysis module are respectively connected to the data processing module.

Images