Overhead transmission line point cloud classification method and system based on rule improved RandLA-Net, and storage medium

Abstract

The invention discloses an overhead transmission line point cloud classification method and system based on rule improved RandLA-Net, and a storage medium, and belongs to the field of three-dimensional laser point cloud processing and power system automation. The classification method comprises the following steps: collecting original point cloud data of the overhead transmission line, and performing preprocessing after data conversion and data classification to obtain a standardized point cloud data set; constructing and training an improved RandLA-Net point cloud classification model, performing feature enhancement through a channel attention mechanism, and outputting a preliminary classification result; performing optimization processing on the preliminary classification result based on a prior rule to obtain a final classification result; and outputting the final classification result and displaying the final classification result through a visual tool for supporting power transmission line inspection. Through improvement of a channel attention mechanism, data preprocessing optimization, combination of four types of prior rules and the like, high-precision classification of power transmission equipment is realized, and reliable data support is provided for power inspection.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of three-dimensional laser point cloud processing and power system automation, and particularly relates to an overhead transmission line point cloud classification method, system and storage medium based on a rule-improved RandLA-Net. BACKGROUND

[0002] As the core infrastructure of the power system, the safe operation of the overhead transmission line is directly related to the stability of the power grid and the reliability of power supply. The traditional inspection method mainly relies on manual climbing or helicopter-mounted visible light equipment operation, which has significant defects such as low efficiency, high safety risk and high cost, and has limited ability to identify subtle defects such as insulator damage and conductor breakage, which is difficult to meet the needs of modern intelligent power grid inspection.

[0003] With the development of laser radar technology, unmanned aerial vehicles can obtain high-precision three-dimensional point cloud data of the transmission line, which provides the possibility for automated inspection. However, point cloud data has characteristics such as disorder, uneven density and class imbalance, and in addition to the complex structure of equipment and diverse spatial relationships in the transmission scene, traditional algorithms face difficulties in feature extraction and insufficient modeling of topological relationships when processing. Although existing deep learning methods have made breakthroughs in point cloud processing efficiency, they still have obvious shortcomings in key device identification accuracy and adaptability to complex environments, which cannot meet the actual needs of power inspection for high precision and high reliability. SUMMARY

[0004] In view of the problems of existing overhead transmission line point cloud classification methods, such as insufficient feature extraction capability, difficulty in processing topological relationships of transmission equipment in complex scenes, and low classification accuracy, the application provides an overhead transmission line point cloud classification method based on a rule-improved RandLA-Net, which realizes high-precision classification of transmission equipment by combining channel attention mechanisms, data preprocessing optimization and four types of prior rules, and provides reliable data support for power inspection.

[0005] According to a first aspect of the application, an overhead transmission line point cloud classification method based on a rule-improved RandLA-Net is provided, comprising:

[0006] S1: Collecting original point cloud data of the overhead transmission line, preprocessing after data conversion and data classification to obtain a standardized point cloud data set;

[0007] S2: Constructing and training an improved RandLA-Net point cloud classification model, enhancing features through a channel attention mechanism, and outputting a preliminary classification result;

[0008] S3: Optimizing the preliminary classification result based on prior rules to obtain a final classification result;

[0009] S4: output the final classification result and display it through a visualization tool to support power line inspection.

[0010] Further, the S1 specifically comprises:

[0011] S11: data acquisition: using a UAV equipped with a laser radar to scan the overhead power transmission line, and collecting the original point cloud data of the overhead power transmission line to store in a las file format;

[0012] S12: data conversion: converting the las file into a txt format, with each line in each txt file corresponding to one point cloud data;

[0013] S13: data classification: system labeling each point cloud data and dividing it into multiple categories, with each category corresponding to a separate label;

[0014] S14: data preprocessing: hierarchical voxel downsampling and KD-Tree index construction for each point cloud data to obtain a standardized point cloud dataset.

[0015] Further, in the S12, each point cloud data includes the three-dimensional coordinates, reflectivity intensity, and color information of the points.

[0016] Further, in the S13, each point cloud data is system labeled and divided into seventeen categories: {jumper, street lamp or street sign, ground wire, building, road, optical cable, insulator_straight line, vehicle, low vegetation, high vegetation, conductor, tower, insulator_strain, insulator_Vstring, dancing spacer, distribution tower, distribution conductor}, with each category corresponding to a label from 0 to 16.

[0017] Further, the S14 specifically comprises:

[0018] S141: hierarchical voxel downsampling: using a coarse and fine-grained combined voxelization strategy to downsample each point cloud data, setting the fine-grained voxel size V fine = 0.01m to retain key detail features, and the coarse-grained voxel size V coarse = 0.06m to quickly reduce the data volume;

[0019] S142: KD-Tree index construction: based on hierarchical voxelization, constructing a KD-Tree spatial index based on coarse-grained voxel point cloud, achieving efficient K-nearest neighbor query through recursive spatial division, realizing accurate mapping of fine-grained point cloud to coarse-grained point cloud, and thus obtaining a standardized point cloud dataset.

[0020] Further, the S141 is specifically implemented as follows:

[0021] Fine-grained voxelization: V fine=0.01m, dividing the three-dimensional space into a cubic grid with a side length of 0.01 meters, each voxel ( (The j-th fine-grained voxel) corresponds to a spatial region; for the set of points falling into this voxel... Its centroid coordinates The calculation is as follows: p j,k =(x j,k ,y j,k ,z j,k ) represents fine-grained voxels The three-dimensional coordinates of the k-th point within the matrix, where n j The number of points within a voxel;

[0022] Coarse-grained voxelization: V coarse =0.06m, the fine-grained voxelization results are downsampled again, the voxel size is increased to 0.06 meters, and the centroid is calculated as follows: Where m j This represents the number of fine voxels contained within a bold voxel.

[0023] Furthermore, the specific steps of S142 are as follows:

[0024] Dimension selection and spatial partitioning: For the point set P within the current node, calculate the variance of the coordinates in each dimension: d∈{x,y,z} where For point p i The coordinate values ​​in dimension d; The mean of the coordinates of dimension d; select the dimension d with the largest variance. max As the dividing axis, the median along this dimension Divide the point set into left and right child nodes:

[0025] Point cloud projection mapping and storage; achieving accurate mapping from fine-grained point clouds to coarse-grained point clouds using KD-Tree, and employing the nearest neighbor query algorithm to map point p in the fine-grained point cloud. i In a coarse-grained point cloud, find its nearest neighbor p. j The mathematical expression is: p k ∈ coarse-grained point cloud, where ||·||2 is the Euclidean distance;

[0026] Projection index and label association: Map index i of the fine-grained point cloud to index j of the coarse-grained point cloud to generate a projection matrix. (n is the number of fine-grained point cloud points); the original label y i∈{0,1,…,16} is mapped to a coarse-grained point cloud according to the projection relationship to ensure the consistency of the category labels of the point cloud after downsampling: The majority(·) option represents the majority category.

[0027] Furthermore, step 1 also includes: dividing the standardized point cloud dataset into a training set, a validation set, and a test set using a stratified sampling strategy at a ratio of 9:1:1.

[0028] Furthermore, S2 specifically includes:

[0029] S21: Organize the standardized point cloud dataset into a shape of N×(3+d) in A matrix, where N is the number of points, 3 represents the spatial coordinates, and d in It is the input feature dimension, the original features used as input to the model;

[0030] S22: For the original features of the model input, the improved RandLA-Net point cloud classification model encoder obtains enhanced features through random sampling, local feature aggregation, and channel attention module;

[0031] S23, for the enhanced features, the decoder of the improved RandLA-Net point cloud classification model enhances the detailed features of the power transmission equipment by upsampling and feature fusion, combined with the channel attention module;

[0032] S24. Based on the features obtained in S23, the classification head maps the feature dimension to the number of categories through a multilayer perceptron, uses the Softmax function to output the category probability of each point, and determines the predicted classification category of each point by taking the index of the category with the highest probability in the probability distribution of each point.

[0033] S25 utilizes the cross-entropy loss function and Adam optimizer to adaptively adjust the learning rate based on the gradient and update the parameters of the improved RandLA-Net point cloud classification model to optimize the model's classification performance.

[0034] S26. The improved RandLA-Net point cloud classification model, after training, is used to output preliminary classification results.

[0035] Furthermore, S22 specifically includes:

[0036] Random sampling: In each layer, the number of point clouds is gradually compressed from the initial N points to 1 / 4, 1 / 16, 1 / 64 and 1 / 256 of the original size;

[0037] Local feature aggregation: Spatial encoding and attention pooling operations are performed to construct multi-level feature representations. Residual connections are used to merge the original features and multi-level feature representations to output the first feature map.

[0038] Channel Attention Module: For the first feature map, the channel attention module performs global average pooling and global max pooling to compress the spatial dimension and obtain global channel information; it learns channel attention weights through two layers of multilayer perceptron; and it multiplies the channel attention weights with the first feature map channel by channel to output the enhanced features.

[0039] Furthermore, S23 specifically includes:

[0040] Upsampling: The point cloud size is gradually restored from N / 256 to N / 64, N / 16, N / 4 and N, and the feature dimension is gradually reduced from 512 to 8.

[0041] Feature fusion: The features of the current layer of the decoder are concatenated and fused with the enhanced features of the corresponding layer of the encoder;

[0042] Channel Attention Module: The concatenated and fused features are transposed and convolved to form a second feature map. Global channel information is extracted by local average pooling and global max pooling. After learning weights through a multilayer perceptron, the weights are multiplied with the second feature map channel by channel to enhance the detailed features of the power transmission equipment.

[0043] Furthermore, in step S25, the loss value between the predicted classification category and the true category label of each point is calculated based on the cross-entropy loss function, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the learning rate is adaptively adjusted and the model parameters are updated and adjusted using the Adam optimizer based on the gradient to optimize the model classification performance.

[0044] Furthermore, S3 specifically includes: based on the preliminary classification results output by the improved RandLA-Net point cloud classification model, enabling four serial rule post-processors to perform step-by-step optimization processing on the preliminary classification results of jumpers, insulator tension, insulator V-strings, and insulator straight lines according to their respective prior rules, and outputting the final classification results.

[0045] Furthermore, in S3, the prior rules include the jumper bidirectional growth rule, the insulator tension string linear constraint rule, the insulator V-string symmetric completion rule, and the insulator straight string vertical continuity rule.

[0046] Furthermore, the prior rule is specifically as follows:

[0047] The jumper bidirectional growth rule, through DBSCAN clustering and PCA principal direction analysis, expands bidirectionally by 0.2 meters along the principal direction to repair jumper breakage issues;

[0048] The linear constraint rules for tension insulator strings ensure the integrity of the linear structure of tension insulators through linearity evaluation and constraint extension;

[0049] The insulator V-string symmetry completion rule restores the V-shaped structure by calculating the symmetry center and mirroring the completion, expanding by 0.2 meters in the vertical direction;

[0050] The vertical continuity rule of the insulator straight string is improved by extending the Z-axis bidirectionally by 0.3 meters and adjusting the dynamic threshold to ensure the integrity of the vertical structure.

[0051] According to a second aspect of the present invention, a point cloud classification system for overhead transmission lines based on rule-improved RandLA-Net is provided. The system includes a processor and a memory for storing executable instructions. The processor is configured to execute the executable instructions to perform the point cloud classification method for overhead transmission lines based on rule-improved RandLA-Net as described in any of the preceding aspects.

[0052] According to a third aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the rule-based improved RandLA-Net overhead transmission line point cloud classification method as described in any of the preceding aspects.

[0053] The beneficial effects of this invention are:

[0054] This invention proposes a rule-based improved RandLA-Net point cloud segmentation method for overhead transmission lines. Combining deep learning with power industry knowledge, the model significantly improves the accuracy of identifying key equipment such as jumpers and insulators through the synergistic effect of channel attention mechanism and four types of prior rules. This effectively addresses the issues of missed detection of small targets and structural discontinuities, ensuring reliable classification accuracy in complex scenarios. Regarding data processing efficiency, hierarchical voxel downsampling and KD-Tree indexing techniques reduce data redundancy while preserving equipment geometric features. The inference speed meets the actual needs of UAV inspections, significantly improving work efficiency and reducing safety risks compared to traditional manual inspections. Furthermore, prior rules designed based on the installation characteristics of transmission equipment can effectively handle complex situations such as vegetation obstruction and equipment overlap, supplementing equipment structural details and outputting classification results that better reflect engineering realities, providing a precise data foundation for the operation and maintenance and condition assessment of power equipment. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0056] Figure 1 This is a flowchart of the overhead transmission line point cloud segmentation method based on rule-based improved RandLA-Net proposed by our laboratory;

[0057] Figure 2 This is a network structure diagram of the improved model based on rule-based RandLA-Net of this invention;

[0058] Figure 3 This is a 3D visualization of the point cloud classification results of the power transmission equipment according to the present invention.

[0059] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0060] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0061] The terms "first," "second," etc., used in this disclosure are for distinguishing similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein.

[0062] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0063] Multiple, including two or more.

[0064] And / or, it should be understood that, for the purposes of this disclosure, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. This reduces manpower input and facilitates business automation, featuring versatility, efficiency, and high precision.

[0065] The present invention provides a point cloud classification method, system, and storage medium for overhead transmission lines based on rule-based improved RandLA-Net, such as...Figure 1 As shown, the method includes:

[0066] S101: Point cloud data acquisition and preprocessing. Through UAV LiDAR scanning, data format conversion, system annotation, and hierarchical preprocessing, a point cloud dataset suitable for deep learning was constructed for overhead transmission lines. Specific steps are as follows:

[0067] 1) Data collection and classification of drone LiDAR: Using a DJI M300RTK drone equipped with a DJI lens, the drone was used to scan overhead power transmission lines with LiDAR. The raw point cloud data collected was stored in LAS file format.

[0068] 2) Data Conversion: The LAS files are converted to TXT format using a data parsing program. Each line in the TXT file corresponds to a point cloud data point. Point cloud data of the power transmission line is acquired using a drone equipped with a lidar system. Where p i =(x i ,y i ,z i ,r i ,g i ,b i ) represent the three-dimensional coordinates, reflection intensity, and color information of the point, respectively.

[0069] 3) Data Classification: The collected data is systematically labeled and divided into seventeen categories: {jump wire, street light or road sign, ground wire, building, road, optical cable, insulator_straight line, vehicle, low vegetation, high vegetation, conductor, pole, insulator_tension, insulator_V string, galloping spacer, distribution pole, distribution conductor}. Each category corresponds to a label from 0 to 16, realizing a precise mapping between data categories and labels, and providing a reliable data foundation for subsequent data processing and model training.

[0070] 4) Data Preprocessing: To reduce data redundancy and build an efficient index, the original point cloud data needs to be subjected to hierarchical voxel downsampling and KD-Tree index construction. The specific steps are as follows:

[0071] a) Layered voxel downsampling: A voxelization strategy combining coarse and fine granularity is used to downsample the original point cloud, setting the fine-grained voxel size V. fine =0.01m is used to preserve key detail features, and the coarse-grained voxel size V coarse =0.06m is used to quickly reduce the amount of data, and the specific implementation is as follows:

[0072] Fine-grained voxelization (V fine =0.01m). The three-dimensional space is divided into a cubic grid with a side length of 0.01 meters, and each voxel... This corresponds to a spatial region. For the set of points falling into this voxel... Its centroid coordinates The calculation is as follows: Where n j This represents the number of points within the voxel. This step preserves the geometric details of small targets such as insulators and jumpers through centroid approximation.

[0073] Coarse-grained voxelization (V coarse =0.06m). The fine-grained voxelization results were downsampled again, increasing the voxel size to 0.06 meters. The centroid was calculated as follows: Where m j This represents the number of fine voxels contained within a coarse voxel. This operation reduces the point cloud density, decreasing the original point cloud data volume by 70%-80%, preserving macroscopic structural features while significantly reducing subsequent computational complexity. Through a hierarchical voxel downsampling strategy, the geometric details of small targets such as conductors and insulators can be preserved while reducing the data volume.

[0074] b) KD-Tree Index Construction: Based on hierarchical voxelization, a KD-Tree spatial index is constructed based on coarse-grained voxel point clouds. Efficient K-nearest neighbor queries are achieved through recursive spatial partitioning. The KD-Tree enables accurate mapping from fine-grained point clouds to coarse-grained point clouds. The specific steps are as follows:

[0075] Dimension selection and spatial partitioning. For the point set P within the current node, calculate the variance of the coordinates in each dimension: d∈{x,y,z} where Let be the mean of the coordinates of dimension d. Select the dimension d with the largest variance. max As the dividing axis, the median along this dimension Divide the point set into left and right child nodes: This strategy ensures that the number of point clouds in the two child nodes is as balanced as possible after each partition, improving indexing efficiency. When a node satisfies the condition that the number of point clouds within the node is less than a threshold T... leaf =50 or the tree depth reaches the preset maximum value D max The recursion terminates when the value reaches 30. The final constructed KD-Tree supports K-nearest neighbor queries with a time complexity of O(logn).

[0076] Point cloud projection mapping and storage. A precise mapping from fine-grained point clouds to coarse-grained point clouds is achieved using a KD-Tree. The nearest neighbor query algorithm is employed to map points p in the fine-grained point cloud. i In a coarse-grained point cloud, find its nearest neighbor p. j The mathematical expression is: p kThe query is implemented using a level-order traversal of a KD-Tree, prioritizing subtrees closer to the query point to accelerate convergence. The query is performed within a coarse-grained point cloud, where ||·||² represents the Euclidean distance.

[0077] The projection index is associated with the label. The index i of the fine-grained point cloud is mapped to the index j of the coarse-grained point cloud, generating a projection matrix. (n is the number of fine-grained point cloud points). Simultaneously, the original annotation label y... i ∈{0,1,…,16} is mapped to a coarse-grained point cloud according to the projection relationship to ensure the consistency of the category labels of the point cloud after downsampling: The majority(·) option represents the majority category, which solves the problem of inconsistent point cloud categories within a voxel.

[0078] The above steps generate a KD-Tree index file (.pkl format), enabling efficient K-nearest neighbor queries on coarse-grained point clouds. The average query time for K=10 on 100,000 point clouds is ≤1.5ms. Simultaneously, a projection index file (.pkl format) is generated to ensure accurate correspondence between the labels of the downsampled coarse-grained point clouds and the original fine-grained point clouds, providing reliable category annotations for model training and establishing an effective association between the "details-structure" features of the point cloud data. This facilitates subsequent optimization of classification results by combining prior rules such as insulator linear constraints.

[0079] 5) Constructing the training dataset: The preprocessed point cloud data is divided into a training set, a validation set, and a test set in a 9:1:1 ratio using a stratified sampling strategy. The training set is used for iterative optimization of model parameters, the validation set assists in adjusting hyperparameters and preventing overfitting during training, and the test set independently evaluates the model's final generalization performance.

[0080] S102: Construction and Training of a Point Cloud Classification Model Based on Rule-Based Improved RandLA-Net. The model architecture of this invention is based on RandLA-Net, with improvements to its channel attention mechanism to enhance the model's ability to extract features from power transmission equipment. Rule-based improvements are also introduced during post-processing. (Refer to...) Figure 2 The improved model network structure diagram is shown below. The specific construction process of the model is as follows:

[0081] 1) Input Layer Data Organization: After preprocessing, the raw point cloud data is fed into the input layer. The input layer receives the preprocessed point cloud data, each point cloud data containing information such as 3D coordinates, reflection intensity, and color. The input data is organized into a shape of N×(3+d) in A matrix, where N is the number of points, 3 represents the spatial coordinates, and d in This refers to the input feature dimensions (including reflectance intensity and color information, etc.). The processed data serves as the model input, laying the foundation for subsequent feature extraction.

[0082] 2) Encoder Feature Extraction: The encoder consists of a four-layer structure, each layer containing random sampling (RS), local feature aggregation (LFA), and channel attention (SE) modules. Data is progressively sampled in the encoder, while higher-level features are extracted simultaneously. In the random sampling stage, a subset of points is uniformly selected from N points as input points for the next layer, making computational efficiency independent of the total number of input points. Specifically, in each layer, the number of point clouds is progressively compressed from the initial N points to 1 / 4, 1 / 16, 1 / 64, and 1 / 256 of the original size. Then, the local feature aggregation (LFA) stage begins, first performing local spatial encoding (LocSE) with a shape of N×(3+d). in Taking point cloud data as input, the K-nearest neighbor algorithm is used to construct a local region containing K neighboring points for each point. The relative positional offset and Euclidean distance between the point and its neighbors are calculated. This spatial relationship information is concatenated with the original features and then subjected to nonlinear transformation by a multilayer perceptron to capture the local geometric structure of the point cloud. Next, attention pooling is performed. The spatially encoded features are concatenated with the original features and then input into the multilayer perceptron. The attention weights are generated by the Softmax function and then performed as dot products with the features. After pooling and dimension adjustment, a shape of (1, d) is generated. out The feature vectors of the encoding layer are used to highlight key features and suppress redundant information. Then, by expanding the residual module, a multi-level feature representation is constructed through two consecutive spatial encoding and attention pooling operations. Residual connections are used to merge the original and enhanced features into the output, expanding the effective receptive field of each point and alleviating the gradient vanishing problem in deep network training. Finally, after each expansion of the residual module output in the encoding layer, the feature map is input into the channel attention module (Squeeze-and-Excitation, SE), at which point the feature map shape is [B, N]. i ,1,C i (B is the batch size, N) i (where Ci is the number of points in the current layer and Ci is the number of channels). The attention module first performs global average pooling (GAP) and global max pooling (GMP) on the feature map to compress the spatial dimensions to obtain [B,C]. i The global channel information is then processed through two layers of MLP (the first layer output dimension C). i / / 4, The activation function is ReLU; the output dimension of the second layer is C. i The channel attention weights are learned using the sigmoid activation function. Finally, the weights are multiplied channel by channel with the original feature map to output the enhanced features. These features are then downsampled by random sampling and passed to the next layer, while also being stored in the feature list for the decoder to obtain the enhanced features.

[0083] 3) Decoder Feature Recovery Process: The decoder also consists of a four-layer structure, each layer containing upsampling (US) and feature fusion operations, while embedding a channel attention mechanism. Data is progressively upsampled in the decoder, while simultaneously fusing features from the encoder to recover fine-grained structural information. In the upsampling stage, the point cloud scale is progressively restored from N / 256 to N / 64, N / 16, N / 4, and N using interpolation and other methods, while the feature dimension is progressively reduced from 512 dimensions to 8 dimensions. Next, feature fusion is performed, concatenating the features of the current layer of the decoder with the skip connection features of the corresponding layer of the encoder to compensate for fine-grained structural information that may be lost during deep feature extraction. Then, after upsampling and feature fusion at each layer of the decoder, the SE module is applied to the features. Specifically, the fused features are transposed and convolved to form a shape [B, N]. j ,1,C j ](N j C represents the number of points in the current layer. j The feature map (with the number of channels) is also processed by GAP and GMP to extract global channel information. After learning the weights through MLP, it is multiplied with the feature map channel by channel to enhance the detailed features of the power transmission equipment (such as connection parts and edge textures). The processed features are fed into the next layer decoder and finally output through the classification head. The channel weights are dynamically adjusted to enhance the ability to recover detailed features.

[0084] 4) Classification Head Design: The classification head receives the features output from the last layer of the decoder, maps the feature dimensions to the number of categories M through a multilayer perceptron, and then uses the Softmax function to obtain the probability of each point belonging to each category. Finally, the final classification category of each point is determined by taking the index of the category with the highest probability in the probability distribution of each point.

[0085] 5) Model Training: During model training, the preprocessed training set point cloud data, stratified and sampled in a 9:1:1 ratio, is input into the constructed rule-based improved RandLA-Net model. The data undergoes progressive downsampling and feature extraction through random sampling by the encoder, local feature aggregation, and channel attention mechanisms. It then enters the decoder, where upsampling, feature fusion, and the SE module achieve feature recovery and detail enhancement. The classification head then uses a multilayer perceptron to map the feature dimensions of the decoder's last layer output to the number of categories M. The Softmax function is then used to obtain the probability distribution of each point belonging to each category. The cross-entropy loss function is then used to calculate the loss between the predicted value and the true label. Backpropagation is used to calculate the gradient of the loss function with respect to the model parameters. The Adam optimizer adaptively adjusts the learning rate and updates the model parameters based on the gradient. During training, the model performance is periodically evaluated using a validation set. Hyperparameters are dynamically adjusted based on the evaluation results to prevent overfitting or underfitting. Training stops when the model's performance on the validation set no longer improves or reaches the preset maximum number of training rounds. This process ultimately yields a model with good classification performance, providing strong support for subsequent point cloud classification tasks for power transmission equipment.

[0086] 6) Rule Post-processing: The rule post-processing stage focuses on optimizing the initial classification results of the model. The model weights with the best performance are selected. After the model inferences to derive the initial classification labels and probabilities, this stage uses four serial rule post-processors to perform step-by-step optimization processing based on their respective prior rules for the initial classification results of jumpers, insulator tension, insulator V-strings, and insulator straight lines, ultimately outputting classification results with higher accuracy and completeness.

[0087] S103: Classification Result Optimization Based on Four Types of Prior Rules. To further improve the accuracy and completeness of point cloud segmentation, this invention designs four types of prior rules to perform post-processing optimization on the initial segmentation results of the model. These rules combine the geometric characteristics and topological constraints of power transmission equipment, and correct model misjudgments through structured analysis. The following is a detailed description of the four types of rules:

[0088] 1) Two-way growth rule for jumpers: To address the breakage and discontinuity issues that easily occur in jumper classification, a two-way expansion strategy is adopted to restore their complete linear structure and ensure the continuous representation of jumper regions in point cloud data. First, spatial clustering is performed using the DBSCAN density clustering algorithm to process point clouds initially classified as jumpers. By setting appropriate neighborhood radii ∈ and minimum number of points MinPts, discretely distributed jumper points are aggregated into spatially continuous clusters, effectively removing noise and isolated points. Then, the main direction is modeled. For each jumper cluster, principal component analysis (PCA) is used to extract its main direction vector. The eigenvector corresponding to the largest eigenvalue of the point cloud covariance matrix within the cluster is calculated to determine the extension trend of the jumper. For example, if the main direction vector is (x, y, z), it represents the main direction of the jumper in three-dimensional space. Then, bi-directional expansion is performed, with step-by-step growth along the main direction and its opposite direction, with each step expansion distance fixed at 0.2 meters. During the expansion process, the neighboring tags of newly added points are monitored in real time. If points other than jumper types, such as conductors or tension insulators, are encountered, growth is immediately terminated. This mechanism ensures that jumpers only expand within areas that conform to their physical connection relationships, preventing incorrect connections to other devices.

[0089] This rule significantly improves the continuity of the jumper structure, effectively restoring the true form of the jumper in the transmission line.

[0090] 2) Linear Constraint Rules for Insulator Tension Strings: This addresses the structural distortion, discontinuity, and overexpansion issues encountered in insulator tension string classification. Based on its linear installation characteristics, it enhances the completeness and geometric consistency of the classification results. Candidate region screening utilizes the DBSCAN algorithm to cluster candidate point clouds of tension insulators, forming preliminary candidate regions. This step effectively eliminates discrete points irrelevant to tension insulators, narrowing the scope of subsequent analysis. Linearity evaluation is then performed. PCA analysis is conducted on each candidate cluster to calculate the linearity index L. L is defined as the ratio of the largest eigenvalue to the sum of all eigenvalues ​​in the principal component analysis. When L is below the threshold of 0.75, the cluster is deemed to lack linearity and discarded; only candidate regions with high linearity are retained for further processing. Constraint-based expansion: Candidate clusters that pass the linearity screening are expanded in a stepwise manner along the main direction and its opposite direction, with a step size set at 0.2 meters. During expansion, the linearity and neighborhood labels of newly added points are continuously monitored. If the linearity falls below the threshold or a non-insulator point is encountered, growth is stopped. This mechanism ensures that tension insulators extend only in the direction that conforms to their linear installation pattern, avoiding accidental expansion into other equipment areas.

[0091] After processing with this rule, the structural integrity of tension insulators is significantly improved, the misclassification rate of linear structures is reduced, and the geometric accuracy of the classification results is effectively improved.

[0092] 3) Insulator V-string Symmetry Completion Rules: Addressing the common structural asymmetry and missing structures in V-shaped insulator string classification, this method utilizes mirror completion and multi-directional expansion strategies based on the insulator's symmetrical installation characteristics to restore the complete V-shaped structure. First, candidate regions are clustered using the DBSCAN algorithm to obtain preliminary V-string regions. Then, the symmetry center is calculated by determining the geometric center and lowest point (minimum Z-coordinate point) of each cluster to establish a reference point. This reference point serves as the benchmark for mirror transformation, ensuring structural symmetry after completion. Mirror completion is then performed based on this reference point. For each point within a cluster, its mirror image with respect to the symmetry center is calculated. If an unclassified point exists in the neighborhood of the mirror image point, it is classified as a V-string, thus completing the symmetrical structure. The completed point clusters are then extended in multiple directions, growing upwards and downwards from the symmetry center with a step size of 0.2 meters. During the expansion process, neighborhood labels are used to determine whether to continue growth, ensuring the integrity and continuity of the V-string structure.

[0093] This rule improves the symmetrical structural integrity of V-string insulators, effectively solving the structural defects caused by shading and viewing angle issues.

[0094] 4) Vertical Continuity Rule for Straight Insulator Strings: Addressing the issue of missing sections in the vertical direction of straight insulators, this rule leverages the vertical installation characteristics of straight insulators to achieve continuous region expansion through dynamic threshold adjustment, ensuring consistency between classification results and actual installation configurations. First, initial region clustering is performed using the DBSCAN algorithm to cluster candidate point clouds of straight insulators, forming initial candidate regions and separating point cloud sets related to straight insulators. Then, the clusters are expanded vertically, using each candidate cluster as a base, with incremental expansion along the positive and negative Z-axis directions, each step set at 0.3 meters. During expansion, the proportion of non-insulator points in the neighborhood of newly added points is calculated in real-time. Dynamic threshold adjustment is then implemented, dynamically adjusting the expansion termination threshold based on the neighborhood label distribution. For example, an initial threshold of 30% is set; if the proportion of non-insulators exceeds this threshold during expansion, growth stops. Simultaneously, as the expansion distance increases, the threshold can be appropriately reduced to accommodate slight deviations that may occur during actual installation.

[0095] This rule effectively solves the problem of vertical breakage in linear insulators, improves the integrity of vertical structures, and significantly improves the classification accuracy in complex terrain.

[0096] Four types of rules, employing a process of structural clustering, directional analysis, constraint growth, and dynamic termination, respectively, refine the optimization of jumpers, insulator tension strings, insulator V-strings, and insulator straight lines. The core idea of ​​the rule post-processor is to combine the geometric characteristics (such as linearity and symmetry) and topological constraints (such as neighborhood label relationships) of the equipment, and correct misjudgments in the model output through mathematical modeling and dynamic adjustment. Specific improvement results are shown in Table 1. Experimental results demonstrate that the four types of prior rules do not act in isolation, but rather form a complete post-processing optimization system through complementary synergy. Their combined effect is significantly better than that of a single rule or a combination of some rules. Specifically, the jumper bidirectional growth rule ensures the continuous expression of linear equipment, the insulator tension string linear constraint rule strengthens the geometric consistency of linear structures, the insulator V-string symmetry completion rule repairs the morphological deficiencies of symmetrical structures, and the insulator straight line string vertical continuity rule perfects the complete form of vertical equipment—these four rules address typical classification defects of different types of components in transmission equipment, forming a comprehensive optimization network covering "linear continuity, geometric constraints, structural symmetry, and vertical integrity." When the four types of rules work together, they not only solve the classification problem of their respective devices individually, but also achieve secondary reinforcement of error correction through cross-validation of topological relationships. For example, the bidirectional growth termination condition of jumpers depends on the accurate classification result of tension insulators, while the symmetrical completion boundary of the V-string of insulators is limited by the vertical expansion range of the straight string. This mutually constraining mechanism significantly reduces the risk of overcorrection that a single rule might produce. Whether by adding channel attention or rule optimization, the segmentation accuracy of key devices such as insulators and jumpers has been improved, significantly enhancing the classification integrity in complex scenarios. These rules provide more accurate and reliable point cloud segmentation results for power line inspection.

[0097] Table 1

[0098]

[0099] S104: Accurate Classification and Output of Power Transmission Equipment. Real-time collected and preprocessed point cloud data is input into a trained classification model. The model predicts the category of each point by fusing local geometric features and global structural features. Then, the prediction results are optimized using the four prior rules mentioned above to obtain the final classification result of the power transmission equipment. The output classification result is point cloud data mapped according to the category labels, and its data format is consistent with the input data for easy subsequent processing and display.

[0100] S105: Through 3D visualization software, point cloud classification results can be displayed intuitively, enabling power inspection personnel to quickly and clearly understand the distribution and status of various devices in transmission lines, providing strong data support for the operation and maintenance of power systems. Furthermore, by achieving end-to-end processing from point cloud acquisition to point cloud classification, the efficiency and accuracy of power inspection are greatly improved, demonstrating significant engineering application value.

[0101] Model prediction results are as follows Figure 3 As shown in the figure, based on 3D point cloud data, the algorithm can accurately identify and segment power facilities and scene elements, such as ground wires, poles, conductors, key components such as jumpers and insulator tension wires in the power system, as well as streetlights or road signs in the scene, and categories such as high and low vegetation that present obvious layers in the natural environment. The point cloud segmentation boundaries of each category are clear, providing reliable data classification support for applications such as power inspection.

[0102] In summary, this invention utilizes a drone equipped with a lidar to collect real-time point cloud data of transmission lines, acquiring information such as the three-dimensional coordinates, reflection intensity, and color of conductors, towers, and insulators. After preprocessing including voxel downsampling and KD-Tree indexing, a standardized dataset is generated. A point cloud classification model based on a rule-improved RandLA-Net is constructed, introducing a channel attention mechanism to enhance feature extraction capabilities. The classification results are optimized using four types of prior rules: a jumper bidirectional spatial growth algorithm based on DBSCAN clustering and PCA principal direction analysis; linear structure analysis and bidirectional constraint growth of insulator tension strings; mirror completion and multi-directional growth of the symmetrical structure of insulator V-strings; and bidirectional growth of insulator straight lines along the Z-axis combined with dynamic threshold adjustment and cylindrical fitting verification. The real-time collected point cloud data is input into the trained classification model, extracting and fusing local geometric features and global structural features to achieve accurate classification of transmission equipment and output equipment categories. This end-to-end processing from point cloud acquisition to point cloud classification allows for intuitive display of point cloud classification results through 3D visualization software, providing intuitive data support for power line inspection.

[0103] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0104] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0105] Through the above description of the embodiments, those skilled in the art can clearly understand that the above implementation methods can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0106] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A point cloud classification method for overhead transmission lines based on rule-based improved RandLA-Net, characterized in that, include: S1: Collect raw point cloud data of overhead transmission lines, and after data conversion and classification, perform preprocessing to obtain a standardized point cloud dataset; S2: Construct and train an improved RandLA-Net point cloud classification model, enhance features through channel attention mechanism, and output preliminary classification results; S3: Optimize the preliminary classification results based on prior rules to obtain the final classification results; S4: Output the final classification results and display them through visualization tools to support power transmission line inspection.

2. The point cloud classification method for overhead transmission lines based on rule-based improved RandLA-Net according to claim 1, characterized in that, S1 specifically includes: S11: Data Acquisition: Use a drone equipped with a lidar to scan overhead power transmission lines, and collect raw point cloud data of the overhead power transmission lines, which is stored in LASIK file format; S12: Data conversion: Convert the LAS file to TXT format, with each line in the TXT file corresponding to one point cloud data; S13: Data Classification: System annotation is performed on each point cloud data, which is divided into multiple categories, and each category corresponds to a separate label; S14: Data preprocessing: Perform hierarchical voxel downsampling and KD-Tree index construction for each point cloud data to obtain a standardized point cloud dataset.

3. The overhead transmission line point cloud classification method based on rule-based improved RandLA-Net according to claim 2, characterized in that, S14 specifically includes: S141: Hierarchical Voxel Downsampling: A voxelization strategy combining coarse and fine granularity is used to downsample each point cloud data point, setting the fine-grained voxel size V. fine =0.01m is used to preserve key detail features, and the coarse-grained voxel size V coarse =0.06m is used to quickly reduce the amount of data; S142: KD-Tree Index Construction: Based on hierarchical voxelization, a KD-Tree spatial index is constructed based on coarse-grained voxel point clouds. Efficient K-nearest neighbor queries are achieved through recursive spatial partitioning, realizing accurate mapping from fine-grained point clouds to coarse-grained point clouds, thereby obtaining a standardized point cloud dataset.

4. The point cloud classification method for overhead transmission lines based on rule-based improved RandLA-Net according to claim 1, characterized in that, S2 specifically includes: S21: Organize the standardized point cloud dataset into a shape of N×(3+d) in A matrix, where N is the number of points, 3 represents the spatial coordinates, and d in It is the input feature dimension, the original features used as input to the model; S22: For the original features of the model input, the improved RandLA-Net point cloud classification model encoder obtains enhanced features through random sampling, local feature aggregation, and channel attention module; S23, for the enhanced features, the decoder of the improved RandLA-Net point cloud classification model enhances the detailed features of the power transmission equipment by upsampling and feature fusion, combined with the channel attention module; S24. Based on the features obtained in S23, the classification head maps the feature dimension to the number of categories through a multilayer perceptron, uses the Softmax function to output the category probability of each point, and determines the predicted classification category of each point by taking the index of the category with the highest probability in the probability distribution of each point. S25 utilizes the cross-entropy loss function and Adam optimizer to adaptively adjust the learning rate based on the gradient and update the parameters of the improved RandLA-Net point cloud classification model to optimize the model's classification performance. S26. The improved RandLA-Net point cloud classification model, after training, is used to output preliminary classification results.

5. The overhead transmission line point cloud classification method based on rule-improved RandLA-Net according to claim 4, characterized in that, S22 specifically includes: Random sampling: In each layer, the number of point clouds is gradually compressed from the initial N points to 1 / 4, 1 / 16, 1 / 64 and 1 / 256 of the original size; Local feature aggregation: Spatial encoding and attention pooling operations are performed to construct multi-level feature representations. Residual connections are used to merge the original features and multi-level feature representations to output the first feature map. Channel Attention Module: For the first feature map, the channel attention module performs global average pooling and global max pooling to compress the spatial dimension and obtain global channel information; it learns channel attention weights through two layers of multilayer perceptron; and it multiplies the channel attention weights with the first feature map channel by channel to output the enhanced features.

6. The overhead transmission line point cloud classification method based on rule-improved RandLA-Net according to claim 4, characterized in that, S23 specifically includes: Upsampling: The point cloud size is gradually restored from N / 256 to N / 64, N / 16, N / 4 and N, and the feature dimension is gradually reduced from 512 to 8. Feature fusion: The features of the current layer of the decoder are concatenated and fused with the enhanced features of the corresponding layer of the encoder; Channel Attention Module: The concatenated and fused features are transposed and convolved to form a second feature map. Global channel information is extracted by local average pooling and global max pooling. After learning weights through a multilayer perceptron, the weights are multiplied with the second feature map channel by channel to enhance the detailed features of the power transmission equipment.

7. The overhead transmission line point cloud classification method based on rule-based improved RandLA-Net according to claim 1, characterized in that, S3 specifically includes: based on the preliminary classification results output by the improved RandLA-Net point cloud classification model, enabling four serial rule post-processors to perform step-by-step optimization processing on the preliminary classification results of jumpers, insulator tension, insulator V-strings, and insulator straight lines according to their respective prior rules, and outputting the final classification results.

8. The point cloud classification method for overhead transmission lines based on rule-based improved RandLA-Net according to claim 1, characterized in that, In S3, the prior rule is specifically as follows: The jumper bidirectional growth rule, through DBSCAN clustering and PCA principal direction analysis, expands bidirectionally by 0.2 meters along the principal direction to repair jumper breakage issues; The linear constraint rules for tension insulator strings ensure the integrity of the linear structure of tension insulators through linearity evaluation and constraint extension; The insulator V-string symmetry completion rule restores the V-shaped structure by calculating the symmetry center and mirroring the completion, expanding by 0.2 meters in the vertical direction; The vertical continuity rule of the insulator straight string is improved by extending the Z-axis bidirectionally by 0.3 meters and adjusting the dynamic threshold to ensure the integrity of the vertical structure.

9. A point cloud classification system for overhead transmission lines based on rule-based improved RandLA-Net, the system comprising: A processor and a memory for storing executable instructions; characterized in that the processor is configured to execute the executable instructions to perform the rule-based improved RandLA-Net overhead transmission line point cloud classification method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the rule-based improved RandLA-Net point cloud classification method for overhead transmission lines as described in any one of claims 1 to 8.

Images