Cable automatic point selection method in live working scene

Abstract

The application provides a cable automatic point selection method in a live working scene, mainly comprising the following steps: point cloud acquisition and preprocessing: collecting three-dimensional point cloud data of the working scene by a laser radar carried by a live working robot, and performing preprocessing operation; main line and branch line point cloud segmentation: based on a deep learning point cloud semantic segmentation model, the preprocessed scene point cloud is segmented and processed to automatically distinguish the corresponding point cloud data; working point automatic generation and screening: based on the completion of main line and branch line point cloud recognition, combined with preset working rules and space constraint conditions, the target working point meeting the live working safety and operation requirements is screened out. Through the point cloud segmentation and structure analysis based on deep learning, the automatic selection of the main line and branch line working points is realized, the inconsistency problem caused by the experience difference and subjective judgment factors of the operators in the manual point selection process is avoided, and the accuracy and stability of the working point selection are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of robotic live-line working technology, specifically a method for automatic cable selection in live-line working scenarios. Background Technology

[0002] Live-line working robots are a type of specialized robot capable of performing high-risk tasks such as splicing and dismantling high-voltage cables in high-altitude environments, replacing manual labor. Compared to traditional manual methods, these robots significantly improve operational safety and efficiency. In actual operation, live-line working robots need to accurately obtain key operational information such as the root location and direction of the main line, as well as the root points and gripping points of branch lines. The accuracy of the selected work points directly affects subsequent robotic arm path planning, end-effector quality, and overall operational safety.

[0003] However, in current engineering applications, the selection of work points for live-line working robots still relies mainly on manual labor, resulting in a low level of automation and the lack of a stable and reliable automatic point selection technology solution. In existing technologies, work points are typically manually designated by operators through a host computer interface. After the work begins, the operator observes the spatial distribution of the main line and branches based on the point cloud data acquired by the robot's onboard LiDAR, and identifies and judges the work target based on personal experience. During the manual point selection phase, the operator needs to manually mark the work point location on the 3D display interface through clicking or touch.

[0004] Due to the complex and ever-changing on-site working environment, the work points initially selected manually often fail to directly meet the actual operational needs. Operators typically need to continuously switch observation perspectives during the manual correction phase, adjusting and confirming the work point location multiple times until they deem it meets the requirements for subsequent operations. While the aforementioned manual point selection method is simple to implement and highly versatile, it has significant shortcomings in practical applications, including a high degree of reliance on operator experience, low efficiency in work point selection, and difficulty in guaranteeing the accuracy and consistency of point selection.

[0005] In summary, existing live-line working robots still heavily rely on manual operation in the work point selection stage, lacking sufficient automation and struggling to adapt to complex and ever-changing live-line working environments. Therefore, there is an urgent need to propose an automated work point selection method that can effectively reduce human intervention and improve the accuracy and consistency of work point selection, to meet the practical needs of live-line working robots developing towards intelligence and automation. Live-line working robots are a type of special robot capable of replacing manual labor in high-altitude environments to complete high-risk operations such as high-voltage cable splicing and dismantling. Compared to traditional manual operations, these robots can significantly improve operational safety and efficiency. In actual operation, live-line working robots need to accurately obtain key operational information such as the root position and cable direction of the main line, as well as the root point and gripping point of branch lines. The accuracy of work point selection directly affects subsequent robotic arm path planning, end-effector operation quality, and overall operational safety.

[0006] However, in current engineering applications, the selection of work points for live-line working robots still relies mainly on manual labor, resulting in a low level of automation and the lack of a stable and reliable automatic point selection technology solution. In existing technologies, work points are typically manually designated by operators through a host computer interface. After the work begins, the operator observes the spatial distribution of the main line and branches based on the point cloud data acquired by the robot's onboard LiDAR, and identifies and judges the work target based on personal experience. During the manual point selection phase, the operator needs to manually mark the work point location on the 3D display interface through clicking or touch.

[0007] Due to the complex and ever-changing on-site working environment, the work points initially selected manually often fail to directly meet the actual operational needs. Operators typically need to continuously switch observation perspectives during the manual correction phase, adjusting and confirming the work point location multiple times until they deem it meets the requirements for subsequent operations. While the aforementioned manual point selection method is simple to implement and highly versatile, it has significant shortcomings in practical applications, including a high degree of reliance on operator experience, low efficiency in work point selection, and difficulty in guaranteeing the accuracy and consistency of point selection.

[0008] In summary, existing live-line working robots still heavily rely on manual operation in the work site selection stage, lacking sufficient automation and struggling to adapt to complex and ever-changing live-line working environments. Therefore, there is an urgent need to propose an automated site selection method that can effectively reduce manual intervention and improve the accuracy and consistency of work site selection, in order to meet the practical needs of the development of live-line working robots towards intelligence and automation. Summary of the Invention

[0009] To address the problems of existing technologies, this invention provides an automatic cable selection method for live-line working scenarios. Through automation, it achieves accurate and stable determination of main and branch line work points, reduces manual intervention, and improves the operational reliability and intelligence level of live-line working robots.

[0010] This invention provides an automatic cable selection method for live-line working scenarios, comprising the following steps:

[0011] 1) Point Cloud Acquisition and Preprocessing: 3D point cloud data of the work scene is acquired using the lidar mounted on the live-line working robot. The raw point cloud is preprocessed to generate standardized point cloud data that meets the requirements of subsequent analysis. ;

[0012] 2) Point cloud segmentation of main lines and branches: Based on a deep learning-based point cloud semantic segmentation model, the preprocessed scene point cloud is segmented to automatically distinguish the point cloud data corresponding to main lines, branches, crossarms, and other non-operational target structures. The specific process is as follows:

[0013] 2.1) The preprocessed standardized point cloud data Input the local spatial feature encoding module to encode the geometric distribution of the point cloud within a local neighborhood;

[0014] 2.2) A graph attention pooling module is used to perform weighted aggregation of neighborhood point features;

[0015] 2.3) Further expand the receptive field of the network by introducing an extended residual module;

[0016] 3) Automatic generation and selection of work points: Based on the identification of main line and branch line point clouds, and combined with preset work rules and spatial constraints, the location of candidate work points is automatically calculated from the point cloud corresponding to the main line or branch line. The spatial rationality and work feasibility of the candidate work points are analyzed, and the target work points that meet the safety and operation requirements of live work are selected, thus realizing the automatic selection of work points.

[0017] The root point of the main line and its extension direction, as well as the root point and snap point of the branch line. Step 1) The preprocessing process includes voxel filtering and outlier removal in sequence, and the specific process is as follows:

[0018] 1.1) Voxel filtering is used to downsample the point cloud, and the voxel size is set to... Reduce point cloud density while maintaining the geometric shape of the point cloud;

[0019] 1.2) Remove interfering points from the existing point cloud and perform statistical analysis on the point cloud. Each point cloud Number of point clouds within a radius ,when At that time, the point was considered an outlier. Remove the point from the middle, where This is a hyperparameter.

[0020] The encoding format described in step 2.1) for the root point of the main line and its extension direction, as well as the root point and snap point of the branch line, is as follows:

[0021]

[0022] in, For point Geometric encoding information, It is a multilayer perceptron. For the location information of the center point, The first point corresponding to the center point Location information of each neighboring point This indicates a feature cascade operation, which extracts local geometric structure features of the point cloud in the neighborhood by encoding the relative displacement and distance information between the center point and its neighboring points.

[0023] The graph attention pooling module described in step 2.2) mainly includes the following steps: (The root point of the main line and its extension direction, as well as the root points and snapping points of the branches.)

[0024] 2.21) Attention Score Calculation: Using Convolution calculates the attention score of a node to its neighboring nodes;

[0025] 2.22) Normalization: Normalizing the attention scores The function is normalized to ensure that the fractions are within the range specified in the original text. Within the range;

[0026] 2.23) Weighted feature summation: Multiply the local features by the corresponding attention scores to obtain weighted features, and sum them over the neighborhood dimension;

[0027] 2.24) Feature mapping: through Convolutional mapping enhances feature representation capabilities.

[0028] The root point of the main line and its extension direction, as well as the root point and grasping point of the branch line, are described in step 2.3). The expanded residual module introduces an expanded sampling strategy while maintaining the sparsity of the point cloud. Each expanded residual module performs local spatial feature encoding and graph attention pooling operations twice in sequence to enhance the feature representation ability of the point cloud in a larger neighborhood. At the same time, a residual connection structure is introduced between the module input and output to realize the cross-layer transfer of features, avoid gradient vanishing and feature degradation during deep network training, and improve the overall perception ability of long strip-shaped structural targets.

[0029] The specific process for selecting target work points that meet the safety and operational requirements for live-line work, as described in step 3), including the root point of the main line and its extension direction, as well as the root point and gripping point of the branch line, is as follows:

[0030] 3.1) Based on the semantic segmentation results of the point cloud, the point cloud is filtered according to the semantic category, and the main line cable point cloud set and the branch line cable point cloud set are extracted respectively.

[0031] 3.2) For the main point cloud set and the branch point cloud set, respectively adopt... Clustering algorithms perform spatial clustering; by setting radius, The minimum number of cluster points is used to aggregate the point clouds corresponding to the same cable into the same cluster, thereby effectively avoiding the overlap of point clouds between different cables;

[0032] 3.3) For each cluster of point clouds, geometric features are calculated to extract the root points and extension directions of the main lines, as well as the root points and snap points of the branches. The extraction process is as follows:

[0033] 3.31) Combining the robot coordinate system Directional information is obtained by comparing the position of each point in the point cloud cluster with that of the central point in the point cloud cluster. The positional relationship in terms of direction is determined by selecting the root points of the main line and the branch line;

[0034] 3.32) A three-dimensional straight line fitting method based on principal component analysis is adopted. By statistically analyzing the point cloud of the main line, the first principal direction of its point cloud distribution is extracted, and this principal direction is used as the extension direction of the main line cable in space.

[0035] 3.3) Using the root point of the branch as a reference, select points in the corresponding branch point cluster that are adjacent to the root point. The distance in the direction satisfies the threshold condition The point is used as the branch line grab point.

[0036] The beneficial effects of this invention are as follows: by using point cloud segmentation and structural analysis based on deep learning, the main line and branch line operation points are automatically selected, avoiding the inconsistency caused by differences in operator experience and subjective judgment factors during manual point selection, and significantly improving the accuracy and stability of operation point selection. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments 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 these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the overall process of point cloud semantic segmentation and mainline localization based on RandLA-Net;

[0039] Figure 2 This is a preprocessed point cloud data map;

[0040] Figure 3 This is a schematic diagram of the point cloud semantic segmentation results;

[0041] Figure 4 This is a schematic diagram of the automatic point selection results. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] In this invention, a live-line working robot is equipped with a lidar for collecting three-dimensional point cloud data within the working area. The sensor is mounted on the robot body. In actual live-line working scenarios, the working area typically contains multiple targets simultaneously, including main cables, branch cables, and background structures, resulting in a complex spatial distribution of point clouds. Traditional methods based on geometric features or geometric fitting are insufficient to reliably distinguish between main cables and branch cables.

[0044] To address the aforementioned issues, this invention proposes an automatic selection method for cable work points based on point cloud semantic segmentation. By performing semantic understanding on the point cloud of the work scene, the method achieves accurate identification of the main line and branch lines, and on this basis, reliably determines the work points.

[0045] like Figure 1 As shown, Figure 1 This is a schematic diagram of the overall process of point cloud semantic segmentation and main line localization based on RandLA-Net according to the present invention. The method of the present invention mainly includes the following steps:

[0046] Initially, a 3D depth sensor was used to acquire raw 3D point cloud data within the work area. To reduce noise interference and computational load, the original point cloud is first processed. Preprocessing is performed, which includes voxel filtering and outlier removal. Specifically, voxel filtering is used to downsample the point cloud, with the voxel size set to... By reducing the point cloud density while maintaining the basic geometric shape of the point cloud, the efficiency of subsequent point cloud segmentation and feature extraction algorithms can be improved.

[0047] The next step is to remove interfering points from the existing point cloud and perform statistical analysis on the point cloud. Each point cloud Number of point clouds within a radius ,when At that time, the point was considered an outlier. Remove this point from the middle (Note: (for hyperparameters) Figure 2 This is the preprocessed point cloud data.

[0048] Next, the preprocessed point cloud is input into the local spatial feature encoding module to encode the geometric distribution of the point cloud within a local neighborhood. The encoding format is as follows:

[0049]

[0050] in, For point Geometric encoding information, It is a multilayer perceptron. For the location information of the center point, The first point corresponding to the center point Location information of each neighboring point This indicates a feature cascade operation, which can effectively extract local geometric structure features of the point cloud in the neighborhood by encoding the relative displacement and distance information between the center point and its neighboring points.

[0051] Next, a graph attention pooling module is used to weight and aggregate the features of neighboring points to enhance the expressive power of key structures and suppress the influence of noise points. The graph attention pooling module mainly includes the following steps: (1) attention score calculation; (2) Normalization; (3) Weighted feature summation; (4) Feature mapping. Specifically, first use The convolution calculates the attention score of a node with respect to its neighboring nodes, and then the attention score is processed... The function is normalized to ensure that the attention score is within the range of normal values. Within the specified range, the local features are multiplied by the corresponding attention scores to obtain weighted features, which are then summed along the neighborhood dimension. Finally, the summation is performed... Convolutional mapping features to further enhance feature representation capabilities.

[0052] Next, the receptive field of the network is further expanded by introducing an extended residual module. This extended residual module introduces an extended sampling strategy while maintaining the sparsity of the point cloud. Each extended residual module sequentially performs two local spatial feature encoding and graph attention pooling operations to enhance the feature representation capability of the point cloud over a larger neighborhood. Simultaneously, a residual connection structure is introduced between the module input and output to achieve cross-layer feature transfer, avoiding gradient vanishing and feature degradation problems during deep network training, thereby improving the overall perception capability of long, strip-shaped targets such as cables.

[0053] Next, based on the point cloud semantic segmentation results, such as Figure 3 As shown, point clouds are filtered according to semantic categories, and main line cable point cloud sets and branch line cable point cloud sets are extracted separately. For the main line point cloud set and the branch line point cloud set, respectively... Clustering algorithms perform spatial clustering. This is achieved by setting... radius, Minimum clustering points group point clouds corresponding to the same cable into the same cluster, thereby effectively avoiding point cloud overlap between different cables.

[0054] Next, geometric feature calculations are performed on the point cloud clusters obtained from each clustering, extracting the root points of the main lines and their extension directions, as well as the root points and gripping points of the branches. Since the robot faces the work scene during operation, the root points of the main lines and branches can be combined with the robot coordinate system. Directional information is obtained by comparing the position of each point in the point cloud cluster with that of the central point in the point cloud cluster. The selection of directional positional relationships is based on the following: The extension direction of the main line is determined using a three-dimensional straight-line fitting method based on Principal Component Analysis (PCA). Statistical analysis of the main line point cloud is performed to extract the first principal direction of its point cloud distribution, which is then used as the extension direction of the main line cable in space. The selection of branch line capture points is based on the root point of the branch line; points are selected from the corresponding branch line point cloud cluster that are closest to the root point. The distance in the direction satisfies the threshold condition The point is used as the target line capture point, and the result is as follows: Figure 4 As shown.

[0055] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, for the device embodiments, the above descriptions are merely preferred embodiments of the present invention. Since they are fundamentally similar to the method embodiments, the descriptions are relatively simple, and relevant parts can be referred to the descriptions of the method embodiments. The above descriptions are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention, without departing from the principle of the present invention, should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for automatic cable selection in live-line working scenarios, characterized in that... Includes the following steps: 1) Point Cloud Acquisition and Preprocessing: 3D point cloud data of the work scene is acquired using the lidar mounted on the live-line working robot. The raw point cloud is preprocessed to generate standardized point cloud data that meets the requirements of subsequent analysis. ; 2) Point cloud segmentation of main lines and branches: Based on a deep learning-based point cloud semantic segmentation model, the preprocessed scene point cloud is segmented to automatically distinguish the point cloud data corresponding to main lines, branches, crossarms, and other non-operational target structures. The specific process is as follows: 2.1) The preprocessed standardized point cloud data Input the local spatial feature encoding module to encode the geometric distribution of the point cloud within a local neighborhood; 2.2) A graph attention pooling module is used to perform weighted aggregation of neighborhood point features; 2.3) Further expand the receptive field of the network by introducing an extended residual module; 3) Automatic generation and selection of work points: Based on the identification of main line and branch line point clouds, and combined with preset work rules and spatial constraints, the location of candidate work points is automatically calculated from the point cloud corresponding to the main line or branch line. The spatial rationality and work feasibility of the candidate work points are analyzed, and the target work points that meet the safety and operation requirements of live work are selected, thus realizing the automatic selection of work points.

2. The automatic cable selection method for live-line working scenarios according to claim 1, characterized in that: Step 1) The preprocessing process includes voxel filtering and outlier removal in sequence, and the specific process is as follows: 1.1) Voxel filtering is used to downsample the point cloud, and the voxel size is set to... Reduce point cloud density while maintaining the geometric shape of the point cloud; 1.2) Remove interfering points from the existing point cloud and perform statistical analysis on the point cloud. Each point cloud Number of point clouds within a radius ,when At that time, the point was considered an outlier. Remove the point from the middle, where This is a hyperparameter.

3. The automatic cable selection method for live-line working scenarios according to claim 1, characterized in that: The encoding format described in step 2.1) is as follows: in, For point Geometric encoding information, It is a multilayer perceptron. For the location information of the center point, The first point corresponding to the center point Location information of each neighboring point This indicates a feature cascade operation, which extracts local geometric structure features of the point cloud in the neighborhood by encoding the relative displacement and distance information between the center point and its neighboring points.

4. The automatic cable selection method for live-line working scenarios according to claim 1, characterized in that: Step 2.2) describes the graph attention pooling module, which mainly includes the following steps: 2.21) Attention Score Calculation: Using Convolution calculates the attention score of a node to its neighboring nodes; 2.22) Normalization: Normalizing the attention scores The function is normalized to ensure that the fractions are within the range specified in the original text. Within the range; 2.23) Weighted feature summation: Multiply the local features by the corresponding attention scores to obtain weighted features, and sum them over the neighborhood dimension; 2.24) Feature mapping: through Convolutional mapping enhances feature representation capabilities.

5. The automatic cable selection method for live-line working scenarios according to claim 1, characterized in that: Step 2.3) The expanded residual module introduces an expanded sampling strategy while maintaining the sparsity of the point cloud. Each expanded residual module performs local spatial feature encoding and graph attention pooling operations twice in sequence to enhance the feature representation ability of the point cloud in a larger neighborhood. At the same time, a residual connection structure is introduced between the module input and output to realize cross-layer feature transfer, avoid gradient vanishing and feature degradation during deep network training, and improve the overall perception ability of long strip-shaped targets.

6. The automatic cable selection method for live-line working scenarios according to claim 1, characterized in that: Step 3) The process of selecting target work points that meet the safety and operational requirements for live-line work is as follows: 3.1) Based on the semantic segmentation results of the point cloud, the point cloud is filtered according to the semantic category, and the main line cable point cloud set and the branch line cable point cloud set are extracted respectively. 3.2) For the main point cloud set and the branch point cloud set, respectively adopt... Clustering algorithms perform spatial clustering; by setting radius, The minimum number of cluster points is used to aggregate the point clouds corresponding to the same cable into the same cluster, thereby effectively avoiding the overlap of point clouds between different cables; 3.3) Perform geometric feature calculations on the point cloud clusters obtained from each clustering, and extract the root points of the main line and its extension direction, as well as the root points and capture points of the branches.

7. The automatic cable selection method for live-line working scenarios according to claim 6, characterized in that: The extraction process described in step 3.3) is as follows: 3.31) Combining the robot coordinate system Directional information is obtained by comparing the position of each point in the point cloud cluster with that of the central point in the point cloud cluster. The positional relationship in terms of direction is determined by selecting the root points of the main line and the branch line; 3.32) A three-dimensional straight line fitting method based on principal component analysis is adopted. By statistically analyzing the point cloud of the main line, the first principal direction of its point cloud distribution is extracted, and this principal direction is used as the extension direction of the main line cable in space. 3.3) Using the root point of the branch as a reference, select points in the corresponding branch point cluster that are adjacent to the root point. The distance in the direction satisfies the threshold condition The point is used as the branch line grab point.

Images